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 int 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 * Minimum page order among possible hugepage sizes, set to a proper value
47 static unsigned int minimum_order __read_mostly = UINT_MAX;
49 __initdata LIST_HEAD(huge_boot_pages);
51 /* for command line parsing */
52 static struct hstate * __initdata parsed_hstate;
53 static unsigned long __initdata default_hstate_max_huge_pages;
54 static unsigned long __initdata default_hstate_size;
57 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
58 * free_huge_pages, and surplus_huge_pages.
60 DEFINE_SPINLOCK(hugetlb_lock);
63 * Serializes faults on the same logical page. This is used to
64 * prevent spurious OOMs when the hugepage pool is fully utilized.
66 static int num_fault_mutexes;
67 static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp;
69 /* Forward declaration */
70 static int hugetlb_acct_memory(struct hstate *h, long delta);
72 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
74 bool free = (spool->count == 0) && (spool->used_hpages == 0);
76 spin_unlock(&spool->lock);
78 /* If no pages are used, and no other handles to the subpool
79 * remain, give up any reservations mased on minimum size and
82 if (spool->min_hpages != -1)
83 hugetlb_acct_memory(spool->hstate,
89 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
92 struct hugepage_subpool *spool;
94 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
98 spin_lock_init(&spool->lock);
100 spool->max_hpages = max_hpages;
102 spool->min_hpages = min_hpages;
104 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
108 spool->rsv_hpages = min_hpages;
113 void hugepage_put_subpool(struct hugepage_subpool *spool)
115 spin_lock(&spool->lock);
116 BUG_ON(!spool->count);
118 unlock_or_release_subpool(spool);
122 * Subpool accounting for allocating and reserving pages.
123 * Return -ENOMEM if there are not enough resources to satisfy the
124 * the request. Otherwise, return the number of pages by which the
125 * global pools must be adjusted (upward). The returned value may
126 * only be different than the passed value (delta) in the case where
127 * a subpool minimum size must be manitained.
129 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
137 spin_lock(&spool->lock);
139 if (spool->max_hpages != -1) { /* maximum size accounting */
140 if ((spool->used_hpages + delta) <= spool->max_hpages)
141 spool->used_hpages += delta;
148 if (spool->min_hpages != -1) { /* minimum size accounting */
149 if (delta > spool->rsv_hpages) {
151 * Asking for more reserves than those already taken on
152 * behalf of subpool. Return difference.
154 ret = delta - spool->rsv_hpages;
155 spool->rsv_hpages = 0;
157 ret = 0; /* reserves already accounted for */
158 spool->rsv_hpages -= delta;
163 spin_unlock(&spool->lock);
168 * Subpool accounting for freeing and unreserving pages.
169 * Return the number of global page reservations that must be dropped.
170 * The return value may only be different than the passed value (delta)
171 * in the case where a subpool minimum size must be maintained.
173 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
181 spin_lock(&spool->lock);
183 if (spool->max_hpages != -1) /* maximum size accounting */
184 spool->used_hpages -= delta;
186 if (spool->min_hpages != -1) { /* minimum size accounting */
187 if (spool->rsv_hpages + delta <= spool->min_hpages)
190 ret = spool->rsv_hpages + delta - spool->min_hpages;
192 spool->rsv_hpages += delta;
193 if (spool->rsv_hpages > spool->min_hpages)
194 spool->rsv_hpages = spool->min_hpages;
198 * If hugetlbfs_put_super couldn't free spool due to an outstanding
199 * quota reference, free it now.
201 unlock_or_release_subpool(spool);
206 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
208 return HUGETLBFS_SB(inode->i_sb)->spool;
211 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
213 return subpool_inode(file_inode(vma->vm_file));
217 * Region tracking -- allows tracking of reservations and instantiated pages
218 * across the pages in a mapping.
220 * The region data structures are embedded into a resv_map and
221 * protected by a resv_map's lock
224 struct list_head link;
229 static long region_add(struct resv_map *resv, long f, long t)
231 struct list_head *head = &resv->regions;
232 struct file_region *rg, *nrg, *trg;
234 spin_lock(&resv->lock);
235 /* Locate the region we are either in or before. */
236 list_for_each_entry(rg, head, link)
240 /* Round our left edge to the current segment if it encloses us. */
244 /* Check for and consume any regions we now overlap with. */
246 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
247 if (&rg->link == head)
252 /* If this area reaches higher then extend our area to
253 * include it completely. If this is not the first area
254 * which we intend to reuse, free it. */
264 spin_unlock(&resv->lock);
268 static long region_chg(struct resv_map *resv, long f, long t)
270 struct list_head *head = &resv->regions;
271 struct file_region *rg, *nrg = NULL;
275 spin_lock(&resv->lock);
276 /* Locate the region we are before or in. */
277 list_for_each_entry(rg, head, link)
281 /* If we are below the current region then a new region is required.
282 * Subtle, allocate a new region at the position but make it zero
283 * size such that we can guarantee to record the reservation. */
284 if (&rg->link == head || t < rg->from) {
286 spin_unlock(&resv->lock);
287 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
293 INIT_LIST_HEAD(&nrg->link);
297 list_add(&nrg->link, rg->link.prev);
302 /* Round our left edge to the current segment if it encloses us. */
307 /* Check for and consume any regions we now overlap with. */
308 list_for_each_entry(rg, rg->link.prev, link) {
309 if (&rg->link == head)
314 /* We overlap with this area, if it extends further than
315 * us then we must extend ourselves. Account for its
316 * existing reservation. */
321 chg -= rg->to - rg->from;
325 spin_unlock(&resv->lock);
326 /* We already know we raced and no longer need the new region */
330 spin_unlock(&resv->lock);
334 static long region_truncate(struct resv_map *resv, long end)
336 struct list_head *head = &resv->regions;
337 struct file_region *rg, *trg;
340 spin_lock(&resv->lock);
341 /* Locate the region we are either in or before. */
342 list_for_each_entry(rg, head, link)
345 if (&rg->link == head)
348 /* If we are in the middle of a region then adjust it. */
349 if (end > rg->from) {
352 rg = list_entry(rg->link.next, typeof(*rg), link);
355 /* Drop any remaining regions. */
356 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
357 if (&rg->link == head)
359 chg += rg->to - rg->from;
365 spin_unlock(&resv->lock);
369 static long region_count(struct resv_map *resv, long f, long t)
371 struct list_head *head = &resv->regions;
372 struct file_region *rg;
375 spin_lock(&resv->lock);
376 /* Locate each segment we overlap with, and count that overlap. */
377 list_for_each_entry(rg, head, link) {
386 seg_from = max(rg->from, f);
387 seg_to = min(rg->to, t);
389 chg += seg_to - seg_from;
391 spin_unlock(&resv->lock);
397 * Convert the address within this vma to the page offset within
398 * the mapping, in pagecache page units; huge pages here.
400 static pgoff_t vma_hugecache_offset(struct hstate *h,
401 struct vm_area_struct *vma, unsigned long address)
403 return ((address - vma->vm_start) >> huge_page_shift(h)) +
404 (vma->vm_pgoff >> huge_page_order(h));
407 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
408 unsigned long address)
410 return vma_hugecache_offset(hstate_vma(vma), vma, address);
414 * Return the size of the pages allocated when backing a VMA. In the majority
415 * cases this will be same size as used by the page table entries.
417 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
419 struct hstate *hstate;
421 if (!is_vm_hugetlb_page(vma))
424 hstate = hstate_vma(vma);
426 return 1UL << huge_page_shift(hstate);
428 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
431 * Return the page size being used by the MMU to back a VMA. In the majority
432 * of cases, the page size used by the kernel matches the MMU size. On
433 * architectures where it differs, an architecture-specific version of this
434 * function is required.
436 #ifndef vma_mmu_pagesize
437 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
439 return vma_kernel_pagesize(vma);
444 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
445 * bits of the reservation map pointer, which are always clear due to
448 #define HPAGE_RESV_OWNER (1UL << 0)
449 #define HPAGE_RESV_UNMAPPED (1UL << 1)
450 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
453 * These helpers are used to track how many pages are reserved for
454 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
455 * is guaranteed to have their future faults succeed.
457 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
458 * the reserve counters are updated with the hugetlb_lock held. It is safe
459 * to reset the VMA at fork() time as it is not in use yet and there is no
460 * chance of the global counters getting corrupted as a result of the values.
462 * The private mapping reservation is represented in a subtly different
463 * manner to a shared mapping. A shared mapping has a region map associated
464 * with the underlying file, this region map represents the backing file
465 * pages which have ever had a reservation assigned which this persists even
466 * after the page is instantiated. A private mapping has a region map
467 * associated with the original mmap which is attached to all VMAs which
468 * reference it, this region map represents those offsets which have consumed
469 * reservation ie. where pages have been instantiated.
471 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
473 return (unsigned long)vma->vm_private_data;
476 static void set_vma_private_data(struct vm_area_struct *vma,
479 vma->vm_private_data = (void *)value;
482 struct resv_map *resv_map_alloc(void)
484 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
488 kref_init(&resv_map->refs);
489 spin_lock_init(&resv_map->lock);
490 INIT_LIST_HEAD(&resv_map->regions);
495 void resv_map_release(struct kref *ref)
497 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
499 /* Clear out any active regions before we release the map. */
500 region_truncate(resv_map, 0);
504 static inline struct resv_map *inode_resv_map(struct inode *inode)
506 return inode->i_mapping->private_data;
509 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
511 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
512 if (vma->vm_flags & VM_MAYSHARE) {
513 struct address_space *mapping = vma->vm_file->f_mapping;
514 struct inode *inode = mapping->host;
516 return inode_resv_map(inode);
519 return (struct resv_map *)(get_vma_private_data(vma) &
524 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
526 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
527 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
529 set_vma_private_data(vma, (get_vma_private_data(vma) &
530 HPAGE_RESV_MASK) | (unsigned long)map);
533 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
535 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
536 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
538 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
541 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
543 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
545 return (get_vma_private_data(vma) & flag) != 0;
548 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
549 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
551 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
552 if (!(vma->vm_flags & VM_MAYSHARE))
553 vma->vm_private_data = (void *)0;
556 /* Returns true if the VMA has associated reserve pages */
557 static int vma_has_reserves(struct vm_area_struct *vma, long chg)
559 if (vma->vm_flags & VM_NORESERVE) {
561 * This address is already reserved by other process(chg == 0),
562 * so, we should decrement reserved count. Without decrementing,
563 * reserve count remains after releasing inode, because this
564 * allocated page will go into page cache and is regarded as
565 * coming from reserved pool in releasing step. Currently, we
566 * don't have any other solution to deal with this situation
567 * properly, so add work-around here.
569 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
575 /* Shared mappings always use reserves */
576 if (vma->vm_flags & VM_MAYSHARE)
580 * Only the process that called mmap() has reserves for
583 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
589 static void enqueue_huge_page(struct hstate *h, struct page *page)
591 int nid = page_to_nid(page);
592 list_move(&page->lru, &h->hugepage_freelists[nid]);
593 h->free_huge_pages++;
594 h->free_huge_pages_node[nid]++;
597 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
601 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
602 if (!is_migrate_isolate_page(page))
605 * if 'non-isolated free hugepage' not found on the list,
606 * the allocation fails.
608 if (&h->hugepage_freelists[nid] == &page->lru)
610 list_move(&page->lru, &h->hugepage_activelist);
611 set_page_refcounted(page);
612 h->free_huge_pages--;
613 h->free_huge_pages_node[nid]--;
617 /* Movability of hugepages depends on migration support. */
618 static inline gfp_t htlb_alloc_mask(struct hstate *h)
620 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
621 return GFP_HIGHUSER_MOVABLE;
626 static struct page *dequeue_huge_page_vma(struct hstate *h,
627 struct vm_area_struct *vma,
628 unsigned long address, int avoid_reserve,
631 struct page *page = NULL;
632 struct mempolicy *mpol;
633 nodemask_t *nodemask;
634 struct zonelist *zonelist;
637 unsigned int cpuset_mems_cookie;
640 * A child process with MAP_PRIVATE mappings created by their parent
641 * have no page reserves. This check ensures that reservations are
642 * not "stolen". The child may still get SIGKILLed
644 if (!vma_has_reserves(vma, chg) &&
645 h->free_huge_pages - h->resv_huge_pages == 0)
648 /* If reserves cannot be used, ensure enough pages are in the pool */
649 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
653 cpuset_mems_cookie = read_mems_allowed_begin();
654 zonelist = huge_zonelist(vma, address,
655 htlb_alloc_mask(h), &mpol, &nodemask);
657 for_each_zone_zonelist_nodemask(zone, z, zonelist,
658 MAX_NR_ZONES - 1, nodemask) {
659 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
660 page = dequeue_huge_page_node(h, zone_to_nid(zone));
664 if (!vma_has_reserves(vma, chg))
667 SetPagePrivate(page);
668 h->resv_huge_pages--;
675 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
684 * common helper functions for hstate_next_node_to_{alloc|free}.
685 * We may have allocated or freed a huge page based on a different
686 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
687 * be outside of *nodes_allowed. Ensure that we use an allowed
688 * node for alloc or free.
690 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
692 nid = next_node(nid, *nodes_allowed);
693 if (nid == MAX_NUMNODES)
694 nid = first_node(*nodes_allowed);
695 VM_BUG_ON(nid >= MAX_NUMNODES);
700 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
702 if (!node_isset(nid, *nodes_allowed))
703 nid = next_node_allowed(nid, nodes_allowed);
708 * returns the previously saved node ["this node"] from which to
709 * allocate a persistent huge page for the pool and advance the
710 * next node from which to allocate, handling wrap at end of node
713 static int hstate_next_node_to_alloc(struct hstate *h,
714 nodemask_t *nodes_allowed)
718 VM_BUG_ON(!nodes_allowed);
720 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
721 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
727 * helper for free_pool_huge_page() - return the previously saved
728 * node ["this node"] from which to free a huge page. Advance the
729 * next node id whether or not we find a free huge page to free so
730 * that the next attempt to free addresses the next node.
732 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
736 VM_BUG_ON(!nodes_allowed);
738 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
739 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
744 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
745 for (nr_nodes = nodes_weight(*mask); \
747 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
750 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
751 for (nr_nodes = nodes_weight(*mask); \
753 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
756 #if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
757 static void destroy_compound_gigantic_page(struct page *page,
761 int nr_pages = 1 << order;
762 struct page *p = page + 1;
764 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
766 set_page_refcounted(p);
767 p->first_page = NULL;
770 set_compound_order(page, 0);
771 __ClearPageHead(page);
774 static void free_gigantic_page(struct page *page, unsigned order)
776 free_contig_range(page_to_pfn(page), 1 << order);
779 static int __alloc_gigantic_page(unsigned long start_pfn,
780 unsigned long nr_pages)
782 unsigned long end_pfn = start_pfn + nr_pages;
783 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
786 static bool pfn_range_valid_gigantic(unsigned long start_pfn,
787 unsigned long nr_pages)
789 unsigned long i, end_pfn = start_pfn + nr_pages;
792 for (i = start_pfn; i < end_pfn; i++) {
796 page = pfn_to_page(i);
798 if (PageReserved(page))
801 if (page_count(page) > 0)
811 static bool zone_spans_last_pfn(const struct zone *zone,
812 unsigned long start_pfn, unsigned long nr_pages)
814 unsigned long last_pfn = start_pfn + nr_pages - 1;
815 return zone_spans_pfn(zone, last_pfn);
818 static struct page *alloc_gigantic_page(int nid, unsigned order)
820 unsigned long nr_pages = 1 << order;
821 unsigned long ret, pfn, flags;
824 z = NODE_DATA(nid)->node_zones;
825 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
826 spin_lock_irqsave(&z->lock, flags);
828 pfn = ALIGN(z->zone_start_pfn, nr_pages);
829 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
830 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
832 * We release the zone lock here because
833 * alloc_contig_range() will also lock the zone
834 * at some point. If there's an allocation
835 * spinning on this lock, it may win the race
836 * and cause alloc_contig_range() to fail...
838 spin_unlock_irqrestore(&z->lock, flags);
839 ret = __alloc_gigantic_page(pfn, nr_pages);
841 return pfn_to_page(pfn);
842 spin_lock_irqsave(&z->lock, flags);
847 spin_unlock_irqrestore(&z->lock, flags);
853 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
854 static void prep_compound_gigantic_page(struct page *page, unsigned long order);
856 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
860 page = alloc_gigantic_page(nid, huge_page_order(h));
862 prep_compound_gigantic_page(page, huge_page_order(h));
863 prep_new_huge_page(h, page, nid);
869 static int alloc_fresh_gigantic_page(struct hstate *h,
870 nodemask_t *nodes_allowed)
872 struct page *page = NULL;
875 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
876 page = alloc_fresh_gigantic_page_node(h, node);
884 static inline bool gigantic_page_supported(void) { return true; }
886 static inline bool gigantic_page_supported(void) { return false; }
887 static inline void free_gigantic_page(struct page *page, unsigned order) { }
888 static inline void destroy_compound_gigantic_page(struct page *page,
889 unsigned long order) { }
890 static inline int alloc_fresh_gigantic_page(struct hstate *h,
891 nodemask_t *nodes_allowed) { return 0; }
894 static void update_and_free_page(struct hstate *h, struct page *page)
898 if (hstate_is_gigantic(h) && !gigantic_page_supported())
902 h->nr_huge_pages_node[page_to_nid(page)]--;
903 for (i = 0; i < pages_per_huge_page(h); i++) {
904 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
905 1 << PG_referenced | 1 << PG_dirty |
906 1 << PG_active | 1 << PG_private |
909 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
910 set_compound_page_dtor(page, NULL);
911 set_page_refcounted(page);
912 if (hstate_is_gigantic(h)) {
913 destroy_compound_gigantic_page(page, huge_page_order(h));
914 free_gigantic_page(page, huge_page_order(h));
916 arch_release_hugepage(page);
917 __free_pages(page, huge_page_order(h));
921 struct hstate *size_to_hstate(unsigned long size)
926 if (huge_page_size(h) == size)
933 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
934 * to hstate->hugepage_activelist.)
936 * This function can be called for tail pages, but never returns true for them.
938 bool page_huge_active(struct page *page)
940 VM_BUG_ON_PAGE(!PageHuge(page), page);
941 return PageHead(page) && PagePrivate(&page[1]);
944 /* never called for tail page */
945 static void set_page_huge_active(struct page *page)
947 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
948 SetPagePrivate(&page[1]);
951 static void clear_page_huge_active(struct page *page)
953 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
954 ClearPagePrivate(&page[1]);
957 void free_huge_page(struct page *page)
960 * Can't pass hstate in here because it is called from the
961 * compound page destructor.
963 struct hstate *h = page_hstate(page);
964 int nid = page_to_nid(page);
965 struct hugepage_subpool *spool =
966 (struct hugepage_subpool *)page_private(page);
967 bool restore_reserve;
969 set_page_private(page, 0);
970 page->mapping = NULL;
971 BUG_ON(page_count(page));
972 BUG_ON(page_mapcount(page));
973 restore_reserve = PagePrivate(page);
974 ClearPagePrivate(page);
977 * A return code of zero implies that the subpool will be under its
978 * minimum size if the reservation is not restored after page is free.
979 * Therefore, force restore_reserve operation.
981 if (hugepage_subpool_put_pages(spool, 1) == 0)
982 restore_reserve = true;
984 spin_lock(&hugetlb_lock);
985 clear_page_huge_active(page);
986 hugetlb_cgroup_uncharge_page(hstate_index(h),
987 pages_per_huge_page(h), page);
989 h->resv_huge_pages++;
991 if (h->surplus_huge_pages_node[nid]) {
992 /* remove the page from active list */
993 list_del(&page->lru);
994 update_and_free_page(h, page);
995 h->surplus_huge_pages--;
996 h->surplus_huge_pages_node[nid]--;
998 arch_clear_hugepage_flags(page);
999 enqueue_huge_page(h, page);
1001 spin_unlock(&hugetlb_lock);
1004 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1006 INIT_LIST_HEAD(&page->lru);
1007 set_compound_page_dtor(page, free_huge_page);
1008 spin_lock(&hugetlb_lock);
1009 set_hugetlb_cgroup(page, NULL);
1011 h->nr_huge_pages_node[nid]++;
1012 spin_unlock(&hugetlb_lock);
1013 put_page(page); /* free it into the hugepage allocator */
1016 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
1019 int nr_pages = 1 << order;
1020 struct page *p = page + 1;
1022 /* we rely on prep_new_huge_page to set the destructor */
1023 set_compound_order(page, order);
1024 __SetPageHead(page);
1025 __ClearPageReserved(page);
1026 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1028 * For gigantic hugepages allocated through bootmem at
1029 * boot, it's safer to be consistent with the not-gigantic
1030 * hugepages and clear the PG_reserved bit from all tail pages
1031 * too. Otherwse drivers using get_user_pages() to access tail
1032 * pages may get the reference counting wrong if they see
1033 * PG_reserved set on a tail page (despite the head page not
1034 * having PG_reserved set). Enforcing this consistency between
1035 * head and tail pages allows drivers to optimize away a check
1036 * on the head page when they need know if put_page() is needed
1037 * after get_user_pages().
1039 __ClearPageReserved(p);
1040 set_page_count(p, 0);
1041 p->first_page = page;
1042 /* Make sure p->first_page is always valid for PageTail() */
1049 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1050 * transparent huge pages. See the PageTransHuge() documentation for more
1053 int PageHuge(struct page *page)
1055 if (!PageCompound(page))
1058 page = compound_head(page);
1059 return get_compound_page_dtor(page) == free_huge_page;
1061 EXPORT_SYMBOL_GPL(PageHuge);
1064 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1065 * normal or transparent huge pages.
1067 int PageHeadHuge(struct page *page_head)
1069 if (!PageHead(page_head))
1072 return get_compound_page_dtor(page_head) == free_huge_page;
1075 pgoff_t __basepage_index(struct page *page)
1077 struct page *page_head = compound_head(page);
1078 pgoff_t index = page_index(page_head);
1079 unsigned long compound_idx;
1081 if (!PageHuge(page_head))
1082 return page_index(page);
1084 if (compound_order(page_head) >= MAX_ORDER)
1085 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1087 compound_idx = page - page_head;
1089 return (index << compound_order(page_head)) + compound_idx;
1092 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1096 page = alloc_pages_exact_node(nid,
1097 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1098 __GFP_REPEAT|__GFP_NOWARN,
1099 huge_page_order(h));
1101 if (arch_prepare_hugepage(page)) {
1102 __free_pages(page, huge_page_order(h));
1105 prep_new_huge_page(h, page, nid);
1111 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1117 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1118 page = alloc_fresh_huge_page_node(h, node);
1126 count_vm_event(HTLB_BUDDY_PGALLOC);
1128 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1134 * Free huge page from pool from next node to free.
1135 * Attempt to keep persistent huge pages more or less
1136 * balanced over allowed nodes.
1137 * Called with hugetlb_lock locked.
1139 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1145 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1147 * If we're returning unused surplus pages, only examine
1148 * nodes with surplus pages.
1150 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1151 !list_empty(&h->hugepage_freelists[node])) {
1153 list_entry(h->hugepage_freelists[node].next,
1155 list_del(&page->lru);
1156 h->free_huge_pages--;
1157 h->free_huge_pages_node[node]--;
1159 h->surplus_huge_pages--;
1160 h->surplus_huge_pages_node[node]--;
1162 update_and_free_page(h, page);
1172 * Dissolve a given free hugepage into free buddy pages. This function does
1173 * nothing for in-use (including surplus) hugepages.
1175 static void dissolve_free_huge_page(struct page *page)
1177 spin_lock(&hugetlb_lock);
1178 if (PageHuge(page) && !page_count(page)) {
1179 struct hstate *h = page_hstate(page);
1180 int nid = page_to_nid(page);
1181 list_del(&page->lru);
1182 h->free_huge_pages--;
1183 h->free_huge_pages_node[nid]--;
1184 update_and_free_page(h, page);
1186 spin_unlock(&hugetlb_lock);
1190 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1191 * make specified memory blocks removable from the system.
1192 * Note that start_pfn should aligned with (minimum) hugepage size.
1194 void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1198 if (!hugepages_supported())
1201 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order));
1202 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
1203 dissolve_free_huge_page(pfn_to_page(pfn));
1206 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
1211 if (hstate_is_gigantic(h))
1215 * Assume we will successfully allocate the surplus page to
1216 * prevent racing processes from causing the surplus to exceed
1219 * This however introduces a different race, where a process B
1220 * tries to grow the static hugepage pool while alloc_pages() is
1221 * called by process A. B will only examine the per-node
1222 * counters in determining if surplus huge pages can be
1223 * converted to normal huge pages in adjust_pool_surplus(). A
1224 * won't be able to increment the per-node counter, until the
1225 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1226 * no more huge pages can be converted from surplus to normal
1227 * state (and doesn't try to convert again). Thus, we have a
1228 * case where a surplus huge page exists, the pool is grown, and
1229 * the surplus huge page still exists after, even though it
1230 * should just have been converted to a normal huge page. This
1231 * does not leak memory, though, as the hugepage will be freed
1232 * once it is out of use. It also does not allow the counters to
1233 * go out of whack in adjust_pool_surplus() as we don't modify
1234 * the node values until we've gotten the hugepage and only the
1235 * per-node value is checked there.
1237 spin_lock(&hugetlb_lock);
1238 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1239 spin_unlock(&hugetlb_lock);
1243 h->surplus_huge_pages++;
1245 spin_unlock(&hugetlb_lock);
1247 if (nid == NUMA_NO_NODE)
1248 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
1249 __GFP_REPEAT|__GFP_NOWARN,
1250 huge_page_order(h));
1252 page = alloc_pages_exact_node(nid,
1253 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1254 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
1256 if (page && arch_prepare_hugepage(page)) {
1257 __free_pages(page, huge_page_order(h));
1261 spin_lock(&hugetlb_lock);
1263 INIT_LIST_HEAD(&page->lru);
1264 r_nid = page_to_nid(page);
1265 set_compound_page_dtor(page, free_huge_page);
1266 set_hugetlb_cgroup(page, NULL);
1268 * We incremented the global counters already
1270 h->nr_huge_pages_node[r_nid]++;
1271 h->surplus_huge_pages_node[r_nid]++;
1272 __count_vm_event(HTLB_BUDDY_PGALLOC);
1275 h->surplus_huge_pages--;
1276 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1278 spin_unlock(&hugetlb_lock);
1284 * This allocation function is useful in the context where vma is irrelevant.
1285 * E.g. soft-offlining uses this function because it only cares physical
1286 * address of error page.
1288 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1290 struct page *page = NULL;
1292 spin_lock(&hugetlb_lock);
1293 if (h->free_huge_pages - h->resv_huge_pages > 0)
1294 page = dequeue_huge_page_node(h, nid);
1295 spin_unlock(&hugetlb_lock);
1298 page = alloc_buddy_huge_page(h, nid);
1304 * Increase the hugetlb pool such that it can accommodate a reservation
1307 static int gather_surplus_pages(struct hstate *h, int delta)
1309 struct list_head surplus_list;
1310 struct page *page, *tmp;
1312 int needed, allocated;
1313 bool alloc_ok = true;
1315 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1317 h->resv_huge_pages += delta;
1322 INIT_LIST_HEAD(&surplus_list);
1326 spin_unlock(&hugetlb_lock);
1327 for (i = 0; i < needed; i++) {
1328 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1333 list_add(&page->lru, &surplus_list);
1338 * After retaking hugetlb_lock, we need to recalculate 'needed'
1339 * because either resv_huge_pages or free_huge_pages may have changed.
1341 spin_lock(&hugetlb_lock);
1342 needed = (h->resv_huge_pages + delta) -
1343 (h->free_huge_pages + allocated);
1348 * We were not able to allocate enough pages to
1349 * satisfy the entire reservation so we free what
1350 * we've allocated so far.
1355 * The surplus_list now contains _at_least_ the number of extra pages
1356 * needed to accommodate the reservation. Add the appropriate number
1357 * of pages to the hugetlb pool and free the extras back to the buddy
1358 * allocator. Commit the entire reservation here to prevent another
1359 * process from stealing the pages as they are added to the pool but
1360 * before they are reserved.
1362 needed += allocated;
1363 h->resv_huge_pages += delta;
1366 /* Free the needed pages to the hugetlb pool */
1367 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1371 * This page is now managed by the hugetlb allocator and has
1372 * no users -- drop the buddy allocator's reference.
1374 put_page_testzero(page);
1375 VM_BUG_ON_PAGE(page_count(page), page);
1376 enqueue_huge_page(h, page);
1379 spin_unlock(&hugetlb_lock);
1381 /* Free unnecessary surplus pages to the buddy allocator */
1382 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1384 spin_lock(&hugetlb_lock);
1390 * When releasing a hugetlb pool reservation, any surplus pages that were
1391 * allocated to satisfy the reservation must be explicitly freed if they were
1393 * Called with hugetlb_lock held.
1395 static void return_unused_surplus_pages(struct hstate *h,
1396 unsigned long unused_resv_pages)
1398 unsigned long nr_pages;
1400 /* Uncommit the reservation */
1401 h->resv_huge_pages -= unused_resv_pages;
1403 /* Cannot return gigantic pages currently */
1404 if (hstate_is_gigantic(h))
1407 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1410 * We want to release as many surplus pages as possible, spread
1411 * evenly across all nodes with memory. Iterate across these nodes
1412 * until we can no longer free unreserved surplus pages. This occurs
1413 * when the nodes with surplus pages have no free pages.
1414 * free_pool_huge_page() will balance the the freed pages across the
1415 * on-line nodes with memory and will handle the hstate accounting.
1417 while (nr_pages--) {
1418 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1420 cond_resched_lock(&hugetlb_lock);
1425 * Determine if the huge page at addr within the vma has an associated
1426 * reservation. Where it does not we will need to logically increase
1427 * reservation and actually increase subpool usage before an allocation
1428 * can occur. Where any new reservation would be required the
1429 * reservation change is prepared, but not committed. Once the page
1430 * has been allocated from the subpool and instantiated the change should
1431 * be committed via vma_commit_reservation. No action is required on
1434 static long vma_needs_reservation(struct hstate *h,
1435 struct vm_area_struct *vma, unsigned long addr)
1437 struct resv_map *resv;
1441 resv = vma_resv_map(vma);
1445 idx = vma_hugecache_offset(h, vma, addr);
1446 chg = region_chg(resv, idx, idx + 1);
1448 if (vma->vm_flags & VM_MAYSHARE)
1451 return chg < 0 ? chg : 0;
1453 static void vma_commit_reservation(struct hstate *h,
1454 struct vm_area_struct *vma, unsigned long addr)
1456 struct resv_map *resv;
1459 resv = vma_resv_map(vma);
1463 idx = vma_hugecache_offset(h, vma, addr);
1464 region_add(resv, idx, idx + 1);
1467 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1468 unsigned long addr, int avoid_reserve)
1470 struct hugepage_subpool *spool = subpool_vma(vma);
1471 struct hstate *h = hstate_vma(vma);
1475 struct hugetlb_cgroup *h_cg;
1477 idx = hstate_index(h);
1479 * Processes that did not create the mapping will have no
1480 * reserves and will not have accounted against subpool
1481 * limit. Check that the subpool limit can be made before
1482 * satisfying the allocation MAP_NORESERVE mappings may also
1483 * need pages and subpool limit allocated allocated if no reserve
1486 chg = vma_needs_reservation(h, vma, addr);
1488 return ERR_PTR(-ENOMEM);
1489 if (chg || avoid_reserve)
1490 if (hugepage_subpool_get_pages(spool, 1) < 0)
1491 return ERR_PTR(-ENOSPC);
1493 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1495 goto out_subpool_put;
1497 spin_lock(&hugetlb_lock);
1498 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
1500 spin_unlock(&hugetlb_lock);
1501 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1503 goto out_uncharge_cgroup;
1505 spin_lock(&hugetlb_lock);
1506 list_move(&page->lru, &h->hugepage_activelist);
1509 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1510 spin_unlock(&hugetlb_lock);
1512 set_page_private(page, (unsigned long)spool);
1514 vma_commit_reservation(h, vma, addr);
1517 out_uncharge_cgroup:
1518 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1520 if (chg || avoid_reserve)
1521 hugepage_subpool_put_pages(spool, 1);
1522 return ERR_PTR(-ENOSPC);
1526 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1527 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1528 * where no ERR_VALUE is expected to be returned.
1530 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1531 unsigned long addr, int avoid_reserve)
1533 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1539 int __weak alloc_bootmem_huge_page(struct hstate *h)
1541 struct huge_bootmem_page *m;
1544 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1547 addr = memblock_virt_alloc_try_nid_nopanic(
1548 huge_page_size(h), huge_page_size(h),
1549 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1552 * Use the beginning of the huge page to store the
1553 * huge_bootmem_page struct (until gather_bootmem
1554 * puts them into the mem_map).
1563 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
1564 /* Put them into a private list first because mem_map is not up yet */
1565 list_add(&m->list, &huge_boot_pages);
1570 static void __init prep_compound_huge_page(struct page *page, int order)
1572 if (unlikely(order > (MAX_ORDER - 1)))
1573 prep_compound_gigantic_page(page, order);
1575 prep_compound_page(page, order);
1578 /* Put bootmem huge pages into the standard lists after mem_map is up */
1579 static void __init gather_bootmem_prealloc(void)
1581 struct huge_bootmem_page *m;
1583 list_for_each_entry(m, &huge_boot_pages, list) {
1584 struct hstate *h = m->hstate;
1587 #ifdef CONFIG_HIGHMEM
1588 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1589 memblock_free_late(__pa(m),
1590 sizeof(struct huge_bootmem_page));
1592 page = virt_to_page(m);
1594 WARN_ON(page_count(page) != 1);
1595 prep_compound_huge_page(page, h->order);
1596 WARN_ON(PageReserved(page));
1597 prep_new_huge_page(h, page, page_to_nid(page));
1599 * If we had gigantic hugepages allocated at boot time, we need
1600 * to restore the 'stolen' pages to totalram_pages in order to
1601 * fix confusing memory reports from free(1) and another
1602 * side-effects, like CommitLimit going negative.
1604 if (hstate_is_gigantic(h))
1605 adjust_managed_page_count(page, 1 << h->order);
1609 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1613 for (i = 0; i < h->max_huge_pages; ++i) {
1614 if (hstate_is_gigantic(h)) {
1615 if (!alloc_bootmem_huge_page(h))
1617 } else if (!alloc_fresh_huge_page(h,
1618 &node_states[N_MEMORY]))
1621 h->max_huge_pages = i;
1624 static void __init hugetlb_init_hstates(void)
1628 for_each_hstate(h) {
1629 if (minimum_order > huge_page_order(h))
1630 minimum_order = huge_page_order(h);
1632 /* oversize hugepages were init'ed in early boot */
1633 if (!hstate_is_gigantic(h))
1634 hugetlb_hstate_alloc_pages(h);
1636 VM_BUG_ON(minimum_order == UINT_MAX);
1639 static char * __init memfmt(char *buf, unsigned long n)
1641 if (n >= (1UL << 30))
1642 sprintf(buf, "%lu GB", n >> 30);
1643 else if (n >= (1UL << 20))
1644 sprintf(buf, "%lu MB", n >> 20);
1646 sprintf(buf, "%lu KB", n >> 10);
1650 static void __init report_hugepages(void)
1654 for_each_hstate(h) {
1656 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1657 memfmt(buf, huge_page_size(h)),
1658 h->free_huge_pages);
1662 #ifdef CONFIG_HIGHMEM
1663 static void try_to_free_low(struct hstate *h, unsigned long count,
1664 nodemask_t *nodes_allowed)
1668 if (hstate_is_gigantic(h))
1671 for_each_node_mask(i, *nodes_allowed) {
1672 struct page *page, *next;
1673 struct list_head *freel = &h->hugepage_freelists[i];
1674 list_for_each_entry_safe(page, next, freel, lru) {
1675 if (count >= h->nr_huge_pages)
1677 if (PageHighMem(page))
1679 list_del(&page->lru);
1680 update_and_free_page(h, page);
1681 h->free_huge_pages--;
1682 h->free_huge_pages_node[page_to_nid(page)]--;
1687 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1688 nodemask_t *nodes_allowed)
1694 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1695 * balanced by operating on them in a round-robin fashion.
1696 * Returns 1 if an adjustment was made.
1698 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1703 VM_BUG_ON(delta != -1 && delta != 1);
1706 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1707 if (h->surplus_huge_pages_node[node])
1711 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1712 if (h->surplus_huge_pages_node[node] <
1713 h->nr_huge_pages_node[node])
1720 h->surplus_huge_pages += delta;
1721 h->surplus_huge_pages_node[node] += delta;
1725 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1726 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1727 nodemask_t *nodes_allowed)
1729 unsigned long min_count, ret;
1731 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1732 return h->max_huge_pages;
1735 * Increase the pool size
1736 * First take pages out of surplus state. Then make up the
1737 * remaining difference by allocating fresh huge pages.
1739 * We might race with alloc_buddy_huge_page() here and be unable
1740 * to convert a surplus huge page to a normal huge page. That is
1741 * not critical, though, it just means the overall size of the
1742 * pool might be one hugepage larger than it needs to be, but
1743 * within all the constraints specified by the sysctls.
1745 spin_lock(&hugetlb_lock);
1746 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1747 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1751 while (count > persistent_huge_pages(h)) {
1753 * If this allocation races such that we no longer need the
1754 * page, free_huge_page will handle it by freeing the page
1755 * and reducing the surplus.
1757 spin_unlock(&hugetlb_lock);
1758 if (hstate_is_gigantic(h))
1759 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
1761 ret = alloc_fresh_huge_page(h, nodes_allowed);
1762 spin_lock(&hugetlb_lock);
1766 /* Bail for signals. Probably ctrl-c from user */
1767 if (signal_pending(current))
1772 * Decrease the pool size
1773 * First return free pages to the buddy allocator (being careful
1774 * to keep enough around to satisfy reservations). Then place
1775 * pages into surplus state as needed so the pool will shrink
1776 * to the desired size as pages become free.
1778 * By placing pages into the surplus state independent of the
1779 * overcommit value, we are allowing the surplus pool size to
1780 * exceed overcommit. There are few sane options here. Since
1781 * alloc_buddy_huge_page() is checking the global counter,
1782 * though, we'll note that we're not allowed to exceed surplus
1783 * and won't grow the pool anywhere else. Not until one of the
1784 * sysctls are changed, or the surplus pages go out of use.
1786 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1787 min_count = max(count, min_count);
1788 try_to_free_low(h, min_count, nodes_allowed);
1789 while (min_count < persistent_huge_pages(h)) {
1790 if (!free_pool_huge_page(h, nodes_allowed, 0))
1792 cond_resched_lock(&hugetlb_lock);
1794 while (count < persistent_huge_pages(h)) {
1795 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1799 ret = persistent_huge_pages(h);
1800 spin_unlock(&hugetlb_lock);
1804 #define HSTATE_ATTR_RO(_name) \
1805 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1807 #define HSTATE_ATTR(_name) \
1808 static struct kobj_attribute _name##_attr = \
1809 __ATTR(_name, 0644, _name##_show, _name##_store)
1811 static struct kobject *hugepages_kobj;
1812 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1814 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1816 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1820 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1821 if (hstate_kobjs[i] == kobj) {
1823 *nidp = NUMA_NO_NODE;
1827 return kobj_to_node_hstate(kobj, nidp);
1830 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1831 struct kobj_attribute *attr, char *buf)
1834 unsigned long nr_huge_pages;
1837 h = kobj_to_hstate(kobj, &nid);
1838 if (nid == NUMA_NO_NODE)
1839 nr_huge_pages = h->nr_huge_pages;
1841 nr_huge_pages = h->nr_huge_pages_node[nid];
1843 return sprintf(buf, "%lu\n", nr_huge_pages);
1846 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
1847 struct hstate *h, int nid,
1848 unsigned long count, size_t len)
1851 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1853 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
1858 if (nid == NUMA_NO_NODE) {
1860 * global hstate attribute
1862 if (!(obey_mempolicy &&
1863 init_nodemask_of_mempolicy(nodes_allowed))) {
1864 NODEMASK_FREE(nodes_allowed);
1865 nodes_allowed = &node_states[N_MEMORY];
1867 } else if (nodes_allowed) {
1869 * per node hstate attribute: adjust count to global,
1870 * but restrict alloc/free to the specified node.
1872 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1873 init_nodemask_of_node(nodes_allowed, nid);
1875 nodes_allowed = &node_states[N_MEMORY];
1877 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1879 if (nodes_allowed != &node_states[N_MEMORY])
1880 NODEMASK_FREE(nodes_allowed);
1884 NODEMASK_FREE(nodes_allowed);
1888 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1889 struct kobject *kobj, const char *buf,
1893 unsigned long count;
1897 err = kstrtoul(buf, 10, &count);
1901 h = kobj_to_hstate(kobj, &nid);
1902 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
1905 static ssize_t nr_hugepages_show(struct kobject *kobj,
1906 struct kobj_attribute *attr, char *buf)
1908 return nr_hugepages_show_common(kobj, attr, buf);
1911 static ssize_t nr_hugepages_store(struct kobject *kobj,
1912 struct kobj_attribute *attr, const char *buf, size_t len)
1914 return nr_hugepages_store_common(false, kobj, buf, len);
1916 HSTATE_ATTR(nr_hugepages);
1921 * hstate attribute for optionally mempolicy-based constraint on persistent
1922 * huge page alloc/free.
1924 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1925 struct kobj_attribute *attr, char *buf)
1927 return nr_hugepages_show_common(kobj, attr, buf);
1930 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1931 struct kobj_attribute *attr, const char *buf, size_t len)
1933 return nr_hugepages_store_common(true, kobj, buf, len);
1935 HSTATE_ATTR(nr_hugepages_mempolicy);
1939 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1940 struct kobj_attribute *attr, char *buf)
1942 struct hstate *h = kobj_to_hstate(kobj, NULL);
1943 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1946 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1947 struct kobj_attribute *attr, const char *buf, size_t count)
1950 unsigned long input;
1951 struct hstate *h = kobj_to_hstate(kobj, NULL);
1953 if (hstate_is_gigantic(h))
1956 err = kstrtoul(buf, 10, &input);
1960 spin_lock(&hugetlb_lock);
1961 h->nr_overcommit_huge_pages = input;
1962 spin_unlock(&hugetlb_lock);
1966 HSTATE_ATTR(nr_overcommit_hugepages);
1968 static ssize_t free_hugepages_show(struct kobject *kobj,
1969 struct kobj_attribute *attr, char *buf)
1972 unsigned long free_huge_pages;
1975 h = kobj_to_hstate(kobj, &nid);
1976 if (nid == NUMA_NO_NODE)
1977 free_huge_pages = h->free_huge_pages;
1979 free_huge_pages = h->free_huge_pages_node[nid];
1981 return sprintf(buf, "%lu\n", free_huge_pages);
1983 HSTATE_ATTR_RO(free_hugepages);
1985 static ssize_t resv_hugepages_show(struct kobject *kobj,
1986 struct kobj_attribute *attr, char *buf)
1988 struct hstate *h = kobj_to_hstate(kobj, NULL);
1989 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1991 HSTATE_ATTR_RO(resv_hugepages);
1993 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1994 struct kobj_attribute *attr, char *buf)
1997 unsigned long surplus_huge_pages;
2000 h = kobj_to_hstate(kobj, &nid);
2001 if (nid == NUMA_NO_NODE)
2002 surplus_huge_pages = h->surplus_huge_pages;
2004 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2006 return sprintf(buf, "%lu\n", surplus_huge_pages);
2008 HSTATE_ATTR_RO(surplus_hugepages);
2010 static struct attribute *hstate_attrs[] = {
2011 &nr_hugepages_attr.attr,
2012 &nr_overcommit_hugepages_attr.attr,
2013 &free_hugepages_attr.attr,
2014 &resv_hugepages_attr.attr,
2015 &surplus_hugepages_attr.attr,
2017 &nr_hugepages_mempolicy_attr.attr,
2022 static struct attribute_group hstate_attr_group = {
2023 .attrs = hstate_attrs,
2026 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2027 struct kobject **hstate_kobjs,
2028 struct attribute_group *hstate_attr_group)
2031 int hi = hstate_index(h);
2033 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2034 if (!hstate_kobjs[hi])
2037 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2039 kobject_put(hstate_kobjs[hi]);
2044 static void __init hugetlb_sysfs_init(void)
2049 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2050 if (!hugepages_kobj)
2053 for_each_hstate(h) {
2054 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2055 hstate_kobjs, &hstate_attr_group);
2057 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2064 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2065 * with node devices in node_devices[] using a parallel array. The array
2066 * index of a node device or _hstate == node id.
2067 * This is here to avoid any static dependency of the node device driver, in
2068 * the base kernel, on the hugetlb module.
2070 struct node_hstate {
2071 struct kobject *hugepages_kobj;
2072 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2074 struct node_hstate node_hstates[MAX_NUMNODES];
2077 * A subset of global hstate attributes for node devices
2079 static struct attribute *per_node_hstate_attrs[] = {
2080 &nr_hugepages_attr.attr,
2081 &free_hugepages_attr.attr,
2082 &surplus_hugepages_attr.attr,
2086 static struct attribute_group per_node_hstate_attr_group = {
2087 .attrs = per_node_hstate_attrs,
2091 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2092 * Returns node id via non-NULL nidp.
2094 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2098 for (nid = 0; nid < nr_node_ids; nid++) {
2099 struct node_hstate *nhs = &node_hstates[nid];
2101 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2102 if (nhs->hstate_kobjs[i] == kobj) {
2114 * Unregister hstate attributes from a single node device.
2115 * No-op if no hstate attributes attached.
2117 static void hugetlb_unregister_node(struct node *node)
2120 struct node_hstate *nhs = &node_hstates[node->dev.id];
2122 if (!nhs->hugepages_kobj)
2123 return; /* no hstate attributes */
2125 for_each_hstate(h) {
2126 int idx = hstate_index(h);
2127 if (nhs->hstate_kobjs[idx]) {
2128 kobject_put(nhs->hstate_kobjs[idx]);
2129 nhs->hstate_kobjs[idx] = NULL;
2133 kobject_put(nhs->hugepages_kobj);
2134 nhs->hugepages_kobj = NULL;
2138 * hugetlb module exit: unregister hstate attributes from node devices
2141 static void hugetlb_unregister_all_nodes(void)
2146 * disable node device registrations.
2148 register_hugetlbfs_with_node(NULL, NULL);
2151 * remove hstate attributes from any nodes that have them.
2153 for (nid = 0; nid < nr_node_ids; nid++)
2154 hugetlb_unregister_node(node_devices[nid]);
2158 * Register hstate attributes for a single node device.
2159 * No-op if attributes already registered.
2161 static void hugetlb_register_node(struct node *node)
2164 struct node_hstate *nhs = &node_hstates[node->dev.id];
2167 if (nhs->hugepages_kobj)
2168 return; /* already allocated */
2170 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2172 if (!nhs->hugepages_kobj)
2175 for_each_hstate(h) {
2176 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2178 &per_node_hstate_attr_group);
2180 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2181 h->name, node->dev.id);
2182 hugetlb_unregister_node(node);
2189 * hugetlb init time: register hstate attributes for all registered node
2190 * devices of nodes that have memory. All on-line nodes should have
2191 * registered their associated device by this time.
2193 static void __init hugetlb_register_all_nodes(void)
2197 for_each_node_state(nid, N_MEMORY) {
2198 struct node *node = node_devices[nid];
2199 if (node->dev.id == nid)
2200 hugetlb_register_node(node);
2204 * Let the node device driver know we're here so it can
2205 * [un]register hstate attributes on node hotplug.
2207 register_hugetlbfs_with_node(hugetlb_register_node,
2208 hugetlb_unregister_node);
2210 #else /* !CONFIG_NUMA */
2212 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2220 static void hugetlb_unregister_all_nodes(void) { }
2222 static void hugetlb_register_all_nodes(void) { }
2226 static void __exit hugetlb_exit(void)
2230 hugetlb_unregister_all_nodes();
2232 for_each_hstate(h) {
2233 kobject_put(hstate_kobjs[hstate_index(h)]);
2236 kobject_put(hugepages_kobj);
2237 kfree(htlb_fault_mutex_table);
2239 module_exit(hugetlb_exit);
2241 static int __init hugetlb_init(void)
2245 if (!hugepages_supported())
2248 if (!size_to_hstate(default_hstate_size)) {
2249 default_hstate_size = HPAGE_SIZE;
2250 if (!size_to_hstate(default_hstate_size))
2251 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2253 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2254 if (default_hstate_max_huge_pages)
2255 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2257 hugetlb_init_hstates();
2258 gather_bootmem_prealloc();
2261 hugetlb_sysfs_init();
2262 hugetlb_register_all_nodes();
2263 hugetlb_cgroup_file_init();
2266 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2268 num_fault_mutexes = 1;
2270 htlb_fault_mutex_table =
2271 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2272 BUG_ON(!htlb_fault_mutex_table);
2274 for (i = 0; i < num_fault_mutexes; i++)
2275 mutex_init(&htlb_fault_mutex_table[i]);
2278 module_init(hugetlb_init);
2280 /* Should be called on processing a hugepagesz=... option */
2281 void __init hugetlb_add_hstate(unsigned order)
2286 if (size_to_hstate(PAGE_SIZE << order)) {
2287 pr_warning("hugepagesz= specified twice, ignoring\n");
2290 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2292 h = &hstates[hugetlb_max_hstate++];
2294 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2295 h->nr_huge_pages = 0;
2296 h->free_huge_pages = 0;
2297 for (i = 0; i < MAX_NUMNODES; ++i)
2298 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2299 INIT_LIST_HEAD(&h->hugepage_activelist);
2300 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2301 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2302 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2303 huge_page_size(h)/1024);
2308 static int __init hugetlb_nrpages_setup(char *s)
2311 static unsigned long *last_mhp;
2314 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2315 * so this hugepages= parameter goes to the "default hstate".
2317 if (!hugetlb_max_hstate)
2318 mhp = &default_hstate_max_huge_pages;
2320 mhp = &parsed_hstate->max_huge_pages;
2322 if (mhp == last_mhp) {
2323 pr_warning("hugepages= specified twice without "
2324 "interleaving hugepagesz=, ignoring\n");
2328 if (sscanf(s, "%lu", mhp) <= 0)
2332 * Global state is always initialized later in hugetlb_init.
2333 * But we need to allocate >= MAX_ORDER hstates here early to still
2334 * use the bootmem allocator.
2336 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2337 hugetlb_hstate_alloc_pages(parsed_hstate);
2343 __setup("hugepages=", hugetlb_nrpages_setup);
2345 static int __init hugetlb_default_setup(char *s)
2347 default_hstate_size = memparse(s, &s);
2350 __setup("default_hugepagesz=", hugetlb_default_setup);
2352 static unsigned int cpuset_mems_nr(unsigned int *array)
2355 unsigned int nr = 0;
2357 for_each_node_mask(node, cpuset_current_mems_allowed)
2363 #ifdef CONFIG_SYSCTL
2364 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2365 struct ctl_table *table, int write,
2366 void __user *buffer, size_t *length, loff_t *ppos)
2368 struct hstate *h = &default_hstate;
2369 unsigned long tmp = h->max_huge_pages;
2372 if (!hugepages_supported())
2376 table->maxlen = sizeof(unsigned long);
2377 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2382 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2383 NUMA_NO_NODE, tmp, *length);
2388 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2389 void __user *buffer, size_t *length, loff_t *ppos)
2392 return hugetlb_sysctl_handler_common(false, table, write,
2393 buffer, length, ppos);
2397 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2398 void __user *buffer, size_t *length, loff_t *ppos)
2400 return hugetlb_sysctl_handler_common(true, table, write,
2401 buffer, length, ppos);
2403 #endif /* CONFIG_NUMA */
2405 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2406 void __user *buffer,
2407 size_t *length, loff_t *ppos)
2409 struct hstate *h = &default_hstate;
2413 if (!hugepages_supported())
2416 tmp = h->nr_overcommit_huge_pages;
2418 if (write && hstate_is_gigantic(h))
2422 table->maxlen = sizeof(unsigned long);
2423 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2428 spin_lock(&hugetlb_lock);
2429 h->nr_overcommit_huge_pages = tmp;
2430 spin_unlock(&hugetlb_lock);
2436 #endif /* CONFIG_SYSCTL */
2438 void hugetlb_report_meminfo(struct seq_file *m)
2440 struct hstate *h = &default_hstate;
2441 if (!hugepages_supported())
2444 "HugePages_Total: %5lu\n"
2445 "HugePages_Free: %5lu\n"
2446 "HugePages_Rsvd: %5lu\n"
2447 "HugePages_Surp: %5lu\n"
2448 "Hugepagesize: %8lu kB\n",
2452 h->surplus_huge_pages,
2453 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2456 int hugetlb_report_node_meminfo(int nid, char *buf)
2458 struct hstate *h = &default_hstate;
2459 if (!hugepages_supported())
2462 "Node %d HugePages_Total: %5u\n"
2463 "Node %d HugePages_Free: %5u\n"
2464 "Node %d HugePages_Surp: %5u\n",
2465 nid, h->nr_huge_pages_node[nid],
2466 nid, h->free_huge_pages_node[nid],
2467 nid, h->surplus_huge_pages_node[nid]);
2470 void hugetlb_show_meminfo(void)
2475 if (!hugepages_supported())
2478 for_each_node_state(nid, N_MEMORY)
2480 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2482 h->nr_huge_pages_node[nid],
2483 h->free_huge_pages_node[nid],
2484 h->surplus_huge_pages_node[nid],
2485 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2488 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2489 unsigned long hugetlb_total_pages(void)
2492 unsigned long nr_total_pages = 0;
2495 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2496 return nr_total_pages;
2499 static int hugetlb_acct_memory(struct hstate *h, long delta)
2503 spin_lock(&hugetlb_lock);
2505 * When cpuset is configured, it breaks the strict hugetlb page
2506 * reservation as the accounting is done on a global variable. Such
2507 * reservation is completely rubbish in the presence of cpuset because
2508 * the reservation is not checked against page availability for the
2509 * current cpuset. Application can still potentially OOM'ed by kernel
2510 * with lack of free htlb page in cpuset that the task is in.
2511 * Attempt to enforce strict accounting with cpuset is almost
2512 * impossible (or too ugly) because cpuset is too fluid that
2513 * task or memory node can be dynamically moved between cpusets.
2515 * The change of semantics for shared hugetlb mapping with cpuset is
2516 * undesirable. However, in order to preserve some of the semantics,
2517 * we fall back to check against current free page availability as
2518 * a best attempt and hopefully to minimize the impact of changing
2519 * semantics that cpuset has.
2522 if (gather_surplus_pages(h, delta) < 0)
2525 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2526 return_unused_surplus_pages(h, delta);
2533 return_unused_surplus_pages(h, (unsigned long) -delta);
2536 spin_unlock(&hugetlb_lock);
2540 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2542 struct resv_map *resv = vma_resv_map(vma);
2545 * This new VMA should share its siblings reservation map if present.
2546 * The VMA will only ever have a valid reservation map pointer where
2547 * it is being copied for another still existing VMA. As that VMA
2548 * has a reference to the reservation map it cannot disappear until
2549 * after this open call completes. It is therefore safe to take a
2550 * new reference here without additional locking.
2552 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2553 kref_get(&resv->refs);
2556 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2558 struct hstate *h = hstate_vma(vma);
2559 struct resv_map *resv = vma_resv_map(vma);
2560 struct hugepage_subpool *spool = subpool_vma(vma);
2561 unsigned long reserve, start, end;
2564 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2567 start = vma_hugecache_offset(h, vma, vma->vm_start);
2568 end = vma_hugecache_offset(h, vma, vma->vm_end);
2570 reserve = (end - start) - region_count(resv, start, end);
2572 kref_put(&resv->refs, resv_map_release);
2576 * Decrement reserve counts. The global reserve count may be
2577 * adjusted if the subpool has a minimum size.
2579 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
2580 hugetlb_acct_memory(h, -gbl_reserve);
2585 * We cannot handle pagefaults against hugetlb pages at all. They cause
2586 * handle_mm_fault() to try to instantiate regular-sized pages in the
2587 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2590 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2596 const struct vm_operations_struct hugetlb_vm_ops = {
2597 .fault = hugetlb_vm_op_fault,
2598 .open = hugetlb_vm_op_open,
2599 .close = hugetlb_vm_op_close,
2602 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2608 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2609 vma->vm_page_prot)));
2611 entry = huge_pte_wrprotect(mk_huge_pte(page,
2612 vma->vm_page_prot));
2614 entry = pte_mkyoung(entry);
2615 entry = pte_mkhuge(entry);
2616 entry = arch_make_huge_pte(entry, vma, page, writable);
2621 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2622 unsigned long address, pte_t *ptep)
2626 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2627 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2628 update_mmu_cache(vma, address, ptep);
2631 static int is_hugetlb_entry_migration(pte_t pte)
2635 if (huge_pte_none(pte) || pte_present(pte))
2637 swp = pte_to_swp_entry(pte);
2638 if (non_swap_entry(swp) && is_migration_entry(swp))
2644 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2648 if (huge_pte_none(pte) || pte_present(pte))
2650 swp = pte_to_swp_entry(pte);
2651 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2657 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2658 struct vm_area_struct *vma)
2660 pte_t *src_pte, *dst_pte, entry;
2661 struct page *ptepage;
2664 struct hstate *h = hstate_vma(vma);
2665 unsigned long sz = huge_page_size(h);
2666 unsigned long mmun_start; /* For mmu_notifiers */
2667 unsigned long mmun_end; /* For mmu_notifiers */
2670 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2672 mmun_start = vma->vm_start;
2673 mmun_end = vma->vm_end;
2675 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2677 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2678 spinlock_t *src_ptl, *dst_ptl;
2679 src_pte = huge_pte_offset(src, addr);
2682 dst_pte = huge_pte_alloc(dst, addr, sz);
2688 /* If the pagetables are shared don't copy or take references */
2689 if (dst_pte == src_pte)
2692 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2693 src_ptl = huge_pte_lockptr(h, src, src_pte);
2694 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
2695 entry = huge_ptep_get(src_pte);
2696 if (huge_pte_none(entry)) { /* skip none entry */
2698 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2699 is_hugetlb_entry_hwpoisoned(entry))) {
2700 swp_entry_t swp_entry = pte_to_swp_entry(entry);
2702 if (is_write_migration_entry(swp_entry) && cow) {
2704 * COW mappings require pages in both
2705 * parent and child to be set to read.
2707 make_migration_entry_read(&swp_entry);
2708 entry = swp_entry_to_pte(swp_entry);
2709 set_huge_pte_at(src, addr, src_pte, entry);
2711 set_huge_pte_at(dst, addr, dst_pte, entry);
2714 huge_ptep_set_wrprotect(src, addr, src_pte);
2715 mmu_notifier_invalidate_range(src, mmun_start,
2718 entry = huge_ptep_get(src_pte);
2719 ptepage = pte_page(entry);
2721 page_dup_rmap(ptepage);
2722 set_huge_pte_at(dst, addr, dst_pte, entry);
2724 spin_unlock(src_ptl);
2725 spin_unlock(dst_ptl);
2729 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
2734 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2735 unsigned long start, unsigned long end,
2736 struct page *ref_page)
2738 int force_flush = 0;
2739 struct mm_struct *mm = vma->vm_mm;
2740 unsigned long address;
2745 struct hstate *h = hstate_vma(vma);
2746 unsigned long sz = huge_page_size(h);
2747 const unsigned long mmun_start = start; /* For mmu_notifiers */
2748 const unsigned long mmun_end = end; /* For mmu_notifiers */
2750 WARN_ON(!is_vm_hugetlb_page(vma));
2751 BUG_ON(start & ~huge_page_mask(h));
2752 BUG_ON(end & ~huge_page_mask(h));
2754 tlb_start_vma(tlb, vma);
2755 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2758 for (; address < end; address += sz) {
2759 ptep = huge_pte_offset(mm, address);
2763 ptl = huge_pte_lock(h, mm, ptep);
2764 if (huge_pmd_unshare(mm, &address, ptep))
2767 pte = huge_ptep_get(ptep);
2768 if (huge_pte_none(pte))
2772 * Migrating hugepage or HWPoisoned hugepage is already
2773 * unmapped and its refcount is dropped, so just clear pte here.
2775 if (unlikely(!pte_present(pte))) {
2776 huge_pte_clear(mm, address, ptep);
2780 page = pte_page(pte);
2782 * If a reference page is supplied, it is because a specific
2783 * page is being unmapped, not a range. Ensure the page we
2784 * are about to unmap is the actual page of interest.
2787 if (page != ref_page)
2791 * Mark the VMA as having unmapped its page so that
2792 * future faults in this VMA will fail rather than
2793 * looking like data was lost
2795 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2798 pte = huge_ptep_get_and_clear(mm, address, ptep);
2799 tlb_remove_tlb_entry(tlb, ptep, address);
2800 if (huge_pte_dirty(pte))
2801 set_page_dirty(page);
2803 page_remove_rmap(page);
2804 force_flush = !__tlb_remove_page(tlb, page);
2810 /* Bail out after unmapping reference page if supplied */
2819 * mmu_gather ran out of room to batch pages, we break out of
2820 * the PTE lock to avoid doing the potential expensive TLB invalidate
2821 * and page-free while holding it.
2826 if (address < end && !ref_page)
2829 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2830 tlb_end_vma(tlb, vma);
2833 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2834 struct vm_area_struct *vma, unsigned long start,
2835 unsigned long end, struct page *ref_page)
2837 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2840 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2841 * test will fail on a vma being torn down, and not grab a page table
2842 * on its way out. We're lucky that the flag has such an appropriate
2843 * name, and can in fact be safely cleared here. We could clear it
2844 * before the __unmap_hugepage_range above, but all that's necessary
2845 * is to clear it before releasing the i_mmap_rwsem. This works
2846 * because in the context this is called, the VMA is about to be
2847 * destroyed and the i_mmap_rwsem is held.
2849 vma->vm_flags &= ~VM_MAYSHARE;
2852 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2853 unsigned long end, struct page *ref_page)
2855 struct mm_struct *mm;
2856 struct mmu_gather tlb;
2860 tlb_gather_mmu(&tlb, mm, start, end);
2861 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2862 tlb_finish_mmu(&tlb, start, end);
2866 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2867 * mappping it owns the reserve page for. The intention is to unmap the page
2868 * from other VMAs and let the children be SIGKILLed if they are faulting the
2871 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2872 struct page *page, unsigned long address)
2874 struct hstate *h = hstate_vma(vma);
2875 struct vm_area_struct *iter_vma;
2876 struct address_space *mapping;
2880 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2881 * from page cache lookup which is in HPAGE_SIZE units.
2883 address = address & huge_page_mask(h);
2884 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2886 mapping = file_inode(vma->vm_file)->i_mapping;
2889 * Take the mapping lock for the duration of the table walk. As
2890 * this mapping should be shared between all the VMAs,
2891 * __unmap_hugepage_range() is called as the lock is already held
2893 i_mmap_lock_write(mapping);
2894 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2895 /* Do not unmap the current VMA */
2896 if (iter_vma == vma)
2900 * Unmap the page from other VMAs without their own reserves.
2901 * They get marked to be SIGKILLed if they fault in these
2902 * areas. This is because a future no-page fault on this VMA
2903 * could insert a zeroed page instead of the data existing
2904 * from the time of fork. This would look like data corruption
2906 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2907 unmap_hugepage_range(iter_vma, address,
2908 address + huge_page_size(h), page);
2910 i_mmap_unlock_write(mapping);
2914 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2915 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2916 * cannot race with other handlers or page migration.
2917 * Keep the pte_same checks anyway to make transition from the mutex easier.
2919 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2920 unsigned long address, pte_t *ptep, pte_t pte,
2921 struct page *pagecache_page, spinlock_t *ptl)
2923 struct hstate *h = hstate_vma(vma);
2924 struct page *old_page, *new_page;
2925 int ret = 0, outside_reserve = 0;
2926 unsigned long mmun_start; /* For mmu_notifiers */
2927 unsigned long mmun_end; /* For mmu_notifiers */
2929 old_page = pte_page(pte);
2932 /* If no-one else is actually using this page, avoid the copy
2933 * and just make the page writable */
2934 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2935 page_move_anon_rmap(old_page, vma, address);
2936 set_huge_ptep_writable(vma, address, ptep);
2941 * If the process that created a MAP_PRIVATE mapping is about to
2942 * perform a COW due to a shared page count, attempt to satisfy
2943 * the allocation without using the existing reserves. The pagecache
2944 * page is used to determine if the reserve at this address was
2945 * consumed or not. If reserves were used, a partial faulted mapping
2946 * at the time of fork() could consume its reserves on COW instead
2947 * of the full address range.
2949 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2950 old_page != pagecache_page)
2951 outside_reserve = 1;
2953 page_cache_get(old_page);
2956 * Drop page table lock as buddy allocator may be called. It will
2957 * be acquired again before returning to the caller, as expected.
2960 new_page = alloc_huge_page(vma, address, outside_reserve);
2962 if (IS_ERR(new_page)) {
2964 * If a process owning a MAP_PRIVATE mapping fails to COW,
2965 * it is due to references held by a child and an insufficient
2966 * huge page pool. To guarantee the original mappers
2967 * reliability, unmap the page from child processes. The child
2968 * may get SIGKILLed if it later faults.
2970 if (outside_reserve) {
2971 page_cache_release(old_page);
2972 BUG_ON(huge_pte_none(pte));
2973 unmap_ref_private(mm, vma, old_page, address);
2974 BUG_ON(huge_pte_none(pte));
2976 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2978 pte_same(huge_ptep_get(ptep), pte)))
2979 goto retry_avoidcopy;
2981 * race occurs while re-acquiring page table
2982 * lock, and our job is done.
2987 ret = (PTR_ERR(new_page) == -ENOMEM) ?
2988 VM_FAULT_OOM : VM_FAULT_SIGBUS;
2989 goto out_release_old;
2993 * When the original hugepage is shared one, it does not have
2994 * anon_vma prepared.
2996 if (unlikely(anon_vma_prepare(vma))) {
2998 goto out_release_all;
3001 copy_user_huge_page(new_page, old_page, address, vma,
3002 pages_per_huge_page(h));
3003 __SetPageUptodate(new_page);
3004 set_page_huge_active(new_page);
3006 mmun_start = address & huge_page_mask(h);
3007 mmun_end = mmun_start + huge_page_size(h);
3008 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3011 * Retake the page table lock to check for racing updates
3012 * before the page tables are altered
3015 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3016 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3017 ClearPagePrivate(new_page);
3020 huge_ptep_clear_flush(vma, address, ptep);
3021 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3022 set_huge_pte_at(mm, address, ptep,
3023 make_huge_pte(vma, new_page, 1));
3024 page_remove_rmap(old_page);
3025 hugepage_add_new_anon_rmap(new_page, vma, address);
3026 /* Make the old page be freed below */
3027 new_page = old_page;
3030 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3032 page_cache_release(new_page);
3034 page_cache_release(old_page);
3036 spin_lock(ptl); /* Caller expects lock to be held */
3040 /* Return the pagecache page at a given address within a VMA */
3041 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3042 struct vm_area_struct *vma, unsigned long address)
3044 struct address_space *mapping;
3047 mapping = vma->vm_file->f_mapping;
3048 idx = vma_hugecache_offset(h, vma, address);
3050 return find_lock_page(mapping, idx);
3054 * Return whether there is a pagecache page to back given address within VMA.
3055 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3057 static bool hugetlbfs_pagecache_present(struct hstate *h,
3058 struct vm_area_struct *vma, unsigned long address)
3060 struct address_space *mapping;
3064 mapping = vma->vm_file->f_mapping;
3065 idx = vma_hugecache_offset(h, vma, address);
3067 page = find_get_page(mapping, idx);
3070 return page != NULL;
3073 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3074 struct address_space *mapping, pgoff_t idx,
3075 unsigned long address, pte_t *ptep, unsigned int flags)
3077 struct hstate *h = hstate_vma(vma);
3078 int ret = VM_FAULT_SIGBUS;
3086 * Currently, we are forced to kill the process in the event the
3087 * original mapper has unmapped pages from the child due to a failed
3088 * COW. Warn that such a situation has occurred as it may not be obvious
3090 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3091 pr_warning("PID %d killed due to inadequate hugepage pool\n",
3097 * Use page lock to guard against racing truncation
3098 * before we get page_table_lock.
3101 page = find_lock_page(mapping, idx);
3103 size = i_size_read(mapping->host) >> huge_page_shift(h);
3106 page = alloc_huge_page(vma, address, 0);
3108 ret = PTR_ERR(page);
3112 ret = VM_FAULT_SIGBUS;
3115 clear_huge_page(page, address, pages_per_huge_page(h));
3116 __SetPageUptodate(page);
3117 set_page_huge_active(page);
3119 if (vma->vm_flags & VM_MAYSHARE) {
3121 struct inode *inode = mapping->host;
3123 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3130 ClearPagePrivate(page);
3132 spin_lock(&inode->i_lock);
3133 inode->i_blocks += blocks_per_huge_page(h);
3134 spin_unlock(&inode->i_lock);
3137 if (unlikely(anon_vma_prepare(vma))) {
3139 goto backout_unlocked;
3145 * If memory error occurs between mmap() and fault, some process
3146 * don't have hwpoisoned swap entry for errored virtual address.
3147 * So we need to block hugepage fault by PG_hwpoison bit check.
3149 if (unlikely(PageHWPoison(page))) {
3150 ret = VM_FAULT_HWPOISON |
3151 VM_FAULT_SET_HINDEX(hstate_index(h));
3152 goto backout_unlocked;
3157 * If we are going to COW a private mapping later, we examine the
3158 * pending reservations for this page now. This will ensure that
3159 * any allocations necessary to record that reservation occur outside
3162 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
3163 if (vma_needs_reservation(h, vma, address) < 0) {
3165 goto backout_unlocked;
3168 ptl = huge_pte_lockptr(h, mm, ptep);
3170 size = i_size_read(mapping->host) >> huge_page_shift(h);
3175 if (!huge_pte_none(huge_ptep_get(ptep)))
3179 ClearPagePrivate(page);
3180 hugepage_add_new_anon_rmap(page, vma, address);
3182 page_dup_rmap(page);
3183 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3184 && (vma->vm_flags & VM_SHARED)));
3185 set_huge_pte_at(mm, address, ptep, new_pte);
3187 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3188 /* Optimization, do the COW without a second fault */
3189 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
3206 static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3207 struct vm_area_struct *vma,
3208 struct address_space *mapping,
3209 pgoff_t idx, unsigned long address)
3211 unsigned long key[2];
3214 if (vma->vm_flags & VM_SHARED) {
3215 key[0] = (unsigned long) mapping;
3218 key[0] = (unsigned long) mm;
3219 key[1] = address >> huge_page_shift(h);
3222 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3224 return hash & (num_fault_mutexes - 1);
3228 * For uniprocesor systems we always use a single mutex, so just
3229 * return 0 and avoid the hashing overhead.
3231 static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3232 struct vm_area_struct *vma,
3233 struct address_space *mapping,
3234 pgoff_t idx, unsigned long address)
3240 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3241 unsigned long address, unsigned int flags)
3248 struct page *page = NULL;
3249 struct page *pagecache_page = NULL;
3250 struct hstate *h = hstate_vma(vma);
3251 struct address_space *mapping;
3252 int need_wait_lock = 0;
3254 address &= huge_page_mask(h);
3256 ptep = huge_pte_offset(mm, address);
3258 entry = huge_ptep_get(ptep);
3259 if (unlikely(is_hugetlb_entry_migration(entry))) {
3260 migration_entry_wait_huge(vma, mm, ptep);
3262 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3263 return VM_FAULT_HWPOISON_LARGE |
3264 VM_FAULT_SET_HINDEX(hstate_index(h));
3267 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3269 return VM_FAULT_OOM;
3271 mapping = vma->vm_file->f_mapping;
3272 idx = vma_hugecache_offset(h, vma, address);
3275 * Serialize hugepage allocation and instantiation, so that we don't
3276 * get spurious allocation failures if two CPUs race to instantiate
3277 * the same page in the page cache.
3279 hash = fault_mutex_hash(h, mm, vma, mapping, idx, address);
3280 mutex_lock(&htlb_fault_mutex_table[hash]);
3282 entry = huge_ptep_get(ptep);
3283 if (huge_pte_none(entry)) {
3284 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3291 * entry could be a migration/hwpoison entry at this point, so this
3292 * check prevents the kernel from going below assuming that we have
3293 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3294 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3297 if (!pte_present(entry))
3301 * If we are going to COW the mapping later, we examine the pending
3302 * reservations for this page now. This will ensure that any
3303 * allocations necessary to record that reservation occur outside the
3304 * spinlock. For private mappings, we also lookup the pagecache
3305 * page now as it is used to determine if a reservation has been
3308 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3309 if (vma_needs_reservation(h, vma, address) < 0) {
3314 if (!(vma->vm_flags & VM_MAYSHARE))
3315 pagecache_page = hugetlbfs_pagecache_page(h,
3319 ptl = huge_pte_lock(h, mm, ptep);
3321 /* Check for a racing update before calling hugetlb_cow */
3322 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3326 * hugetlb_cow() requires page locks of pte_page(entry) and
3327 * pagecache_page, so here we need take the former one
3328 * when page != pagecache_page or !pagecache_page.
3330 page = pte_page(entry);
3331 if (page != pagecache_page)
3332 if (!trylock_page(page)) {
3339 if (flags & FAULT_FLAG_WRITE) {
3340 if (!huge_pte_write(entry)) {
3341 ret = hugetlb_cow(mm, vma, address, ptep, entry,
3342 pagecache_page, ptl);
3345 entry = huge_pte_mkdirty(entry);
3347 entry = pte_mkyoung(entry);
3348 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3349 flags & FAULT_FLAG_WRITE))
3350 update_mmu_cache(vma, address, ptep);
3352 if (page != pagecache_page)
3358 if (pagecache_page) {
3359 unlock_page(pagecache_page);
3360 put_page(pagecache_page);
3363 mutex_unlock(&htlb_fault_mutex_table[hash]);
3365 * Generally it's safe to hold refcount during waiting page lock. But
3366 * here we just wait to defer the next page fault to avoid busy loop and
3367 * the page is not used after unlocked before returning from the current
3368 * page fault. So we are safe from accessing freed page, even if we wait
3369 * here without taking refcount.
3372 wait_on_page_locked(page);
3376 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3377 struct page **pages, struct vm_area_struct **vmas,
3378 unsigned long *position, unsigned long *nr_pages,
3379 long i, unsigned int flags)
3381 unsigned long pfn_offset;
3382 unsigned long vaddr = *position;
3383 unsigned long remainder = *nr_pages;
3384 struct hstate *h = hstate_vma(vma);
3386 while (vaddr < vma->vm_end && remainder) {
3388 spinlock_t *ptl = NULL;
3393 * If we have a pending SIGKILL, don't keep faulting pages and
3394 * potentially allocating memory.
3396 if (unlikely(fatal_signal_pending(current))) {
3402 * Some archs (sparc64, sh*) have multiple pte_ts to
3403 * each hugepage. We have to make sure we get the
3404 * first, for the page indexing below to work.
3406 * Note that page table lock is not held when pte is null.
3408 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3410 ptl = huge_pte_lock(h, mm, pte);
3411 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3414 * When coredumping, it suits get_dump_page if we just return
3415 * an error where there's an empty slot with no huge pagecache
3416 * to back it. This way, we avoid allocating a hugepage, and
3417 * the sparse dumpfile avoids allocating disk blocks, but its
3418 * huge holes still show up with zeroes where they need to be.
3420 if (absent && (flags & FOLL_DUMP) &&
3421 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3429 * We need call hugetlb_fault for both hugepages under migration
3430 * (in which case hugetlb_fault waits for the migration,) and
3431 * hwpoisoned hugepages (in which case we need to prevent the
3432 * caller from accessing to them.) In order to do this, we use
3433 * here is_swap_pte instead of is_hugetlb_entry_migration and
3434 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3435 * both cases, and because we can't follow correct pages
3436 * directly from any kind of swap entries.
3438 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3439 ((flags & FOLL_WRITE) &&
3440 !huge_pte_write(huge_ptep_get(pte)))) {
3445 ret = hugetlb_fault(mm, vma, vaddr,
3446 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3447 if (!(ret & VM_FAULT_ERROR))
3454 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3455 page = pte_page(huge_ptep_get(pte));
3458 pages[i] = mem_map_offset(page, pfn_offset);
3459 get_page_foll(pages[i]);
3469 if (vaddr < vma->vm_end && remainder &&
3470 pfn_offset < pages_per_huge_page(h)) {
3472 * We use pfn_offset to avoid touching the pageframes
3473 * of this compound page.
3479 *nr_pages = remainder;
3482 return i ? i : -EFAULT;
3485 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3486 unsigned long address, unsigned long end, pgprot_t newprot)
3488 struct mm_struct *mm = vma->vm_mm;
3489 unsigned long start = address;
3492 struct hstate *h = hstate_vma(vma);
3493 unsigned long pages = 0;
3495 BUG_ON(address >= end);
3496 flush_cache_range(vma, address, end);
3498 mmu_notifier_invalidate_range_start(mm, start, end);
3499 i_mmap_lock_write(vma->vm_file->f_mapping);
3500 for (; address < end; address += huge_page_size(h)) {
3502 ptep = huge_pte_offset(mm, address);
3505 ptl = huge_pte_lock(h, mm, ptep);
3506 if (huge_pmd_unshare(mm, &address, ptep)) {
3511 pte = huge_ptep_get(ptep);
3512 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3516 if (unlikely(is_hugetlb_entry_migration(pte))) {
3517 swp_entry_t entry = pte_to_swp_entry(pte);
3519 if (is_write_migration_entry(entry)) {
3522 make_migration_entry_read(&entry);
3523 newpte = swp_entry_to_pte(entry);
3524 set_huge_pte_at(mm, address, ptep, newpte);
3530 if (!huge_pte_none(pte)) {
3531 pte = huge_ptep_get_and_clear(mm, address, ptep);
3532 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3533 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3534 set_huge_pte_at(mm, address, ptep, pte);
3540 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
3541 * may have cleared our pud entry and done put_page on the page table:
3542 * once we release i_mmap_rwsem, another task can do the final put_page
3543 * and that page table be reused and filled with junk.
3545 flush_tlb_range(vma, start, end);
3546 mmu_notifier_invalidate_range(mm, start, end);
3547 i_mmap_unlock_write(vma->vm_file->f_mapping);
3548 mmu_notifier_invalidate_range_end(mm, start, end);
3550 return pages << h->order;
3553 int hugetlb_reserve_pages(struct inode *inode,
3555 struct vm_area_struct *vma,
3556 vm_flags_t vm_flags)
3559 struct hstate *h = hstate_inode(inode);
3560 struct hugepage_subpool *spool = subpool_inode(inode);
3561 struct resv_map *resv_map;
3565 * Only apply hugepage reservation if asked. At fault time, an
3566 * attempt will be made for VM_NORESERVE to allocate a page
3567 * without using reserves
3569 if (vm_flags & VM_NORESERVE)
3573 * Shared mappings base their reservation on the number of pages that
3574 * are already allocated on behalf of the file. Private mappings need
3575 * to reserve the full area even if read-only as mprotect() may be
3576 * called to make the mapping read-write. Assume !vma is a shm mapping
3578 if (!vma || vma->vm_flags & VM_MAYSHARE) {
3579 resv_map = inode_resv_map(inode);
3581 chg = region_chg(resv_map, from, to);
3584 resv_map = resv_map_alloc();
3590 set_vma_resv_map(vma, resv_map);
3591 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3600 * There must be enough pages in the subpool for the mapping. If
3601 * the subpool has a minimum size, there may be some global
3602 * reservations already in place (gbl_reserve).
3604 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
3605 if (gbl_reserve < 0) {
3611 * Check enough hugepages are available for the reservation.
3612 * Hand the pages back to the subpool if there are not
3614 ret = hugetlb_acct_memory(h, gbl_reserve);
3616 /* put back original number of pages, chg */
3617 (void)hugepage_subpool_put_pages(spool, chg);
3622 * Account for the reservations made. Shared mappings record regions
3623 * that have reservations as they are shared by multiple VMAs.
3624 * When the last VMA disappears, the region map says how much
3625 * the reservation was and the page cache tells how much of
3626 * the reservation was consumed. Private mappings are per-VMA and
3627 * only the consumed reservations are tracked. When the VMA
3628 * disappears, the original reservation is the VMA size and the
3629 * consumed reservations are stored in the map. Hence, nothing
3630 * else has to be done for private mappings here
3632 if (!vma || vma->vm_flags & VM_MAYSHARE)
3633 region_add(resv_map, from, to);
3636 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3637 kref_put(&resv_map->refs, resv_map_release);
3641 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3643 struct hstate *h = hstate_inode(inode);
3644 struct resv_map *resv_map = inode_resv_map(inode);
3646 struct hugepage_subpool *spool = subpool_inode(inode);
3650 chg = region_truncate(resv_map, offset);
3651 spin_lock(&inode->i_lock);
3652 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3653 spin_unlock(&inode->i_lock);
3656 * If the subpool has a minimum size, the number of global
3657 * reservations to be released may be adjusted.
3659 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
3660 hugetlb_acct_memory(h, -gbl_reserve);
3663 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3664 static unsigned long page_table_shareable(struct vm_area_struct *svma,
3665 struct vm_area_struct *vma,
3666 unsigned long addr, pgoff_t idx)
3668 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3670 unsigned long sbase = saddr & PUD_MASK;
3671 unsigned long s_end = sbase + PUD_SIZE;
3673 /* Allow segments to share if only one is marked locked */
3674 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3675 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3678 * match the virtual addresses, permission and the alignment of the
3681 if (pmd_index(addr) != pmd_index(saddr) ||
3682 vm_flags != svm_flags ||
3683 sbase < svma->vm_start || svma->vm_end < s_end)
3689 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3691 unsigned long base = addr & PUD_MASK;
3692 unsigned long end = base + PUD_SIZE;
3695 * check on proper vm_flags and page table alignment
3697 if (vma->vm_flags & VM_MAYSHARE &&
3698 vma->vm_start <= base && end <= vma->vm_end)
3704 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3705 * and returns the corresponding pte. While this is not necessary for the
3706 * !shared pmd case because we can allocate the pmd later as well, it makes the
3707 * code much cleaner. pmd allocation is essential for the shared case because
3708 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
3709 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3710 * bad pmd for sharing.
3712 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3714 struct vm_area_struct *vma = find_vma(mm, addr);
3715 struct address_space *mapping = vma->vm_file->f_mapping;
3716 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3718 struct vm_area_struct *svma;
3719 unsigned long saddr;
3724 if (!vma_shareable(vma, addr))
3725 return (pte_t *)pmd_alloc(mm, pud, addr);
3727 i_mmap_lock_write(mapping);
3728 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3732 saddr = page_table_shareable(svma, vma, addr, idx);
3734 spte = huge_pte_offset(svma->vm_mm, saddr);
3737 get_page(virt_to_page(spte));
3746 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
3748 if (pud_none(*pud)) {
3749 pud_populate(mm, pud,
3750 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3752 put_page(virt_to_page(spte));
3757 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3758 i_mmap_unlock_write(mapping);
3763 * unmap huge page backed by shared pte.
3765 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3766 * indicated by page_count > 1, unmap is achieved by clearing pud and
3767 * decrementing the ref count. If count == 1, the pte page is not shared.
3769 * called with page table lock held.
3771 * returns: 1 successfully unmapped a shared pte page
3772 * 0 the underlying pte page is not shared, or it is the last user
3774 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3776 pgd_t *pgd = pgd_offset(mm, *addr);
3777 pud_t *pud = pud_offset(pgd, *addr);
3779 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3780 if (page_count(virt_to_page(ptep)) == 1)
3784 put_page(virt_to_page(ptep));
3786 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3789 #define want_pmd_share() (1)
3790 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3791 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3796 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3800 #define want_pmd_share() (0)
3801 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3803 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3804 pte_t *huge_pte_alloc(struct mm_struct *mm,
3805 unsigned long addr, unsigned long sz)
3811 pgd = pgd_offset(mm, addr);
3812 pud = pud_alloc(mm, pgd, addr);
3814 if (sz == PUD_SIZE) {
3817 BUG_ON(sz != PMD_SIZE);
3818 if (want_pmd_share() && pud_none(*pud))
3819 pte = huge_pmd_share(mm, addr, pud);
3821 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3824 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3829 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3835 pgd = pgd_offset(mm, addr);
3836 if (pgd_present(*pgd)) {
3837 pud = pud_offset(pgd, addr);
3838 if (pud_present(*pud)) {
3840 return (pte_t *)pud;
3841 pmd = pmd_offset(pud, addr);
3844 return (pte_t *) pmd;
3847 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3850 * These functions are overwritable if your architecture needs its own
3853 struct page * __weak
3854 follow_huge_addr(struct mm_struct *mm, unsigned long address,
3857 return ERR_PTR(-EINVAL);
3860 struct page * __weak
3861 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3862 pmd_t *pmd, int flags)
3864 struct page *page = NULL;
3867 ptl = pmd_lockptr(mm, pmd);
3870 * make sure that the address range covered by this pmd is not
3871 * unmapped from other threads.
3873 if (!pmd_huge(*pmd))
3875 if (pmd_present(*pmd)) {
3876 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
3877 if (flags & FOLL_GET)
3880 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
3882 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
3886 * hwpoisoned entry is treated as no_page_table in
3887 * follow_page_mask().
3895 struct page * __weak
3896 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3897 pud_t *pud, int flags)
3899 if (flags & FOLL_GET)
3902 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
3905 #ifdef CONFIG_MEMORY_FAILURE
3908 * This function is called from memory failure code.
3909 * Assume the caller holds page lock of the head page.
3911 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3913 struct hstate *h = page_hstate(hpage);
3914 int nid = page_to_nid(hpage);
3917 spin_lock(&hugetlb_lock);
3919 * Just checking !page_huge_active is not enough, because that could be
3920 * an isolated/hwpoisoned hugepage (which have >0 refcount).
3922 if (!page_huge_active(hpage) && !page_count(hpage)) {
3924 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3925 * but dangling hpage->lru can trigger list-debug warnings
3926 * (this happens when we call unpoison_memory() on it),
3927 * so let it point to itself with list_del_init().
3929 list_del_init(&hpage->lru);
3930 set_page_refcounted(hpage);
3931 h->free_huge_pages--;
3932 h->free_huge_pages_node[nid]--;
3935 spin_unlock(&hugetlb_lock);
3940 bool isolate_huge_page(struct page *page, struct list_head *list)
3944 VM_BUG_ON_PAGE(!PageHead(page), page);
3945 spin_lock(&hugetlb_lock);
3946 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
3950 clear_page_huge_active(page);
3951 list_move_tail(&page->lru, list);
3953 spin_unlock(&hugetlb_lock);
3957 void putback_active_hugepage(struct page *page)
3959 VM_BUG_ON_PAGE(!PageHead(page), page);
3960 spin_lock(&hugetlb_lock);
3961 set_page_huge_active(page);
3962 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
3963 spin_unlock(&hugetlb_lock);