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, nr_huge_pages, and free_huge_pages
53 DEFINE_SPINLOCK(hugetlb_lock);
55 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
57 bool free = (spool->count == 0) && (spool->used_hpages == 0);
59 spin_unlock(&spool->lock);
61 /* If no pages are used, and no other handles to the subpool
62 * remain, free the subpool the subpool remain */
67 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
69 struct hugepage_subpool *spool;
71 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
75 spin_lock_init(&spool->lock);
77 spool->max_hpages = nr_blocks;
78 spool->used_hpages = 0;
83 void hugepage_put_subpool(struct hugepage_subpool *spool)
85 spin_lock(&spool->lock);
86 BUG_ON(!spool->count);
88 unlock_or_release_subpool(spool);
91 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
99 spin_lock(&spool->lock);
100 if ((spool->used_hpages + delta) <= spool->max_hpages) {
101 spool->used_hpages += delta;
105 spin_unlock(&spool->lock);
110 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
116 spin_lock(&spool->lock);
117 spool->used_hpages -= delta;
118 /* If hugetlbfs_put_super couldn't free spool due to
119 * an outstanding quota reference, free it now. */
120 unlock_or_release_subpool(spool);
123 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
125 return HUGETLBFS_SB(inode->i_sb)->spool;
128 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
130 return subpool_inode(file_inode(vma->vm_file));
134 * Region tracking -- allows tracking of reservations and instantiated pages
135 * across the pages in a mapping.
137 * The region data structures are protected by a combination of the mmap_sem
138 * and the hugetlb_instantion_mutex. To access or modify a region the caller
139 * must either hold the mmap_sem for write, or the mmap_sem for read and
140 * the hugetlb_instantiation mutex:
142 * down_write(&mm->mmap_sem);
144 * down_read(&mm->mmap_sem);
145 * mutex_lock(&hugetlb_instantiation_mutex);
148 struct list_head link;
153 static long region_add(struct list_head *head, long f, long t)
155 struct file_region *rg, *nrg, *trg;
157 /* Locate the region we are either in or before. */
158 list_for_each_entry(rg, head, link)
162 /* Round our left edge to the current segment if it encloses us. */
166 /* Check for and consume any regions we now overlap with. */
168 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
169 if (&rg->link == head)
174 /* If this area reaches higher then extend our area to
175 * include it completely. If this is not the first area
176 * which we intend to reuse, free it. */
189 static long region_chg(struct list_head *head, long f, long t)
191 struct file_region *rg, *nrg;
194 /* Locate the region we are before or in. */
195 list_for_each_entry(rg, head, link)
199 /* If we are below the current region then a new region is required.
200 * Subtle, allocate a new region at the position but make it zero
201 * size such that we can guarantee to record the reservation. */
202 if (&rg->link == head || t < rg->from) {
203 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
208 INIT_LIST_HEAD(&nrg->link);
209 list_add(&nrg->link, rg->link.prev);
214 /* Round our left edge to the current segment if it encloses us. */
219 /* Check for and consume any regions we now overlap with. */
220 list_for_each_entry(rg, rg->link.prev, link) {
221 if (&rg->link == head)
226 /* We overlap with this area, if it extends further than
227 * us then we must extend ourselves. Account for its
228 * existing reservation. */
233 chg -= rg->to - rg->from;
238 static long region_truncate(struct list_head *head, long end)
240 struct file_region *rg, *trg;
243 /* Locate the region we are either in or before. */
244 list_for_each_entry(rg, head, link)
247 if (&rg->link == head)
250 /* If we are in the middle of a region then adjust it. */
251 if (end > rg->from) {
254 rg = list_entry(rg->link.next, typeof(*rg), link);
257 /* Drop any remaining regions. */
258 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
259 if (&rg->link == head)
261 chg += rg->to - rg->from;
268 static long region_count(struct list_head *head, long f, long t)
270 struct file_region *rg;
273 /* Locate each segment we overlap with, and count that overlap. */
274 list_for_each_entry(rg, head, link) {
283 seg_from = max(rg->from, f);
284 seg_to = min(rg->to, t);
286 chg += seg_to - seg_from;
293 * Convert the address within this vma to the page offset within
294 * the mapping, in pagecache page units; huge pages here.
296 static pgoff_t vma_hugecache_offset(struct hstate *h,
297 struct vm_area_struct *vma, unsigned long address)
299 return ((address - vma->vm_start) >> huge_page_shift(h)) +
300 (vma->vm_pgoff >> huge_page_order(h));
303 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
304 unsigned long address)
306 return vma_hugecache_offset(hstate_vma(vma), vma, address);
310 * Return the size of the pages allocated when backing a VMA. In the majority
311 * cases this will be same size as used by the page table entries.
313 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
315 struct hstate *hstate;
317 if (!is_vm_hugetlb_page(vma))
320 hstate = hstate_vma(vma);
322 return 1UL << (hstate->order + PAGE_SHIFT);
324 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
327 * Return the page size being used by the MMU to back a VMA. In the majority
328 * of cases, the page size used by the kernel matches the MMU size. On
329 * architectures where it differs, an architecture-specific version of this
330 * function is required.
332 #ifndef vma_mmu_pagesize
333 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
335 return vma_kernel_pagesize(vma);
340 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
341 * bits of the reservation map pointer, which are always clear due to
344 #define HPAGE_RESV_OWNER (1UL << 0)
345 #define HPAGE_RESV_UNMAPPED (1UL << 1)
346 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
349 * These helpers are used to track how many pages are reserved for
350 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
351 * is guaranteed to have their future faults succeed.
353 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
354 * the reserve counters are updated with the hugetlb_lock held. It is safe
355 * to reset the VMA at fork() time as it is not in use yet and there is no
356 * chance of the global counters getting corrupted as a result of the values.
358 * The private mapping reservation is represented in a subtly different
359 * manner to a shared mapping. A shared mapping has a region map associated
360 * with the underlying file, this region map represents the backing file
361 * pages which have ever had a reservation assigned which this persists even
362 * after the page is instantiated. A private mapping has a region map
363 * associated with the original mmap which is attached to all VMAs which
364 * reference it, this region map represents those offsets which have consumed
365 * reservation ie. where pages have been instantiated.
367 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
369 return (unsigned long)vma->vm_private_data;
372 static void set_vma_private_data(struct vm_area_struct *vma,
375 vma->vm_private_data = (void *)value;
380 struct list_head regions;
383 static struct resv_map *resv_map_alloc(void)
385 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
389 kref_init(&resv_map->refs);
390 INIT_LIST_HEAD(&resv_map->regions);
395 static void resv_map_release(struct kref *ref)
397 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
399 /* Clear out any active regions before we release the map. */
400 region_truncate(&resv_map->regions, 0);
404 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
406 VM_BUG_ON(!is_vm_hugetlb_page(vma));
407 if (!(vma->vm_flags & VM_MAYSHARE))
408 return (struct resv_map *)(get_vma_private_data(vma) &
413 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
415 VM_BUG_ON(!is_vm_hugetlb_page(vma));
416 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
418 set_vma_private_data(vma, (get_vma_private_data(vma) &
419 HPAGE_RESV_MASK) | (unsigned long)map);
422 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
424 VM_BUG_ON(!is_vm_hugetlb_page(vma));
425 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
427 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
430 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
432 VM_BUG_ON(!is_vm_hugetlb_page(vma));
434 return (get_vma_private_data(vma) & flag) != 0;
437 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
438 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
440 VM_BUG_ON(!is_vm_hugetlb_page(vma));
441 if (!(vma->vm_flags & VM_MAYSHARE))
442 vma->vm_private_data = (void *)0;
445 /* Returns true if the VMA has associated reserve pages */
446 static int vma_has_reserves(struct vm_area_struct *vma, long chg)
448 if (vma->vm_flags & VM_NORESERVE) {
450 * This address is already reserved by other process(chg == 0),
451 * so, we should decrement reserved count. Without decrementing,
452 * reserve count remains after releasing inode, because this
453 * allocated page will go into page cache and is regarded as
454 * coming from reserved pool in releasing step. Currently, we
455 * don't have any other solution to deal with this situation
456 * properly, so add work-around here.
458 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
464 /* Shared mappings always use reserves */
465 if (vma->vm_flags & VM_MAYSHARE)
469 * Only the process that called mmap() has reserves for
472 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
478 static void copy_gigantic_page(struct page *dst, struct page *src)
481 struct hstate *h = page_hstate(src);
482 struct page *dst_base = dst;
483 struct page *src_base = src;
485 for (i = 0; i < pages_per_huge_page(h); ) {
487 copy_highpage(dst, src);
490 dst = mem_map_next(dst, dst_base, i);
491 src = mem_map_next(src, src_base, i);
495 void copy_huge_page(struct page *dst, struct page *src)
498 struct hstate *h = page_hstate(src);
500 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
501 copy_gigantic_page(dst, src);
506 for (i = 0; i < pages_per_huge_page(h); i++) {
508 copy_highpage(dst + i, src + i);
512 static void enqueue_huge_page(struct hstate *h, struct page *page)
514 int nid = page_to_nid(page);
515 list_move(&page->lru, &h->hugepage_freelists[nid]);
516 h->free_huge_pages++;
517 h->free_huge_pages_node[nid]++;
520 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
524 if (list_empty(&h->hugepage_freelists[nid]))
526 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
527 list_move(&page->lru, &h->hugepage_activelist);
528 set_page_refcounted(page);
529 h->free_huge_pages--;
530 h->free_huge_pages_node[nid]--;
534 static struct page *dequeue_huge_page_vma(struct hstate *h,
535 struct vm_area_struct *vma,
536 unsigned long address, int avoid_reserve,
539 struct page *page = NULL;
540 struct mempolicy *mpol;
541 nodemask_t *nodemask;
542 struct zonelist *zonelist;
545 unsigned int cpuset_mems_cookie;
548 cpuset_mems_cookie = get_mems_allowed();
549 zonelist = huge_zonelist(vma, address,
550 htlb_alloc_mask, &mpol, &nodemask);
552 * A child process with MAP_PRIVATE mappings created by their parent
553 * have no page reserves. This check ensures that reservations are
554 * not "stolen". The child may still get SIGKILLed
556 if (!vma_has_reserves(vma, chg) &&
557 h->free_huge_pages - h->resv_huge_pages == 0)
560 /* If reserves cannot be used, ensure enough pages are in the pool */
561 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
564 for_each_zone_zonelist_nodemask(zone, z, zonelist,
565 MAX_NR_ZONES - 1, nodemask) {
566 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
567 page = dequeue_huge_page_node(h, zone_to_nid(zone));
571 if (!vma_has_reserves(vma, chg))
574 h->resv_huge_pages--;
581 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
590 static void update_and_free_page(struct hstate *h, struct page *page)
594 VM_BUG_ON(h->order >= MAX_ORDER);
597 h->nr_huge_pages_node[page_to_nid(page)]--;
598 for (i = 0; i < pages_per_huge_page(h); i++) {
599 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
600 1 << PG_referenced | 1 << PG_dirty |
601 1 << PG_active | 1 << PG_reserved |
602 1 << PG_private | 1 << PG_writeback);
604 VM_BUG_ON(hugetlb_cgroup_from_page(page));
605 set_compound_page_dtor(page, NULL);
606 set_page_refcounted(page);
607 arch_release_hugepage(page);
608 __free_pages(page, huge_page_order(h));
611 struct hstate *size_to_hstate(unsigned long size)
616 if (huge_page_size(h) == size)
622 static void free_huge_page(struct page *page)
625 * Can't pass hstate in here because it is called from the
626 * compound page destructor.
628 struct hstate *h = page_hstate(page);
629 int nid = page_to_nid(page);
630 struct hugepage_subpool *spool =
631 (struct hugepage_subpool *)page_private(page);
633 set_page_private(page, 0);
634 page->mapping = NULL;
635 BUG_ON(page_count(page));
636 BUG_ON(page_mapcount(page));
638 spin_lock(&hugetlb_lock);
639 hugetlb_cgroup_uncharge_page(hstate_index(h),
640 pages_per_huge_page(h), page);
641 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
642 /* remove the page from active list */
643 list_del(&page->lru);
644 update_and_free_page(h, page);
645 h->surplus_huge_pages--;
646 h->surplus_huge_pages_node[nid]--;
648 arch_clear_hugepage_flags(page);
649 enqueue_huge_page(h, page);
651 spin_unlock(&hugetlb_lock);
652 hugepage_subpool_put_pages(spool, 1);
655 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
657 INIT_LIST_HEAD(&page->lru);
658 set_compound_page_dtor(page, free_huge_page);
659 spin_lock(&hugetlb_lock);
660 set_hugetlb_cgroup(page, NULL);
662 h->nr_huge_pages_node[nid]++;
663 spin_unlock(&hugetlb_lock);
664 put_page(page); /* free it into the hugepage allocator */
667 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
670 int nr_pages = 1 << order;
671 struct page *p = page + 1;
673 /* we rely on prep_new_huge_page to set the destructor */
674 set_compound_order(page, order);
676 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
678 set_page_count(p, 0);
679 p->first_page = page;
684 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
685 * transparent huge pages. See the PageTransHuge() documentation for more
688 int PageHuge(struct page *page)
690 compound_page_dtor *dtor;
692 if (!PageCompound(page))
695 page = compound_head(page);
696 dtor = get_compound_page_dtor(page);
698 return dtor == free_huge_page;
700 EXPORT_SYMBOL_GPL(PageHuge);
702 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
706 if (h->order >= MAX_ORDER)
709 page = alloc_pages_exact_node(nid,
710 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
711 __GFP_REPEAT|__GFP_NOWARN,
714 if (arch_prepare_hugepage(page)) {
715 __free_pages(page, huge_page_order(h));
718 prep_new_huge_page(h, page, nid);
725 * common helper functions for hstate_next_node_to_{alloc|free}.
726 * We may have allocated or freed a huge page based on a different
727 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
728 * be outside of *nodes_allowed. Ensure that we use an allowed
729 * node for alloc or free.
731 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
733 nid = next_node(nid, *nodes_allowed);
734 if (nid == MAX_NUMNODES)
735 nid = first_node(*nodes_allowed);
736 VM_BUG_ON(nid >= MAX_NUMNODES);
741 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
743 if (!node_isset(nid, *nodes_allowed))
744 nid = next_node_allowed(nid, nodes_allowed);
749 * returns the previously saved node ["this node"] from which to
750 * allocate a persistent huge page for the pool and advance the
751 * next node from which to allocate, handling wrap at end of node
754 static int hstate_next_node_to_alloc(struct hstate *h,
755 nodemask_t *nodes_allowed)
759 VM_BUG_ON(!nodes_allowed);
761 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
762 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
768 * helper for free_pool_huge_page() - return the previously saved
769 * node ["this node"] from which to free a huge page. Advance the
770 * next node id whether or not we find a free huge page to free so
771 * that the next attempt to free addresses the next node.
773 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
777 VM_BUG_ON(!nodes_allowed);
779 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
780 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
785 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
786 for (nr_nodes = nodes_weight(*mask); \
788 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
791 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
792 for (nr_nodes = nodes_weight(*mask); \
794 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
797 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
803 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
804 page = alloc_fresh_huge_page_node(h, node);
812 count_vm_event(HTLB_BUDDY_PGALLOC);
814 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
820 * Free huge page from pool from next node to free.
821 * Attempt to keep persistent huge pages more or less
822 * balanced over allowed nodes.
823 * Called with hugetlb_lock locked.
825 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
831 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
833 * If we're returning unused surplus pages, only examine
834 * nodes with surplus pages.
836 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
837 !list_empty(&h->hugepage_freelists[node])) {
839 list_entry(h->hugepage_freelists[node].next,
841 list_del(&page->lru);
842 h->free_huge_pages--;
843 h->free_huge_pages_node[node]--;
845 h->surplus_huge_pages--;
846 h->surplus_huge_pages_node[node]--;
848 update_and_free_page(h, page);
857 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
862 if (h->order >= MAX_ORDER)
866 * Assume we will successfully allocate the surplus page to
867 * prevent racing processes from causing the surplus to exceed
870 * This however introduces a different race, where a process B
871 * tries to grow the static hugepage pool while alloc_pages() is
872 * called by process A. B will only examine the per-node
873 * counters in determining if surplus huge pages can be
874 * converted to normal huge pages in adjust_pool_surplus(). A
875 * won't be able to increment the per-node counter, until the
876 * lock is dropped by B, but B doesn't drop hugetlb_lock until
877 * no more huge pages can be converted from surplus to normal
878 * state (and doesn't try to convert again). Thus, we have a
879 * case where a surplus huge page exists, the pool is grown, and
880 * the surplus huge page still exists after, even though it
881 * should just have been converted to a normal huge page. This
882 * does not leak memory, though, as the hugepage will be freed
883 * once it is out of use. It also does not allow the counters to
884 * go out of whack in adjust_pool_surplus() as we don't modify
885 * the node values until we've gotten the hugepage and only the
886 * per-node value is checked there.
888 spin_lock(&hugetlb_lock);
889 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
890 spin_unlock(&hugetlb_lock);
894 h->surplus_huge_pages++;
896 spin_unlock(&hugetlb_lock);
898 if (nid == NUMA_NO_NODE)
899 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
900 __GFP_REPEAT|__GFP_NOWARN,
903 page = alloc_pages_exact_node(nid,
904 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
905 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
907 if (page && arch_prepare_hugepage(page)) {
908 __free_pages(page, huge_page_order(h));
912 spin_lock(&hugetlb_lock);
914 INIT_LIST_HEAD(&page->lru);
915 r_nid = page_to_nid(page);
916 set_compound_page_dtor(page, free_huge_page);
917 set_hugetlb_cgroup(page, NULL);
919 * We incremented the global counters already
921 h->nr_huge_pages_node[r_nid]++;
922 h->surplus_huge_pages_node[r_nid]++;
923 __count_vm_event(HTLB_BUDDY_PGALLOC);
926 h->surplus_huge_pages--;
927 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
929 spin_unlock(&hugetlb_lock);
935 * This allocation function is useful in the context where vma is irrelevant.
936 * E.g. soft-offlining uses this function because it only cares physical
937 * address of error page.
939 struct page *alloc_huge_page_node(struct hstate *h, int nid)
941 struct page *page = NULL;
943 spin_lock(&hugetlb_lock);
944 if (h->free_huge_pages - h->resv_huge_pages > 0)
945 page = dequeue_huge_page_node(h, nid);
946 spin_unlock(&hugetlb_lock);
949 page = alloc_buddy_huge_page(h, nid);
955 * Increase the hugetlb pool such that it can accommodate a reservation
958 static int gather_surplus_pages(struct hstate *h, int delta)
960 struct list_head surplus_list;
961 struct page *page, *tmp;
963 int needed, allocated;
964 bool alloc_ok = true;
966 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
968 h->resv_huge_pages += delta;
973 INIT_LIST_HEAD(&surplus_list);
977 spin_unlock(&hugetlb_lock);
978 for (i = 0; i < needed; i++) {
979 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
984 list_add(&page->lru, &surplus_list);
989 * After retaking hugetlb_lock, we need to recalculate 'needed'
990 * because either resv_huge_pages or free_huge_pages may have changed.
992 spin_lock(&hugetlb_lock);
993 needed = (h->resv_huge_pages + delta) -
994 (h->free_huge_pages + allocated);
999 * We were not able to allocate enough pages to
1000 * satisfy the entire reservation so we free what
1001 * we've allocated so far.
1006 * The surplus_list now contains _at_least_ the number of extra pages
1007 * needed to accommodate the reservation. Add the appropriate number
1008 * of pages to the hugetlb pool and free the extras back to the buddy
1009 * allocator. Commit the entire reservation here to prevent another
1010 * process from stealing the pages as they are added to the pool but
1011 * before they are reserved.
1013 needed += allocated;
1014 h->resv_huge_pages += delta;
1017 /* Free the needed pages to the hugetlb pool */
1018 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1022 * This page is now managed by the hugetlb allocator and has
1023 * no users -- drop the buddy allocator's reference.
1025 put_page_testzero(page);
1026 VM_BUG_ON(page_count(page));
1027 enqueue_huge_page(h, page);
1030 spin_unlock(&hugetlb_lock);
1032 /* Free unnecessary surplus pages to the buddy allocator */
1033 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1035 spin_lock(&hugetlb_lock);
1041 * When releasing a hugetlb pool reservation, any surplus pages that were
1042 * allocated to satisfy the reservation must be explicitly freed if they were
1044 * Called with hugetlb_lock held.
1046 static void return_unused_surplus_pages(struct hstate *h,
1047 unsigned long unused_resv_pages)
1049 unsigned long nr_pages;
1051 /* Uncommit the reservation */
1052 h->resv_huge_pages -= unused_resv_pages;
1054 /* Cannot return gigantic pages currently */
1055 if (h->order >= MAX_ORDER)
1058 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1061 * We want to release as many surplus pages as possible, spread
1062 * evenly across all nodes with memory. Iterate across these nodes
1063 * until we can no longer free unreserved surplus pages. This occurs
1064 * when the nodes with surplus pages have no free pages.
1065 * free_pool_huge_page() will balance the the freed pages across the
1066 * on-line nodes with memory and will handle the hstate accounting.
1068 while (nr_pages--) {
1069 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1075 * Determine if the huge page at addr within the vma has an associated
1076 * reservation. Where it does not we will need to logically increase
1077 * reservation and actually increase subpool usage before an allocation
1078 * can occur. Where any new reservation would be required the
1079 * reservation change is prepared, but not committed. Once the page
1080 * has been allocated from the subpool and instantiated the change should
1081 * be committed via vma_commit_reservation. No action is required on
1084 static long vma_needs_reservation(struct hstate *h,
1085 struct vm_area_struct *vma, unsigned long addr)
1087 struct address_space *mapping = vma->vm_file->f_mapping;
1088 struct inode *inode = mapping->host;
1090 if (vma->vm_flags & VM_MAYSHARE) {
1091 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1092 return region_chg(&inode->i_mapping->private_list,
1095 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1100 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1101 struct resv_map *reservations = vma_resv_map(vma);
1103 err = region_chg(&reservations->regions, idx, idx + 1);
1109 static void vma_commit_reservation(struct hstate *h,
1110 struct vm_area_struct *vma, unsigned long addr)
1112 struct address_space *mapping = vma->vm_file->f_mapping;
1113 struct inode *inode = mapping->host;
1115 if (vma->vm_flags & VM_MAYSHARE) {
1116 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1117 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1119 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1120 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1121 struct resv_map *reservations = vma_resv_map(vma);
1123 /* Mark this page used in the map. */
1124 region_add(&reservations->regions, idx, idx + 1);
1128 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1129 unsigned long addr, int avoid_reserve)
1131 struct hugepage_subpool *spool = subpool_vma(vma);
1132 struct hstate *h = hstate_vma(vma);
1136 struct hugetlb_cgroup *h_cg;
1138 idx = hstate_index(h);
1140 * Processes that did not create the mapping will have no
1141 * reserves and will not have accounted against subpool
1142 * limit. Check that the subpool limit can be made before
1143 * satisfying the allocation MAP_NORESERVE mappings may also
1144 * need pages and subpool limit allocated allocated if no reserve
1147 chg = vma_needs_reservation(h, vma, addr);
1149 return ERR_PTR(-ENOMEM);
1151 if (hugepage_subpool_get_pages(spool, chg))
1152 return ERR_PTR(-ENOSPC);
1154 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1156 hugepage_subpool_put_pages(spool, chg);
1157 return ERR_PTR(-ENOSPC);
1159 spin_lock(&hugetlb_lock);
1160 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
1162 spin_unlock(&hugetlb_lock);
1163 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1165 hugetlb_cgroup_uncharge_cgroup(idx,
1166 pages_per_huge_page(h),
1168 hugepage_subpool_put_pages(spool, chg);
1169 return ERR_PTR(-ENOSPC);
1171 spin_lock(&hugetlb_lock);
1172 list_move(&page->lru, &h->hugepage_activelist);
1175 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1176 spin_unlock(&hugetlb_lock);
1178 set_page_private(page, (unsigned long)spool);
1180 vma_commit_reservation(h, vma, addr);
1184 int __weak alloc_bootmem_huge_page(struct hstate *h)
1186 struct huge_bootmem_page *m;
1189 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1192 addr = __alloc_bootmem_node_nopanic(NODE_DATA(node),
1193 huge_page_size(h), huge_page_size(h), 0);
1197 * Use the beginning of the huge page to store the
1198 * huge_bootmem_page struct (until gather_bootmem
1199 * puts them into the mem_map).
1208 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1209 /* Put them into a private list first because mem_map is not up yet */
1210 list_add(&m->list, &huge_boot_pages);
1215 static void prep_compound_huge_page(struct page *page, int order)
1217 if (unlikely(order > (MAX_ORDER - 1)))
1218 prep_compound_gigantic_page(page, order);
1220 prep_compound_page(page, order);
1223 /* Put bootmem huge pages into the standard lists after mem_map is up */
1224 static void __init gather_bootmem_prealloc(void)
1226 struct huge_bootmem_page *m;
1228 list_for_each_entry(m, &huge_boot_pages, list) {
1229 struct hstate *h = m->hstate;
1232 #ifdef CONFIG_HIGHMEM
1233 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1234 free_bootmem_late((unsigned long)m,
1235 sizeof(struct huge_bootmem_page));
1237 page = virt_to_page(m);
1239 __ClearPageReserved(page);
1240 WARN_ON(page_count(page) != 1);
1241 prep_compound_huge_page(page, h->order);
1242 prep_new_huge_page(h, page, page_to_nid(page));
1244 * If we had gigantic hugepages allocated at boot time, we need
1245 * to restore the 'stolen' pages to totalram_pages in order to
1246 * fix confusing memory reports from free(1) and another
1247 * side-effects, like CommitLimit going negative.
1249 if (h->order > (MAX_ORDER - 1))
1250 totalram_pages += 1 << h->order;
1254 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1258 for (i = 0; i < h->max_huge_pages; ++i) {
1259 if (h->order >= MAX_ORDER) {
1260 if (!alloc_bootmem_huge_page(h))
1262 } else if (!alloc_fresh_huge_page(h,
1263 &node_states[N_MEMORY]))
1266 h->max_huge_pages = i;
1269 static void __init hugetlb_init_hstates(void)
1273 for_each_hstate(h) {
1274 /* oversize hugepages were init'ed in early boot */
1275 if (h->order < MAX_ORDER)
1276 hugetlb_hstate_alloc_pages(h);
1280 static char * __init memfmt(char *buf, unsigned long n)
1282 if (n >= (1UL << 30))
1283 sprintf(buf, "%lu GB", n >> 30);
1284 else if (n >= (1UL << 20))
1285 sprintf(buf, "%lu MB", n >> 20);
1287 sprintf(buf, "%lu KB", n >> 10);
1291 static void __init report_hugepages(void)
1295 for_each_hstate(h) {
1297 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1298 memfmt(buf, huge_page_size(h)),
1299 h->free_huge_pages);
1303 #ifdef CONFIG_HIGHMEM
1304 static void try_to_free_low(struct hstate *h, unsigned long count,
1305 nodemask_t *nodes_allowed)
1309 if (h->order >= MAX_ORDER)
1312 for_each_node_mask(i, *nodes_allowed) {
1313 struct page *page, *next;
1314 struct list_head *freel = &h->hugepage_freelists[i];
1315 list_for_each_entry_safe(page, next, freel, lru) {
1316 if (count >= h->nr_huge_pages)
1318 if (PageHighMem(page))
1320 list_del(&page->lru);
1321 update_and_free_page(h, page);
1322 h->free_huge_pages--;
1323 h->free_huge_pages_node[page_to_nid(page)]--;
1328 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1329 nodemask_t *nodes_allowed)
1335 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1336 * balanced by operating on them in a round-robin fashion.
1337 * Returns 1 if an adjustment was made.
1339 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1344 VM_BUG_ON(delta != -1 && delta != 1);
1347 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1348 if (h->surplus_huge_pages_node[node])
1352 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1353 if (h->surplus_huge_pages_node[node] <
1354 h->nr_huge_pages_node[node])
1361 h->surplus_huge_pages += delta;
1362 h->surplus_huge_pages_node[node] += delta;
1366 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1367 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1368 nodemask_t *nodes_allowed)
1370 unsigned long min_count, ret;
1372 if (h->order >= MAX_ORDER)
1373 return h->max_huge_pages;
1376 * Increase the pool size
1377 * First take pages out of surplus state. Then make up the
1378 * remaining difference by allocating fresh huge pages.
1380 * We might race with alloc_buddy_huge_page() here and be unable
1381 * to convert a surplus huge page to a normal huge page. That is
1382 * not critical, though, it just means the overall size of the
1383 * pool might be one hugepage larger than it needs to be, but
1384 * within all the constraints specified by the sysctls.
1386 spin_lock(&hugetlb_lock);
1387 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1388 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1392 while (count > persistent_huge_pages(h)) {
1394 * If this allocation races such that we no longer need the
1395 * page, free_huge_page will handle it by freeing the page
1396 * and reducing the surplus.
1398 spin_unlock(&hugetlb_lock);
1399 ret = alloc_fresh_huge_page(h, nodes_allowed);
1400 spin_lock(&hugetlb_lock);
1404 /* Bail for signals. Probably ctrl-c from user */
1405 if (signal_pending(current))
1410 * Decrease the pool size
1411 * First return free pages to the buddy allocator (being careful
1412 * to keep enough around to satisfy reservations). Then place
1413 * pages into surplus state as needed so the pool will shrink
1414 * to the desired size as pages become free.
1416 * By placing pages into the surplus state independent of the
1417 * overcommit value, we are allowing the surplus pool size to
1418 * exceed overcommit. There are few sane options here. Since
1419 * alloc_buddy_huge_page() is checking the global counter,
1420 * though, we'll note that we're not allowed to exceed surplus
1421 * and won't grow the pool anywhere else. Not until one of the
1422 * sysctls are changed, or the surplus pages go out of use.
1424 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1425 min_count = max(count, min_count);
1426 try_to_free_low(h, min_count, nodes_allowed);
1427 while (min_count < persistent_huge_pages(h)) {
1428 if (!free_pool_huge_page(h, nodes_allowed, 0))
1431 while (count < persistent_huge_pages(h)) {
1432 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1436 ret = persistent_huge_pages(h);
1437 spin_unlock(&hugetlb_lock);
1441 #define HSTATE_ATTR_RO(_name) \
1442 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1444 #define HSTATE_ATTR(_name) \
1445 static struct kobj_attribute _name##_attr = \
1446 __ATTR(_name, 0644, _name##_show, _name##_store)
1448 static struct kobject *hugepages_kobj;
1449 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1451 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1453 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1457 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1458 if (hstate_kobjs[i] == kobj) {
1460 *nidp = NUMA_NO_NODE;
1464 return kobj_to_node_hstate(kobj, nidp);
1467 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1468 struct kobj_attribute *attr, char *buf)
1471 unsigned long nr_huge_pages;
1474 h = kobj_to_hstate(kobj, &nid);
1475 if (nid == NUMA_NO_NODE)
1476 nr_huge_pages = h->nr_huge_pages;
1478 nr_huge_pages = h->nr_huge_pages_node[nid];
1480 return sprintf(buf, "%lu\n", nr_huge_pages);
1483 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1484 struct kobject *kobj, struct kobj_attribute *attr,
1485 const char *buf, size_t len)
1489 unsigned long count;
1491 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1493 err = strict_strtoul(buf, 10, &count);
1497 h = kobj_to_hstate(kobj, &nid);
1498 if (h->order >= MAX_ORDER) {
1503 if (nid == NUMA_NO_NODE) {
1505 * global hstate attribute
1507 if (!(obey_mempolicy &&
1508 init_nodemask_of_mempolicy(nodes_allowed))) {
1509 NODEMASK_FREE(nodes_allowed);
1510 nodes_allowed = &node_states[N_MEMORY];
1512 } else if (nodes_allowed) {
1514 * per node hstate attribute: adjust count to global,
1515 * but restrict alloc/free to the specified node.
1517 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1518 init_nodemask_of_node(nodes_allowed, nid);
1520 nodes_allowed = &node_states[N_MEMORY];
1522 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1524 if (nodes_allowed != &node_states[N_MEMORY])
1525 NODEMASK_FREE(nodes_allowed);
1529 NODEMASK_FREE(nodes_allowed);
1533 static ssize_t nr_hugepages_show(struct kobject *kobj,
1534 struct kobj_attribute *attr, char *buf)
1536 return nr_hugepages_show_common(kobj, attr, buf);
1539 static ssize_t nr_hugepages_store(struct kobject *kobj,
1540 struct kobj_attribute *attr, const char *buf, size_t len)
1542 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1544 HSTATE_ATTR(nr_hugepages);
1549 * hstate attribute for optionally mempolicy-based constraint on persistent
1550 * huge page alloc/free.
1552 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1553 struct kobj_attribute *attr, char *buf)
1555 return nr_hugepages_show_common(kobj, attr, buf);
1558 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1559 struct kobj_attribute *attr, const char *buf, size_t len)
1561 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1563 HSTATE_ATTR(nr_hugepages_mempolicy);
1567 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1568 struct kobj_attribute *attr, char *buf)
1570 struct hstate *h = kobj_to_hstate(kobj, NULL);
1571 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1574 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1575 struct kobj_attribute *attr, const char *buf, size_t count)
1578 unsigned long input;
1579 struct hstate *h = kobj_to_hstate(kobj, NULL);
1581 if (h->order >= MAX_ORDER)
1584 err = strict_strtoul(buf, 10, &input);
1588 spin_lock(&hugetlb_lock);
1589 h->nr_overcommit_huge_pages = input;
1590 spin_unlock(&hugetlb_lock);
1594 HSTATE_ATTR(nr_overcommit_hugepages);
1596 static ssize_t free_hugepages_show(struct kobject *kobj,
1597 struct kobj_attribute *attr, char *buf)
1600 unsigned long free_huge_pages;
1603 h = kobj_to_hstate(kobj, &nid);
1604 if (nid == NUMA_NO_NODE)
1605 free_huge_pages = h->free_huge_pages;
1607 free_huge_pages = h->free_huge_pages_node[nid];
1609 return sprintf(buf, "%lu\n", free_huge_pages);
1611 HSTATE_ATTR_RO(free_hugepages);
1613 static ssize_t resv_hugepages_show(struct kobject *kobj,
1614 struct kobj_attribute *attr, char *buf)
1616 struct hstate *h = kobj_to_hstate(kobj, NULL);
1617 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1619 HSTATE_ATTR_RO(resv_hugepages);
1621 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1622 struct kobj_attribute *attr, char *buf)
1625 unsigned long surplus_huge_pages;
1628 h = kobj_to_hstate(kobj, &nid);
1629 if (nid == NUMA_NO_NODE)
1630 surplus_huge_pages = h->surplus_huge_pages;
1632 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1634 return sprintf(buf, "%lu\n", surplus_huge_pages);
1636 HSTATE_ATTR_RO(surplus_hugepages);
1638 static struct attribute *hstate_attrs[] = {
1639 &nr_hugepages_attr.attr,
1640 &nr_overcommit_hugepages_attr.attr,
1641 &free_hugepages_attr.attr,
1642 &resv_hugepages_attr.attr,
1643 &surplus_hugepages_attr.attr,
1645 &nr_hugepages_mempolicy_attr.attr,
1650 static struct attribute_group hstate_attr_group = {
1651 .attrs = hstate_attrs,
1654 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1655 struct kobject **hstate_kobjs,
1656 struct attribute_group *hstate_attr_group)
1659 int hi = hstate_index(h);
1661 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1662 if (!hstate_kobjs[hi])
1665 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1667 kobject_put(hstate_kobjs[hi]);
1672 static void __init hugetlb_sysfs_init(void)
1677 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1678 if (!hugepages_kobj)
1681 for_each_hstate(h) {
1682 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1683 hstate_kobjs, &hstate_attr_group);
1685 pr_err("Hugetlb: Unable to add hstate %s", h->name);
1692 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1693 * with node devices in node_devices[] using a parallel array. The array
1694 * index of a node device or _hstate == node id.
1695 * This is here to avoid any static dependency of the node device driver, in
1696 * the base kernel, on the hugetlb module.
1698 struct node_hstate {
1699 struct kobject *hugepages_kobj;
1700 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1702 struct node_hstate node_hstates[MAX_NUMNODES];
1705 * A subset of global hstate attributes for node devices
1707 static struct attribute *per_node_hstate_attrs[] = {
1708 &nr_hugepages_attr.attr,
1709 &free_hugepages_attr.attr,
1710 &surplus_hugepages_attr.attr,
1714 static struct attribute_group per_node_hstate_attr_group = {
1715 .attrs = per_node_hstate_attrs,
1719 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1720 * Returns node id via non-NULL nidp.
1722 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1726 for (nid = 0; nid < nr_node_ids; nid++) {
1727 struct node_hstate *nhs = &node_hstates[nid];
1729 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1730 if (nhs->hstate_kobjs[i] == kobj) {
1742 * Unregister hstate attributes from a single node device.
1743 * No-op if no hstate attributes attached.
1745 static void hugetlb_unregister_node(struct node *node)
1748 struct node_hstate *nhs = &node_hstates[node->dev.id];
1750 if (!nhs->hugepages_kobj)
1751 return; /* no hstate attributes */
1753 for_each_hstate(h) {
1754 int idx = hstate_index(h);
1755 if (nhs->hstate_kobjs[idx]) {
1756 kobject_put(nhs->hstate_kobjs[idx]);
1757 nhs->hstate_kobjs[idx] = NULL;
1761 kobject_put(nhs->hugepages_kobj);
1762 nhs->hugepages_kobj = NULL;
1766 * hugetlb module exit: unregister hstate attributes from node devices
1769 static void hugetlb_unregister_all_nodes(void)
1774 * disable node device registrations.
1776 register_hugetlbfs_with_node(NULL, NULL);
1779 * remove hstate attributes from any nodes that have them.
1781 for (nid = 0; nid < nr_node_ids; nid++)
1782 hugetlb_unregister_node(node_devices[nid]);
1786 * Register hstate attributes for a single node device.
1787 * No-op if attributes already registered.
1789 static void hugetlb_register_node(struct node *node)
1792 struct node_hstate *nhs = &node_hstates[node->dev.id];
1795 if (nhs->hugepages_kobj)
1796 return; /* already allocated */
1798 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1800 if (!nhs->hugepages_kobj)
1803 for_each_hstate(h) {
1804 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1806 &per_node_hstate_attr_group);
1808 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1809 h->name, node->dev.id);
1810 hugetlb_unregister_node(node);
1817 * hugetlb init time: register hstate attributes for all registered node
1818 * devices of nodes that have memory. All on-line nodes should have
1819 * registered their associated device by this time.
1821 static void hugetlb_register_all_nodes(void)
1825 for_each_node_state(nid, N_MEMORY) {
1826 struct node *node = node_devices[nid];
1827 if (node->dev.id == nid)
1828 hugetlb_register_node(node);
1832 * Let the node device driver know we're here so it can
1833 * [un]register hstate attributes on node hotplug.
1835 register_hugetlbfs_with_node(hugetlb_register_node,
1836 hugetlb_unregister_node);
1838 #else /* !CONFIG_NUMA */
1840 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1848 static void hugetlb_unregister_all_nodes(void) { }
1850 static void hugetlb_register_all_nodes(void) { }
1854 static void __exit hugetlb_exit(void)
1858 hugetlb_unregister_all_nodes();
1860 for_each_hstate(h) {
1861 kobject_put(hstate_kobjs[hstate_index(h)]);
1864 kobject_put(hugepages_kobj);
1866 module_exit(hugetlb_exit);
1868 static int __init hugetlb_init(void)
1870 /* Some platform decide whether they support huge pages at boot
1871 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1872 * there is no such support
1874 if (HPAGE_SHIFT == 0)
1877 if (!size_to_hstate(default_hstate_size)) {
1878 default_hstate_size = HPAGE_SIZE;
1879 if (!size_to_hstate(default_hstate_size))
1880 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1882 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1883 if (default_hstate_max_huge_pages)
1884 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1886 hugetlb_init_hstates();
1887 gather_bootmem_prealloc();
1890 hugetlb_sysfs_init();
1891 hugetlb_register_all_nodes();
1892 hugetlb_cgroup_file_init();
1896 module_init(hugetlb_init);
1898 /* Should be called on processing a hugepagesz=... option */
1899 void __init hugetlb_add_hstate(unsigned order)
1904 if (size_to_hstate(PAGE_SIZE << order)) {
1905 pr_warning("hugepagesz= specified twice, ignoring\n");
1908 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1910 h = &hstates[hugetlb_max_hstate++];
1912 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1913 h->nr_huge_pages = 0;
1914 h->free_huge_pages = 0;
1915 for (i = 0; i < MAX_NUMNODES; ++i)
1916 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1917 INIT_LIST_HEAD(&h->hugepage_activelist);
1918 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
1919 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
1920 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1921 huge_page_size(h)/1024);
1926 static int __init hugetlb_nrpages_setup(char *s)
1929 static unsigned long *last_mhp;
1932 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1933 * so this hugepages= parameter goes to the "default hstate".
1935 if (!hugetlb_max_hstate)
1936 mhp = &default_hstate_max_huge_pages;
1938 mhp = &parsed_hstate->max_huge_pages;
1940 if (mhp == last_mhp) {
1941 pr_warning("hugepages= specified twice without "
1942 "interleaving hugepagesz=, ignoring\n");
1946 if (sscanf(s, "%lu", mhp) <= 0)
1950 * Global state is always initialized later in hugetlb_init.
1951 * But we need to allocate >= MAX_ORDER hstates here early to still
1952 * use the bootmem allocator.
1954 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
1955 hugetlb_hstate_alloc_pages(parsed_hstate);
1961 __setup("hugepages=", hugetlb_nrpages_setup);
1963 static int __init hugetlb_default_setup(char *s)
1965 default_hstate_size = memparse(s, &s);
1968 __setup("default_hugepagesz=", hugetlb_default_setup);
1970 static unsigned int cpuset_mems_nr(unsigned int *array)
1973 unsigned int nr = 0;
1975 for_each_node_mask(node, cpuset_current_mems_allowed)
1981 #ifdef CONFIG_SYSCTL
1982 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1983 struct ctl_table *table, int write,
1984 void __user *buffer, size_t *length, loff_t *ppos)
1986 struct hstate *h = &default_hstate;
1990 tmp = h->max_huge_pages;
1992 if (write && h->order >= MAX_ORDER)
1996 table->maxlen = sizeof(unsigned long);
1997 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2002 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2003 GFP_KERNEL | __GFP_NORETRY);
2004 if (!(obey_mempolicy &&
2005 init_nodemask_of_mempolicy(nodes_allowed))) {
2006 NODEMASK_FREE(nodes_allowed);
2007 nodes_allowed = &node_states[N_MEMORY];
2009 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2011 if (nodes_allowed != &node_states[N_MEMORY])
2012 NODEMASK_FREE(nodes_allowed);
2018 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2019 void __user *buffer, size_t *length, loff_t *ppos)
2022 return hugetlb_sysctl_handler_common(false, table, write,
2023 buffer, length, ppos);
2027 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2028 void __user *buffer, size_t *length, loff_t *ppos)
2030 return hugetlb_sysctl_handler_common(true, table, write,
2031 buffer, length, ppos);
2033 #endif /* CONFIG_NUMA */
2035 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2036 void __user *buffer,
2037 size_t *length, loff_t *ppos)
2039 proc_dointvec(table, write, buffer, length, ppos);
2040 if (hugepages_treat_as_movable)
2041 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2043 htlb_alloc_mask = GFP_HIGHUSER;
2047 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2048 void __user *buffer,
2049 size_t *length, loff_t *ppos)
2051 struct hstate *h = &default_hstate;
2055 tmp = h->nr_overcommit_huge_pages;
2057 if (write && h->order >= MAX_ORDER)
2061 table->maxlen = sizeof(unsigned long);
2062 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2067 spin_lock(&hugetlb_lock);
2068 h->nr_overcommit_huge_pages = tmp;
2069 spin_unlock(&hugetlb_lock);
2075 #endif /* CONFIG_SYSCTL */
2077 void hugetlb_report_meminfo(struct seq_file *m)
2079 struct hstate *h = &default_hstate;
2081 "HugePages_Total: %5lu\n"
2082 "HugePages_Free: %5lu\n"
2083 "HugePages_Rsvd: %5lu\n"
2084 "HugePages_Surp: %5lu\n"
2085 "Hugepagesize: %8lu kB\n",
2089 h->surplus_huge_pages,
2090 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2093 int hugetlb_report_node_meminfo(int nid, char *buf)
2095 struct hstate *h = &default_hstate;
2097 "Node %d HugePages_Total: %5u\n"
2098 "Node %d HugePages_Free: %5u\n"
2099 "Node %d HugePages_Surp: %5u\n",
2100 nid, h->nr_huge_pages_node[nid],
2101 nid, h->free_huge_pages_node[nid],
2102 nid, h->surplus_huge_pages_node[nid]);
2105 void hugetlb_show_meminfo(void)
2110 for_each_node_state(nid, N_MEMORY)
2112 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2114 h->nr_huge_pages_node[nid],
2115 h->free_huge_pages_node[nid],
2116 h->surplus_huge_pages_node[nid],
2117 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2120 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2121 unsigned long hugetlb_total_pages(void)
2124 unsigned long nr_total_pages = 0;
2127 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2128 return nr_total_pages;
2131 static int hugetlb_acct_memory(struct hstate *h, long delta)
2135 spin_lock(&hugetlb_lock);
2137 * When cpuset is configured, it breaks the strict hugetlb page
2138 * reservation as the accounting is done on a global variable. Such
2139 * reservation is completely rubbish in the presence of cpuset because
2140 * the reservation is not checked against page availability for the
2141 * current cpuset. Application can still potentially OOM'ed by kernel
2142 * with lack of free htlb page in cpuset that the task is in.
2143 * Attempt to enforce strict accounting with cpuset is almost
2144 * impossible (or too ugly) because cpuset is too fluid that
2145 * task or memory node can be dynamically moved between cpusets.
2147 * The change of semantics for shared hugetlb mapping with cpuset is
2148 * undesirable. However, in order to preserve some of the semantics,
2149 * we fall back to check against current free page availability as
2150 * a best attempt and hopefully to minimize the impact of changing
2151 * semantics that cpuset has.
2154 if (gather_surplus_pages(h, delta) < 0)
2157 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2158 return_unused_surplus_pages(h, delta);
2165 return_unused_surplus_pages(h, (unsigned long) -delta);
2168 spin_unlock(&hugetlb_lock);
2172 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2174 struct resv_map *reservations = vma_resv_map(vma);
2177 * This new VMA should share its siblings reservation map if present.
2178 * The VMA will only ever have a valid reservation map pointer where
2179 * it is being copied for another still existing VMA. As that VMA
2180 * has a reference to the reservation map it cannot disappear until
2181 * after this open call completes. It is therefore safe to take a
2182 * new reference here without additional locking.
2185 kref_get(&reservations->refs);
2188 static void resv_map_put(struct vm_area_struct *vma)
2190 struct resv_map *reservations = vma_resv_map(vma);
2194 kref_put(&reservations->refs, resv_map_release);
2197 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2199 struct hstate *h = hstate_vma(vma);
2200 struct resv_map *reservations = vma_resv_map(vma);
2201 struct hugepage_subpool *spool = subpool_vma(vma);
2202 unsigned long reserve;
2203 unsigned long start;
2207 start = vma_hugecache_offset(h, vma, vma->vm_start);
2208 end = vma_hugecache_offset(h, vma, vma->vm_end);
2210 reserve = (end - start) -
2211 region_count(&reservations->regions, start, end);
2216 hugetlb_acct_memory(h, -reserve);
2217 hugepage_subpool_put_pages(spool, reserve);
2223 * We cannot handle pagefaults against hugetlb pages at all. They cause
2224 * handle_mm_fault() to try to instantiate regular-sized pages in the
2225 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2228 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2234 const struct vm_operations_struct hugetlb_vm_ops = {
2235 .fault = hugetlb_vm_op_fault,
2236 .open = hugetlb_vm_op_open,
2237 .close = hugetlb_vm_op_close,
2240 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2246 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2247 vma->vm_page_prot)));
2249 entry = huge_pte_wrprotect(mk_huge_pte(page,
2250 vma->vm_page_prot));
2252 entry = pte_mkyoung(entry);
2253 entry = pte_mkhuge(entry);
2254 entry = arch_make_huge_pte(entry, vma, page, writable);
2259 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2260 unsigned long address, pte_t *ptep)
2264 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2265 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2266 update_mmu_cache(vma, address, ptep);
2270 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2271 struct vm_area_struct *vma)
2273 pte_t *src_pte, *dst_pte, entry;
2274 struct page *ptepage;
2277 struct hstate *h = hstate_vma(vma);
2278 unsigned long sz = huge_page_size(h);
2280 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2282 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2283 src_pte = huge_pte_offset(src, addr);
2286 dst_pte = huge_pte_alloc(dst, addr, sz);
2290 /* If the pagetables are shared don't copy or take references */
2291 if (dst_pte == src_pte)
2294 spin_lock(&dst->page_table_lock);
2295 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2296 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2298 huge_ptep_set_wrprotect(src, addr, src_pte);
2299 entry = huge_ptep_get(src_pte);
2300 ptepage = pte_page(entry);
2302 page_dup_rmap(ptepage);
2303 set_huge_pte_at(dst, addr, dst_pte, entry);
2305 spin_unlock(&src->page_table_lock);
2306 spin_unlock(&dst->page_table_lock);
2314 static int is_hugetlb_entry_migration(pte_t pte)
2318 if (huge_pte_none(pte) || pte_present(pte))
2320 swp = pte_to_swp_entry(pte);
2321 if (non_swap_entry(swp) && is_migration_entry(swp))
2327 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2331 if (huge_pte_none(pte) || pte_present(pte))
2333 swp = pte_to_swp_entry(pte);
2334 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2340 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2341 unsigned long start, unsigned long end,
2342 struct page *ref_page)
2344 int force_flush = 0;
2345 struct mm_struct *mm = vma->vm_mm;
2346 unsigned long address;
2350 struct hstate *h = hstate_vma(vma);
2351 unsigned long sz = huge_page_size(h);
2352 const unsigned long mmun_start = start; /* For mmu_notifiers */
2353 const unsigned long mmun_end = end; /* For mmu_notifiers */
2355 WARN_ON(!is_vm_hugetlb_page(vma));
2356 BUG_ON(start & ~huge_page_mask(h));
2357 BUG_ON(end & ~huge_page_mask(h));
2359 tlb_start_vma(tlb, vma);
2360 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2362 spin_lock(&mm->page_table_lock);
2363 for (address = start; address < end; address += sz) {
2364 ptep = huge_pte_offset(mm, address);
2368 if (huge_pmd_unshare(mm, &address, ptep))
2371 pte = huge_ptep_get(ptep);
2372 if (huge_pte_none(pte))
2376 * HWPoisoned hugepage is already unmapped and dropped reference
2378 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
2379 huge_pte_clear(mm, address, ptep);
2383 page = pte_page(pte);
2385 * If a reference page is supplied, it is because a specific
2386 * page is being unmapped, not a range. Ensure the page we
2387 * are about to unmap is the actual page of interest.
2390 if (page != ref_page)
2394 * Mark the VMA as having unmapped its page so that
2395 * future faults in this VMA will fail rather than
2396 * looking like data was lost
2398 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2401 pte = huge_ptep_get_and_clear(mm, address, ptep);
2402 tlb_remove_tlb_entry(tlb, ptep, address);
2403 if (huge_pte_dirty(pte))
2404 set_page_dirty(page);
2406 page_remove_rmap(page);
2407 force_flush = !__tlb_remove_page(tlb, page);
2410 /* Bail out after unmapping reference page if supplied */
2414 spin_unlock(&mm->page_table_lock);
2416 * mmu_gather ran out of room to batch pages, we break out of
2417 * the PTE lock to avoid doing the potential expensive TLB invalidate
2418 * and page-free while holding it.
2423 if (address < end && !ref_page)
2426 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2427 tlb_end_vma(tlb, vma);
2430 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2431 struct vm_area_struct *vma, unsigned long start,
2432 unsigned long end, struct page *ref_page)
2434 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2437 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2438 * test will fail on a vma being torn down, and not grab a page table
2439 * on its way out. We're lucky that the flag has such an appropriate
2440 * name, and can in fact be safely cleared here. We could clear it
2441 * before the __unmap_hugepage_range above, but all that's necessary
2442 * is to clear it before releasing the i_mmap_mutex. This works
2443 * because in the context this is called, the VMA is about to be
2444 * destroyed and the i_mmap_mutex is held.
2446 vma->vm_flags &= ~VM_MAYSHARE;
2449 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2450 unsigned long end, struct page *ref_page)
2452 struct mm_struct *mm;
2453 struct mmu_gather tlb;
2457 tlb_gather_mmu(&tlb, mm, 0);
2458 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2459 tlb_finish_mmu(&tlb, start, end);
2463 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2464 * mappping it owns the reserve page for. The intention is to unmap the page
2465 * from other VMAs and let the children be SIGKILLed if they are faulting the
2468 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2469 struct page *page, unsigned long address)
2471 struct hstate *h = hstate_vma(vma);
2472 struct vm_area_struct *iter_vma;
2473 struct address_space *mapping;
2477 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2478 * from page cache lookup which is in HPAGE_SIZE units.
2480 address = address & huge_page_mask(h);
2481 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2483 mapping = file_inode(vma->vm_file)->i_mapping;
2486 * Take the mapping lock for the duration of the table walk. As
2487 * this mapping should be shared between all the VMAs,
2488 * __unmap_hugepage_range() is called as the lock is already held
2490 mutex_lock(&mapping->i_mmap_mutex);
2491 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2492 /* Do not unmap the current VMA */
2493 if (iter_vma == vma)
2497 * Unmap the page from other VMAs without their own reserves.
2498 * They get marked to be SIGKILLed if they fault in these
2499 * areas. This is because a future no-page fault on this VMA
2500 * could insert a zeroed page instead of the data existing
2501 * from the time of fork. This would look like data corruption
2503 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2504 unmap_hugepage_range(iter_vma, address,
2505 address + huge_page_size(h), page);
2507 mutex_unlock(&mapping->i_mmap_mutex);
2513 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2514 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2515 * cannot race with other handlers or page migration.
2516 * Keep the pte_same checks anyway to make transition from the mutex easier.
2518 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2519 unsigned long address, pte_t *ptep, pte_t pte,
2520 struct page *pagecache_page)
2522 struct hstate *h = hstate_vma(vma);
2523 struct page *old_page, *new_page;
2524 int outside_reserve = 0;
2525 unsigned long mmun_start; /* For mmu_notifiers */
2526 unsigned long mmun_end; /* For mmu_notifiers */
2528 old_page = pte_page(pte);
2531 /* If no-one else is actually using this page, avoid the copy
2532 * and just make the page writable */
2533 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2534 page_move_anon_rmap(old_page, vma, address);
2535 set_huge_ptep_writable(vma, address, ptep);
2540 * If the process that created a MAP_PRIVATE mapping is about to
2541 * perform a COW due to a shared page count, attempt to satisfy
2542 * the allocation without using the existing reserves. The pagecache
2543 * page is used to determine if the reserve at this address was
2544 * consumed or not. If reserves were used, a partial faulted mapping
2545 * at the time of fork() could consume its reserves on COW instead
2546 * of the full address range.
2548 if (!(vma->vm_flags & VM_MAYSHARE) &&
2549 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2550 old_page != pagecache_page)
2551 outside_reserve = 1;
2553 page_cache_get(old_page);
2555 /* Drop page_table_lock as buddy allocator may be called */
2556 spin_unlock(&mm->page_table_lock);
2557 new_page = alloc_huge_page(vma, address, outside_reserve);
2559 if (IS_ERR(new_page)) {
2560 long err = PTR_ERR(new_page);
2561 page_cache_release(old_page);
2564 * If a process owning a MAP_PRIVATE mapping fails to COW,
2565 * it is due to references held by a child and an insufficient
2566 * huge page pool. To guarantee the original mappers
2567 * reliability, unmap the page from child processes. The child
2568 * may get SIGKILLed if it later faults.
2570 if (outside_reserve) {
2571 BUG_ON(huge_pte_none(pte));
2572 if (unmap_ref_private(mm, vma, old_page, address)) {
2573 BUG_ON(huge_pte_none(pte));
2574 spin_lock(&mm->page_table_lock);
2575 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2576 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2577 goto retry_avoidcopy;
2579 * race occurs while re-acquiring page_table_lock, and
2587 /* Caller expects lock to be held */
2588 spin_lock(&mm->page_table_lock);
2590 return VM_FAULT_OOM;
2592 return VM_FAULT_SIGBUS;
2596 * When the original hugepage is shared one, it does not have
2597 * anon_vma prepared.
2599 if (unlikely(anon_vma_prepare(vma))) {
2600 page_cache_release(new_page);
2601 page_cache_release(old_page);
2602 /* Caller expects lock to be held */
2603 spin_lock(&mm->page_table_lock);
2604 return VM_FAULT_OOM;
2607 copy_user_huge_page(new_page, old_page, address, vma,
2608 pages_per_huge_page(h));
2609 __SetPageUptodate(new_page);
2611 mmun_start = address & huge_page_mask(h);
2612 mmun_end = mmun_start + huge_page_size(h);
2613 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2615 * Retake the page_table_lock to check for racing updates
2616 * before the page tables are altered
2618 spin_lock(&mm->page_table_lock);
2619 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2620 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2622 huge_ptep_clear_flush(vma, address, ptep);
2623 set_huge_pte_at(mm, address, ptep,
2624 make_huge_pte(vma, new_page, 1));
2625 page_remove_rmap(old_page);
2626 hugepage_add_new_anon_rmap(new_page, vma, address);
2627 /* Make the old page be freed below */
2628 new_page = old_page;
2630 spin_unlock(&mm->page_table_lock);
2631 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2632 /* Caller expects lock to be held */
2633 spin_lock(&mm->page_table_lock);
2634 page_cache_release(new_page);
2635 page_cache_release(old_page);
2639 /* Return the pagecache page at a given address within a VMA */
2640 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2641 struct vm_area_struct *vma, unsigned long address)
2643 struct address_space *mapping;
2646 mapping = vma->vm_file->f_mapping;
2647 idx = vma_hugecache_offset(h, vma, address);
2649 return find_lock_page(mapping, idx);
2653 * Return whether there is a pagecache page to back given address within VMA.
2654 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2656 static bool hugetlbfs_pagecache_present(struct hstate *h,
2657 struct vm_area_struct *vma, unsigned long address)
2659 struct address_space *mapping;
2663 mapping = vma->vm_file->f_mapping;
2664 idx = vma_hugecache_offset(h, vma, address);
2666 page = find_get_page(mapping, idx);
2669 return page != NULL;
2672 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2673 unsigned long address, pte_t *ptep, unsigned int flags)
2675 struct hstate *h = hstate_vma(vma);
2676 int ret = VM_FAULT_SIGBUS;
2681 struct address_space *mapping;
2685 * Currently, we are forced to kill the process in the event the
2686 * original mapper has unmapped pages from the child due to a failed
2687 * COW. Warn that such a situation has occurred as it may not be obvious
2689 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2690 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2695 mapping = vma->vm_file->f_mapping;
2696 idx = vma_hugecache_offset(h, vma, address);
2699 * Use page lock to guard against racing truncation
2700 * before we get page_table_lock.
2703 page = find_lock_page(mapping, idx);
2705 size = i_size_read(mapping->host) >> huge_page_shift(h);
2708 page = alloc_huge_page(vma, address, 0);
2710 ret = PTR_ERR(page);
2714 ret = VM_FAULT_SIGBUS;
2717 clear_huge_page(page, address, pages_per_huge_page(h));
2718 __SetPageUptodate(page);
2720 if (vma->vm_flags & VM_MAYSHARE) {
2722 struct inode *inode = mapping->host;
2724 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2732 spin_lock(&inode->i_lock);
2733 inode->i_blocks += blocks_per_huge_page(h);
2734 spin_unlock(&inode->i_lock);
2737 if (unlikely(anon_vma_prepare(vma))) {
2739 goto backout_unlocked;
2745 * If memory error occurs between mmap() and fault, some process
2746 * don't have hwpoisoned swap entry for errored virtual address.
2747 * So we need to block hugepage fault by PG_hwpoison bit check.
2749 if (unlikely(PageHWPoison(page))) {
2750 ret = VM_FAULT_HWPOISON |
2751 VM_FAULT_SET_HINDEX(hstate_index(h));
2752 goto backout_unlocked;
2757 * If we are going to COW a private mapping later, we examine the
2758 * pending reservations for this page now. This will ensure that
2759 * any allocations necessary to record that reservation occur outside
2762 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2763 if (vma_needs_reservation(h, vma, address) < 0) {
2765 goto backout_unlocked;
2768 spin_lock(&mm->page_table_lock);
2769 size = i_size_read(mapping->host) >> huge_page_shift(h);
2774 if (!huge_pte_none(huge_ptep_get(ptep)))
2778 hugepage_add_new_anon_rmap(page, vma, address);
2780 page_dup_rmap(page);
2781 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2782 && (vma->vm_flags & VM_SHARED)));
2783 set_huge_pte_at(mm, address, ptep, new_pte);
2785 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2786 /* Optimization, do the COW without a second fault */
2787 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2790 spin_unlock(&mm->page_table_lock);
2796 spin_unlock(&mm->page_table_lock);
2803 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2804 unsigned long address, unsigned int flags)
2809 struct page *page = NULL;
2810 struct page *pagecache_page = NULL;
2811 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2812 struct hstate *h = hstate_vma(vma);
2814 address &= huge_page_mask(h);
2816 ptep = huge_pte_offset(mm, address);
2818 entry = huge_ptep_get(ptep);
2819 if (unlikely(is_hugetlb_entry_migration(entry))) {
2820 migration_entry_wait_huge(mm, ptep);
2822 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2823 return VM_FAULT_HWPOISON_LARGE |
2824 VM_FAULT_SET_HINDEX(hstate_index(h));
2827 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2829 return VM_FAULT_OOM;
2832 * Serialize hugepage allocation and instantiation, so that we don't
2833 * get spurious allocation failures if two CPUs race to instantiate
2834 * the same page in the page cache.
2836 mutex_lock(&hugetlb_instantiation_mutex);
2837 entry = huge_ptep_get(ptep);
2838 if (huge_pte_none(entry)) {
2839 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2846 * If we are going to COW the mapping later, we examine the pending
2847 * reservations for this page now. This will ensure that any
2848 * allocations necessary to record that reservation occur outside the
2849 * spinlock. For private mappings, we also lookup the pagecache
2850 * page now as it is used to determine if a reservation has been
2853 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
2854 if (vma_needs_reservation(h, vma, address) < 0) {
2859 if (!(vma->vm_flags & VM_MAYSHARE))
2860 pagecache_page = hugetlbfs_pagecache_page(h,
2865 * hugetlb_cow() requires page locks of pte_page(entry) and
2866 * pagecache_page, so here we need take the former one
2867 * when page != pagecache_page or !pagecache_page.
2868 * Note that locking order is always pagecache_page -> page,
2869 * so no worry about deadlock.
2871 page = pte_page(entry);
2873 if (page != pagecache_page)
2876 spin_lock(&mm->page_table_lock);
2877 /* Check for a racing update before calling hugetlb_cow */
2878 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2879 goto out_page_table_lock;
2882 if (flags & FAULT_FLAG_WRITE) {
2883 if (!huge_pte_write(entry)) {
2884 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2886 goto out_page_table_lock;
2888 entry = huge_pte_mkdirty(entry);
2890 entry = pte_mkyoung(entry);
2891 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2892 flags & FAULT_FLAG_WRITE))
2893 update_mmu_cache(vma, address, ptep);
2895 out_page_table_lock:
2896 spin_unlock(&mm->page_table_lock);
2898 if (pagecache_page) {
2899 unlock_page(pagecache_page);
2900 put_page(pagecache_page);
2902 if (page != pagecache_page)
2907 mutex_unlock(&hugetlb_instantiation_mutex);
2912 /* Can be overriden by architectures */
2913 __attribute__((weak)) struct page *
2914 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2915 pud_t *pud, int write)
2921 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2922 struct page **pages, struct vm_area_struct **vmas,
2923 unsigned long *position, unsigned long *nr_pages,
2924 long i, unsigned int flags)
2926 unsigned long pfn_offset;
2927 unsigned long vaddr = *position;
2928 unsigned long remainder = *nr_pages;
2929 struct hstate *h = hstate_vma(vma);
2931 spin_lock(&mm->page_table_lock);
2932 while (vaddr < vma->vm_end && remainder) {
2938 * Some archs (sparc64, sh*) have multiple pte_ts to
2939 * each hugepage. We have to make sure we get the
2940 * first, for the page indexing below to work.
2942 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2943 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2946 * When coredumping, it suits get_dump_page if we just return
2947 * an error where there's an empty slot with no huge pagecache
2948 * to back it. This way, we avoid allocating a hugepage, and
2949 * the sparse dumpfile avoids allocating disk blocks, but its
2950 * huge holes still show up with zeroes where they need to be.
2952 if (absent && (flags & FOLL_DUMP) &&
2953 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2959 * We need call hugetlb_fault for both hugepages under migration
2960 * (in which case hugetlb_fault waits for the migration,) and
2961 * hwpoisoned hugepages (in which case we need to prevent the
2962 * caller from accessing to them.) In order to do this, we use
2963 * here is_swap_pte instead of is_hugetlb_entry_migration and
2964 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
2965 * both cases, and because we can't follow correct pages
2966 * directly from any kind of swap entries.
2968 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
2969 ((flags & FOLL_WRITE) &&
2970 !huge_pte_write(huge_ptep_get(pte)))) {
2973 spin_unlock(&mm->page_table_lock);
2974 ret = hugetlb_fault(mm, vma, vaddr,
2975 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2976 spin_lock(&mm->page_table_lock);
2977 if (!(ret & VM_FAULT_ERROR))
2984 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2985 page = pte_page(huge_ptep_get(pte));
2988 pages[i] = mem_map_offset(page, pfn_offset);
2999 if (vaddr < vma->vm_end && remainder &&
3000 pfn_offset < pages_per_huge_page(h)) {
3002 * We use pfn_offset to avoid touching the pageframes
3003 * of this compound page.
3008 spin_unlock(&mm->page_table_lock);
3009 *nr_pages = remainder;
3012 return i ? i : -EFAULT;
3015 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3016 unsigned long address, unsigned long end, pgprot_t newprot)
3018 struct mm_struct *mm = vma->vm_mm;
3019 unsigned long start = address;
3022 struct hstate *h = hstate_vma(vma);
3023 unsigned long pages = 0;
3025 BUG_ON(address >= end);
3026 flush_cache_range(vma, address, end);
3028 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3029 spin_lock(&mm->page_table_lock);
3030 for (; address < end; address += huge_page_size(h)) {
3031 ptep = huge_pte_offset(mm, address);
3034 if (huge_pmd_unshare(mm, &address, ptep)) {
3038 if (!huge_pte_none(huge_ptep_get(ptep))) {
3039 pte = huge_ptep_get_and_clear(mm, address, ptep);
3040 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3041 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3042 set_huge_pte_at(mm, address, ptep, pte);
3046 spin_unlock(&mm->page_table_lock);
3048 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3049 * may have cleared our pud entry and done put_page on the page table:
3050 * once we release i_mmap_mutex, another task can do the final put_page
3051 * and that page table be reused and filled with junk.
3053 flush_tlb_range(vma, start, end);
3054 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3056 return pages << h->order;
3059 int hugetlb_reserve_pages(struct inode *inode,
3061 struct vm_area_struct *vma,
3062 vm_flags_t vm_flags)
3065 struct hstate *h = hstate_inode(inode);
3066 struct hugepage_subpool *spool = subpool_inode(inode);
3069 * Only apply hugepage reservation if asked. At fault time, an
3070 * attempt will be made for VM_NORESERVE to allocate a page
3071 * without using reserves
3073 if (vm_flags & VM_NORESERVE)
3077 * Shared mappings base their reservation on the number of pages that
3078 * are already allocated on behalf of the file. Private mappings need
3079 * to reserve the full area even if read-only as mprotect() may be
3080 * called to make the mapping read-write. Assume !vma is a shm mapping
3082 if (!vma || vma->vm_flags & VM_MAYSHARE)
3083 chg = region_chg(&inode->i_mapping->private_list, from, to);
3085 struct resv_map *resv_map = resv_map_alloc();
3091 set_vma_resv_map(vma, resv_map);
3092 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3100 /* There must be enough pages in the subpool for the mapping */
3101 if (hugepage_subpool_get_pages(spool, chg)) {
3107 * Check enough hugepages are available for the reservation.
3108 * Hand the pages back to the subpool if there are not
3110 ret = hugetlb_acct_memory(h, chg);
3112 hugepage_subpool_put_pages(spool, chg);
3117 * Account for the reservations made. Shared mappings record regions
3118 * that have reservations as they are shared by multiple VMAs.
3119 * When the last VMA disappears, the region map says how much
3120 * the reservation was and the page cache tells how much of
3121 * the reservation was consumed. Private mappings are per-VMA and
3122 * only the consumed reservations are tracked. When the VMA
3123 * disappears, the original reservation is the VMA size and the
3124 * consumed reservations are stored in the map. Hence, nothing
3125 * else has to be done for private mappings here
3127 if (!vma || vma->vm_flags & VM_MAYSHARE)
3128 region_add(&inode->i_mapping->private_list, from, to);
3136 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3138 struct hstate *h = hstate_inode(inode);
3139 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3140 struct hugepage_subpool *spool = subpool_inode(inode);
3142 spin_lock(&inode->i_lock);
3143 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3144 spin_unlock(&inode->i_lock);
3146 hugepage_subpool_put_pages(spool, (chg - freed));
3147 hugetlb_acct_memory(h, -(chg - freed));
3150 #ifdef CONFIG_MEMORY_FAILURE
3152 /* Should be called in hugetlb_lock */
3153 static int is_hugepage_on_freelist(struct page *hpage)
3157 struct hstate *h = page_hstate(hpage);
3158 int nid = page_to_nid(hpage);
3160 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3167 * This function is called from memory failure code.
3168 * Assume the caller holds page lock of the head page.
3170 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3172 struct hstate *h = page_hstate(hpage);
3173 int nid = page_to_nid(hpage);
3176 spin_lock(&hugetlb_lock);
3177 if (is_hugepage_on_freelist(hpage)) {
3179 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3180 * but dangling hpage->lru can trigger list-debug warnings
3181 * (this happens when we call unpoison_memory() on it),
3182 * so let it point to itself with list_del_init().
3184 list_del_init(&hpage->lru);
3185 set_page_refcounted(hpage);
3186 h->free_huge_pages--;
3187 h->free_huge_pages_node[nid]--;
3190 spin_unlock(&hugetlb_lock);