2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/module.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/rmap.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
26 #include <asm/pgtable.h>
30 #include <linux/hugetlb.h>
31 #include <linux/hugetlb_cgroup.h>
32 #include <linux/node.h>
35 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
36 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
37 unsigned long hugepages_treat_as_movable;
39 int hugetlb_max_hstate __read_mostly;
40 unsigned int default_hstate_idx;
41 struct hstate hstates[HUGE_MAX_HSTATE];
43 __initdata LIST_HEAD(huge_boot_pages);
45 /* for command line parsing */
46 static struct hstate * __initdata parsed_hstate;
47 static unsigned long __initdata default_hstate_max_huge_pages;
48 static unsigned long __initdata default_hstate_size;
51 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
52 * free_huge_pages, and surplus_huge_pages.
54 DEFINE_SPINLOCK(hugetlb_lock);
56 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
58 bool free = (spool->count == 0) && (spool->used_hpages == 0);
60 spin_unlock(&spool->lock);
62 /* If no pages are used, and no other handles to the subpool
63 * remain, free the subpool the subpool remain */
68 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
70 struct hugepage_subpool *spool;
72 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
76 spin_lock_init(&spool->lock);
78 spool->max_hpages = nr_blocks;
79 spool->used_hpages = 0;
84 void hugepage_put_subpool(struct hugepage_subpool *spool)
86 spin_lock(&spool->lock);
87 BUG_ON(!spool->count);
89 unlock_or_release_subpool(spool);
92 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
100 spin_lock(&spool->lock);
101 if ((spool->used_hpages + delta) <= spool->max_hpages) {
102 spool->used_hpages += delta;
106 spin_unlock(&spool->lock);
111 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
117 spin_lock(&spool->lock);
118 spool->used_hpages -= delta;
119 /* If hugetlbfs_put_super couldn't free spool due to
120 * an outstanding quota reference, free it now. */
121 unlock_or_release_subpool(spool);
124 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
126 return HUGETLBFS_SB(inode->i_sb)->spool;
129 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
131 return subpool_inode(file_inode(vma->vm_file));
135 * Region tracking -- allows tracking of reservations and instantiated pages
136 * across the pages in a mapping.
138 * The region data structures are protected by a combination of the mmap_sem
139 * and the hugetlb_instantiation_mutex. To access or modify a region the caller
140 * must either hold the mmap_sem for write, or the mmap_sem for read and
141 * the hugetlb_instantiation_mutex:
143 * down_write(&mm->mmap_sem);
145 * down_read(&mm->mmap_sem);
146 * mutex_lock(&hugetlb_instantiation_mutex);
149 struct list_head link;
154 static long region_add(struct list_head *head, long f, long t)
156 struct file_region *rg, *nrg, *trg;
158 /* Locate the region we are either in or before. */
159 list_for_each_entry(rg, head, link)
163 /* Round our left edge to the current segment if it encloses us. */
167 /* Check for and consume any regions we now overlap with. */
169 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
170 if (&rg->link == head)
175 /* If this area reaches higher then extend our area to
176 * include it completely. If this is not the first area
177 * which we intend to reuse, free it. */
190 static long region_chg(struct list_head *head, long f, long t)
192 struct file_region *rg, *nrg;
195 /* Locate the region we are before or in. */
196 list_for_each_entry(rg, head, link)
200 /* If we are below the current region then a new region is required.
201 * Subtle, allocate a new region at the position but make it zero
202 * size such that we can guarantee to record the reservation. */
203 if (&rg->link == head || t < rg->from) {
204 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
209 INIT_LIST_HEAD(&nrg->link);
210 list_add(&nrg->link, rg->link.prev);
215 /* Round our left edge to the current segment if it encloses us. */
220 /* Check for and consume any regions we now overlap with. */
221 list_for_each_entry(rg, rg->link.prev, link) {
222 if (&rg->link == head)
227 /* We overlap with this area, if it extends further than
228 * us then we must extend ourselves. Account for its
229 * existing reservation. */
234 chg -= rg->to - rg->from;
239 static long region_truncate(struct list_head *head, long end)
241 struct file_region *rg, *trg;
244 /* Locate the region we are either in or before. */
245 list_for_each_entry(rg, head, link)
248 if (&rg->link == head)
251 /* If we are in the middle of a region then adjust it. */
252 if (end > rg->from) {
255 rg = list_entry(rg->link.next, typeof(*rg), link);
258 /* Drop any remaining regions. */
259 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
260 if (&rg->link == head)
262 chg += rg->to - rg->from;
269 static long region_count(struct list_head *head, long f, long t)
271 struct file_region *rg;
274 /* Locate each segment we overlap with, and count that overlap. */
275 list_for_each_entry(rg, head, link) {
284 seg_from = max(rg->from, f);
285 seg_to = min(rg->to, t);
287 chg += seg_to - seg_from;
294 * Convert the address within this vma to the page offset within
295 * the mapping, in pagecache page units; huge pages here.
297 static pgoff_t vma_hugecache_offset(struct hstate *h,
298 struct vm_area_struct *vma, unsigned long address)
300 return ((address - vma->vm_start) >> huge_page_shift(h)) +
301 (vma->vm_pgoff >> huge_page_order(h));
304 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
305 unsigned long address)
307 return vma_hugecache_offset(hstate_vma(vma), vma, address);
311 * Return the size of the pages allocated when backing a VMA. In the majority
312 * cases this will be same size as used by the page table entries.
314 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
316 struct hstate *hstate;
318 if (!is_vm_hugetlb_page(vma))
321 hstate = hstate_vma(vma);
323 return 1UL << huge_page_shift(hstate);
325 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
328 * Return the page size being used by the MMU to back a VMA. In the majority
329 * of cases, the page size used by the kernel matches the MMU size. On
330 * architectures where it differs, an architecture-specific version of this
331 * function is required.
333 #ifndef vma_mmu_pagesize
334 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
336 return vma_kernel_pagesize(vma);
341 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
342 * bits of the reservation map pointer, which are always clear due to
345 #define HPAGE_RESV_OWNER (1UL << 0)
346 #define HPAGE_RESV_UNMAPPED (1UL << 1)
347 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
350 * These helpers are used to track how many pages are reserved for
351 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
352 * is guaranteed to have their future faults succeed.
354 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
355 * the reserve counters are updated with the hugetlb_lock held. It is safe
356 * to reset the VMA at fork() time as it is not in use yet and there is no
357 * chance of the global counters getting corrupted as a result of the values.
359 * The private mapping reservation is represented in a subtly different
360 * manner to a shared mapping. A shared mapping has a region map associated
361 * with the underlying file, this region map represents the backing file
362 * pages which have ever had a reservation assigned which this persists even
363 * after the page is instantiated. A private mapping has a region map
364 * associated with the original mmap which is attached to all VMAs which
365 * reference it, this region map represents those offsets which have consumed
366 * reservation ie. where pages have been instantiated.
368 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
370 return (unsigned long)vma->vm_private_data;
373 static void set_vma_private_data(struct vm_area_struct *vma,
376 vma->vm_private_data = (void *)value;
381 struct list_head regions;
384 static struct resv_map *resv_map_alloc(void)
386 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
390 kref_init(&resv_map->refs);
391 INIT_LIST_HEAD(&resv_map->regions);
396 static void resv_map_release(struct kref *ref)
398 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
400 /* Clear out any active regions before we release the map. */
401 region_truncate(&resv_map->regions, 0);
405 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
407 VM_BUG_ON(!is_vm_hugetlb_page(vma));
408 if (!(vma->vm_flags & VM_MAYSHARE))
409 return (struct resv_map *)(get_vma_private_data(vma) &
414 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
416 VM_BUG_ON(!is_vm_hugetlb_page(vma));
417 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
419 set_vma_private_data(vma, (get_vma_private_data(vma) &
420 HPAGE_RESV_MASK) | (unsigned long)map);
423 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
425 VM_BUG_ON(!is_vm_hugetlb_page(vma));
426 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
428 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
431 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
433 VM_BUG_ON(!is_vm_hugetlb_page(vma));
435 return (get_vma_private_data(vma) & flag) != 0;
438 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
439 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
441 VM_BUG_ON(!is_vm_hugetlb_page(vma));
442 if (!(vma->vm_flags & VM_MAYSHARE))
443 vma->vm_private_data = (void *)0;
446 /* Returns true if the VMA has associated reserve pages */
447 static int vma_has_reserves(struct vm_area_struct *vma, long chg)
449 if (vma->vm_flags & VM_NORESERVE) {
451 * This address is already reserved by other process(chg == 0),
452 * so, we should decrement reserved count. Without decrementing,
453 * reserve count remains after releasing inode, because this
454 * allocated page will go into page cache and is regarded as
455 * coming from reserved pool in releasing step. Currently, we
456 * don't have any other solution to deal with this situation
457 * properly, so add work-around here.
459 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
465 /* Shared mappings always use reserves */
466 if (vma->vm_flags & VM_MAYSHARE)
470 * Only the process that called mmap() has reserves for
473 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
479 static void copy_gigantic_page(struct page *dst, struct page *src)
482 struct hstate *h = page_hstate(src);
483 struct page *dst_base = dst;
484 struct page *src_base = src;
486 for (i = 0; i < pages_per_huge_page(h); ) {
488 copy_highpage(dst, src);
491 dst = mem_map_next(dst, dst_base, i);
492 src = mem_map_next(src, src_base, i);
496 void copy_huge_page(struct page *dst, struct page *src)
499 struct hstate *h = page_hstate(src);
501 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
502 copy_gigantic_page(dst, src);
507 for (i = 0; i < pages_per_huge_page(h); i++) {
509 copy_highpage(dst + i, src + i);
513 static void enqueue_huge_page(struct hstate *h, struct page *page)
515 int nid = page_to_nid(page);
516 list_move(&page->lru, &h->hugepage_freelists[nid]);
517 h->free_huge_pages++;
518 h->free_huge_pages_node[nid]++;
521 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
525 if (list_empty(&h->hugepage_freelists[nid]))
527 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
528 list_move(&page->lru, &h->hugepage_activelist);
529 set_page_refcounted(page);
530 h->free_huge_pages--;
531 h->free_huge_pages_node[nid]--;
535 static struct page *dequeue_huge_page_vma(struct hstate *h,
536 struct vm_area_struct *vma,
537 unsigned long address, int avoid_reserve,
540 struct page *page = NULL;
541 struct mempolicy *mpol;
542 nodemask_t *nodemask;
543 struct zonelist *zonelist;
546 unsigned int cpuset_mems_cookie;
549 * A child process with MAP_PRIVATE mappings created by their parent
550 * have no page reserves. This check ensures that reservations are
551 * not "stolen". The child may still get SIGKILLed
553 if (!vma_has_reserves(vma, chg) &&
554 h->free_huge_pages - h->resv_huge_pages == 0)
557 /* If reserves cannot be used, ensure enough pages are in the pool */
558 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
562 cpuset_mems_cookie = get_mems_allowed();
563 zonelist = huge_zonelist(vma, address,
564 htlb_alloc_mask, &mpol, &nodemask);
566 for_each_zone_zonelist_nodemask(zone, z, zonelist,
567 MAX_NR_ZONES - 1, nodemask) {
568 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
569 page = dequeue_huge_page_node(h, zone_to_nid(zone));
573 if (!vma_has_reserves(vma, chg))
576 SetPagePrivate(page);
577 h->resv_huge_pages--;
584 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
592 static void update_and_free_page(struct hstate *h, struct page *page)
596 VM_BUG_ON(h->order >= MAX_ORDER);
599 h->nr_huge_pages_node[page_to_nid(page)]--;
600 for (i = 0; i < pages_per_huge_page(h); i++) {
601 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
602 1 << PG_referenced | 1 << PG_dirty |
603 1 << PG_active | 1 << PG_reserved |
604 1 << PG_private | 1 << PG_writeback);
606 VM_BUG_ON(hugetlb_cgroup_from_page(page));
607 set_compound_page_dtor(page, NULL);
608 set_page_refcounted(page);
609 arch_release_hugepage(page);
610 __free_pages(page, huge_page_order(h));
613 struct hstate *size_to_hstate(unsigned long size)
618 if (huge_page_size(h) == size)
624 static void free_huge_page(struct page *page)
627 * Can't pass hstate in here because it is called from the
628 * compound page destructor.
630 struct hstate *h = page_hstate(page);
631 int nid = page_to_nid(page);
632 struct hugepage_subpool *spool =
633 (struct hugepage_subpool *)page_private(page);
634 bool restore_reserve;
636 set_page_private(page, 0);
637 page->mapping = NULL;
638 BUG_ON(page_count(page));
639 BUG_ON(page_mapcount(page));
640 restore_reserve = PagePrivate(page);
642 spin_lock(&hugetlb_lock);
643 hugetlb_cgroup_uncharge_page(hstate_index(h),
644 pages_per_huge_page(h), page);
646 h->resv_huge_pages++;
648 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
649 /* remove the page from active list */
650 list_del(&page->lru);
651 update_and_free_page(h, page);
652 h->surplus_huge_pages--;
653 h->surplus_huge_pages_node[nid]--;
655 arch_clear_hugepage_flags(page);
656 enqueue_huge_page(h, page);
658 spin_unlock(&hugetlb_lock);
659 hugepage_subpool_put_pages(spool, 1);
662 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
664 INIT_LIST_HEAD(&page->lru);
665 set_compound_page_dtor(page, free_huge_page);
666 spin_lock(&hugetlb_lock);
667 set_hugetlb_cgroup(page, NULL);
669 h->nr_huge_pages_node[nid]++;
670 spin_unlock(&hugetlb_lock);
671 put_page(page); /* free it into the hugepage allocator */
674 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
677 int nr_pages = 1 << order;
678 struct page *p = page + 1;
680 /* we rely on prep_new_huge_page to set the destructor */
681 set_compound_order(page, order);
683 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
685 set_page_count(p, 0);
686 p->first_page = page;
691 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
692 * transparent huge pages. See the PageTransHuge() documentation for more
695 int PageHuge(struct page *page)
697 compound_page_dtor *dtor;
699 if (!PageCompound(page))
702 page = compound_head(page);
703 dtor = get_compound_page_dtor(page);
705 return dtor == free_huge_page;
707 EXPORT_SYMBOL_GPL(PageHuge);
709 pgoff_t __basepage_index(struct page *page)
711 struct page *page_head = compound_head(page);
712 pgoff_t index = page_index(page_head);
713 unsigned long compound_idx;
715 if (!PageHuge(page_head))
716 return page_index(page);
718 if (compound_order(page_head) >= MAX_ORDER)
719 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
721 compound_idx = page - page_head;
723 return (index << compound_order(page_head)) + compound_idx;
726 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
730 if (h->order >= MAX_ORDER)
733 page = alloc_pages_exact_node(nid,
734 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
735 __GFP_REPEAT|__GFP_NOWARN,
738 if (arch_prepare_hugepage(page)) {
739 __free_pages(page, huge_page_order(h));
742 prep_new_huge_page(h, page, nid);
749 * common helper functions for hstate_next_node_to_{alloc|free}.
750 * We may have allocated or freed a huge page based on a different
751 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
752 * be outside of *nodes_allowed. Ensure that we use an allowed
753 * node for alloc or free.
755 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
757 nid = next_node(nid, *nodes_allowed);
758 if (nid == MAX_NUMNODES)
759 nid = first_node(*nodes_allowed);
760 VM_BUG_ON(nid >= MAX_NUMNODES);
765 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
767 if (!node_isset(nid, *nodes_allowed))
768 nid = next_node_allowed(nid, nodes_allowed);
773 * returns the previously saved node ["this node"] from which to
774 * allocate a persistent huge page for the pool and advance the
775 * next node from which to allocate, handling wrap at end of node
778 static int hstate_next_node_to_alloc(struct hstate *h,
779 nodemask_t *nodes_allowed)
783 VM_BUG_ON(!nodes_allowed);
785 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
786 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
792 * helper for free_pool_huge_page() - return the previously saved
793 * node ["this node"] from which to free a huge page. Advance the
794 * next node id whether or not we find a free huge page to free so
795 * that the next attempt to free addresses the next node.
797 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
801 VM_BUG_ON(!nodes_allowed);
803 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
804 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
809 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
810 for (nr_nodes = nodes_weight(*mask); \
812 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
815 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
816 for (nr_nodes = nodes_weight(*mask); \
818 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
821 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
827 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
828 page = alloc_fresh_huge_page_node(h, node);
836 count_vm_event(HTLB_BUDDY_PGALLOC);
838 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
844 * Free huge page from pool from next node to free.
845 * Attempt to keep persistent huge pages more or less
846 * balanced over allowed nodes.
847 * Called with hugetlb_lock locked.
849 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
855 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
857 * If we're returning unused surplus pages, only examine
858 * nodes with surplus pages.
860 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
861 !list_empty(&h->hugepage_freelists[node])) {
863 list_entry(h->hugepage_freelists[node].next,
865 list_del(&page->lru);
866 h->free_huge_pages--;
867 h->free_huge_pages_node[node]--;
869 h->surplus_huge_pages--;
870 h->surplus_huge_pages_node[node]--;
872 update_and_free_page(h, page);
881 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
886 if (h->order >= MAX_ORDER)
890 * Assume we will successfully allocate the surplus page to
891 * prevent racing processes from causing the surplus to exceed
894 * This however introduces a different race, where a process B
895 * tries to grow the static hugepage pool while alloc_pages() is
896 * called by process A. B will only examine the per-node
897 * counters in determining if surplus huge pages can be
898 * converted to normal huge pages in adjust_pool_surplus(). A
899 * won't be able to increment the per-node counter, until the
900 * lock is dropped by B, but B doesn't drop hugetlb_lock until
901 * no more huge pages can be converted from surplus to normal
902 * state (and doesn't try to convert again). Thus, we have a
903 * case where a surplus huge page exists, the pool is grown, and
904 * the surplus huge page still exists after, even though it
905 * should just have been converted to a normal huge page. This
906 * does not leak memory, though, as the hugepage will be freed
907 * once it is out of use. It also does not allow the counters to
908 * go out of whack in adjust_pool_surplus() as we don't modify
909 * the node values until we've gotten the hugepage and only the
910 * per-node value is checked there.
912 spin_lock(&hugetlb_lock);
913 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
914 spin_unlock(&hugetlb_lock);
918 h->surplus_huge_pages++;
920 spin_unlock(&hugetlb_lock);
922 if (nid == NUMA_NO_NODE)
923 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
924 __GFP_REPEAT|__GFP_NOWARN,
927 page = alloc_pages_exact_node(nid,
928 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
929 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
931 if (page && arch_prepare_hugepage(page)) {
932 __free_pages(page, huge_page_order(h));
936 spin_lock(&hugetlb_lock);
938 INIT_LIST_HEAD(&page->lru);
939 r_nid = page_to_nid(page);
940 set_compound_page_dtor(page, free_huge_page);
941 set_hugetlb_cgroup(page, NULL);
943 * We incremented the global counters already
945 h->nr_huge_pages_node[r_nid]++;
946 h->surplus_huge_pages_node[r_nid]++;
947 __count_vm_event(HTLB_BUDDY_PGALLOC);
950 h->surplus_huge_pages--;
951 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
953 spin_unlock(&hugetlb_lock);
959 * This allocation function is useful in the context where vma is irrelevant.
960 * E.g. soft-offlining uses this function because it only cares physical
961 * address of error page.
963 struct page *alloc_huge_page_node(struct hstate *h, int nid)
965 struct page *page = NULL;
967 spin_lock(&hugetlb_lock);
968 if (h->free_huge_pages - h->resv_huge_pages > 0)
969 page = dequeue_huge_page_node(h, nid);
970 spin_unlock(&hugetlb_lock);
973 page = alloc_buddy_huge_page(h, nid);
979 * Increase the hugetlb pool such that it can accommodate a reservation
982 static int gather_surplus_pages(struct hstate *h, int delta)
984 struct list_head surplus_list;
985 struct page *page, *tmp;
987 int needed, allocated;
988 bool alloc_ok = true;
990 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
992 h->resv_huge_pages += delta;
997 INIT_LIST_HEAD(&surplus_list);
1001 spin_unlock(&hugetlb_lock);
1002 for (i = 0; i < needed; i++) {
1003 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1008 list_add(&page->lru, &surplus_list);
1013 * After retaking hugetlb_lock, we need to recalculate 'needed'
1014 * because either resv_huge_pages or free_huge_pages may have changed.
1016 spin_lock(&hugetlb_lock);
1017 needed = (h->resv_huge_pages + delta) -
1018 (h->free_huge_pages + allocated);
1023 * We were not able to allocate enough pages to
1024 * satisfy the entire reservation so we free what
1025 * we've allocated so far.
1030 * The surplus_list now contains _at_least_ the number of extra pages
1031 * needed to accommodate the reservation. Add the appropriate number
1032 * of pages to the hugetlb pool and free the extras back to the buddy
1033 * allocator. Commit the entire reservation here to prevent another
1034 * process from stealing the pages as they are added to the pool but
1035 * before they are reserved.
1037 needed += allocated;
1038 h->resv_huge_pages += delta;
1041 /* Free the needed pages to the hugetlb pool */
1042 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1046 * This page is now managed by the hugetlb allocator and has
1047 * no users -- drop the buddy allocator's reference.
1049 put_page_testzero(page);
1050 VM_BUG_ON(page_count(page));
1051 enqueue_huge_page(h, page);
1054 spin_unlock(&hugetlb_lock);
1056 /* Free unnecessary surplus pages to the buddy allocator */
1057 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1059 spin_lock(&hugetlb_lock);
1065 * When releasing a hugetlb pool reservation, any surplus pages that were
1066 * allocated to satisfy the reservation must be explicitly freed if they were
1068 * Called with hugetlb_lock held.
1070 static void return_unused_surplus_pages(struct hstate *h,
1071 unsigned long unused_resv_pages)
1073 unsigned long nr_pages;
1075 /* Uncommit the reservation */
1076 h->resv_huge_pages -= unused_resv_pages;
1078 /* Cannot return gigantic pages currently */
1079 if (h->order >= MAX_ORDER)
1082 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1085 * We want to release as many surplus pages as possible, spread
1086 * evenly across all nodes with memory. Iterate across these nodes
1087 * until we can no longer free unreserved surplus pages. This occurs
1088 * when the nodes with surplus pages have no free pages.
1089 * free_pool_huge_page() will balance the the freed pages across the
1090 * on-line nodes with memory and will handle the hstate accounting.
1092 while (nr_pages--) {
1093 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1099 * Determine if the huge page at addr within the vma has an associated
1100 * reservation. Where it does not we will need to logically increase
1101 * reservation and actually increase subpool usage before an allocation
1102 * can occur. Where any new reservation would be required the
1103 * reservation change is prepared, but not committed. Once the page
1104 * has been allocated from the subpool and instantiated the change should
1105 * be committed via vma_commit_reservation. No action is required on
1108 static long vma_needs_reservation(struct hstate *h,
1109 struct vm_area_struct *vma, unsigned long addr)
1111 struct address_space *mapping = vma->vm_file->f_mapping;
1112 struct inode *inode = mapping->host;
1114 if (vma->vm_flags & VM_MAYSHARE) {
1115 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1116 return region_chg(&inode->i_mapping->private_list,
1119 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1124 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1125 struct resv_map *resv = vma_resv_map(vma);
1127 err = region_chg(&resv->regions, idx, idx + 1);
1133 static void vma_commit_reservation(struct hstate *h,
1134 struct vm_area_struct *vma, unsigned long addr)
1136 struct address_space *mapping = vma->vm_file->f_mapping;
1137 struct inode *inode = mapping->host;
1139 if (vma->vm_flags & VM_MAYSHARE) {
1140 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1141 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1143 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1144 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1145 struct resv_map *resv = vma_resv_map(vma);
1147 /* Mark this page used in the map. */
1148 region_add(&resv->regions, idx, idx + 1);
1152 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1153 unsigned long addr, int avoid_reserve)
1155 struct hugepage_subpool *spool = subpool_vma(vma);
1156 struct hstate *h = hstate_vma(vma);
1160 struct hugetlb_cgroup *h_cg;
1162 idx = hstate_index(h);
1164 * Processes that did not create the mapping will have no
1165 * reserves and will not have accounted against subpool
1166 * limit. Check that the subpool limit can be made before
1167 * satisfying the allocation MAP_NORESERVE mappings may also
1168 * need pages and subpool limit allocated allocated if no reserve
1171 chg = vma_needs_reservation(h, vma, addr);
1173 return ERR_PTR(-ENOMEM);
1174 if (chg || avoid_reserve)
1175 if (hugepage_subpool_get_pages(spool, 1))
1176 return ERR_PTR(-ENOSPC);
1178 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1180 if (chg || avoid_reserve)
1181 hugepage_subpool_put_pages(spool, 1);
1182 return ERR_PTR(-ENOSPC);
1184 spin_lock(&hugetlb_lock);
1185 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
1187 spin_unlock(&hugetlb_lock);
1188 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1190 hugetlb_cgroup_uncharge_cgroup(idx,
1191 pages_per_huge_page(h),
1193 if (chg || avoid_reserve)
1194 hugepage_subpool_put_pages(spool, 1);
1195 return ERR_PTR(-ENOSPC);
1197 spin_lock(&hugetlb_lock);
1198 list_move(&page->lru, &h->hugepage_activelist);
1201 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1202 spin_unlock(&hugetlb_lock);
1204 set_page_private(page, (unsigned long)spool);
1206 vma_commit_reservation(h, vma, addr);
1210 int __weak alloc_bootmem_huge_page(struct hstate *h)
1212 struct huge_bootmem_page *m;
1215 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1218 addr = __alloc_bootmem_node_nopanic(NODE_DATA(node),
1219 huge_page_size(h), huge_page_size(h), 0);
1223 * Use the beginning of the huge page to store the
1224 * huge_bootmem_page struct (until gather_bootmem
1225 * puts them into the mem_map).
1234 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1235 /* Put them into a private list first because mem_map is not up yet */
1236 list_add(&m->list, &huge_boot_pages);
1241 static void prep_compound_huge_page(struct page *page, int order)
1243 if (unlikely(order > (MAX_ORDER - 1)))
1244 prep_compound_gigantic_page(page, order);
1246 prep_compound_page(page, order);
1249 /* Put bootmem huge pages into the standard lists after mem_map is up */
1250 static void __init gather_bootmem_prealloc(void)
1252 struct huge_bootmem_page *m;
1254 list_for_each_entry(m, &huge_boot_pages, list) {
1255 struct hstate *h = m->hstate;
1258 #ifdef CONFIG_HIGHMEM
1259 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1260 free_bootmem_late((unsigned long)m,
1261 sizeof(struct huge_bootmem_page));
1263 page = virt_to_page(m);
1265 __ClearPageReserved(page);
1266 WARN_ON(page_count(page) != 1);
1267 prep_compound_huge_page(page, h->order);
1268 prep_new_huge_page(h, page, page_to_nid(page));
1270 * If we had gigantic hugepages allocated at boot time, we need
1271 * to restore the 'stolen' pages to totalram_pages in order to
1272 * fix confusing memory reports from free(1) and another
1273 * side-effects, like CommitLimit going negative.
1275 if (h->order > (MAX_ORDER - 1))
1276 adjust_managed_page_count(page, 1 << h->order);
1280 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1284 for (i = 0; i < h->max_huge_pages; ++i) {
1285 if (h->order >= MAX_ORDER) {
1286 if (!alloc_bootmem_huge_page(h))
1288 } else if (!alloc_fresh_huge_page(h,
1289 &node_states[N_MEMORY]))
1292 h->max_huge_pages = i;
1295 static void __init hugetlb_init_hstates(void)
1299 for_each_hstate(h) {
1300 /* oversize hugepages were init'ed in early boot */
1301 if (h->order < MAX_ORDER)
1302 hugetlb_hstate_alloc_pages(h);
1306 static char * __init memfmt(char *buf, unsigned long n)
1308 if (n >= (1UL << 30))
1309 sprintf(buf, "%lu GB", n >> 30);
1310 else if (n >= (1UL << 20))
1311 sprintf(buf, "%lu MB", n >> 20);
1313 sprintf(buf, "%lu KB", n >> 10);
1317 static void __init report_hugepages(void)
1321 for_each_hstate(h) {
1323 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1324 memfmt(buf, huge_page_size(h)),
1325 h->free_huge_pages);
1329 #ifdef CONFIG_HIGHMEM
1330 static void try_to_free_low(struct hstate *h, unsigned long count,
1331 nodemask_t *nodes_allowed)
1335 if (h->order >= MAX_ORDER)
1338 for_each_node_mask(i, *nodes_allowed) {
1339 struct page *page, *next;
1340 struct list_head *freel = &h->hugepage_freelists[i];
1341 list_for_each_entry_safe(page, next, freel, lru) {
1342 if (count >= h->nr_huge_pages)
1344 if (PageHighMem(page))
1346 list_del(&page->lru);
1347 update_and_free_page(h, page);
1348 h->free_huge_pages--;
1349 h->free_huge_pages_node[page_to_nid(page)]--;
1354 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1355 nodemask_t *nodes_allowed)
1361 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1362 * balanced by operating on them in a round-robin fashion.
1363 * Returns 1 if an adjustment was made.
1365 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1370 VM_BUG_ON(delta != -1 && delta != 1);
1373 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1374 if (h->surplus_huge_pages_node[node])
1378 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1379 if (h->surplus_huge_pages_node[node] <
1380 h->nr_huge_pages_node[node])
1387 h->surplus_huge_pages += delta;
1388 h->surplus_huge_pages_node[node] += delta;
1392 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1393 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1394 nodemask_t *nodes_allowed)
1396 unsigned long min_count, ret;
1398 if (h->order >= MAX_ORDER)
1399 return h->max_huge_pages;
1402 * Increase the pool size
1403 * First take pages out of surplus state. Then make up the
1404 * remaining difference by allocating fresh huge pages.
1406 * We might race with alloc_buddy_huge_page() here and be unable
1407 * to convert a surplus huge page to a normal huge page. That is
1408 * not critical, though, it just means the overall size of the
1409 * pool might be one hugepage larger than it needs to be, but
1410 * within all the constraints specified by the sysctls.
1412 spin_lock(&hugetlb_lock);
1413 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1414 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1418 while (count > persistent_huge_pages(h)) {
1420 * If this allocation races such that we no longer need the
1421 * page, free_huge_page will handle it by freeing the page
1422 * and reducing the surplus.
1424 spin_unlock(&hugetlb_lock);
1425 ret = alloc_fresh_huge_page(h, nodes_allowed);
1426 spin_lock(&hugetlb_lock);
1430 /* Bail for signals. Probably ctrl-c from user */
1431 if (signal_pending(current))
1436 * Decrease the pool size
1437 * First return free pages to the buddy allocator (being careful
1438 * to keep enough around to satisfy reservations). Then place
1439 * pages into surplus state as needed so the pool will shrink
1440 * to the desired size as pages become free.
1442 * By placing pages into the surplus state independent of the
1443 * overcommit value, we are allowing the surplus pool size to
1444 * exceed overcommit. There are few sane options here. Since
1445 * alloc_buddy_huge_page() is checking the global counter,
1446 * though, we'll note that we're not allowed to exceed surplus
1447 * and won't grow the pool anywhere else. Not until one of the
1448 * sysctls are changed, or the surplus pages go out of use.
1450 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1451 min_count = max(count, min_count);
1452 try_to_free_low(h, min_count, nodes_allowed);
1453 while (min_count < persistent_huge_pages(h)) {
1454 if (!free_pool_huge_page(h, nodes_allowed, 0))
1457 while (count < persistent_huge_pages(h)) {
1458 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1462 ret = persistent_huge_pages(h);
1463 spin_unlock(&hugetlb_lock);
1467 #define HSTATE_ATTR_RO(_name) \
1468 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1470 #define HSTATE_ATTR(_name) \
1471 static struct kobj_attribute _name##_attr = \
1472 __ATTR(_name, 0644, _name##_show, _name##_store)
1474 static struct kobject *hugepages_kobj;
1475 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1477 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1479 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1483 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1484 if (hstate_kobjs[i] == kobj) {
1486 *nidp = NUMA_NO_NODE;
1490 return kobj_to_node_hstate(kobj, nidp);
1493 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1494 struct kobj_attribute *attr, char *buf)
1497 unsigned long nr_huge_pages;
1500 h = kobj_to_hstate(kobj, &nid);
1501 if (nid == NUMA_NO_NODE)
1502 nr_huge_pages = h->nr_huge_pages;
1504 nr_huge_pages = h->nr_huge_pages_node[nid];
1506 return sprintf(buf, "%lu\n", nr_huge_pages);
1509 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1510 struct kobject *kobj, struct kobj_attribute *attr,
1511 const char *buf, size_t len)
1515 unsigned long count;
1517 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1519 err = kstrtoul(buf, 10, &count);
1523 h = kobj_to_hstate(kobj, &nid);
1524 if (h->order >= MAX_ORDER) {
1529 if (nid == NUMA_NO_NODE) {
1531 * global hstate attribute
1533 if (!(obey_mempolicy &&
1534 init_nodemask_of_mempolicy(nodes_allowed))) {
1535 NODEMASK_FREE(nodes_allowed);
1536 nodes_allowed = &node_states[N_MEMORY];
1538 } else if (nodes_allowed) {
1540 * per node hstate attribute: adjust count to global,
1541 * but restrict alloc/free to the specified node.
1543 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1544 init_nodemask_of_node(nodes_allowed, nid);
1546 nodes_allowed = &node_states[N_MEMORY];
1548 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1550 if (nodes_allowed != &node_states[N_MEMORY])
1551 NODEMASK_FREE(nodes_allowed);
1555 NODEMASK_FREE(nodes_allowed);
1559 static ssize_t nr_hugepages_show(struct kobject *kobj,
1560 struct kobj_attribute *attr, char *buf)
1562 return nr_hugepages_show_common(kobj, attr, buf);
1565 static ssize_t nr_hugepages_store(struct kobject *kobj,
1566 struct kobj_attribute *attr, const char *buf, size_t len)
1568 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1570 HSTATE_ATTR(nr_hugepages);
1575 * hstate attribute for optionally mempolicy-based constraint on persistent
1576 * huge page alloc/free.
1578 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1579 struct kobj_attribute *attr, char *buf)
1581 return nr_hugepages_show_common(kobj, attr, buf);
1584 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1585 struct kobj_attribute *attr, const char *buf, size_t len)
1587 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1589 HSTATE_ATTR(nr_hugepages_mempolicy);
1593 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1594 struct kobj_attribute *attr, char *buf)
1596 struct hstate *h = kobj_to_hstate(kobj, NULL);
1597 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1600 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1601 struct kobj_attribute *attr, const char *buf, size_t count)
1604 unsigned long input;
1605 struct hstate *h = kobj_to_hstate(kobj, NULL);
1607 if (h->order >= MAX_ORDER)
1610 err = kstrtoul(buf, 10, &input);
1614 spin_lock(&hugetlb_lock);
1615 h->nr_overcommit_huge_pages = input;
1616 spin_unlock(&hugetlb_lock);
1620 HSTATE_ATTR(nr_overcommit_hugepages);
1622 static ssize_t free_hugepages_show(struct kobject *kobj,
1623 struct kobj_attribute *attr, char *buf)
1626 unsigned long free_huge_pages;
1629 h = kobj_to_hstate(kobj, &nid);
1630 if (nid == NUMA_NO_NODE)
1631 free_huge_pages = h->free_huge_pages;
1633 free_huge_pages = h->free_huge_pages_node[nid];
1635 return sprintf(buf, "%lu\n", free_huge_pages);
1637 HSTATE_ATTR_RO(free_hugepages);
1639 static ssize_t resv_hugepages_show(struct kobject *kobj,
1640 struct kobj_attribute *attr, char *buf)
1642 struct hstate *h = kobj_to_hstate(kobj, NULL);
1643 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1645 HSTATE_ATTR_RO(resv_hugepages);
1647 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1648 struct kobj_attribute *attr, char *buf)
1651 unsigned long surplus_huge_pages;
1654 h = kobj_to_hstate(kobj, &nid);
1655 if (nid == NUMA_NO_NODE)
1656 surplus_huge_pages = h->surplus_huge_pages;
1658 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1660 return sprintf(buf, "%lu\n", surplus_huge_pages);
1662 HSTATE_ATTR_RO(surplus_hugepages);
1664 static struct attribute *hstate_attrs[] = {
1665 &nr_hugepages_attr.attr,
1666 &nr_overcommit_hugepages_attr.attr,
1667 &free_hugepages_attr.attr,
1668 &resv_hugepages_attr.attr,
1669 &surplus_hugepages_attr.attr,
1671 &nr_hugepages_mempolicy_attr.attr,
1676 static struct attribute_group hstate_attr_group = {
1677 .attrs = hstate_attrs,
1680 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1681 struct kobject **hstate_kobjs,
1682 struct attribute_group *hstate_attr_group)
1685 int hi = hstate_index(h);
1687 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1688 if (!hstate_kobjs[hi])
1691 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1693 kobject_put(hstate_kobjs[hi]);
1698 static void __init hugetlb_sysfs_init(void)
1703 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1704 if (!hugepages_kobj)
1707 for_each_hstate(h) {
1708 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1709 hstate_kobjs, &hstate_attr_group);
1711 pr_err("Hugetlb: Unable to add hstate %s", h->name);
1718 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1719 * with node devices in node_devices[] using a parallel array. The array
1720 * index of a node device or _hstate == node id.
1721 * This is here to avoid any static dependency of the node device driver, in
1722 * the base kernel, on the hugetlb module.
1724 struct node_hstate {
1725 struct kobject *hugepages_kobj;
1726 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1728 struct node_hstate node_hstates[MAX_NUMNODES];
1731 * A subset of global hstate attributes for node devices
1733 static struct attribute *per_node_hstate_attrs[] = {
1734 &nr_hugepages_attr.attr,
1735 &free_hugepages_attr.attr,
1736 &surplus_hugepages_attr.attr,
1740 static struct attribute_group per_node_hstate_attr_group = {
1741 .attrs = per_node_hstate_attrs,
1745 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1746 * Returns node id via non-NULL nidp.
1748 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1752 for (nid = 0; nid < nr_node_ids; nid++) {
1753 struct node_hstate *nhs = &node_hstates[nid];
1755 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1756 if (nhs->hstate_kobjs[i] == kobj) {
1768 * Unregister hstate attributes from a single node device.
1769 * No-op if no hstate attributes attached.
1771 static void hugetlb_unregister_node(struct node *node)
1774 struct node_hstate *nhs = &node_hstates[node->dev.id];
1776 if (!nhs->hugepages_kobj)
1777 return; /* no hstate attributes */
1779 for_each_hstate(h) {
1780 int idx = hstate_index(h);
1781 if (nhs->hstate_kobjs[idx]) {
1782 kobject_put(nhs->hstate_kobjs[idx]);
1783 nhs->hstate_kobjs[idx] = NULL;
1787 kobject_put(nhs->hugepages_kobj);
1788 nhs->hugepages_kobj = NULL;
1792 * hugetlb module exit: unregister hstate attributes from node devices
1795 static void hugetlb_unregister_all_nodes(void)
1800 * disable node device registrations.
1802 register_hugetlbfs_with_node(NULL, NULL);
1805 * remove hstate attributes from any nodes that have them.
1807 for (nid = 0; nid < nr_node_ids; nid++)
1808 hugetlb_unregister_node(node_devices[nid]);
1812 * Register hstate attributes for a single node device.
1813 * No-op if attributes already registered.
1815 static void hugetlb_register_node(struct node *node)
1818 struct node_hstate *nhs = &node_hstates[node->dev.id];
1821 if (nhs->hugepages_kobj)
1822 return; /* already allocated */
1824 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1826 if (!nhs->hugepages_kobj)
1829 for_each_hstate(h) {
1830 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1832 &per_node_hstate_attr_group);
1834 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1835 h->name, node->dev.id);
1836 hugetlb_unregister_node(node);
1843 * hugetlb init time: register hstate attributes for all registered node
1844 * devices of nodes that have memory. All on-line nodes should have
1845 * registered their associated device by this time.
1847 static void hugetlb_register_all_nodes(void)
1851 for_each_node_state(nid, N_MEMORY) {
1852 struct node *node = node_devices[nid];
1853 if (node->dev.id == nid)
1854 hugetlb_register_node(node);
1858 * Let the node device driver know we're here so it can
1859 * [un]register hstate attributes on node hotplug.
1861 register_hugetlbfs_with_node(hugetlb_register_node,
1862 hugetlb_unregister_node);
1864 #else /* !CONFIG_NUMA */
1866 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1874 static void hugetlb_unregister_all_nodes(void) { }
1876 static void hugetlb_register_all_nodes(void) { }
1880 static void __exit hugetlb_exit(void)
1884 hugetlb_unregister_all_nodes();
1886 for_each_hstate(h) {
1887 kobject_put(hstate_kobjs[hstate_index(h)]);
1890 kobject_put(hugepages_kobj);
1892 module_exit(hugetlb_exit);
1894 static int __init hugetlb_init(void)
1896 /* Some platform decide whether they support huge pages at boot
1897 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1898 * there is no such support
1900 if (HPAGE_SHIFT == 0)
1903 if (!size_to_hstate(default_hstate_size)) {
1904 default_hstate_size = HPAGE_SIZE;
1905 if (!size_to_hstate(default_hstate_size))
1906 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1908 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1909 if (default_hstate_max_huge_pages)
1910 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1912 hugetlb_init_hstates();
1913 gather_bootmem_prealloc();
1916 hugetlb_sysfs_init();
1917 hugetlb_register_all_nodes();
1918 hugetlb_cgroup_file_init();
1922 module_init(hugetlb_init);
1924 /* Should be called on processing a hugepagesz=... option */
1925 void __init hugetlb_add_hstate(unsigned order)
1930 if (size_to_hstate(PAGE_SIZE << order)) {
1931 pr_warning("hugepagesz= specified twice, ignoring\n");
1934 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1936 h = &hstates[hugetlb_max_hstate++];
1938 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1939 h->nr_huge_pages = 0;
1940 h->free_huge_pages = 0;
1941 for (i = 0; i < MAX_NUMNODES; ++i)
1942 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1943 INIT_LIST_HEAD(&h->hugepage_activelist);
1944 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
1945 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
1946 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1947 huge_page_size(h)/1024);
1952 static int __init hugetlb_nrpages_setup(char *s)
1955 static unsigned long *last_mhp;
1958 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1959 * so this hugepages= parameter goes to the "default hstate".
1961 if (!hugetlb_max_hstate)
1962 mhp = &default_hstate_max_huge_pages;
1964 mhp = &parsed_hstate->max_huge_pages;
1966 if (mhp == last_mhp) {
1967 pr_warning("hugepages= specified twice without "
1968 "interleaving hugepagesz=, ignoring\n");
1972 if (sscanf(s, "%lu", mhp) <= 0)
1976 * Global state is always initialized later in hugetlb_init.
1977 * But we need to allocate >= MAX_ORDER hstates here early to still
1978 * use the bootmem allocator.
1980 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
1981 hugetlb_hstate_alloc_pages(parsed_hstate);
1987 __setup("hugepages=", hugetlb_nrpages_setup);
1989 static int __init hugetlb_default_setup(char *s)
1991 default_hstate_size = memparse(s, &s);
1994 __setup("default_hugepagesz=", hugetlb_default_setup);
1996 static unsigned int cpuset_mems_nr(unsigned int *array)
1999 unsigned int nr = 0;
2001 for_each_node_mask(node, cpuset_current_mems_allowed)
2007 #ifdef CONFIG_SYSCTL
2008 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2009 struct ctl_table *table, int write,
2010 void __user *buffer, size_t *length, loff_t *ppos)
2012 struct hstate *h = &default_hstate;
2016 tmp = h->max_huge_pages;
2018 if (write && h->order >= MAX_ORDER)
2022 table->maxlen = sizeof(unsigned long);
2023 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2028 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2029 GFP_KERNEL | __GFP_NORETRY);
2030 if (!(obey_mempolicy &&
2031 init_nodemask_of_mempolicy(nodes_allowed))) {
2032 NODEMASK_FREE(nodes_allowed);
2033 nodes_allowed = &node_states[N_MEMORY];
2035 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2037 if (nodes_allowed != &node_states[N_MEMORY])
2038 NODEMASK_FREE(nodes_allowed);
2044 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2045 void __user *buffer, size_t *length, loff_t *ppos)
2048 return hugetlb_sysctl_handler_common(false, table, write,
2049 buffer, length, ppos);
2053 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2054 void __user *buffer, size_t *length, loff_t *ppos)
2056 return hugetlb_sysctl_handler_common(true, table, write,
2057 buffer, length, ppos);
2059 #endif /* CONFIG_NUMA */
2061 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2062 void __user *buffer,
2063 size_t *length, loff_t *ppos)
2065 proc_dointvec(table, write, buffer, length, ppos);
2066 if (hugepages_treat_as_movable)
2067 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2069 htlb_alloc_mask = GFP_HIGHUSER;
2073 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2074 void __user *buffer,
2075 size_t *length, loff_t *ppos)
2077 struct hstate *h = &default_hstate;
2081 tmp = h->nr_overcommit_huge_pages;
2083 if (write && h->order >= MAX_ORDER)
2087 table->maxlen = sizeof(unsigned long);
2088 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2093 spin_lock(&hugetlb_lock);
2094 h->nr_overcommit_huge_pages = tmp;
2095 spin_unlock(&hugetlb_lock);
2101 #endif /* CONFIG_SYSCTL */
2103 void hugetlb_report_meminfo(struct seq_file *m)
2105 struct hstate *h = &default_hstate;
2107 "HugePages_Total: %5lu\n"
2108 "HugePages_Free: %5lu\n"
2109 "HugePages_Rsvd: %5lu\n"
2110 "HugePages_Surp: %5lu\n"
2111 "Hugepagesize: %8lu kB\n",
2115 h->surplus_huge_pages,
2116 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2119 int hugetlb_report_node_meminfo(int nid, char *buf)
2121 struct hstate *h = &default_hstate;
2123 "Node %d HugePages_Total: %5u\n"
2124 "Node %d HugePages_Free: %5u\n"
2125 "Node %d HugePages_Surp: %5u\n",
2126 nid, h->nr_huge_pages_node[nid],
2127 nid, h->free_huge_pages_node[nid],
2128 nid, h->surplus_huge_pages_node[nid]);
2131 void hugetlb_show_meminfo(void)
2136 for_each_node_state(nid, N_MEMORY)
2138 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2140 h->nr_huge_pages_node[nid],
2141 h->free_huge_pages_node[nid],
2142 h->surplus_huge_pages_node[nid],
2143 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2146 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2147 unsigned long hugetlb_total_pages(void)
2150 unsigned long nr_total_pages = 0;
2153 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2154 return nr_total_pages;
2157 static int hugetlb_acct_memory(struct hstate *h, long delta)
2161 spin_lock(&hugetlb_lock);
2163 * When cpuset is configured, it breaks the strict hugetlb page
2164 * reservation as the accounting is done on a global variable. Such
2165 * reservation is completely rubbish in the presence of cpuset because
2166 * the reservation is not checked against page availability for the
2167 * current cpuset. Application can still potentially OOM'ed by kernel
2168 * with lack of free htlb page in cpuset that the task is in.
2169 * Attempt to enforce strict accounting with cpuset is almost
2170 * impossible (or too ugly) because cpuset is too fluid that
2171 * task or memory node can be dynamically moved between cpusets.
2173 * The change of semantics for shared hugetlb mapping with cpuset is
2174 * undesirable. However, in order to preserve some of the semantics,
2175 * we fall back to check against current free page availability as
2176 * a best attempt and hopefully to minimize the impact of changing
2177 * semantics that cpuset has.
2180 if (gather_surplus_pages(h, delta) < 0)
2183 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2184 return_unused_surplus_pages(h, delta);
2191 return_unused_surplus_pages(h, (unsigned long) -delta);
2194 spin_unlock(&hugetlb_lock);
2198 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2200 struct resv_map *resv = vma_resv_map(vma);
2203 * This new VMA should share its siblings reservation map if present.
2204 * The VMA will only ever have a valid reservation map pointer where
2205 * it is being copied for another still existing VMA. As that VMA
2206 * has a reference to the reservation map it cannot disappear until
2207 * after this open call completes. It is therefore safe to take a
2208 * new reference here without additional locking.
2211 kref_get(&resv->refs);
2214 static void resv_map_put(struct vm_area_struct *vma)
2216 struct resv_map *resv = vma_resv_map(vma);
2220 kref_put(&resv->refs, resv_map_release);
2223 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2225 struct hstate *h = hstate_vma(vma);
2226 struct resv_map *resv = vma_resv_map(vma);
2227 struct hugepage_subpool *spool = subpool_vma(vma);
2228 unsigned long reserve;
2229 unsigned long start;
2233 start = vma_hugecache_offset(h, vma, vma->vm_start);
2234 end = vma_hugecache_offset(h, vma, vma->vm_end);
2236 reserve = (end - start) -
2237 region_count(&resv->regions, start, end);
2242 hugetlb_acct_memory(h, -reserve);
2243 hugepage_subpool_put_pages(spool, reserve);
2249 * We cannot handle pagefaults against hugetlb pages at all. They cause
2250 * handle_mm_fault() to try to instantiate regular-sized pages in the
2251 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2254 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2260 const struct vm_operations_struct hugetlb_vm_ops = {
2261 .fault = hugetlb_vm_op_fault,
2262 .open = hugetlb_vm_op_open,
2263 .close = hugetlb_vm_op_close,
2266 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2272 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2273 vma->vm_page_prot)));
2275 entry = huge_pte_wrprotect(mk_huge_pte(page,
2276 vma->vm_page_prot));
2278 entry = pte_mkyoung(entry);
2279 entry = pte_mkhuge(entry);
2280 entry = arch_make_huge_pte(entry, vma, page, writable);
2285 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2286 unsigned long address, pte_t *ptep)
2290 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2291 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2292 update_mmu_cache(vma, address, ptep);
2296 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2297 struct vm_area_struct *vma)
2299 pte_t *src_pte, *dst_pte, entry;
2300 struct page *ptepage;
2303 struct hstate *h = hstate_vma(vma);
2304 unsigned long sz = huge_page_size(h);
2306 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2308 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2309 src_pte = huge_pte_offset(src, addr);
2312 dst_pte = huge_pte_alloc(dst, addr, sz);
2316 /* If the pagetables are shared don't copy or take references */
2317 if (dst_pte == src_pte)
2320 spin_lock(&dst->page_table_lock);
2321 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2322 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2324 huge_ptep_set_wrprotect(src, addr, src_pte);
2325 entry = huge_ptep_get(src_pte);
2326 ptepage = pte_page(entry);
2328 page_dup_rmap(ptepage);
2329 set_huge_pte_at(dst, addr, dst_pte, entry);
2331 spin_unlock(&src->page_table_lock);
2332 spin_unlock(&dst->page_table_lock);
2340 static int is_hugetlb_entry_migration(pte_t pte)
2344 if (huge_pte_none(pte) || pte_present(pte))
2346 swp = pte_to_swp_entry(pte);
2347 if (non_swap_entry(swp) && is_migration_entry(swp))
2353 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2357 if (huge_pte_none(pte) || pte_present(pte))
2359 swp = pte_to_swp_entry(pte);
2360 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2366 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2367 unsigned long start, unsigned long end,
2368 struct page *ref_page)
2370 int force_flush = 0;
2371 struct mm_struct *mm = vma->vm_mm;
2372 unsigned long address;
2376 struct hstate *h = hstate_vma(vma);
2377 unsigned long sz = huge_page_size(h);
2378 const unsigned long mmun_start = start; /* For mmu_notifiers */
2379 const unsigned long mmun_end = end; /* For mmu_notifiers */
2381 WARN_ON(!is_vm_hugetlb_page(vma));
2382 BUG_ON(start & ~huge_page_mask(h));
2383 BUG_ON(end & ~huge_page_mask(h));
2385 tlb_start_vma(tlb, vma);
2386 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2388 spin_lock(&mm->page_table_lock);
2389 for (address = start; address < end; address += sz) {
2390 ptep = huge_pte_offset(mm, address);
2394 if (huge_pmd_unshare(mm, &address, ptep))
2397 pte = huge_ptep_get(ptep);
2398 if (huge_pte_none(pte))
2402 * HWPoisoned hugepage is already unmapped and dropped reference
2404 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
2405 huge_pte_clear(mm, address, ptep);
2409 page = pte_page(pte);
2411 * If a reference page is supplied, it is because a specific
2412 * page is being unmapped, not a range. Ensure the page we
2413 * are about to unmap is the actual page of interest.
2416 if (page != ref_page)
2420 * Mark the VMA as having unmapped its page so that
2421 * future faults in this VMA will fail rather than
2422 * looking like data was lost
2424 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2427 pte = huge_ptep_get_and_clear(mm, address, ptep);
2428 tlb_remove_tlb_entry(tlb, ptep, address);
2429 if (huge_pte_dirty(pte))
2430 set_page_dirty(page);
2432 page_remove_rmap(page);
2433 force_flush = !__tlb_remove_page(tlb, page);
2436 /* Bail out after unmapping reference page if supplied */
2440 spin_unlock(&mm->page_table_lock);
2442 * mmu_gather ran out of room to batch pages, we break out of
2443 * the PTE lock to avoid doing the potential expensive TLB invalidate
2444 * and page-free while holding it.
2449 if (address < end && !ref_page)
2452 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2453 tlb_end_vma(tlb, vma);
2456 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2457 struct vm_area_struct *vma, unsigned long start,
2458 unsigned long end, struct page *ref_page)
2460 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2463 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2464 * test will fail on a vma being torn down, and not grab a page table
2465 * on its way out. We're lucky that the flag has such an appropriate
2466 * name, and can in fact be safely cleared here. We could clear it
2467 * before the __unmap_hugepage_range above, but all that's necessary
2468 * is to clear it before releasing the i_mmap_mutex. This works
2469 * because in the context this is called, the VMA is about to be
2470 * destroyed and the i_mmap_mutex is held.
2472 vma->vm_flags &= ~VM_MAYSHARE;
2475 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2476 unsigned long end, struct page *ref_page)
2478 struct mm_struct *mm;
2479 struct mmu_gather tlb;
2483 tlb_gather_mmu(&tlb, mm, start, end);
2484 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2485 tlb_finish_mmu(&tlb, start, end);
2489 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2490 * mappping it owns the reserve page for. The intention is to unmap the page
2491 * from other VMAs and let the children be SIGKILLed if they are faulting the
2494 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2495 struct page *page, unsigned long address)
2497 struct hstate *h = hstate_vma(vma);
2498 struct vm_area_struct *iter_vma;
2499 struct address_space *mapping;
2503 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2504 * from page cache lookup which is in HPAGE_SIZE units.
2506 address = address & huge_page_mask(h);
2507 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2509 mapping = file_inode(vma->vm_file)->i_mapping;
2512 * Take the mapping lock for the duration of the table walk. As
2513 * this mapping should be shared between all the VMAs,
2514 * __unmap_hugepage_range() is called as the lock is already held
2516 mutex_lock(&mapping->i_mmap_mutex);
2517 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2518 /* Do not unmap the current VMA */
2519 if (iter_vma == vma)
2523 * Unmap the page from other VMAs without their own reserves.
2524 * They get marked to be SIGKILLed if they fault in these
2525 * areas. This is because a future no-page fault on this VMA
2526 * could insert a zeroed page instead of the data existing
2527 * from the time of fork. This would look like data corruption
2529 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2530 unmap_hugepage_range(iter_vma, address,
2531 address + huge_page_size(h), page);
2533 mutex_unlock(&mapping->i_mmap_mutex);
2539 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2540 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2541 * cannot race with other handlers or page migration.
2542 * Keep the pte_same checks anyway to make transition from the mutex easier.
2544 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2545 unsigned long address, pte_t *ptep, pte_t pte,
2546 struct page *pagecache_page)
2548 struct hstate *h = hstate_vma(vma);
2549 struct page *old_page, *new_page;
2550 int outside_reserve = 0;
2551 unsigned long mmun_start; /* For mmu_notifiers */
2552 unsigned long mmun_end; /* For mmu_notifiers */
2554 old_page = pte_page(pte);
2557 /* If no-one else is actually using this page, avoid the copy
2558 * and just make the page writable */
2559 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2560 page_move_anon_rmap(old_page, vma, address);
2561 set_huge_ptep_writable(vma, address, ptep);
2566 * If the process that created a MAP_PRIVATE mapping is about to
2567 * perform a COW due to a shared page count, attempt to satisfy
2568 * the allocation without using the existing reserves. The pagecache
2569 * page is used to determine if the reserve at this address was
2570 * consumed or not. If reserves were used, a partial faulted mapping
2571 * at the time of fork() could consume its reserves on COW instead
2572 * of the full address range.
2574 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2575 old_page != pagecache_page)
2576 outside_reserve = 1;
2578 page_cache_get(old_page);
2580 /* Drop page_table_lock as buddy allocator may be called */
2581 spin_unlock(&mm->page_table_lock);
2582 new_page = alloc_huge_page(vma, address, outside_reserve);
2584 if (IS_ERR(new_page)) {
2585 long err = PTR_ERR(new_page);
2586 page_cache_release(old_page);
2589 * If a process owning a MAP_PRIVATE mapping fails to COW,
2590 * it is due to references held by a child and an insufficient
2591 * huge page pool. To guarantee the original mappers
2592 * reliability, unmap the page from child processes. The child
2593 * may get SIGKILLed if it later faults.
2595 if (outside_reserve) {
2596 BUG_ON(huge_pte_none(pte));
2597 if (unmap_ref_private(mm, vma, old_page, address)) {
2598 BUG_ON(huge_pte_none(pte));
2599 spin_lock(&mm->page_table_lock);
2600 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2601 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2602 goto retry_avoidcopy;
2604 * race occurs while re-acquiring page_table_lock, and
2612 /* Caller expects lock to be held */
2613 spin_lock(&mm->page_table_lock);
2615 return VM_FAULT_OOM;
2617 return VM_FAULT_SIGBUS;
2621 * When the original hugepage is shared one, it does not have
2622 * anon_vma prepared.
2624 if (unlikely(anon_vma_prepare(vma))) {
2625 page_cache_release(new_page);
2626 page_cache_release(old_page);
2627 /* Caller expects lock to be held */
2628 spin_lock(&mm->page_table_lock);
2629 return VM_FAULT_OOM;
2632 copy_user_huge_page(new_page, old_page, address, vma,
2633 pages_per_huge_page(h));
2634 __SetPageUptodate(new_page);
2636 mmun_start = address & huge_page_mask(h);
2637 mmun_end = mmun_start + huge_page_size(h);
2638 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2640 * Retake the page_table_lock to check for racing updates
2641 * before the page tables are altered
2643 spin_lock(&mm->page_table_lock);
2644 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2645 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2646 ClearPagePrivate(new_page);
2649 huge_ptep_clear_flush(vma, address, ptep);
2650 set_huge_pte_at(mm, address, ptep,
2651 make_huge_pte(vma, new_page, 1));
2652 page_remove_rmap(old_page);
2653 hugepage_add_new_anon_rmap(new_page, vma, address);
2654 /* Make the old page be freed below */
2655 new_page = old_page;
2657 spin_unlock(&mm->page_table_lock);
2658 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2659 page_cache_release(new_page);
2660 page_cache_release(old_page);
2662 /* Caller expects lock to be held */
2663 spin_lock(&mm->page_table_lock);
2667 /* Return the pagecache page at a given address within a VMA */
2668 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2669 struct vm_area_struct *vma, unsigned long address)
2671 struct address_space *mapping;
2674 mapping = vma->vm_file->f_mapping;
2675 idx = vma_hugecache_offset(h, vma, address);
2677 return find_lock_page(mapping, idx);
2681 * Return whether there is a pagecache page to back given address within VMA.
2682 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2684 static bool hugetlbfs_pagecache_present(struct hstate *h,
2685 struct vm_area_struct *vma, unsigned long address)
2687 struct address_space *mapping;
2691 mapping = vma->vm_file->f_mapping;
2692 idx = vma_hugecache_offset(h, vma, address);
2694 page = find_get_page(mapping, idx);
2697 return page != NULL;
2700 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2701 unsigned long address, pte_t *ptep, unsigned int flags)
2703 struct hstate *h = hstate_vma(vma);
2704 int ret = VM_FAULT_SIGBUS;
2709 struct address_space *mapping;
2713 * Currently, we are forced to kill the process in the event the
2714 * original mapper has unmapped pages from the child due to a failed
2715 * COW. Warn that such a situation has occurred as it may not be obvious
2717 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2718 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2723 mapping = vma->vm_file->f_mapping;
2724 idx = vma_hugecache_offset(h, vma, address);
2727 * Use page lock to guard against racing truncation
2728 * before we get page_table_lock.
2731 page = find_lock_page(mapping, idx);
2733 size = i_size_read(mapping->host) >> huge_page_shift(h);
2736 page = alloc_huge_page(vma, address, 0);
2738 ret = PTR_ERR(page);
2742 ret = VM_FAULT_SIGBUS;
2745 clear_huge_page(page, address, pages_per_huge_page(h));
2746 __SetPageUptodate(page);
2748 if (vma->vm_flags & VM_MAYSHARE) {
2750 struct inode *inode = mapping->host;
2752 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2759 ClearPagePrivate(page);
2761 spin_lock(&inode->i_lock);
2762 inode->i_blocks += blocks_per_huge_page(h);
2763 spin_unlock(&inode->i_lock);
2766 if (unlikely(anon_vma_prepare(vma))) {
2768 goto backout_unlocked;
2774 * If memory error occurs between mmap() and fault, some process
2775 * don't have hwpoisoned swap entry for errored virtual address.
2776 * So we need to block hugepage fault by PG_hwpoison bit check.
2778 if (unlikely(PageHWPoison(page))) {
2779 ret = VM_FAULT_HWPOISON |
2780 VM_FAULT_SET_HINDEX(hstate_index(h));
2781 goto backout_unlocked;
2786 * If we are going to COW a private mapping later, we examine the
2787 * pending reservations for this page now. This will ensure that
2788 * any allocations necessary to record that reservation occur outside
2791 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2792 if (vma_needs_reservation(h, vma, address) < 0) {
2794 goto backout_unlocked;
2797 spin_lock(&mm->page_table_lock);
2798 size = i_size_read(mapping->host) >> huge_page_shift(h);
2803 if (!huge_pte_none(huge_ptep_get(ptep)))
2807 ClearPagePrivate(page);
2808 hugepage_add_new_anon_rmap(page, vma, address);
2811 page_dup_rmap(page);
2812 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2813 && (vma->vm_flags & VM_SHARED)));
2814 set_huge_pte_at(mm, address, ptep, new_pte);
2816 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2817 /* Optimization, do the COW without a second fault */
2818 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2821 spin_unlock(&mm->page_table_lock);
2827 spin_unlock(&mm->page_table_lock);
2834 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2835 unsigned long address, unsigned int flags)
2840 struct page *page = NULL;
2841 struct page *pagecache_page = NULL;
2842 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2843 struct hstate *h = hstate_vma(vma);
2845 address &= huge_page_mask(h);
2847 ptep = huge_pte_offset(mm, address);
2849 entry = huge_ptep_get(ptep);
2850 if (unlikely(is_hugetlb_entry_migration(entry))) {
2851 migration_entry_wait_huge(mm, ptep);
2853 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2854 return VM_FAULT_HWPOISON_LARGE |
2855 VM_FAULT_SET_HINDEX(hstate_index(h));
2858 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2860 return VM_FAULT_OOM;
2863 * Serialize hugepage allocation and instantiation, so that we don't
2864 * get spurious allocation failures if two CPUs race to instantiate
2865 * the same page in the page cache.
2867 mutex_lock(&hugetlb_instantiation_mutex);
2868 entry = huge_ptep_get(ptep);
2869 if (huge_pte_none(entry)) {
2870 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2877 * If we are going to COW the mapping later, we examine the pending
2878 * reservations for this page now. This will ensure that any
2879 * allocations necessary to record that reservation occur outside the
2880 * spinlock. For private mappings, we also lookup the pagecache
2881 * page now as it is used to determine if a reservation has been
2884 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
2885 if (vma_needs_reservation(h, vma, address) < 0) {
2890 if (!(vma->vm_flags & VM_MAYSHARE))
2891 pagecache_page = hugetlbfs_pagecache_page(h,
2896 * hugetlb_cow() requires page locks of pte_page(entry) and
2897 * pagecache_page, so here we need take the former one
2898 * when page != pagecache_page or !pagecache_page.
2899 * Note that locking order is always pagecache_page -> page,
2900 * so no worry about deadlock.
2902 page = pte_page(entry);
2904 if (page != pagecache_page)
2907 spin_lock(&mm->page_table_lock);
2908 /* Check for a racing update before calling hugetlb_cow */
2909 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2910 goto out_page_table_lock;
2913 if (flags & FAULT_FLAG_WRITE) {
2914 if (!huge_pte_write(entry)) {
2915 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2917 goto out_page_table_lock;
2919 entry = huge_pte_mkdirty(entry);
2921 entry = pte_mkyoung(entry);
2922 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2923 flags & FAULT_FLAG_WRITE))
2924 update_mmu_cache(vma, address, ptep);
2926 out_page_table_lock:
2927 spin_unlock(&mm->page_table_lock);
2929 if (pagecache_page) {
2930 unlock_page(pagecache_page);
2931 put_page(pagecache_page);
2933 if (page != pagecache_page)
2938 mutex_unlock(&hugetlb_instantiation_mutex);
2943 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2944 struct page **pages, struct vm_area_struct **vmas,
2945 unsigned long *position, unsigned long *nr_pages,
2946 long i, unsigned int flags)
2948 unsigned long pfn_offset;
2949 unsigned long vaddr = *position;
2950 unsigned long remainder = *nr_pages;
2951 struct hstate *h = hstate_vma(vma);
2953 spin_lock(&mm->page_table_lock);
2954 while (vaddr < vma->vm_end && remainder) {
2960 * Some archs (sparc64, sh*) have multiple pte_ts to
2961 * each hugepage. We have to make sure we get the
2962 * first, for the page indexing below to work.
2964 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2965 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2968 * When coredumping, it suits get_dump_page if we just return
2969 * an error where there's an empty slot with no huge pagecache
2970 * to back it. This way, we avoid allocating a hugepage, and
2971 * the sparse dumpfile avoids allocating disk blocks, but its
2972 * huge holes still show up with zeroes where they need to be.
2974 if (absent && (flags & FOLL_DUMP) &&
2975 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2981 * We need call hugetlb_fault for both hugepages under migration
2982 * (in which case hugetlb_fault waits for the migration,) and
2983 * hwpoisoned hugepages (in which case we need to prevent the
2984 * caller from accessing to them.) In order to do this, we use
2985 * here is_swap_pte instead of is_hugetlb_entry_migration and
2986 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
2987 * both cases, and because we can't follow correct pages
2988 * directly from any kind of swap entries.
2990 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
2991 ((flags & FOLL_WRITE) &&
2992 !huge_pte_write(huge_ptep_get(pte)))) {
2995 spin_unlock(&mm->page_table_lock);
2996 ret = hugetlb_fault(mm, vma, vaddr,
2997 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2998 spin_lock(&mm->page_table_lock);
2999 if (!(ret & VM_FAULT_ERROR))
3006 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3007 page = pte_page(huge_ptep_get(pte));
3010 pages[i] = mem_map_offset(page, pfn_offset);
3021 if (vaddr < vma->vm_end && remainder &&
3022 pfn_offset < pages_per_huge_page(h)) {
3024 * We use pfn_offset to avoid touching the pageframes
3025 * of this compound page.
3030 spin_unlock(&mm->page_table_lock);
3031 *nr_pages = remainder;
3034 return i ? i : -EFAULT;
3037 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3038 unsigned long address, unsigned long end, pgprot_t newprot)
3040 struct mm_struct *mm = vma->vm_mm;
3041 unsigned long start = address;
3044 struct hstate *h = hstate_vma(vma);
3045 unsigned long pages = 0;
3047 BUG_ON(address >= end);
3048 flush_cache_range(vma, address, end);
3050 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3051 spin_lock(&mm->page_table_lock);
3052 for (; address < end; address += huge_page_size(h)) {
3053 ptep = huge_pte_offset(mm, address);
3056 if (huge_pmd_unshare(mm, &address, ptep)) {
3060 if (!huge_pte_none(huge_ptep_get(ptep))) {
3061 pte = huge_ptep_get_and_clear(mm, address, ptep);
3062 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3063 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3064 set_huge_pte_at(mm, address, ptep, pte);
3068 spin_unlock(&mm->page_table_lock);
3070 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3071 * may have cleared our pud entry and done put_page on the page table:
3072 * once we release i_mmap_mutex, another task can do the final put_page
3073 * and that page table be reused and filled with junk.
3075 flush_tlb_range(vma, start, end);
3076 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3078 return pages << h->order;
3081 int hugetlb_reserve_pages(struct inode *inode,
3083 struct vm_area_struct *vma,
3084 vm_flags_t vm_flags)
3087 struct hstate *h = hstate_inode(inode);
3088 struct hugepage_subpool *spool = subpool_inode(inode);
3091 * Only apply hugepage reservation if asked. At fault time, an
3092 * attempt will be made for VM_NORESERVE to allocate a page
3093 * without using reserves
3095 if (vm_flags & VM_NORESERVE)
3099 * Shared mappings base their reservation on the number of pages that
3100 * are already allocated on behalf of the file. Private mappings need
3101 * to reserve the full area even if read-only as mprotect() may be
3102 * called to make the mapping read-write. Assume !vma is a shm mapping
3104 if (!vma || vma->vm_flags & VM_MAYSHARE)
3105 chg = region_chg(&inode->i_mapping->private_list, from, to);
3107 struct resv_map *resv_map = resv_map_alloc();
3113 set_vma_resv_map(vma, resv_map);
3114 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3122 /* There must be enough pages in the subpool for the mapping */
3123 if (hugepage_subpool_get_pages(spool, chg)) {
3129 * Check enough hugepages are available for the reservation.
3130 * Hand the pages back to the subpool if there are not
3132 ret = hugetlb_acct_memory(h, chg);
3134 hugepage_subpool_put_pages(spool, chg);
3139 * Account for the reservations made. Shared mappings record regions
3140 * that have reservations as they are shared by multiple VMAs.
3141 * When the last VMA disappears, the region map says how much
3142 * the reservation was and the page cache tells how much of
3143 * the reservation was consumed. Private mappings are per-VMA and
3144 * only the consumed reservations are tracked. When the VMA
3145 * disappears, the original reservation is the VMA size and the
3146 * consumed reservations are stored in the map. Hence, nothing
3147 * else has to be done for private mappings here
3149 if (!vma || vma->vm_flags & VM_MAYSHARE)
3150 region_add(&inode->i_mapping->private_list, from, to);
3158 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3160 struct hstate *h = hstate_inode(inode);
3161 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3162 struct hugepage_subpool *spool = subpool_inode(inode);
3164 spin_lock(&inode->i_lock);
3165 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3166 spin_unlock(&inode->i_lock);
3168 hugepage_subpool_put_pages(spool, (chg - freed));
3169 hugetlb_acct_memory(h, -(chg - freed));
3172 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3173 static unsigned long page_table_shareable(struct vm_area_struct *svma,
3174 struct vm_area_struct *vma,
3175 unsigned long addr, pgoff_t idx)
3177 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3179 unsigned long sbase = saddr & PUD_MASK;
3180 unsigned long s_end = sbase + PUD_SIZE;
3182 /* Allow segments to share if only one is marked locked */
3183 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3184 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3187 * match the virtual addresses, permission and the alignment of the
3190 if (pmd_index(addr) != pmd_index(saddr) ||
3191 vm_flags != svm_flags ||
3192 sbase < svma->vm_start || svma->vm_end < s_end)
3198 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3200 unsigned long base = addr & PUD_MASK;
3201 unsigned long end = base + PUD_SIZE;
3204 * check on proper vm_flags and page table alignment
3206 if (vma->vm_flags & VM_MAYSHARE &&
3207 vma->vm_start <= base && end <= vma->vm_end)
3213 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3214 * and returns the corresponding pte. While this is not necessary for the
3215 * !shared pmd case because we can allocate the pmd later as well, it makes the
3216 * code much cleaner. pmd allocation is essential for the shared case because
3217 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3218 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3219 * bad pmd for sharing.
3221 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3223 struct vm_area_struct *vma = find_vma(mm, addr);
3224 struct address_space *mapping = vma->vm_file->f_mapping;
3225 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3227 struct vm_area_struct *svma;
3228 unsigned long saddr;
3232 if (!vma_shareable(vma, addr))
3233 return (pte_t *)pmd_alloc(mm, pud, addr);
3235 mutex_lock(&mapping->i_mmap_mutex);
3236 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3240 saddr = page_table_shareable(svma, vma, addr, idx);
3242 spte = huge_pte_offset(svma->vm_mm, saddr);
3244 get_page(virt_to_page(spte));
3253 spin_lock(&mm->page_table_lock);
3255 pud_populate(mm, pud,
3256 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3258 put_page(virt_to_page(spte));
3259 spin_unlock(&mm->page_table_lock);
3261 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3262 mutex_unlock(&mapping->i_mmap_mutex);
3267 * unmap huge page backed by shared pte.
3269 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3270 * indicated by page_count > 1, unmap is achieved by clearing pud and
3271 * decrementing the ref count. If count == 1, the pte page is not shared.
3273 * called with vma->vm_mm->page_table_lock held.
3275 * returns: 1 successfully unmapped a shared pte page
3276 * 0 the underlying pte page is not shared, or it is the last user
3278 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3280 pgd_t *pgd = pgd_offset(mm, *addr);
3281 pud_t *pud = pud_offset(pgd, *addr);
3283 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3284 if (page_count(virt_to_page(ptep)) == 1)
3288 put_page(virt_to_page(ptep));
3289 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3292 #define want_pmd_share() (1)
3293 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3294 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3298 #define want_pmd_share() (0)
3299 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3301 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3302 pte_t *huge_pte_alloc(struct mm_struct *mm,
3303 unsigned long addr, unsigned long sz)
3309 pgd = pgd_offset(mm, addr);
3310 pud = pud_alloc(mm, pgd, addr);
3312 if (sz == PUD_SIZE) {
3315 BUG_ON(sz != PMD_SIZE);
3316 if (want_pmd_share() && pud_none(*pud))
3317 pte = huge_pmd_share(mm, addr, pud);
3319 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3322 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3327 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3333 pgd = pgd_offset(mm, addr);
3334 if (pgd_present(*pgd)) {
3335 pud = pud_offset(pgd, addr);
3336 if (pud_present(*pud)) {
3338 return (pte_t *)pud;
3339 pmd = pmd_offset(pud, addr);
3342 return (pte_t *) pmd;
3346 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3347 pmd_t *pmd, int write)
3351 page = pte_page(*(pte_t *)pmd);
3353 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
3358 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3359 pud_t *pud, int write)
3363 page = pte_page(*(pte_t *)pud);
3365 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
3369 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3371 /* Can be overriden by architectures */
3372 __attribute__((weak)) struct page *
3373 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3374 pud_t *pud, int write)
3380 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3382 #ifdef CONFIG_MEMORY_FAILURE
3384 /* Should be called in hugetlb_lock */
3385 static int is_hugepage_on_freelist(struct page *hpage)
3389 struct hstate *h = page_hstate(hpage);
3390 int nid = page_to_nid(hpage);
3392 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3399 * This function is called from memory failure code.
3400 * Assume the caller holds page lock of the head page.
3402 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3404 struct hstate *h = page_hstate(hpage);
3405 int nid = page_to_nid(hpage);
3408 spin_lock(&hugetlb_lock);
3409 if (is_hugepage_on_freelist(hpage)) {
3411 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3412 * but dangling hpage->lru can trigger list-debug warnings
3413 * (this happens when we call unpoison_memory() on it),
3414 * so let it point to itself with list_del_init().
3416 list_del_init(&hpage->lru);
3417 set_page_refcounted(hpage);
3418 h->free_huge_pages--;
3419 h->free_huge_pages_node[nid]--;
3422 spin_unlock(&hugetlb_lock);
3427 bool isolate_huge_page(struct page *page, struct list_head *list)
3429 VM_BUG_ON(!PageHead(page));
3430 if (!get_page_unless_zero(page))
3432 spin_lock(&hugetlb_lock);
3433 list_move_tail(&page->lru, list);
3434 spin_unlock(&hugetlb_lock);
3438 void putback_active_hugepage(struct page *page)
3440 VM_BUG_ON(!PageHead(page));
3441 spin_lock(&hugetlb_lock);
3442 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
3443 spin_unlock(&hugetlb_lock);