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 /* Decrement the reserved pages in the hugepage pool by one */
438 static void decrement_hugepage_resv_vma(struct hstate *h,
439 struct vm_area_struct *vma)
441 if (vma->vm_flags & VM_NORESERVE)
444 if (vma->vm_flags & VM_MAYSHARE) {
445 /* Shared mappings always use reserves */
446 h->resv_huge_pages--;
447 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
449 * Only the process that called mmap() has reserves for
452 h->resv_huge_pages--;
456 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
457 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
459 VM_BUG_ON(!is_vm_hugetlb_page(vma));
460 if (!(vma->vm_flags & VM_MAYSHARE))
461 vma->vm_private_data = (void *)0;
464 /* Returns true if the VMA has associated reserve pages */
465 static int vma_has_reserves(struct vm_area_struct *vma)
467 if (vma->vm_flags & VM_MAYSHARE)
469 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
474 static void copy_gigantic_page(struct page *dst, struct page *src)
477 struct hstate *h = page_hstate(src);
478 struct page *dst_base = dst;
479 struct page *src_base = src;
481 for (i = 0; i < pages_per_huge_page(h); ) {
483 copy_highpage(dst, src);
486 dst = mem_map_next(dst, dst_base, i);
487 src = mem_map_next(src, src_base, i);
491 void copy_huge_page(struct page *dst, struct page *src)
494 struct hstate *h = page_hstate(src);
496 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
497 copy_gigantic_page(dst, src);
502 for (i = 0; i < pages_per_huge_page(h); i++) {
504 copy_highpage(dst + i, src + i);
508 static void enqueue_huge_page(struct hstate *h, struct page *page)
510 int nid = page_to_nid(page);
511 list_move(&page->lru, &h->hugepage_freelists[nid]);
512 h->free_huge_pages++;
513 h->free_huge_pages_node[nid]++;
516 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
520 if (list_empty(&h->hugepage_freelists[nid]))
522 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
523 list_move(&page->lru, &h->hugepage_activelist);
524 set_page_refcounted(page);
525 h->free_huge_pages--;
526 h->free_huge_pages_node[nid]--;
530 static struct page *dequeue_huge_page_vma(struct hstate *h,
531 struct vm_area_struct *vma,
532 unsigned long address, int avoid_reserve)
534 struct page *page = NULL;
535 struct mempolicy *mpol;
536 nodemask_t *nodemask;
537 struct zonelist *zonelist;
540 unsigned int cpuset_mems_cookie;
543 cpuset_mems_cookie = get_mems_allowed();
544 zonelist = huge_zonelist(vma, address,
545 htlb_alloc_mask, &mpol, &nodemask);
547 * A child process with MAP_PRIVATE mappings created by their parent
548 * have no page reserves. This check ensures that reservations are
549 * not "stolen". The child may still get SIGKILLed
551 if (!vma_has_reserves(vma) &&
552 h->free_huge_pages - h->resv_huge_pages == 0)
555 /* If reserves cannot be used, ensure enough pages are in the pool */
556 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
559 for_each_zone_zonelist_nodemask(zone, z, zonelist,
560 MAX_NR_ZONES - 1, nodemask) {
561 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
562 page = dequeue_huge_page_node(h, zone_to_nid(zone));
565 decrement_hugepage_resv_vma(h, vma);
572 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
581 static void update_and_free_page(struct hstate *h, struct page *page)
585 VM_BUG_ON(h->order >= MAX_ORDER);
588 h->nr_huge_pages_node[page_to_nid(page)]--;
589 for (i = 0; i < pages_per_huge_page(h); i++) {
590 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
591 1 << PG_referenced | 1 << PG_dirty |
592 1 << PG_active | 1 << PG_reserved |
593 1 << PG_private | 1 << PG_writeback);
595 VM_BUG_ON(hugetlb_cgroup_from_page(page));
596 set_compound_page_dtor(page, NULL);
597 set_page_refcounted(page);
598 arch_release_hugepage(page);
599 __free_pages(page, huge_page_order(h));
602 struct hstate *size_to_hstate(unsigned long size)
607 if (huge_page_size(h) == size)
613 static void free_huge_page(struct page *page)
616 * Can't pass hstate in here because it is called from the
617 * compound page destructor.
619 struct hstate *h = page_hstate(page);
620 int nid = page_to_nid(page);
621 struct hugepage_subpool *spool =
622 (struct hugepage_subpool *)page_private(page);
624 set_page_private(page, 0);
625 page->mapping = NULL;
626 BUG_ON(page_count(page));
627 BUG_ON(page_mapcount(page));
629 spin_lock(&hugetlb_lock);
630 hugetlb_cgroup_uncharge_page(hstate_index(h),
631 pages_per_huge_page(h), page);
632 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
633 /* remove the page from active list */
634 list_del(&page->lru);
635 update_and_free_page(h, page);
636 h->surplus_huge_pages--;
637 h->surplus_huge_pages_node[nid]--;
639 arch_clear_hugepage_flags(page);
640 enqueue_huge_page(h, page);
642 spin_unlock(&hugetlb_lock);
643 hugepage_subpool_put_pages(spool, 1);
646 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
648 INIT_LIST_HEAD(&page->lru);
649 set_compound_page_dtor(page, free_huge_page);
650 spin_lock(&hugetlb_lock);
651 set_hugetlb_cgroup(page, NULL);
653 h->nr_huge_pages_node[nid]++;
654 spin_unlock(&hugetlb_lock);
655 put_page(page); /* free it into the hugepage allocator */
658 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
661 int nr_pages = 1 << order;
662 struct page *p = page + 1;
664 /* we rely on prep_new_huge_page to set the destructor */
665 set_compound_order(page, order);
667 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
669 set_page_count(p, 0);
670 p->first_page = page;
675 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
676 * transparent huge pages. See the PageTransHuge() documentation for more
679 int PageHuge(struct page *page)
681 compound_page_dtor *dtor;
683 if (!PageCompound(page))
686 page = compound_head(page);
687 dtor = get_compound_page_dtor(page);
689 return dtor == free_huge_page;
691 EXPORT_SYMBOL_GPL(PageHuge);
694 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
695 * normal or transparent huge pages.
697 int PageHeadHuge(struct page *page_head)
699 compound_page_dtor *dtor;
701 if (!PageHead(page_head))
704 dtor = get_compound_page_dtor(page_head);
706 return dtor == free_huge_page;
708 EXPORT_SYMBOL_GPL(PageHeadHuge);
710 pgoff_t __basepage_index(struct page *page)
712 struct page *page_head = compound_head(page);
713 pgoff_t index = page_index(page_head);
714 unsigned long compound_idx;
716 if (!PageHuge(page_head))
717 return page_index(page);
719 if (compound_order(page_head) >= MAX_ORDER)
720 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
722 compound_idx = page - page_head;
724 return (index << compound_order(page_head)) + compound_idx;
727 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
731 if (h->order >= MAX_ORDER)
734 page = alloc_pages_exact_node(nid,
735 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
736 __GFP_REPEAT|__GFP_NOWARN,
739 if (arch_prepare_hugepage(page)) {
740 __free_pages(page, huge_page_order(h));
743 prep_new_huge_page(h, page, nid);
750 * common helper functions for hstate_next_node_to_{alloc|free}.
751 * We may have allocated or freed a huge page based on a different
752 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
753 * be outside of *nodes_allowed. Ensure that we use an allowed
754 * node for alloc or free.
756 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
758 nid = next_node(nid, *nodes_allowed);
759 if (nid == MAX_NUMNODES)
760 nid = first_node(*nodes_allowed);
761 VM_BUG_ON(nid >= MAX_NUMNODES);
766 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
768 if (!node_isset(nid, *nodes_allowed))
769 nid = next_node_allowed(nid, nodes_allowed);
774 * returns the previously saved node ["this node"] from which to
775 * allocate a persistent huge page for the pool and advance the
776 * next node from which to allocate, handling wrap at end of node
779 static int hstate_next_node_to_alloc(struct hstate *h,
780 nodemask_t *nodes_allowed)
784 VM_BUG_ON(!nodes_allowed);
786 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
787 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
792 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
799 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
800 next_nid = start_nid;
803 page = alloc_fresh_huge_page_node(h, next_nid);
808 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
809 } while (next_nid != start_nid);
812 count_vm_event(HTLB_BUDDY_PGALLOC);
814 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
820 * helper for free_pool_huge_page() - return the previously saved
821 * node ["this node"] from which to free a huge page. Advance the
822 * next node id whether or not we find a free huge page to free so
823 * that the next attempt to free addresses the next node.
825 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
829 VM_BUG_ON(!nodes_allowed);
831 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
832 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
838 * Free huge page from pool from next node to free.
839 * Attempt to keep persistent huge pages more or less
840 * balanced over allowed nodes.
841 * Called with hugetlb_lock locked.
843 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
850 start_nid = hstate_next_node_to_free(h, nodes_allowed);
851 next_nid = start_nid;
855 * If we're returning unused surplus pages, only examine
856 * nodes with surplus pages.
858 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
859 !list_empty(&h->hugepage_freelists[next_nid])) {
861 list_entry(h->hugepage_freelists[next_nid].next,
863 list_del(&page->lru);
864 h->free_huge_pages--;
865 h->free_huge_pages_node[next_nid]--;
867 h->surplus_huge_pages--;
868 h->surplus_huge_pages_node[next_nid]--;
870 update_and_free_page(h, page);
874 next_nid = hstate_next_node_to_free(h, nodes_allowed);
875 } while (next_nid != start_nid);
880 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
885 if (h->order >= MAX_ORDER)
889 * Assume we will successfully allocate the surplus page to
890 * prevent racing processes from causing the surplus to exceed
893 * This however introduces a different race, where a process B
894 * tries to grow the static hugepage pool while alloc_pages() is
895 * called by process A. B will only examine the per-node
896 * counters in determining if surplus huge pages can be
897 * converted to normal huge pages in adjust_pool_surplus(). A
898 * won't be able to increment the per-node counter, until the
899 * lock is dropped by B, but B doesn't drop hugetlb_lock until
900 * no more huge pages can be converted from surplus to normal
901 * state (and doesn't try to convert again). Thus, we have a
902 * case where a surplus huge page exists, the pool is grown, and
903 * the surplus huge page still exists after, even though it
904 * should just have been converted to a normal huge page. This
905 * does not leak memory, though, as the hugepage will be freed
906 * once it is out of use. It also does not allow the counters to
907 * go out of whack in adjust_pool_surplus() as we don't modify
908 * the node values until we've gotten the hugepage and only the
909 * per-node value is checked there.
911 spin_lock(&hugetlb_lock);
912 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
913 spin_unlock(&hugetlb_lock);
917 h->surplus_huge_pages++;
919 spin_unlock(&hugetlb_lock);
921 if (nid == NUMA_NO_NODE)
922 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
923 __GFP_REPEAT|__GFP_NOWARN,
926 page = alloc_pages_exact_node(nid,
927 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
928 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
930 if (page && arch_prepare_hugepage(page)) {
931 __free_pages(page, huge_page_order(h));
935 spin_lock(&hugetlb_lock);
937 INIT_LIST_HEAD(&page->lru);
938 r_nid = page_to_nid(page);
939 set_compound_page_dtor(page, free_huge_page);
940 set_hugetlb_cgroup(page, NULL);
942 * We incremented the global counters already
944 h->nr_huge_pages_node[r_nid]++;
945 h->surplus_huge_pages_node[r_nid]++;
946 __count_vm_event(HTLB_BUDDY_PGALLOC);
949 h->surplus_huge_pages--;
950 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
952 spin_unlock(&hugetlb_lock);
958 * This allocation function is useful in the context where vma is irrelevant.
959 * E.g. soft-offlining uses this function because it only cares physical
960 * address of error page.
962 struct page *alloc_huge_page_node(struct hstate *h, int nid)
966 spin_lock(&hugetlb_lock);
967 page = dequeue_huge_page_node(h, nid);
968 spin_unlock(&hugetlb_lock);
971 page = alloc_buddy_huge_page(h, nid);
977 * Increase the hugetlb pool such that it can accommodate a reservation
980 static int gather_surplus_pages(struct hstate *h, int delta)
982 struct list_head surplus_list;
983 struct page *page, *tmp;
985 int needed, allocated;
986 bool alloc_ok = true;
988 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
990 h->resv_huge_pages += delta;
995 INIT_LIST_HEAD(&surplus_list);
999 spin_unlock(&hugetlb_lock);
1000 for (i = 0; i < needed; i++) {
1001 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1006 list_add(&page->lru, &surplus_list);
1011 * After retaking hugetlb_lock, we need to recalculate 'needed'
1012 * because either resv_huge_pages or free_huge_pages may have changed.
1014 spin_lock(&hugetlb_lock);
1015 needed = (h->resv_huge_pages + delta) -
1016 (h->free_huge_pages + allocated);
1021 * We were not able to allocate enough pages to
1022 * satisfy the entire reservation so we free what
1023 * we've allocated so far.
1028 * The surplus_list now contains _at_least_ the number of extra pages
1029 * needed to accommodate the reservation. Add the appropriate number
1030 * of pages to the hugetlb pool and free the extras back to the buddy
1031 * allocator. Commit the entire reservation here to prevent another
1032 * process from stealing the pages as they are added to the pool but
1033 * before they are reserved.
1035 needed += allocated;
1036 h->resv_huge_pages += delta;
1039 /* Free the needed pages to the hugetlb pool */
1040 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1044 * This page is now managed by the hugetlb allocator and has
1045 * no users -- drop the buddy allocator's reference.
1047 put_page_testzero(page);
1048 VM_BUG_ON(page_count(page));
1049 enqueue_huge_page(h, page);
1052 spin_unlock(&hugetlb_lock);
1054 /* Free unnecessary surplus pages to the buddy allocator */
1055 if (!list_empty(&surplus_list)) {
1056 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1060 spin_lock(&hugetlb_lock);
1066 * When releasing a hugetlb pool reservation, any surplus pages that were
1067 * allocated to satisfy the reservation must be explicitly freed if they were
1069 * Called with hugetlb_lock held.
1071 static void return_unused_surplus_pages(struct hstate *h,
1072 unsigned long unused_resv_pages)
1074 unsigned long nr_pages;
1076 /* Uncommit the reservation */
1077 h->resv_huge_pages -= unused_resv_pages;
1079 /* Cannot return gigantic pages currently */
1080 if (h->order >= MAX_ORDER)
1083 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1086 * We want to release as many surplus pages as possible, spread
1087 * evenly across all nodes with memory. Iterate across these nodes
1088 * until we can no longer free unreserved surplus pages. This occurs
1089 * when the nodes with surplus pages have no free pages.
1090 * free_pool_huge_page() will balance the the freed pages across the
1091 * on-line nodes with memory and will handle the hstate accounting.
1093 while (nr_pages--) {
1094 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1100 * Determine if the huge page at addr within the vma has an associated
1101 * reservation. Where it does not we will need to logically increase
1102 * reservation and actually increase subpool usage before an allocation
1103 * can occur. Where any new reservation would be required the
1104 * reservation change is prepared, but not committed. Once the page
1105 * has been allocated from the subpool and instantiated the change should
1106 * be committed via vma_commit_reservation. No action is required on
1109 static long vma_needs_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 return region_chg(&inode->i_mapping->private_list,
1120 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1125 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1126 struct resv_map *reservations = vma_resv_map(vma);
1128 err = region_chg(&reservations->regions, idx, idx + 1);
1134 static void vma_commit_reservation(struct hstate *h,
1135 struct vm_area_struct *vma, unsigned long addr)
1137 struct address_space *mapping = vma->vm_file->f_mapping;
1138 struct inode *inode = mapping->host;
1140 if (vma->vm_flags & VM_MAYSHARE) {
1141 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1142 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1144 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1145 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1146 struct resv_map *reservations = vma_resv_map(vma);
1148 /* Mark this page used in the map. */
1149 region_add(&reservations->regions, idx, idx + 1);
1153 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1154 unsigned long addr, int avoid_reserve)
1156 struct hugepage_subpool *spool = subpool_vma(vma);
1157 struct hstate *h = hstate_vma(vma);
1161 struct hugetlb_cgroup *h_cg;
1163 idx = hstate_index(h);
1165 * Processes that did not create the mapping will have no
1166 * reserves and will not have accounted against subpool
1167 * limit. Check that the subpool limit can be made before
1168 * satisfying the allocation MAP_NORESERVE mappings may also
1169 * need pages and subpool limit allocated allocated if no reserve
1172 chg = vma_needs_reservation(h, vma, addr);
1174 return ERR_PTR(-ENOMEM);
1176 if (hugepage_subpool_get_pages(spool, chg))
1177 return ERR_PTR(-ENOSPC);
1179 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1181 hugepage_subpool_put_pages(spool, chg);
1182 return ERR_PTR(-ENOSPC);
1184 spin_lock(&hugetlb_lock);
1185 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1187 /* update page cgroup details */
1188 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h),
1190 spin_unlock(&hugetlb_lock);
1192 spin_unlock(&hugetlb_lock);
1193 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1195 hugetlb_cgroup_uncharge_cgroup(idx,
1196 pages_per_huge_page(h),
1198 hugepage_subpool_put_pages(spool, chg);
1199 return ERR_PTR(-ENOSPC);
1201 spin_lock(&hugetlb_lock);
1202 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h),
1204 list_move(&page->lru, &h->hugepage_activelist);
1205 spin_unlock(&hugetlb_lock);
1208 set_page_private(page, (unsigned long)spool);
1210 vma_commit_reservation(h, vma, addr);
1214 int __weak alloc_bootmem_huge_page(struct hstate *h)
1216 struct huge_bootmem_page *m;
1217 int nr_nodes = nodes_weight(node_states[N_MEMORY]);
1222 addr = __alloc_bootmem_node_nopanic(
1223 NODE_DATA(hstate_next_node_to_alloc(h,
1224 &node_states[N_MEMORY])),
1225 huge_page_size(h), huge_page_size(h), 0);
1229 * Use the beginning of the huge page to store the
1230 * huge_bootmem_page struct (until gather_bootmem
1231 * puts them into the mem_map).
1241 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1242 /* Put them into a private list first because mem_map is not up yet */
1243 list_add(&m->list, &huge_boot_pages);
1248 static void prep_compound_huge_page(struct page *page, int order)
1250 if (unlikely(order > (MAX_ORDER - 1)))
1251 prep_compound_gigantic_page(page, order);
1253 prep_compound_page(page, order);
1256 /* Put bootmem huge pages into the standard lists after mem_map is up */
1257 static void __init gather_bootmem_prealloc(void)
1259 struct huge_bootmem_page *m;
1261 list_for_each_entry(m, &huge_boot_pages, list) {
1262 struct hstate *h = m->hstate;
1265 #ifdef CONFIG_HIGHMEM
1266 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1267 free_bootmem_late((unsigned long)m,
1268 sizeof(struct huge_bootmem_page));
1270 page = virt_to_page(m);
1272 __ClearPageReserved(page);
1273 WARN_ON(page_count(page) != 1);
1274 prep_compound_huge_page(page, h->order);
1275 prep_new_huge_page(h, page, page_to_nid(page));
1277 * If we had gigantic hugepages allocated at boot time, we need
1278 * to restore the 'stolen' pages to totalram_pages in order to
1279 * fix confusing memory reports from free(1) and another
1280 * side-effects, like CommitLimit going negative.
1282 if (h->order > (MAX_ORDER - 1))
1283 totalram_pages += 1 << h->order;
1287 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1291 for (i = 0; i < h->max_huge_pages; ++i) {
1292 if (h->order >= MAX_ORDER) {
1293 if (!alloc_bootmem_huge_page(h))
1295 } else if (!alloc_fresh_huge_page(h,
1296 &node_states[N_MEMORY]))
1299 h->max_huge_pages = i;
1302 static void __init hugetlb_init_hstates(void)
1306 for_each_hstate(h) {
1307 /* oversize hugepages were init'ed in early boot */
1308 if (h->order < MAX_ORDER)
1309 hugetlb_hstate_alloc_pages(h);
1313 static char * __init memfmt(char *buf, unsigned long n)
1315 if (n >= (1UL << 30))
1316 sprintf(buf, "%lu GB", n >> 30);
1317 else if (n >= (1UL << 20))
1318 sprintf(buf, "%lu MB", n >> 20);
1320 sprintf(buf, "%lu KB", n >> 10);
1324 static void __init report_hugepages(void)
1328 for_each_hstate(h) {
1330 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1331 memfmt(buf, huge_page_size(h)),
1332 h->free_huge_pages);
1336 #ifdef CONFIG_HIGHMEM
1337 static void try_to_free_low(struct hstate *h, unsigned long count,
1338 nodemask_t *nodes_allowed)
1342 if (h->order >= MAX_ORDER)
1345 for_each_node_mask(i, *nodes_allowed) {
1346 struct page *page, *next;
1347 struct list_head *freel = &h->hugepage_freelists[i];
1348 list_for_each_entry_safe(page, next, freel, lru) {
1349 if (count >= h->nr_huge_pages)
1351 if (PageHighMem(page))
1353 list_del(&page->lru);
1354 update_and_free_page(h, page);
1355 h->free_huge_pages--;
1356 h->free_huge_pages_node[page_to_nid(page)]--;
1361 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1362 nodemask_t *nodes_allowed)
1368 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1369 * balanced by operating on them in a round-robin fashion.
1370 * Returns 1 if an adjustment was made.
1372 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1375 int start_nid, next_nid;
1378 VM_BUG_ON(delta != -1 && delta != 1);
1381 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1383 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1384 next_nid = start_nid;
1390 * To shrink on this node, there must be a surplus page
1392 if (!h->surplus_huge_pages_node[nid]) {
1393 next_nid = hstate_next_node_to_alloc(h,
1400 * Surplus cannot exceed the total number of pages
1402 if (h->surplus_huge_pages_node[nid] >=
1403 h->nr_huge_pages_node[nid]) {
1404 next_nid = hstate_next_node_to_free(h,
1410 h->surplus_huge_pages += delta;
1411 h->surplus_huge_pages_node[nid] += delta;
1414 } while (next_nid != start_nid);
1419 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1420 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1421 nodemask_t *nodes_allowed)
1423 unsigned long min_count, ret;
1425 if (h->order >= MAX_ORDER)
1426 return h->max_huge_pages;
1429 * Increase the pool size
1430 * First take pages out of surplus state. Then make up the
1431 * remaining difference by allocating fresh huge pages.
1433 * We might race with alloc_buddy_huge_page() here and be unable
1434 * to convert a surplus huge page to a normal huge page. That is
1435 * not critical, though, it just means the overall size of the
1436 * pool might be one hugepage larger than it needs to be, but
1437 * within all the constraints specified by the sysctls.
1439 spin_lock(&hugetlb_lock);
1440 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1441 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1445 while (count > persistent_huge_pages(h)) {
1447 * If this allocation races such that we no longer need the
1448 * page, free_huge_page will handle it by freeing the page
1449 * and reducing the surplus.
1451 spin_unlock(&hugetlb_lock);
1452 ret = alloc_fresh_huge_page(h, nodes_allowed);
1453 spin_lock(&hugetlb_lock);
1457 /* Bail for signals. Probably ctrl-c from user */
1458 if (signal_pending(current))
1463 * Decrease the pool size
1464 * First return free pages to the buddy allocator (being careful
1465 * to keep enough around to satisfy reservations). Then place
1466 * pages into surplus state as needed so the pool will shrink
1467 * to the desired size as pages become free.
1469 * By placing pages into the surplus state independent of the
1470 * overcommit value, we are allowing the surplus pool size to
1471 * exceed overcommit. There are few sane options here. Since
1472 * alloc_buddy_huge_page() is checking the global counter,
1473 * though, we'll note that we're not allowed to exceed surplus
1474 * and won't grow the pool anywhere else. Not until one of the
1475 * sysctls are changed, or the surplus pages go out of use.
1477 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1478 min_count = max(count, min_count);
1479 try_to_free_low(h, min_count, nodes_allowed);
1480 while (min_count < persistent_huge_pages(h)) {
1481 if (!free_pool_huge_page(h, nodes_allowed, 0))
1484 while (count < persistent_huge_pages(h)) {
1485 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1489 ret = persistent_huge_pages(h);
1490 spin_unlock(&hugetlb_lock);
1494 #define HSTATE_ATTR_RO(_name) \
1495 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1497 #define HSTATE_ATTR(_name) \
1498 static struct kobj_attribute _name##_attr = \
1499 __ATTR(_name, 0644, _name##_show, _name##_store)
1501 static struct kobject *hugepages_kobj;
1502 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1504 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1506 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1510 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1511 if (hstate_kobjs[i] == kobj) {
1513 *nidp = NUMA_NO_NODE;
1517 return kobj_to_node_hstate(kobj, nidp);
1520 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1521 struct kobj_attribute *attr, char *buf)
1524 unsigned long nr_huge_pages;
1527 h = kobj_to_hstate(kobj, &nid);
1528 if (nid == NUMA_NO_NODE)
1529 nr_huge_pages = h->nr_huge_pages;
1531 nr_huge_pages = h->nr_huge_pages_node[nid];
1533 return sprintf(buf, "%lu\n", nr_huge_pages);
1536 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1537 struct kobject *kobj, struct kobj_attribute *attr,
1538 const char *buf, size_t len)
1542 unsigned long count;
1544 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1546 err = strict_strtoul(buf, 10, &count);
1550 h = kobj_to_hstate(kobj, &nid);
1551 if (h->order >= MAX_ORDER) {
1556 if (nid == NUMA_NO_NODE) {
1558 * global hstate attribute
1560 if (!(obey_mempolicy &&
1561 init_nodemask_of_mempolicy(nodes_allowed))) {
1562 NODEMASK_FREE(nodes_allowed);
1563 nodes_allowed = &node_states[N_MEMORY];
1565 } else if (nodes_allowed) {
1567 * per node hstate attribute: adjust count to global,
1568 * but restrict alloc/free to the specified node.
1570 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1571 init_nodemask_of_node(nodes_allowed, nid);
1573 nodes_allowed = &node_states[N_MEMORY];
1575 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1577 if (nodes_allowed != &node_states[N_MEMORY])
1578 NODEMASK_FREE(nodes_allowed);
1582 NODEMASK_FREE(nodes_allowed);
1586 static ssize_t nr_hugepages_show(struct kobject *kobj,
1587 struct kobj_attribute *attr, char *buf)
1589 return nr_hugepages_show_common(kobj, attr, buf);
1592 static ssize_t nr_hugepages_store(struct kobject *kobj,
1593 struct kobj_attribute *attr, const char *buf, size_t len)
1595 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1597 HSTATE_ATTR(nr_hugepages);
1602 * hstate attribute for optionally mempolicy-based constraint on persistent
1603 * huge page alloc/free.
1605 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1606 struct kobj_attribute *attr, char *buf)
1608 return nr_hugepages_show_common(kobj, attr, buf);
1611 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1612 struct kobj_attribute *attr, const char *buf, size_t len)
1614 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1616 HSTATE_ATTR(nr_hugepages_mempolicy);
1620 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1621 struct kobj_attribute *attr, char *buf)
1623 struct hstate *h = kobj_to_hstate(kobj, NULL);
1624 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1627 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1628 struct kobj_attribute *attr, const char *buf, size_t count)
1631 unsigned long input;
1632 struct hstate *h = kobj_to_hstate(kobj, NULL);
1634 if (h->order >= MAX_ORDER)
1637 err = strict_strtoul(buf, 10, &input);
1641 spin_lock(&hugetlb_lock);
1642 h->nr_overcommit_huge_pages = input;
1643 spin_unlock(&hugetlb_lock);
1647 HSTATE_ATTR(nr_overcommit_hugepages);
1649 static ssize_t free_hugepages_show(struct kobject *kobj,
1650 struct kobj_attribute *attr, char *buf)
1653 unsigned long free_huge_pages;
1656 h = kobj_to_hstate(kobj, &nid);
1657 if (nid == NUMA_NO_NODE)
1658 free_huge_pages = h->free_huge_pages;
1660 free_huge_pages = h->free_huge_pages_node[nid];
1662 return sprintf(buf, "%lu\n", free_huge_pages);
1664 HSTATE_ATTR_RO(free_hugepages);
1666 static ssize_t resv_hugepages_show(struct kobject *kobj,
1667 struct kobj_attribute *attr, char *buf)
1669 struct hstate *h = kobj_to_hstate(kobj, NULL);
1670 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1672 HSTATE_ATTR_RO(resv_hugepages);
1674 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1675 struct kobj_attribute *attr, char *buf)
1678 unsigned long surplus_huge_pages;
1681 h = kobj_to_hstate(kobj, &nid);
1682 if (nid == NUMA_NO_NODE)
1683 surplus_huge_pages = h->surplus_huge_pages;
1685 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1687 return sprintf(buf, "%lu\n", surplus_huge_pages);
1689 HSTATE_ATTR_RO(surplus_hugepages);
1691 static struct attribute *hstate_attrs[] = {
1692 &nr_hugepages_attr.attr,
1693 &nr_overcommit_hugepages_attr.attr,
1694 &free_hugepages_attr.attr,
1695 &resv_hugepages_attr.attr,
1696 &surplus_hugepages_attr.attr,
1698 &nr_hugepages_mempolicy_attr.attr,
1703 static struct attribute_group hstate_attr_group = {
1704 .attrs = hstate_attrs,
1707 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1708 struct kobject **hstate_kobjs,
1709 struct attribute_group *hstate_attr_group)
1712 int hi = hstate_index(h);
1714 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1715 if (!hstate_kobjs[hi])
1718 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1720 kobject_put(hstate_kobjs[hi]);
1725 static void __init hugetlb_sysfs_init(void)
1730 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1731 if (!hugepages_kobj)
1734 for_each_hstate(h) {
1735 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1736 hstate_kobjs, &hstate_attr_group);
1738 pr_err("Hugetlb: Unable to add hstate %s", h->name);
1745 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1746 * with node devices in node_devices[] using a parallel array. The array
1747 * index of a node device or _hstate == node id.
1748 * This is here to avoid any static dependency of the node device driver, in
1749 * the base kernel, on the hugetlb module.
1751 struct node_hstate {
1752 struct kobject *hugepages_kobj;
1753 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1755 struct node_hstate node_hstates[MAX_NUMNODES];
1758 * A subset of global hstate attributes for node devices
1760 static struct attribute *per_node_hstate_attrs[] = {
1761 &nr_hugepages_attr.attr,
1762 &free_hugepages_attr.attr,
1763 &surplus_hugepages_attr.attr,
1767 static struct attribute_group per_node_hstate_attr_group = {
1768 .attrs = per_node_hstate_attrs,
1772 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1773 * Returns node id via non-NULL nidp.
1775 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1779 for (nid = 0; nid < nr_node_ids; nid++) {
1780 struct node_hstate *nhs = &node_hstates[nid];
1782 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1783 if (nhs->hstate_kobjs[i] == kobj) {
1795 * Unregister hstate attributes from a single node device.
1796 * No-op if no hstate attributes attached.
1798 static void hugetlb_unregister_node(struct node *node)
1801 struct node_hstate *nhs = &node_hstates[node->dev.id];
1803 if (!nhs->hugepages_kobj)
1804 return; /* no hstate attributes */
1806 for_each_hstate(h) {
1807 int idx = hstate_index(h);
1808 if (nhs->hstate_kobjs[idx]) {
1809 kobject_put(nhs->hstate_kobjs[idx]);
1810 nhs->hstate_kobjs[idx] = NULL;
1814 kobject_put(nhs->hugepages_kobj);
1815 nhs->hugepages_kobj = NULL;
1819 * hugetlb module exit: unregister hstate attributes from node devices
1822 static void hugetlb_unregister_all_nodes(void)
1827 * disable node device registrations.
1829 register_hugetlbfs_with_node(NULL, NULL);
1832 * remove hstate attributes from any nodes that have them.
1834 for (nid = 0; nid < nr_node_ids; nid++)
1835 hugetlb_unregister_node(node_devices[nid]);
1839 * Register hstate attributes for a single node device.
1840 * No-op if attributes already registered.
1842 static void hugetlb_register_node(struct node *node)
1845 struct node_hstate *nhs = &node_hstates[node->dev.id];
1848 if (nhs->hugepages_kobj)
1849 return; /* already allocated */
1851 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1853 if (!nhs->hugepages_kobj)
1856 for_each_hstate(h) {
1857 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1859 &per_node_hstate_attr_group);
1861 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1862 h->name, node->dev.id);
1863 hugetlb_unregister_node(node);
1870 * hugetlb init time: register hstate attributes for all registered node
1871 * devices of nodes that have memory. All on-line nodes should have
1872 * registered their associated device by this time.
1874 static void hugetlb_register_all_nodes(void)
1878 for_each_node_state(nid, N_MEMORY) {
1879 struct node *node = node_devices[nid];
1880 if (node->dev.id == nid)
1881 hugetlb_register_node(node);
1885 * Let the node device driver know we're here so it can
1886 * [un]register hstate attributes on node hotplug.
1888 register_hugetlbfs_with_node(hugetlb_register_node,
1889 hugetlb_unregister_node);
1891 #else /* !CONFIG_NUMA */
1893 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1901 static void hugetlb_unregister_all_nodes(void) { }
1903 static void hugetlb_register_all_nodes(void) { }
1907 static void __exit hugetlb_exit(void)
1911 hugetlb_unregister_all_nodes();
1913 for_each_hstate(h) {
1914 kobject_put(hstate_kobjs[hstate_index(h)]);
1917 kobject_put(hugepages_kobj);
1919 module_exit(hugetlb_exit);
1921 static int __init hugetlb_init(void)
1923 /* Some platform decide whether they support huge pages at boot
1924 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1925 * there is no such support
1927 if (HPAGE_SHIFT == 0)
1930 if (!size_to_hstate(default_hstate_size)) {
1931 default_hstate_size = HPAGE_SIZE;
1932 if (!size_to_hstate(default_hstate_size))
1933 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1935 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1936 if (default_hstate_max_huge_pages)
1937 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1939 hugetlb_init_hstates();
1940 gather_bootmem_prealloc();
1943 hugetlb_sysfs_init();
1944 hugetlb_register_all_nodes();
1945 hugetlb_cgroup_file_init();
1949 module_init(hugetlb_init);
1951 /* Should be called on processing a hugepagesz=... option */
1952 void __init hugetlb_add_hstate(unsigned order)
1957 if (size_to_hstate(PAGE_SIZE << order)) {
1958 pr_warning("hugepagesz= specified twice, ignoring\n");
1961 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1963 h = &hstates[hugetlb_max_hstate++];
1965 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1966 h->nr_huge_pages = 0;
1967 h->free_huge_pages = 0;
1968 for (i = 0; i < MAX_NUMNODES; ++i)
1969 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1970 INIT_LIST_HEAD(&h->hugepage_activelist);
1971 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
1972 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
1973 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1974 huge_page_size(h)/1024);
1979 static int __init hugetlb_nrpages_setup(char *s)
1982 static unsigned long *last_mhp;
1985 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1986 * so this hugepages= parameter goes to the "default hstate".
1988 if (!hugetlb_max_hstate)
1989 mhp = &default_hstate_max_huge_pages;
1991 mhp = &parsed_hstate->max_huge_pages;
1993 if (mhp == last_mhp) {
1994 pr_warning("hugepages= specified twice without "
1995 "interleaving hugepagesz=, ignoring\n");
1999 if (sscanf(s, "%lu", mhp) <= 0)
2003 * Global state is always initialized later in hugetlb_init.
2004 * But we need to allocate >= MAX_ORDER hstates here early to still
2005 * use the bootmem allocator.
2007 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2008 hugetlb_hstate_alloc_pages(parsed_hstate);
2014 __setup("hugepages=", hugetlb_nrpages_setup);
2016 static int __init hugetlb_default_setup(char *s)
2018 default_hstate_size = memparse(s, &s);
2021 __setup("default_hugepagesz=", hugetlb_default_setup);
2023 static unsigned int cpuset_mems_nr(unsigned int *array)
2026 unsigned int nr = 0;
2028 for_each_node_mask(node, cpuset_current_mems_allowed)
2034 #ifdef CONFIG_SYSCTL
2035 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2036 struct ctl_table *table, int write,
2037 void __user *buffer, size_t *length, loff_t *ppos)
2039 struct hstate *h = &default_hstate;
2043 tmp = h->max_huge_pages;
2045 if (write && h->order >= MAX_ORDER)
2049 table->maxlen = sizeof(unsigned long);
2050 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2055 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2056 GFP_KERNEL | __GFP_NORETRY);
2057 if (!(obey_mempolicy &&
2058 init_nodemask_of_mempolicy(nodes_allowed))) {
2059 NODEMASK_FREE(nodes_allowed);
2060 nodes_allowed = &node_states[N_MEMORY];
2062 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2064 if (nodes_allowed != &node_states[N_MEMORY])
2065 NODEMASK_FREE(nodes_allowed);
2071 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2072 void __user *buffer, size_t *length, loff_t *ppos)
2075 return hugetlb_sysctl_handler_common(false, table, write,
2076 buffer, length, ppos);
2080 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2081 void __user *buffer, size_t *length, loff_t *ppos)
2083 return hugetlb_sysctl_handler_common(true, table, write,
2084 buffer, length, ppos);
2086 #endif /* CONFIG_NUMA */
2088 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2089 void __user *buffer,
2090 size_t *length, loff_t *ppos)
2092 proc_dointvec(table, write, buffer, length, ppos);
2093 if (hugepages_treat_as_movable)
2094 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2096 htlb_alloc_mask = GFP_HIGHUSER;
2100 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2101 void __user *buffer,
2102 size_t *length, loff_t *ppos)
2104 struct hstate *h = &default_hstate;
2108 tmp = h->nr_overcommit_huge_pages;
2110 if (write && h->order >= MAX_ORDER)
2114 table->maxlen = sizeof(unsigned long);
2115 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2120 spin_lock(&hugetlb_lock);
2121 h->nr_overcommit_huge_pages = tmp;
2122 spin_unlock(&hugetlb_lock);
2128 #endif /* CONFIG_SYSCTL */
2130 void hugetlb_report_meminfo(struct seq_file *m)
2132 struct hstate *h = &default_hstate;
2134 "HugePages_Total: %5lu\n"
2135 "HugePages_Free: %5lu\n"
2136 "HugePages_Rsvd: %5lu\n"
2137 "HugePages_Surp: %5lu\n"
2138 "Hugepagesize: %8lu kB\n",
2142 h->surplus_huge_pages,
2143 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2146 int hugetlb_report_node_meminfo(int nid, char *buf)
2148 struct hstate *h = &default_hstate;
2150 "Node %d HugePages_Total: %5u\n"
2151 "Node %d HugePages_Free: %5u\n"
2152 "Node %d HugePages_Surp: %5u\n",
2153 nid, h->nr_huge_pages_node[nid],
2154 nid, h->free_huge_pages_node[nid],
2155 nid, h->surplus_huge_pages_node[nid]);
2158 void hugetlb_show_meminfo(void)
2163 for_each_node_state(nid, N_MEMORY)
2165 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2167 h->nr_huge_pages_node[nid],
2168 h->free_huge_pages_node[nid],
2169 h->surplus_huge_pages_node[nid],
2170 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2173 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2174 unsigned long hugetlb_total_pages(void)
2177 unsigned long nr_total_pages = 0;
2180 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2181 return nr_total_pages;
2184 static int hugetlb_acct_memory(struct hstate *h, long delta)
2188 spin_lock(&hugetlb_lock);
2190 * When cpuset is configured, it breaks the strict hugetlb page
2191 * reservation as the accounting is done on a global variable. Such
2192 * reservation is completely rubbish in the presence of cpuset because
2193 * the reservation is not checked against page availability for the
2194 * current cpuset. Application can still potentially OOM'ed by kernel
2195 * with lack of free htlb page in cpuset that the task is in.
2196 * Attempt to enforce strict accounting with cpuset is almost
2197 * impossible (or too ugly) because cpuset is too fluid that
2198 * task or memory node can be dynamically moved between cpusets.
2200 * The change of semantics for shared hugetlb mapping with cpuset is
2201 * undesirable. However, in order to preserve some of the semantics,
2202 * we fall back to check against current free page availability as
2203 * a best attempt and hopefully to minimize the impact of changing
2204 * semantics that cpuset has.
2207 if (gather_surplus_pages(h, delta) < 0)
2210 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2211 return_unused_surplus_pages(h, delta);
2218 return_unused_surplus_pages(h, (unsigned long) -delta);
2221 spin_unlock(&hugetlb_lock);
2225 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2227 struct resv_map *reservations = vma_resv_map(vma);
2230 * This new VMA should share its siblings reservation map if present.
2231 * The VMA will only ever have a valid reservation map pointer where
2232 * it is being copied for another still existing VMA. As that VMA
2233 * has a reference to the reservation map it cannot disappear until
2234 * after this open call completes. It is therefore safe to take a
2235 * new reference here without additional locking.
2238 kref_get(&reservations->refs);
2241 static void resv_map_put(struct vm_area_struct *vma)
2243 struct resv_map *reservations = vma_resv_map(vma);
2247 kref_put(&reservations->refs, resv_map_release);
2250 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2252 struct hstate *h = hstate_vma(vma);
2253 struct resv_map *reservations = vma_resv_map(vma);
2254 struct hugepage_subpool *spool = subpool_vma(vma);
2255 unsigned long reserve;
2256 unsigned long start;
2260 start = vma_hugecache_offset(h, vma, vma->vm_start);
2261 end = vma_hugecache_offset(h, vma, vma->vm_end);
2263 reserve = (end - start) -
2264 region_count(&reservations->regions, start, end);
2269 hugetlb_acct_memory(h, -reserve);
2270 hugepage_subpool_put_pages(spool, reserve);
2276 * We cannot handle pagefaults against hugetlb pages at all. They cause
2277 * handle_mm_fault() to try to instantiate regular-sized pages in the
2278 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2281 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2287 const struct vm_operations_struct hugetlb_vm_ops = {
2288 .fault = hugetlb_vm_op_fault,
2289 .open = hugetlb_vm_op_open,
2290 .close = hugetlb_vm_op_close,
2293 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2299 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2300 vma->vm_page_prot)));
2302 entry = huge_pte_wrprotect(mk_huge_pte(page,
2303 vma->vm_page_prot));
2305 entry = pte_mkyoung(entry);
2306 entry = pte_mkhuge(entry);
2307 entry = arch_make_huge_pte(entry, vma, page, writable);
2312 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2313 unsigned long address, pte_t *ptep)
2317 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2318 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2319 update_mmu_cache(vma, address, ptep);
2323 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2324 struct vm_area_struct *vma)
2326 pte_t *src_pte, *dst_pte, entry;
2327 struct page *ptepage;
2330 struct hstate *h = hstate_vma(vma);
2331 unsigned long sz = huge_page_size(h);
2333 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2335 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2336 src_pte = huge_pte_offset(src, addr);
2339 dst_pte = huge_pte_alloc(dst, addr, sz);
2343 /* If the pagetables are shared don't copy or take references */
2344 if (dst_pte == src_pte)
2347 spin_lock(&dst->page_table_lock);
2348 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2349 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2351 huge_ptep_set_wrprotect(src, addr, src_pte);
2352 entry = huge_ptep_get(src_pte);
2353 ptepage = pte_page(entry);
2355 page_dup_rmap(ptepage);
2356 set_huge_pte_at(dst, addr, dst_pte, entry);
2358 spin_unlock(&src->page_table_lock);
2359 spin_unlock(&dst->page_table_lock);
2367 static int is_hugetlb_entry_migration(pte_t pte)
2371 if (huge_pte_none(pte) || pte_present(pte))
2373 swp = pte_to_swp_entry(pte);
2374 if (non_swap_entry(swp) && is_migration_entry(swp))
2380 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2384 if (huge_pte_none(pte) || pte_present(pte))
2386 swp = pte_to_swp_entry(pte);
2387 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2393 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2394 unsigned long start, unsigned long end,
2395 struct page *ref_page)
2397 int force_flush = 0;
2398 struct mm_struct *mm = vma->vm_mm;
2399 unsigned long address;
2403 struct hstate *h = hstate_vma(vma);
2404 unsigned long sz = huge_page_size(h);
2405 const unsigned long mmun_start = start; /* For mmu_notifiers */
2406 const unsigned long mmun_end = end; /* For mmu_notifiers */
2408 WARN_ON(!is_vm_hugetlb_page(vma));
2409 BUG_ON(start & ~huge_page_mask(h));
2410 BUG_ON(end & ~huge_page_mask(h));
2412 tlb_start_vma(tlb, vma);
2413 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2415 spin_lock(&mm->page_table_lock);
2416 for (address = start; address < end; address += sz) {
2417 ptep = huge_pte_offset(mm, address);
2421 if (huge_pmd_unshare(mm, &address, ptep))
2424 pte = huge_ptep_get(ptep);
2425 if (huge_pte_none(pte))
2429 * HWPoisoned hugepage is already unmapped and dropped reference
2431 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
2432 huge_pte_clear(mm, address, ptep);
2436 page = pte_page(pte);
2438 * If a reference page is supplied, it is because a specific
2439 * page is being unmapped, not a range. Ensure the page we
2440 * are about to unmap is the actual page of interest.
2443 if (page != ref_page)
2447 * Mark the VMA as having unmapped its page so that
2448 * future faults in this VMA will fail rather than
2449 * looking like data was lost
2451 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2454 pte = huge_ptep_get_and_clear(mm, address, ptep);
2455 tlb_remove_tlb_entry(tlb, ptep, address);
2456 if (huge_pte_dirty(pte))
2457 set_page_dirty(page);
2459 page_remove_rmap(page);
2460 force_flush = !__tlb_remove_page(tlb, page);
2463 /* Bail out after unmapping reference page if supplied */
2467 spin_unlock(&mm->page_table_lock);
2469 * mmu_gather ran out of room to batch pages, we break out of
2470 * the PTE lock to avoid doing the potential expensive TLB invalidate
2471 * and page-free while holding it.
2476 if (address < end && !ref_page)
2479 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2480 tlb_end_vma(tlb, vma);
2483 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2484 struct vm_area_struct *vma, unsigned long start,
2485 unsigned long end, struct page *ref_page)
2487 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2490 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2491 * test will fail on a vma being torn down, and not grab a page table
2492 * on its way out. We're lucky that the flag has such an appropriate
2493 * name, and can in fact be safely cleared here. We could clear it
2494 * before the __unmap_hugepage_range above, but all that's necessary
2495 * is to clear it before releasing the i_mmap_mutex. This works
2496 * because in the context this is called, the VMA is about to be
2497 * destroyed and the i_mmap_mutex is held.
2499 vma->vm_flags &= ~VM_MAYSHARE;
2502 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2503 unsigned long end, struct page *ref_page)
2505 struct mm_struct *mm;
2506 struct mmu_gather tlb;
2510 tlb_gather_mmu(&tlb, mm, start, end);
2511 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2512 tlb_finish_mmu(&tlb, start, end);
2516 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2517 * mappping it owns the reserve page for. The intention is to unmap the page
2518 * from other VMAs and let the children be SIGKILLed if they are faulting the
2521 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2522 struct page *page, unsigned long address)
2524 struct hstate *h = hstate_vma(vma);
2525 struct vm_area_struct *iter_vma;
2526 struct address_space *mapping;
2530 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2531 * from page cache lookup which is in HPAGE_SIZE units.
2533 address = address & huge_page_mask(h);
2534 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2536 mapping = file_inode(vma->vm_file)->i_mapping;
2539 * Take the mapping lock for the duration of the table walk. As
2540 * this mapping should be shared between all the VMAs,
2541 * __unmap_hugepage_range() is called as the lock is already held
2543 mutex_lock(&mapping->i_mmap_mutex);
2544 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2545 /* Do not unmap the current VMA */
2546 if (iter_vma == vma)
2550 * Unmap the page from other VMAs without their own reserves.
2551 * They get marked to be SIGKILLed if they fault in these
2552 * areas. This is because a future no-page fault on this VMA
2553 * could insert a zeroed page instead of the data existing
2554 * from the time of fork. This would look like data corruption
2556 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2557 unmap_hugepage_range(iter_vma, address,
2558 address + huge_page_size(h), page);
2560 mutex_unlock(&mapping->i_mmap_mutex);
2566 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2567 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2568 * cannot race with other handlers or page migration.
2569 * Keep the pte_same checks anyway to make transition from the mutex easier.
2571 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2572 unsigned long address, pte_t *ptep, pte_t pte,
2573 struct page *pagecache_page)
2575 struct hstate *h = hstate_vma(vma);
2576 struct page *old_page, *new_page;
2578 int outside_reserve = 0;
2579 unsigned long mmun_start; /* For mmu_notifiers */
2580 unsigned long mmun_end; /* For mmu_notifiers */
2582 old_page = pte_page(pte);
2585 /* If no-one else is actually using this page, avoid the copy
2586 * and just make the page writable */
2587 avoidcopy = (page_mapcount(old_page) == 1);
2589 if (PageAnon(old_page))
2590 page_move_anon_rmap(old_page, vma, address);
2591 set_huge_ptep_writable(vma, address, ptep);
2596 * If the process that created a MAP_PRIVATE mapping is about to
2597 * perform a COW due to a shared page count, attempt to satisfy
2598 * the allocation without using the existing reserves. The pagecache
2599 * page is used to determine if the reserve at this address was
2600 * consumed or not. If reserves were used, a partial faulted mapping
2601 * at the time of fork() could consume its reserves on COW instead
2602 * of the full address range.
2604 if (!(vma->vm_flags & VM_MAYSHARE) &&
2605 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2606 old_page != pagecache_page)
2607 outside_reserve = 1;
2609 page_cache_get(old_page);
2611 /* Drop page_table_lock as buddy allocator may be called */
2612 spin_unlock(&mm->page_table_lock);
2613 new_page = alloc_huge_page(vma, address, outside_reserve);
2615 if (IS_ERR(new_page)) {
2616 long err = PTR_ERR(new_page);
2617 page_cache_release(old_page);
2620 * If a process owning a MAP_PRIVATE mapping fails to COW,
2621 * it is due to references held by a child and an insufficient
2622 * huge page pool. To guarantee the original mappers
2623 * reliability, unmap the page from child processes. The child
2624 * may get SIGKILLed if it later faults.
2626 if (outside_reserve) {
2627 BUG_ON(huge_pte_none(pte));
2628 if (unmap_ref_private(mm, vma, old_page, address)) {
2629 BUG_ON(huge_pte_none(pte));
2630 spin_lock(&mm->page_table_lock);
2631 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2632 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2633 goto retry_avoidcopy;
2635 * race occurs while re-acquiring page_table_lock, and
2643 /* Caller expects lock to be held */
2644 spin_lock(&mm->page_table_lock);
2646 return VM_FAULT_OOM;
2648 return VM_FAULT_SIGBUS;
2652 * When the original hugepage is shared one, it does not have
2653 * anon_vma prepared.
2655 if (unlikely(anon_vma_prepare(vma))) {
2656 page_cache_release(new_page);
2657 page_cache_release(old_page);
2658 /* Caller expects lock to be held */
2659 spin_lock(&mm->page_table_lock);
2660 return VM_FAULT_OOM;
2663 copy_user_huge_page(new_page, old_page, address, vma,
2664 pages_per_huge_page(h));
2665 __SetPageUptodate(new_page);
2667 mmun_start = address & huge_page_mask(h);
2668 mmun_end = mmun_start + huge_page_size(h);
2669 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2671 * Retake the page_table_lock to check for racing updates
2672 * before the page tables are altered
2674 spin_lock(&mm->page_table_lock);
2675 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2676 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2678 huge_ptep_clear_flush(vma, address, ptep);
2679 set_huge_pte_at(mm, address, ptep,
2680 make_huge_pte(vma, new_page, 1));
2681 page_remove_rmap(old_page);
2682 hugepage_add_new_anon_rmap(new_page, vma, address);
2683 /* Make the old page be freed below */
2684 new_page = old_page;
2686 spin_unlock(&mm->page_table_lock);
2687 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2688 /* Caller expects lock to be held */
2689 spin_lock(&mm->page_table_lock);
2690 page_cache_release(new_page);
2691 page_cache_release(old_page);
2695 /* Return the pagecache page at a given address within a VMA */
2696 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2697 struct vm_area_struct *vma, unsigned long address)
2699 struct address_space *mapping;
2702 mapping = vma->vm_file->f_mapping;
2703 idx = vma_hugecache_offset(h, vma, address);
2705 return find_lock_page(mapping, idx);
2709 * Return whether there is a pagecache page to back given address within VMA.
2710 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2712 static bool hugetlbfs_pagecache_present(struct hstate *h,
2713 struct vm_area_struct *vma, unsigned long address)
2715 struct address_space *mapping;
2719 mapping = vma->vm_file->f_mapping;
2720 idx = vma_hugecache_offset(h, vma, address);
2722 page = find_get_page(mapping, idx);
2725 return page != NULL;
2728 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2729 unsigned long address, pte_t *ptep, unsigned int flags)
2731 struct hstate *h = hstate_vma(vma);
2732 int ret = VM_FAULT_SIGBUS;
2737 struct address_space *mapping;
2741 * Currently, we are forced to kill the process in the event the
2742 * original mapper has unmapped pages from the child due to a failed
2743 * COW. Warn that such a situation has occurred as it may not be obvious
2745 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2746 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2751 mapping = vma->vm_file->f_mapping;
2752 idx = vma_hugecache_offset(h, vma, address);
2755 * Use page lock to guard against racing truncation
2756 * before we get page_table_lock.
2759 page = find_lock_page(mapping, idx);
2761 size = i_size_read(mapping->host) >> huge_page_shift(h);
2764 page = alloc_huge_page(vma, address, 0);
2766 ret = PTR_ERR(page);
2770 ret = VM_FAULT_SIGBUS;
2773 clear_huge_page(page, address, pages_per_huge_page(h));
2774 __SetPageUptodate(page);
2776 if (vma->vm_flags & VM_MAYSHARE) {
2778 struct inode *inode = mapping->host;
2780 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2788 spin_lock(&inode->i_lock);
2789 inode->i_blocks += blocks_per_huge_page(h);
2790 spin_unlock(&inode->i_lock);
2793 if (unlikely(anon_vma_prepare(vma))) {
2795 goto backout_unlocked;
2801 * If memory error occurs between mmap() and fault, some process
2802 * don't have hwpoisoned swap entry for errored virtual address.
2803 * So we need to block hugepage fault by PG_hwpoison bit check.
2805 if (unlikely(PageHWPoison(page))) {
2806 ret = VM_FAULT_HWPOISON |
2807 VM_FAULT_SET_HINDEX(hstate_index(h));
2808 goto backout_unlocked;
2813 * If we are going to COW a private mapping later, we examine the
2814 * pending reservations for this page now. This will ensure that
2815 * any allocations necessary to record that reservation occur outside
2818 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2819 if (vma_needs_reservation(h, vma, address) < 0) {
2821 goto backout_unlocked;
2824 spin_lock(&mm->page_table_lock);
2825 size = i_size_read(mapping->host) >> huge_page_shift(h);
2830 if (!huge_pte_none(huge_ptep_get(ptep)))
2834 hugepage_add_new_anon_rmap(page, vma, address);
2836 page_dup_rmap(page);
2837 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2838 && (vma->vm_flags & VM_SHARED)));
2839 set_huge_pte_at(mm, address, ptep, new_pte);
2841 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2842 /* Optimization, do the COW without a second fault */
2843 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2846 spin_unlock(&mm->page_table_lock);
2852 spin_unlock(&mm->page_table_lock);
2859 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2860 unsigned long address, unsigned int flags)
2865 struct page *page = NULL;
2866 struct page *pagecache_page = NULL;
2867 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2868 struct hstate *h = hstate_vma(vma);
2870 address &= huge_page_mask(h);
2872 ptep = huge_pte_offset(mm, address);
2874 entry = huge_ptep_get(ptep);
2875 if (unlikely(is_hugetlb_entry_migration(entry))) {
2876 migration_entry_wait_huge(mm, ptep);
2878 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2879 return VM_FAULT_HWPOISON_LARGE |
2880 VM_FAULT_SET_HINDEX(hstate_index(h));
2883 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2885 return VM_FAULT_OOM;
2888 * Serialize hugepage allocation and instantiation, so that we don't
2889 * get spurious allocation failures if two CPUs race to instantiate
2890 * the same page in the page cache.
2892 mutex_lock(&hugetlb_instantiation_mutex);
2893 entry = huge_ptep_get(ptep);
2894 if (huge_pte_none(entry)) {
2895 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2902 * If we are going to COW the mapping later, we examine the pending
2903 * reservations for this page now. This will ensure that any
2904 * allocations necessary to record that reservation occur outside the
2905 * spinlock. For private mappings, we also lookup the pagecache
2906 * page now as it is used to determine if a reservation has been
2909 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
2910 if (vma_needs_reservation(h, vma, address) < 0) {
2915 if (!(vma->vm_flags & VM_MAYSHARE))
2916 pagecache_page = hugetlbfs_pagecache_page(h,
2921 * hugetlb_cow() requires page locks of pte_page(entry) and
2922 * pagecache_page, so here we need take the former one
2923 * when page != pagecache_page or !pagecache_page.
2924 * Note that locking order is always pagecache_page -> page,
2925 * so no worry about deadlock.
2927 page = pte_page(entry);
2929 if (page != pagecache_page)
2932 spin_lock(&mm->page_table_lock);
2933 /* Check for a racing update before calling hugetlb_cow */
2934 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2935 goto out_page_table_lock;
2938 if (flags & FAULT_FLAG_WRITE) {
2939 if (!huge_pte_write(entry)) {
2940 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2942 goto out_page_table_lock;
2944 entry = huge_pte_mkdirty(entry);
2946 entry = pte_mkyoung(entry);
2947 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2948 flags & FAULT_FLAG_WRITE))
2949 update_mmu_cache(vma, address, ptep);
2951 out_page_table_lock:
2952 spin_unlock(&mm->page_table_lock);
2954 if (pagecache_page) {
2955 unlock_page(pagecache_page);
2956 put_page(pagecache_page);
2958 if (page != pagecache_page)
2963 mutex_unlock(&hugetlb_instantiation_mutex);
2968 /* Can be overriden by architectures */
2969 __attribute__((weak)) struct page *
2970 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2971 pud_t *pud, int write)
2977 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2978 struct page **pages, struct vm_area_struct **vmas,
2979 unsigned long *position, unsigned long *nr_pages,
2980 long i, unsigned int flags)
2982 unsigned long pfn_offset;
2983 unsigned long vaddr = *position;
2984 unsigned long remainder = *nr_pages;
2985 struct hstate *h = hstate_vma(vma);
2987 spin_lock(&mm->page_table_lock);
2988 while (vaddr < vma->vm_end && remainder) {
2994 * Some archs (sparc64, sh*) have multiple pte_ts to
2995 * each hugepage. We have to make sure we get the
2996 * first, for the page indexing below to work.
2998 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2999 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3002 * When coredumping, it suits get_dump_page if we just return
3003 * an error where there's an empty slot with no huge pagecache
3004 * to back it. This way, we avoid allocating a hugepage, and
3005 * the sparse dumpfile avoids allocating disk blocks, but its
3006 * huge holes still show up with zeroes where they need to be.
3008 if (absent && (flags & FOLL_DUMP) &&
3009 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3015 * We need call hugetlb_fault for both hugepages under migration
3016 * (in which case hugetlb_fault waits for the migration,) and
3017 * hwpoisoned hugepages (in which case we need to prevent the
3018 * caller from accessing to them.) In order to do this, we use
3019 * here is_swap_pte instead of is_hugetlb_entry_migration and
3020 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3021 * both cases, and because we can't follow correct pages
3022 * directly from any kind of swap entries.
3024 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3025 ((flags & FOLL_WRITE) &&
3026 !huge_pte_write(huge_ptep_get(pte)))) {
3029 spin_unlock(&mm->page_table_lock);
3030 ret = hugetlb_fault(mm, vma, vaddr,
3031 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3032 spin_lock(&mm->page_table_lock);
3033 if (!(ret & VM_FAULT_ERROR))
3040 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3041 page = pte_page(huge_ptep_get(pte));
3044 pages[i] = mem_map_offset(page, pfn_offset);
3055 if (vaddr < vma->vm_end && remainder &&
3056 pfn_offset < pages_per_huge_page(h)) {
3058 * We use pfn_offset to avoid touching the pageframes
3059 * of this compound page.
3064 spin_unlock(&mm->page_table_lock);
3065 *nr_pages = remainder;
3068 return i ? i : -EFAULT;
3071 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3072 unsigned long address, unsigned long end, pgprot_t newprot)
3074 struct mm_struct *mm = vma->vm_mm;
3075 unsigned long start = address;
3078 struct hstate *h = hstate_vma(vma);
3079 unsigned long pages = 0;
3081 BUG_ON(address >= end);
3082 flush_cache_range(vma, address, end);
3084 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3085 spin_lock(&mm->page_table_lock);
3086 for (; address < end; address += huge_page_size(h)) {
3087 ptep = huge_pte_offset(mm, address);
3090 if (huge_pmd_unshare(mm, &address, ptep)) {
3094 if (!huge_pte_none(huge_ptep_get(ptep))) {
3095 pte = huge_ptep_get_and_clear(mm, address, ptep);
3096 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3097 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3098 set_huge_pte_at(mm, address, ptep, pte);
3102 spin_unlock(&mm->page_table_lock);
3104 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3105 * may have cleared our pud entry and done put_page on the page table:
3106 * once we release i_mmap_mutex, another task can do the final put_page
3107 * and that page table be reused and filled with junk.
3109 flush_tlb_range(vma, start, end);
3110 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3112 return pages << h->order;
3115 int hugetlb_reserve_pages(struct inode *inode,
3117 struct vm_area_struct *vma,
3118 vm_flags_t vm_flags)
3121 struct hstate *h = hstate_inode(inode);
3122 struct hugepage_subpool *spool = subpool_inode(inode);
3125 * Only apply hugepage reservation if asked. At fault time, an
3126 * attempt will be made for VM_NORESERVE to allocate a page
3127 * without using reserves
3129 if (vm_flags & VM_NORESERVE)
3133 * Shared mappings base their reservation on the number of pages that
3134 * are already allocated on behalf of the file. Private mappings need
3135 * to reserve the full area even if read-only as mprotect() may be
3136 * called to make the mapping read-write. Assume !vma is a shm mapping
3138 if (!vma || vma->vm_flags & VM_MAYSHARE)
3139 chg = region_chg(&inode->i_mapping->private_list, from, to);
3141 struct resv_map *resv_map = resv_map_alloc();
3147 set_vma_resv_map(vma, resv_map);
3148 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3156 /* There must be enough pages in the subpool for the mapping */
3157 if (hugepage_subpool_get_pages(spool, chg)) {
3163 * Check enough hugepages are available for the reservation.
3164 * Hand the pages back to the subpool if there are not
3166 ret = hugetlb_acct_memory(h, chg);
3168 hugepage_subpool_put_pages(spool, chg);
3173 * Account for the reservations made. Shared mappings record regions
3174 * that have reservations as they are shared by multiple VMAs.
3175 * When the last VMA disappears, the region map says how much
3176 * the reservation was and the page cache tells how much of
3177 * the reservation was consumed. Private mappings are per-VMA and
3178 * only the consumed reservations are tracked. When the VMA
3179 * disappears, the original reservation is the VMA size and the
3180 * consumed reservations are stored in the map. Hence, nothing
3181 * else has to be done for private mappings here
3183 if (!vma || vma->vm_flags & VM_MAYSHARE)
3184 region_add(&inode->i_mapping->private_list, from, to);
3192 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3194 struct hstate *h = hstate_inode(inode);
3195 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3196 struct hugepage_subpool *spool = subpool_inode(inode);
3198 spin_lock(&inode->i_lock);
3199 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3200 spin_unlock(&inode->i_lock);
3202 hugepage_subpool_put_pages(spool, (chg - freed));
3203 hugetlb_acct_memory(h, -(chg - freed));
3206 #ifdef CONFIG_MEMORY_FAILURE
3208 /* Should be called in hugetlb_lock */
3209 static int is_hugepage_on_freelist(struct page *hpage)
3213 struct hstate *h = page_hstate(hpage);
3214 int nid = page_to_nid(hpage);
3216 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3223 * This function is called from memory failure code.
3224 * Assume the caller holds page lock of the head page.
3226 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3228 struct hstate *h = page_hstate(hpage);
3229 int nid = page_to_nid(hpage);
3232 spin_lock(&hugetlb_lock);
3233 if (is_hugepage_on_freelist(hpage)) {
3235 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3236 * but dangling hpage->lru can trigger list-debug warnings
3237 * (this happens when we call unpoison_memory() on it),
3238 * so let it point to itself with list_del_init().
3240 list_del_init(&hpage->lru);
3241 set_page_refcounted(hpage);
3242 h->free_huge_pages--;
3243 h->free_huge_pages_node[nid]--;
3246 spin_unlock(&hugetlb_lock);