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_instantiation_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 << huge_page_shift(hstate);
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 * A child process with MAP_PRIVATE mappings created by their parent
544 * have no page reserves. This check ensures that reservations are
545 * not "stolen". The child may still get SIGKILLed
547 if (!vma_has_reserves(vma) &&
548 h->free_huge_pages - h->resv_huge_pages == 0)
551 /* If reserves cannot be used, ensure enough pages are in the pool */
552 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
556 cpuset_mems_cookie = get_mems_allowed();
557 zonelist = huge_zonelist(vma, address,
558 htlb_alloc_mask, &mpol, &nodemask);
560 for_each_zone_zonelist_nodemask(zone, z, zonelist,
561 MAX_NR_ZONES - 1, nodemask) {
562 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
563 page = dequeue_huge_page_node(h, zone_to_nid(zone));
566 decrement_hugepage_resv_vma(h, vma);
573 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);
693 pgoff_t __basepage_index(struct page *page)
695 struct page *page_head = compound_head(page);
696 pgoff_t index = page_index(page_head);
697 unsigned long compound_idx;
699 if (!PageHuge(page_head))
700 return page_index(page);
702 if (compound_order(page_head) >= MAX_ORDER)
703 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
705 compound_idx = page - page_head;
707 return (index << compound_order(page_head)) + compound_idx;
710 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
714 if (h->order >= MAX_ORDER)
717 page = alloc_pages_exact_node(nid,
718 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
719 __GFP_REPEAT|__GFP_NOWARN,
722 if (arch_prepare_hugepage(page)) {
723 __free_pages(page, huge_page_order(h));
726 prep_new_huge_page(h, page, nid);
733 * common helper functions for hstate_next_node_to_{alloc|free}.
734 * We may have allocated or freed a huge page based on a different
735 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
736 * be outside of *nodes_allowed. Ensure that we use an allowed
737 * node for alloc or free.
739 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
741 nid = next_node(nid, *nodes_allowed);
742 if (nid == MAX_NUMNODES)
743 nid = first_node(*nodes_allowed);
744 VM_BUG_ON(nid >= MAX_NUMNODES);
749 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
751 if (!node_isset(nid, *nodes_allowed))
752 nid = next_node_allowed(nid, nodes_allowed);
757 * returns the previously saved node ["this node"] from which to
758 * allocate a persistent huge page for the pool and advance the
759 * next node from which to allocate, handling wrap at end of node
762 static int hstate_next_node_to_alloc(struct hstate *h,
763 nodemask_t *nodes_allowed)
767 VM_BUG_ON(!nodes_allowed);
769 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
770 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
775 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
782 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
783 next_nid = start_nid;
786 page = alloc_fresh_huge_page_node(h, next_nid);
791 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
792 } while (next_nid != start_nid);
795 count_vm_event(HTLB_BUDDY_PGALLOC);
797 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
803 * helper for free_pool_huge_page() - return the previously saved
804 * node ["this node"] from which to free a huge page. Advance the
805 * next node id whether or not we find a free huge page to free so
806 * that the next attempt to free addresses the next node.
808 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
812 VM_BUG_ON(!nodes_allowed);
814 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
815 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
821 * Free huge page from pool from next node to free.
822 * Attempt to keep persistent huge pages more or less
823 * balanced over allowed nodes.
824 * Called with hugetlb_lock locked.
826 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
833 start_nid = hstate_next_node_to_free(h, nodes_allowed);
834 next_nid = start_nid;
838 * If we're returning unused surplus pages, only examine
839 * nodes with surplus pages.
841 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
842 !list_empty(&h->hugepage_freelists[next_nid])) {
844 list_entry(h->hugepage_freelists[next_nid].next,
846 list_del(&page->lru);
847 h->free_huge_pages--;
848 h->free_huge_pages_node[next_nid]--;
850 h->surplus_huge_pages--;
851 h->surplus_huge_pages_node[next_nid]--;
853 update_and_free_page(h, page);
857 next_nid = hstate_next_node_to_free(h, nodes_allowed);
858 } while (next_nid != start_nid);
863 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
868 if (h->order >= MAX_ORDER)
872 * Assume we will successfully allocate the surplus page to
873 * prevent racing processes from causing the surplus to exceed
876 * This however introduces a different race, where a process B
877 * tries to grow the static hugepage pool while alloc_pages() is
878 * called by process A. B will only examine the per-node
879 * counters in determining if surplus huge pages can be
880 * converted to normal huge pages in adjust_pool_surplus(). A
881 * won't be able to increment the per-node counter, until the
882 * lock is dropped by B, but B doesn't drop hugetlb_lock until
883 * no more huge pages can be converted from surplus to normal
884 * state (and doesn't try to convert again). Thus, we have a
885 * case where a surplus huge page exists, the pool is grown, and
886 * the surplus huge page still exists after, even though it
887 * should just have been converted to a normal huge page. This
888 * does not leak memory, though, as the hugepage will be freed
889 * once it is out of use. It also does not allow the counters to
890 * go out of whack in adjust_pool_surplus() as we don't modify
891 * the node values until we've gotten the hugepage and only the
892 * per-node value is checked there.
894 spin_lock(&hugetlb_lock);
895 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
896 spin_unlock(&hugetlb_lock);
900 h->surplus_huge_pages++;
902 spin_unlock(&hugetlb_lock);
904 if (nid == NUMA_NO_NODE)
905 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
906 __GFP_REPEAT|__GFP_NOWARN,
909 page = alloc_pages_exact_node(nid,
910 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
911 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
913 if (page && arch_prepare_hugepage(page)) {
914 __free_pages(page, huge_page_order(h));
918 spin_lock(&hugetlb_lock);
920 INIT_LIST_HEAD(&page->lru);
921 r_nid = page_to_nid(page);
922 set_compound_page_dtor(page, free_huge_page);
923 set_hugetlb_cgroup(page, NULL);
925 * We incremented the global counters already
927 h->nr_huge_pages_node[r_nid]++;
928 h->surplus_huge_pages_node[r_nid]++;
929 __count_vm_event(HTLB_BUDDY_PGALLOC);
932 h->surplus_huge_pages--;
933 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
935 spin_unlock(&hugetlb_lock);
941 * This allocation function is useful in the context where vma is irrelevant.
942 * E.g. soft-offlining uses this function because it only cares physical
943 * address of error page.
945 struct page *alloc_huge_page_node(struct hstate *h, int nid)
949 spin_lock(&hugetlb_lock);
950 page = dequeue_huge_page_node(h, nid);
951 spin_unlock(&hugetlb_lock);
954 page = alloc_buddy_huge_page(h, nid);
960 * Increase the hugetlb pool such that it can accommodate a reservation
963 static int gather_surplus_pages(struct hstate *h, int delta)
965 struct list_head surplus_list;
966 struct page *page, *tmp;
968 int needed, allocated;
969 bool alloc_ok = true;
971 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
973 h->resv_huge_pages += delta;
978 INIT_LIST_HEAD(&surplus_list);
982 spin_unlock(&hugetlb_lock);
983 for (i = 0; i < needed; i++) {
984 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
989 list_add(&page->lru, &surplus_list);
994 * After retaking hugetlb_lock, we need to recalculate 'needed'
995 * because either resv_huge_pages or free_huge_pages may have changed.
997 spin_lock(&hugetlb_lock);
998 needed = (h->resv_huge_pages + delta) -
999 (h->free_huge_pages + allocated);
1004 * We were not able to allocate enough pages to
1005 * satisfy the entire reservation so we free what
1006 * we've allocated so far.
1011 * The surplus_list now contains _at_least_ the number of extra pages
1012 * needed to accommodate the reservation. Add the appropriate number
1013 * of pages to the hugetlb pool and free the extras back to the buddy
1014 * allocator. Commit the entire reservation here to prevent another
1015 * process from stealing the pages as they are added to the pool but
1016 * before they are reserved.
1018 needed += allocated;
1019 h->resv_huge_pages += delta;
1022 /* Free the needed pages to the hugetlb pool */
1023 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1027 * This page is now managed by the hugetlb allocator and has
1028 * no users -- drop the buddy allocator's reference.
1030 put_page_testzero(page);
1031 VM_BUG_ON(page_count(page));
1032 enqueue_huge_page(h, page);
1035 spin_unlock(&hugetlb_lock);
1037 /* Free unnecessary surplus pages to the buddy allocator */
1038 if (!list_empty(&surplus_list)) {
1039 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1043 spin_lock(&hugetlb_lock);
1049 * When releasing a hugetlb pool reservation, any surplus pages that were
1050 * allocated to satisfy the reservation must be explicitly freed if they were
1052 * Called with hugetlb_lock held.
1054 static void return_unused_surplus_pages(struct hstate *h,
1055 unsigned long unused_resv_pages)
1057 unsigned long nr_pages;
1059 /* Uncommit the reservation */
1060 h->resv_huge_pages -= unused_resv_pages;
1062 /* Cannot return gigantic pages currently */
1063 if (h->order >= MAX_ORDER)
1066 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1069 * We want to release as many surplus pages as possible, spread
1070 * evenly across all nodes with memory. Iterate across these nodes
1071 * until we can no longer free unreserved surplus pages. This occurs
1072 * when the nodes with surplus pages have no free pages.
1073 * free_pool_huge_page() will balance the the freed pages across the
1074 * on-line nodes with memory and will handle the hstate accounting.
1076 while (nr_pages--) {
1077 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1083 * Determine if the huge page at addr within the vma has an associated
1084 * reservation. Where it does not we will need to logically increase
1085 * reservation and actually increase subpool usage before an allocation
1086 * can occur. Where any new reservation would be required the
1087 * reservation change is prepared, but not committed. Once the page
1088 * has been allocated from the subpool and instantiated the change should
1089 * be committed via vma_commit_reservation. No action is required on
1092 static long vma_needs_reservation(struct hstate *h,
1093 struct vm_area_struct *vma, unsigned long addr)
1095 struct address_space *mapping = vma->vm_file->f_mapping;
1096 struct inode *inode = mapping->host;
1098 if (vma->vm_flags & VM_MAYSHARE) {
1099 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1100 return region_chg(&inode->i_mapping->private_list,
1103 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1108 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1109 struct resv_map *reservations = vma_resv_map(vma);
1111 err = region_chg(&reservations->regions, idx, idx + 1);
1117 static void vma_commit_reservation(struct hstate *h,
1118 struct vm_area_struct *vma, unsigned long addr)
1120 struct address_space *mapping = vma->vm_file->f_mapping;
1121 struct inode *inode = mapping->host;
1123 if (vma->vm_flags & VM_MAYSHARE) {
1124 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1125 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1127 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1128 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1129 struct resv_map *reservations = vma_resv_map(vma);
1131 /* Mark this page used in the map. */
1132 region_add(&reservations->regions, idx, idx + 1);
1136 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1137 unsigned long addr, int avoid_reserve)
1139 struct hugepage_subpool *spool = subpool_vma(vma);
1140 struct hstate *h = hstate_vma(vma);
1144 struct hugetlb_cgroup *h_cg;
1146 idx = hstate_index(h);
1148 * Processes that did not create the mapping will have no
1149 * reserves and will not have accounted against subpool
1150 * limit. Check that the subpool limit can be made before
1151 * satisfying the allocation MAP_NORESERVE mappings may also
1152 * need pages and subpool limit allocated allocated if no reserve
1155 chg = vma_needs_reservation(h, vma, addr);
1157 return ERR_PTR(-ENOMEM);
1159 if (hugepage_subpool_get_pages(spool, chg))
1160 return ERR_PTR(-ENOSPC);
1162 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1164 hugepage_subpool_put_pages(spool, chg);
1165 return ERR_PTR(-ENOSPC);
1167 spin_lock(&hugetlb_lock);
1168 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1170 spin_unlock(&hugetlb_lock);
1171 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1173 hugetlb_cgroup_uncharge_cgroup(idx,
1174 pages_per_huge_page(h),
1176 hugepage_subpool_put_pages(spool, chg);
1177 return ERR_PTR(-ENOSPC);
1179 spin_lock(&hugetlb_lock);
1180 list_move(&page->lru, &h->hugepage_activelist);
1183 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1184 spin_unlock(&hugetlb_lock);
1186 set_page_private(page, (unsigned long)spool);
1188 vma_commit_reservation(h, vma, addr);
1192 int __weak alloc_bootmem_huge_page(struct hstate *h)
1194 struct huge_bootmem_page *m;
1195 int nr_nodes = nodes_weight(node_states[N_MEMORY]);
1200 addr = __alloc_bootmem_node_nopanic(
1201 NODE_DATA(hstate_next_node_to_alloc(h,
1202 &node_states[N_MEMORY])),
1203 huge_page_size(h), huge_page_size(h), 0);
1207 * Use the beginning of the huge page to store the
1208 * huge_bootmem_page struct (until gather_bootmem
1209 * puts them into the mem_map).
1219 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1220 /* Put them into a private list first because mem_map is not up yet */
1221 list_add(&m->list, &huge_boot_pages);
1226 static void prep_compound_huge_page(struct page *page, int order)
1228 if (unlikely(order > (MAX_ORDER - 1)))
1229 prep_compound_gigantic_page(page, order);
1231 prep_compound_page(page, order);
1234 /* Put bootmem huge pages into the standard lists after mem_map is up */
1235 static void __init gather_bootmem_prealloc(void)
1237 struct huge_bootmem_page *m;
1239 list_for_each_entry(m, &huge_boot_pages, list) {
1240 struct hstate *h = m->hstate;
1243 #ifdef CONFIG_HIGHMEM
1244 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1245 free_bootmem_late((unsigned long)m,
1246 sizeof(struct huge_bootmem_page));
1248 page = virt_to_page(m);
1250 __ClearPageReserved(page);
1251 WARN_ON(page_count(page) != 1);
1252 prep_compound_huge_page(page, h->order);
1253 prep_new_huge_page(h, page, page_to_nid(page));
1255 * If we had gigantic hugepages allocated at boot time, we need
1256 * to restore the 'stolen' pages to totalram_pages in order to
1257 * fix confusing memory reports from free(1) and another
1258 * side-effects, like CommitLimit going negative.
1260 if (h->order > (MAX_ORDER - 1))
1261 adjust_managed_page_count(page, 1 << h->order);
1265 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1269 for (i = 0; i < h->max_huge_pages; ++i) {
1270 if (h->order >= MAX_ORDER) {
1271 if (!alloc_bootmem_huge_page(h))
1273 } else if (!alloc_fresh_huge_page(h,
1274 &node_states[N_MEMORY]))
1277 h->max_huge_pages = i;
1280 static void __init hugetlb_init_hstates(void)
1284 for_each_hstate(h) {
1285 /* oversize hugepages were init'ed in early boot */
1286 if (h->order < MAX_ORDER)
1287 hugetlb_hstate_alloc_pages(h);
1291 static char * __init memfmt(char *buf, unsigned long n)
1293 if (n >= (1UL << 30))
1294 sprintf(buf, "%lu GB", n >> 30);
1295 else if (n >= (1UL << 20))
1296 sprintf(buf, "%lu MB", n >> 20);
1298 sprintf(buf, "%lu KB", n >> 10);
1302 static void __init report_hugepages(void)
1306 for_each_hstate(h) {
1308 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1309 memfmt(buf, huge_page_size(h)),
1310 h->free_huge_pages);
1314 #ifdef CONFIG_HIGHMEM
1315 static void try_to_free_low(struct hstate *h, unsigned long count,
1316 nodemask_t *nodes_allowed)
1320 if (h->order >= MAX_ORDER)
1323 for_each_node_mask(i, *nodes_allowed) {
1324 struct page *page, *next;
1325 struct list_head *freel = &h->hugepage_freelists[i];
1326 list_for_each_entry_safe(page, next, freel, lru) {
1327 if (count >= h->nr_huge_pages)
1329 if (PageHighMem(page))
1331 list_del(&page->lru);
1332 update_and_free_page(h, page);
1333 h->free_huge_pages--;
1334 h->free_huge_pages_node[page_to_nid(page)]--;
1339 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1340 nodemask_t *nodes_allowed)
1346 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1347 * balanced by operating on them in a round-robin fashion.
1348 * Returns 1 if an adjustment was made.
1350 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1353 int start_nid, next_nid;
1356 VM_BUG_ON(delta != -1 && delta != 1);
1359 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1361 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1362 next_nid = start_nid;
1368 * To shrink on this node, there must be a surplus page
1370 if (!h->surplus_huge_pages_node[nid]) {
1371 next_nid = hstate_next_node_to_alloc(h,
1378 * Surplus cannot exceed the total number of pages
1380 if (h->surplus_huge_pages_node[nid] >=
1381 h->nr_huge_pages_node[nid]) {
1382 next_nid = hstate_next_node_to_free(h,
1388 h->surplus_huge_pages += delta;
1389 h->surplus_huge_pages_node[nid] += delta;
1392 } while (next_nid != start_nid);
1397 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1398 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1399 nodemask_t *nodes_allowed)
1401 unsigned long min_count, ret;
1403 if (h->order >= MAX_ORDER)
1404 return h->max_huge_pages;
1407 * Increase the pool size
1408 * First take pages out of surplus state. Then make up the
1409 * remaining difference by allocating fresh huge pages.
1411 * We might race with alloc_buddy_huge_page() here and be unable
1412 * to convert a surplus huge page to a normal huge page. That is
1413 * not critical, though, it just means the overall size of the
1414 * pool might be one hugepage larger than it needs to be, but
1415 * within all the constraints specified by the sysctls.
1417 spin_lock(&hugetlb_lock);
1418 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1419 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1423 while (count > persistent_huge_pages(h)) {
1425 * If this allocation races such that we no longer need the
1426 * page, free_huge_page will handle it by freeing the page
1427 * and reducing the surplus.
1429 spin_unlock(&hugetlb_lock);
1430 ret = alloc_fresh_huge_page(h, nodes_allowed);
1431 spin_lock(&hugetlb_lock);
1435 /* Bail for signals. Probably ctrl-c from user */
1436 if (signal_pending(current))
1441 * Decrease the pool size
1442 * First return free pages to the buddy allocator (being careful
1443 * to keep enough around to satisfy reservations). Then place
1444 * pages into surplus state as needed so the pool will shrink
1445 * to the desired size as pages become free.
1447 * By placing pages into the surplus state independent of the
1448 * overcommit value, we are allowing the surplus pool size to
1449 * exceed overcommit. There are few sane options here. Since
1450 * alloc_buddy_huge_page() is checking the global counter,
1451 * though, we'll note that we're not allowed to exceed surplus
1452 * and won't grow the pool anywhere else. Not until one of the
1453 * sysctls are changed, or the surplus pages go out of use.
1455 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1456 min_count = max(count, min_count);
1457 try_to_free_low(h, min_count, nodes_allowed);
1458 while (min_count < persistent_huge_pages(h)) {
1459 if (!free_pool_huge_page(h, nodes_allowed, 0))
1462 while (count < persistent_huge_pages(h)) {
1463 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1467 ret = persistent_huge_pages(h);
1468 spin_unlock(&hugetlb_lock);
1472 #define HSTATE_ATTR_RO(_name) \
1473 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1475 #define HSTATE_ATTR(_name) \
1476 static struct kobj_attribute _name##_attr = \
1477 __ATTR(_name, 0644, _name##_show, _name##_store)
1479 static struct kobject *hugepages_kobj;
1480 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1482 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1484 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1488 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1489 if (hstate_kobjs[i] == kobj) {
1491 *nidp = NUMA_NO_NODE;
1495 return kobj_to_node_hstate(kobj, nidp);
1498 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1499 struct kobj_attribute *attr, char *buf)
1502 unsigned long nr_huge_pages;
1505 h = kobj_to_hstate(kobj, &nid);
1506 if (nid == NUMA_NO_NODE)
1507 nr_huge_pages = h->nr_huge_pages;
1509 nr_huge_pages = h->nr_huge_pages_node[nid];
1511 return sprintf(buf, "%lu\n", nr_huge_pages);
1514 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1515 struct kobject *kobj, struct kobj_attribute *attr,
1516 const char *buf, size_t len)
1520 unsigned long count;
1522 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1524 err = kstrtoul(buf, 10, &count);
1528 h = kobj_to_hstate(kobj, &nid);
1529 if (h->order >= MAX_ORDER) {
1534 if (nid == NUMA_NO_NODE) {
1536 * global hstate attribute
1538 if (!(obey_mempolicy &&
1539 init_nodemask_of_mempolicy(nodes_allowed))) {
1540 NODEMASK_FREE(nodes_allowed);
1541 nodes_allowed = &node_states[N_MEMORY];
1543 } else if (nodes_allowed) {
1545 * per node hstate attribute: adjust count to global,
1546 * but restrict alloc/free to the specified node.
1548 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1549 init_nodemask_of_node(nodes_allowed, nid);
1551 nodes_allowed = &node_states[N_MEMORY];
1553 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1555 if (nodes_allowed != &node_states[N_MEMORY])
1556 NODEMASK_FREE(nodes_allowed);
1560 NODEMASK_FREE(nodes_allowed);
1564 static ssize_t nr_hugepages_show(struct kobject *kobj,
1565 struct kobj_attribute *attr, char *buf)
1567 return nr_hugepages_show_common(kobj, attr, buf);
1570 static ssize_t nr_hugepages_store(struct kobject *kobj,
1571 struct kobj_attribute *attr, const char *buf, size_t len)
1573 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1575 HSTATE_ATTR(nr_hugepages);
1580 * hstate attribute for optionally mempolicy-based constraint on persistent
1581 * huge page alloc/free.
1583 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1584 struct kobj_attribute *attr, char *buf)
1586 return nr_hugepages_show_common(kobj, attr, buf);
1589 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1590 struct kobj_attribute *attr, const char *buf, size_t len)
1592 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1594 HSTATE_ATTR(nr_hugepages_mempolicy);
1598 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1599 struct kobj_attribute *attr, char *buf)
1601 struct hstate *h = kobj_to_hstate(kobj, NULL);
1602 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1605 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1606 struct kobj_attribute *attr, const char *buf, size_t count)
1609 unsigned long input;
1610 struct hstate *h = kobj_to_hstate(kobj, NULL);
1612 if (h->order >= MAX_ORDER)
1615 err = kstrtoul(buf, 10, &input);
1619 spin_lock(&hugetlb_lock);
1620 h->nr_overcommit_huge_pages = input;
1621 spin_unlock(&hugetlb_lock);
1625 HSTATE_ATTR(nr_overcommit_hugepages);
1627 static ssize_t free_hugepages_show(struct kobject *kobj,
1628 struct kobj_attribute *attr, char *buf)
1631 unsigned long free_huge_pages;
1634 h = kobj_to_hstate(kobj, &nid);
1635 if (nid == NUMA_NO_NODE)
1636 free_huge_pages = h->free_huge_pages;
1638 free_huge_pages = h->free_huge_pages_node[nid];
1640 return sprintf(buf, "%lu\n", free_huge_pages);
1642 HSTATE_ATTR_RO(free_hugepages);
1644 static ssize_t resv_hugepages_show(struct kobject *kobj,
1645 struct kobj_attribute *attr, char *buf)
1647 struct hstate *h = kobj_to_hstate(kobj, NULL);
1648 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1650 HSTATE_ATTR_RO(resv_hugepages);
1652 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1653 struct kobj_attribute *attr, char *buf)
1656 unsigned long surplus_huge_pages;
1659 h = kobj_to_hstate(kobj, &nid);
1660 if (nid == NUMA_NO_NODE)
1661 surplus_huge_pages = h->surplus_huge_pages;
1663 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1665 return sprintf(buf, "%lu\n", surplus_huge_pages);
1667 HSTATE_ATTR_RO(surplus_hugepages);
1669 static struct attribute *hstate_attrs[] = {
1670 &nr_hugepages_attr.attr,
1671 &nr_overcommit_hugepages_attr.attr,
1672 &free_hugepages_attr.attr,
1673 &resv_hugepages_attr.attr,
1674 &surplus_hugepages_attr.attr,
1676 &nr_hugepages_mempolicy_attr.attr,
1681 static struct attribute_group hstate_attr_group = {
1682 .attrs = hstate_attrs,
1685 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1686 struct kobject **hstate_kobjs,
1687 struct attribute_group *hstate_attr_group)
1690 int hi = hstate_index(h);
1692 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1693 if (!hstate_kobjs[hi])
1696 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1698 kobject_put(hstate_kobjs[hi]);
1703 static void __init hugetlb_sysfs_init(void)
1708 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1709 if (!hugepages_kobj)
1712 for_each_hstate(h) {
1713 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1714 hstate_kobjs, &hstate_attr_group);
1716 pr_err("Hugetlb: Unable to add hstate %s", h->name);
1723 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1724 * with node devices in node_devices[] using a parallel array. The array
1725 * index of a node device or _hstate == node id.
1726 * This is here to avoid any static dependency of the node device driver, in
1727 * the base kernel, on the hugetlb module.
1729 struct node_hstate {
1730 struct kobject *hugepages_kobj;
1731 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1733 struct node_hstate node_hstates[MAX_NUMNODES];
1736 * A subset of global hstate attributes for node devices
1738 static struct attribute *per_node_hstate_attrs[] = {
1739 &nr_hugepages_attr.attr,
1740 &free_hugepages_attr.attr,
1741 &surplus_hugepages_attr.attr,
1745 static struct attribute_group per_node_hstate_attr_group = {
1746 .attrs = per_node_hstate_attrs,
1750 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1751 * Returns node id via non-NULL nidp.
1753 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1757 for (nid = 0; nid < nr_node_ids; nid++) {
1758 struct node_hstate *nhs = &node_hstates[nid];
1760 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1761 if (nhs->hstate_kobjs[i] == kobj) {
1773 * Unregister hstate attributes from a single node device.
1774 * No-op if no hstate attributes attached.
1776 static void hugetlb_unregister_node(struct node *node)
1779 struct node_hstate *nhs = &node_hstates[node->dev.id];
1781 if (!nhs->hugepages_kobj)
1782 return; /* no hstate attributes */
1784 for_each_hstate(h) {
1785 int idx = hstate_index(h);
1786 if (nhs->hstate_kobjs[idx]) {
1787 kobject_put(nhs->hstate_kobjs[idx]);
1788 nhs->hstate_kobjs[idx] = NULL;
1792 kobject_put(nhs->hugepages_kobj);
1793 nhs->hugepages_kobj = NULL;
1797 * hugetlb module exit: unregister hstate attributes from node devices
1800 static void hugetlb_unregister_all_nodes(void)
1805 * disable node device registrations.
1807 register_hugetlbfs_with_node(NULL, NULL);
1810 * remove hstate attributes from any nodes that have them.
1812 for (nid = 0; nid < nr_node_ids; nid++)
1813 hugetlb_unregister_node(node_devices[nid]);
1817 * Register hstate attributes for a single node device.
1818 * No-op if attributes already registered.
1820 static void hugetlb_register_node(struct node *node)
1823 struct node_hstate *nhs = &node_hstates[node->dev.id];
1826 if (nhs->hugepages_kobj)
1827 return; /* already allocated */
1829 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1831 if (!nhs->hugepages_kobj)
1834 for_each_hstate(h) {
1835 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1837 &per_node_hstate_attr_group);
1839 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1840 h->name, node->dev.id);
1841 hugetlb_unregister_node(node);
1848 * hugetlb init time: register hstate attributes for all registered node
1849 * devices of nodes that have memory. All on-line nodes should have
1850 * registered their associated device by this time.
1852 static void hugetlb_register_all_nodes(void)
1856 for_each_node_state(nid, N_MEMORY) {
1857 struct node *node = node_devices[nid];
1858 if (node->dev.id == nid)
1859 hugetlb_register_node(node);
1863 * Let the node device driver know we're here so it can
1864 * [un]register hstate attributes on node hotplug.
1866 register_hugetlbfs_with_node(hugetlb_register_node,
1867 hugetlb_unregister_node);
1869 #else /* !CONFIG_NUMA */
1871 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1879 static void hugetlb_unregister_all_nodes(void) { }
1881 static void hugetlb_register_all_nodes(void) { }
1885 static void __exit hugetlb_exit(void)
1889 hugetlb_unregister_all_nodes();
1891 for_each_hstate(h) {
1892 kobject_put(hstate_kobjs[hstate_index(h)]);
1895 kobject_put(hugepages_kobj);
1897 module_exit(hugetlb_exit);
1899 static int __init hugetlb_init(void)
1901 /* Some platform decide whether they support huge pages at boot
1902 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1903 * there is no such support
1905 if (HPAGE_SHIFT == 0)
1908 if (!size_to_hstate(default_hstate_size)) {
1909 default_hstate_size = HPAGE_SIZE;
1910 if (!size_to_hstate(default_hstate_size))
1911 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1913 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1914 if (default_hstate_max_huge_pages)
1915 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1917 hugetlb_init_hstates();
1918 gather_bootmem_prealloc();
1921 hugetlb_sysfs_init();
1922 hugetlb_register_all_nodes();
1923 hugetlb_cgroup_file_init();
1927 module_init(hugetlb_init);
1929 /* Should be called on processing a hugepagesz=... option */
1930 void __init hugetlb_add_hstate(unsigned order)
1935 if (size_to_hstate(PAGE_SIZE << order)) {
1936 pr_warning("hugepagesz= specified twice, ignoring\n");
1939 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1941 h = &hstates[hugetlb_max_hstate++];
1943 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1944 h->nr_huge_pages = 0;
1945 h->free_huge_pages = 0;
1946 for (i = 0; i < MAX_NUMNODES; ++i)
1947 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1948 INIT_LIST_HEAD(&h->hugepage_activelist);
1949 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
1950 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
1951 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1952 huge_page_size(h)/1024);
1957 static int __init hugetlb_nrpages_setup(char *s)
1960 static unsigned long *last_mhp;
1963 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1964 * so this hugepages= parameter goes to the "default hstate".
1966 if (!hugetlb_max_hstate)
1967 mhp = &default_hstate_max_huge_pages;
1969 mhp = &parsed_hstate->max_huge_pages;
1971 if (mhp == last_mhp) {
1972 pr_warning("hugepages= specified twice without "
1973 "interleaving hugepagesz=, ignoring\n");
1977 if (sscanf(s, "%lu", mhp) <= 0)
1981 * Global state is always initialized later in hugetlb_init.
1982 * But we need to allocate >= MAX_ORDER hstates here early to still
1983 * use the bootmem allocator.
1985 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
1986 hugetlb_hstate_alloc_pages(parsed_hstate);
1992 __setup("hugepages=", hugetlb_nrpages_setup);
1994 static int __init hugetlb_default_setup(char *s)
1996 default_hstate_size = memparse(s, &s);
1999 __setup("default_hugepagesz=", hugetlb_default_setup);
2001 static unsigned int cpuset_mems_nr(unsigned int *array)
2004 unsigned int nr = 0;
2006 for_each_node_mask(node, cpuset_current_mems_allowed)
2012 #ifdef CONFIG_SYSCTL
2013 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2014 struct ctl_table *table, int write,
2015 void __user *buffer, size_t *length, loff_t *ppos)
2017 struct hstate *h = &default_hstate;
2021 tmp = h->max_huge_pages;
2023 if (write && h->order >= MAX_ORDER)
2027 table->maxlen = sizeof(unsigned long);
2028 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2033 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2034 GFP_KERNEL | __GFP_NORETRY);
2035 if (!(obey_mempolicy &&
2036 init_nodemask_of_mempolicy(nodes_allowed))) {
2037 NODEMASK_FREE(nodes_allowed);
2038 nodes_allowed = &node_states[N_MEMORY];
2040 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2042 if (nodes_allowed != &node_states[N_MEMORY])
2043 NODEMASK_FREE(nodes_allowed);
2049 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2050 void __user *buffer, size_t *length, loff_t *ppos)
2053 return hugetlb_sysctl_handler_common(false, table, write,
2054 buffer, length, ppos);
2058 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2059 void __user *buffer, size_t *length, loff_t *ppos)
2061 return hugetlb_sysctl_handler_common(true, table, write,
2062 buffer, length, ppos);
2064 #endif /* CONFIG_NUMA */
2066 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2067 void __user *buffer,
2068 size_t *length, loff_t *ppos)
2070 proc_dointvec(table, write, buffer, length, ppos);
2071 if (hugepages_treat_as_movable)
2072 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2074 htlb_alloc_mask = GFP_HIGHUSER;
2078 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2079 void __user *buffer,
2080 size_t *length, loff_t *ppos)
2082 struct hstate *h = &default_hstate;
2086 tmp = h->nr_overcommit_huge_pages;
2088 if (write && h->order >= MAX_ORDER)
2092 table->maxlen = sizeof(unsigned long);
2093 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2098 spin_lock(&hugetlb_lock);
2099 h->nr_overcommit_huge_pages = tmp;
2100 spin_unlock(&hugetlb_lock);
2106 #endif /* CONFIG_SYSCTL */
2108 void hugetlb_report_meminfo(struct seq_file *m)
2110 struct hstate *h = &default_hstate;
2112 "HugePages_Total: %5lu\n"
2113 "HugePages_Free: %5lu\n"
2114 "HugePages_Rsvd: %5lu\n"
2115 "HugePages_Surp: %5lu\n"
2116 "Hugepagesize: %8lu kB\n",
2120 h->surplus_huge_pages,
2121 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2124 int hugetlb_report_node_meminfo(int nid, char *buf)
2126 struct hstate *h = &default_hstate;
2128 "Node %d HugePages_Total: %5u\n"
2129 "Node %d HugePages_Free: %5u\n"
2130 "Node %d HugePages_Surp: %5u\n",
2131 nid, h->nr_huge_pages_node[nid],
2132 nid, h->free_huge_pages_node[nid],
2133 nid, h->surplus_huge_pages_node[nid]);
2136 void hugetlb_show_meminfo(void)
2141 for_each_node_state(nid, N_MEMORY)
2143 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2145 h->nr_huge_pages_node[nid],
2146 h->free_huge_pages_node[nid],
2147 h->surplus_huge_pages_node[nid],
2148 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2151 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2152 unsigned long hugetlb_total_pages(void)
2155 unsigned long nr_total_pages = 0;
2158 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2159 return nr_total_pages;
2162 static int hugetlb_acct_memory(struct hstate *h, long delta)
2166 spin_lock(&hugetlb_lock);
2168 * When cpuset is configured, it breaks the strict hugetlb page
2169 * reservation as the accounting is done on a global variable. Such
2170 * reservation is completely rubbish in the presence of cpuset because
2171 * the reservation is not checked against page availability for the
2172 * current cpuset. Application can still potentially OOM'ed by kernel
2173 * with lack of free htlb page in cpuset that the task is in.
2174 * Attempt to enforce strict accounting with cpuset is almost
2175 * impossible (or too ugly) because cpuset is too fluid that
2176 * task or memory node can be dynamically moved between cpusets.
2178 * The change of semantics for shared hugetlb mapping with cpuset is
2179 * undesirable. However, in order to preserve some of the semantics,
2180 * we fall back to check against current free page availability as
2181 * a best attempt and hopefully to minimize the impact of changing
2182 * semantics that cpuset has.
2185 if (gather_surplus_pages(h, delta) < 0)
2188 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2189 return_unused_surplus_pages(h, delta);
2196 return_unused_surplus_pages(h, (unsigned long) -delta);
2199 spin_unlock(&hugetlb_lock);
2203 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2205 struct resv_map *reservations = vma_resv_map(vma);
2208 * This new VMA should share its siblings reservation map if present.
2209 * The VMA will only ever have a valid reservation map pointer where
2210 * it is being copied for another still existing VMA. As that VMA
2211 * has a reference to the reservation map it cannot disappear until
2212 * after this open call completes. It is therefore safe to take a
2213 * new reference here without additional locking.
2216 kref_get(&reservations->refs);
2219 static void resv_map_put(struct vm_area_struct *vma)
2221 struct resv_map *reservations = vma_resv_map(vma);
2225 kref_put(&reservations->refs, resv_map_release);
2228 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2230 struct hstate *h = hstate_vma(vma);
2231 struct resv_map *reservations = vma_resv_map(vma);
2232 struct hugepage_subpool *spool = subpool_vma(vma);
2233 unsigned long reserve;
2234 unsigned long start;
2238 start = vma_hugecache_offset(h, vma, vma->vm_start);
2239 end = vma_hugecache_offset(h, vma, vma->vm_end);
2241 reserve = (end - start) -
2242 region_count(&reservations->regions, start, end);
2247 hugetlb_acct_memory(h, -reserve);
2248 hugepage_subpool_put_pages(spool, reserve);
2254 * We cannot handle pagefaults against hugetlb pages at all. They cause
2255 * handle_mm_fault() to try to instantiate regular-sized pages in the
2256 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2259 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2265 const struct vm_operations_struct hugetlb_vm_ops = {
2266 .fault = hugetlb_vm_op_fault,
2267 .open = hugetlb_vm_op_open,
2268 .close = hugetlb_vm_op_close,
2271 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2277 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2278 vma->vm_page_prot)));
2280 entry = huge_pte_wrprotect(mk_huge_pte(page,
2281 vma->vm_page_prot));
2283 entry = pte_mkyoung(entry);
2284 entry = pte_mkhuge(entry);
2285 entry = arch_make_huge_pte(entry, vma, page, writable);
2290 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2291 unsigned long address, pte_t *ptep)
2295 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2296 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2297 update_mmu_cache(vma, address, ptep);
2301 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2302 struct vm_area_struct *vma)
2304 pte_t *src_pte, *dst_pte, entry;
2305 struct page *ptepage;
2308 struct hstate *h = hstate_vma(vma);
2309 unsigned long sz = huge_page_size(h);
2311 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2313 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2314 src_pte = huge_pte_offset(src, addr);
2317 dst_pte = huge_pte_alloc(dst, addr, sz);
2321 /* If the pagetables are shared don't copy or take references */
2322 if (dst_pte == src_pte)
2325 spin_lock(&dst->page_table_lock);
2326 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2327 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2329 huge_ptep_set_wrprotect(src, addr, src_pte);
2330 entry = huge_ptep_get(src_pte);
2331 ptepage = pte_page(entry);
2333 page_dup_rmap(ptepage);
2334 set_huge_pte_at(dst, addr, dst_pte, entry);
2336 spin_unlock(&src->page_table_lock);
2337 spin_unlock(&dst->page_table_lock);
2345 static int is_hugetlb_entry_migration(pte_t pte)
2349 if (huge_pte_none(pte) || pte_present(pte))
2351 swp = pte_to_swp_entry(pte);
2352 if (non_swap_entry(swp) && is_migration_entry(swp))
2358 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2362 if (huge_pte_none(pte) || pte_present(pte))
2364 swp = pte_to_swp_entry(pte);
2365 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2371 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2372 unsigned long start, unsigned long end,
2373 struct page *ref_page)
2375 int force_flush = 0;
2376 struct mm_struct *mm = vma->vm_mm;
2377 unsigned long address;
2381 struct hstate *h = hstate_vma(vma);
2382 unsigned long sz = huge_page_size(h);
2383 const unsigned long mmun_start = start; /* For mmu_notifiers */
2384 const unsigned long mmun_end = end; /* For mmu_notifiers */
2386 WARN_ON(!is_vm_hugetlb_page(vma));
2387 BUG_ON(start & ~huge_page_mask(h));
2388 BUG_ON(end & ~huge_page_mask(h));
2390 tlb_start_vma(tlb, vma);
2391 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2393 spin_lock(&mm->page_table_lock);
2394 for (address = start; address < end; address += sz) {
2395 ptep = huge_pte_offset(mm, address);
2399 if (huge_pmd_unshare(mm, &address, ptep))
2402 pte = huge_ptep_get(ptep);
2403 if (huge_pte_none(pte))
2407 * HWPoisoned hugepage is already unmapped and dropped reference
2409 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
2410 huge_pte_clear(mm, address, ptep);
2414 page = pte_page(pte);
2416 * If a reference page is supplied, it is because a specific
2417 * page is being unmapped, not a range. Ensure the page we
2418 * are about to unmap is the actual page of interest.
2421 if (page != ref_page)
2425 * Mark the VMA as having unmapped its page so that
2426 * future faults in this VMA will fail rather than
2427 * looking like data was lost
2429 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2432 pte = huge_ptep_get_and_clear(mm, address, ptep);
2433 tlb_remove_tlb_entry(tlb, ptep, address);
2434 if (huge_pte_dirty(pte))
2435 set_page_dirty(page);
2437 page_remove_rmap(page);
2438 force_flush = !__tlb_remove_page(tlb, page);
2441 /* Bail out after unmapping reference page if supplied */
2445 spin_unlock(&mm->page_table_lock);
2447 * mmu_gather ran out of room to batch pages, we break out of
2448 * the PTE lock to avoid doing the potential expensive TLB invalidate
2449 * and page-free while holding it.
2454 if (address < end && !ref_page)
2457 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2458 tlb_end_vma(tlb, vma);
2461 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2462 struct vm_area_struct *vma, unsigned long start,
2463 unsigned long end, struct page *ref_page)
2465 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2468 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2469 * test will fail on a vma being torn down, and not grab a page table
2470 * on its way out. We're lucky that the flag has such an appropriate
2471 * name, and can in fact be safely cleared here. We could clear it
2472 * before the __unmap_hugepage_range above, but all that's necessary
2473 * is to clear it before releasing the i_mmap_mutex. This works
2474 * because in the context this is called, the VMA is about to be
2475 * destroyed and the i_mmap_mutex is held.
2477 vma->vm_flags &= ~VM_MAYSHARE;
2480 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2481 unsigned long end, struct page *ref_page)
2483 struct mm_struct *mm;
2484 struct mmu_gather tlb;
2488 tlb_gather_mmu(&tlb, mm, start, end);
2489 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2490 tlb_finish_mmu(&tlb, start, end);
2494 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2495 * mappping it owns the reserve page for. The intention is to unmap the page
2496 * from other VMAs and let the children be SIGKILLed if they are faulting the
2499 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2500 struct page *page, unsigned long address)
2502 struct hstate *h = hstate_vma(vma);
2503 struct vm_area_struct *iter_vma;
2504 struct address_space *mapping;
2508 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2509 * from page cache lookup which is in HPAGE_SIZE units.
2511 address = address & huge_page_mask(h);
2512 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2514 mapping = file_inode(vma->vm_file)->i_mapping;
2517 * Take the mapping lock for the duration of the table walk. As
2518 * this mapping should be shared between all the VMAs,
2519 * __unmap_hugepage_range() is called as the lock is already held
2521 mutex_lock(&mapping->i_mmap_mutex);
2522 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2523 /* Do not unmap the current VMA */
2524 if (iter_vma == vma)
2528 * Unmap the page from other VMAs without their own reserves.
2529 * They get marked to be SIGKILLed if they fault in these
2530 * areas. This is because a future no-page fault on this VMA
2531 * could insert a zeroed page instead of the data existing
2532 * from the time of fork. This would look like data corruption
2534 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2535 unmap_hugepage_range(iter_vma, address,
2536 address + huge_page_size(h), page);
2538 mutex_unlock(&mapping->i_mmap_mutex);
2544 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2545 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2546 * cannot race with other handlers or page migration.
2547 * Keep the pte_same checks anyway to make transition from the mutex easier.
2549 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2550 unsigned long address, pte_t *ptep, pte_t pte,
2551 struct page *pagecache_page)
2553 struct hstate *h = hstate_vma(vma);
2554 struct page *old_page, *new_page;
2556 int outside_reserve = 0;
2557 unsigned long mmun_start; /* For mmu_notifiers */
2558 unsigned long mmun_end; /* For mmu_notifiers */
2560 old_page = pte_page(pte);
2563 /* If no-one else is actually using this page, avoid the copy
2564 * and just make the page writable */
2565 avoidcopy = (page_mapcount(old_page) == 1);
2567 if (PageAnon(old_page))
2568 page_move_anon_rmap(old_page, vma, address);
2569 set_huge_ptep_writable(vma, address, ptep);
2574 * If the process that created a MAP_PRIVATE mapping is about to
2575 * perform a COW due to a shared page count, attempt to satisfy
2576 * the allocation without using the existing reserves. The pagecache
2577 * page is used to determine if the reserve at this address was
2578 * consumed or not. If reserves were used, a partial faulted mapping
2579 * at the time of fork() could consume its reserves on COW instead
2580 * of the full address range.
2582 if (!(vma->vm_flags & VM_MAYSHARE) &&
2583 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2584 old_page != pagecache_page)
2585 outside_reserve = 1;
2587 page_cache_get(old_page);
2589 /* Drop page_table_lock as buddy allocator may be called */
2590 spin_unlock(&mm->page_table_lock);
2591 new_page = alloc_huge_page(vma, address, outside_reserve);
2593 if (IS_ERR(new_page)) {
2594 long err = PTR_ERR(new_page);
2595 page_cache_release(old_page);
2598 * If a process owning a MAP_PRIVATE mapping fails to COW,
2599 * it is due to references held by a child and an insufficient
2600 * huge page pool. To guarantee the original mappers
2601 * reliability, unmap the page from child processes. The child
2602 * may get SIGKILLed if it later faults.
2604 if (outside_reserve) {
2605 BUG_ON(huge_pte_none(pte));
2606 if (unmap_ref_private(mm, vma, old_page, address)) {
2607 BUG_ON(huge_pte_none(pte));
2608 spin_lock(&mm->page_table_lock);
2609 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2610 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2611 goto retry_avoidcopy;
2613 * race occurs while re-acquiring page_table_lock, and
2621 /* Caller expects lock to be held */
2622 spin_lock(&mm->page_table_lock);
2624 return VM_FAULT_OOM;
2626 return VM_FAULT_SIGBUS;
2630 * When the original hugepage is shared one, it does not have
2631 * anon_vma prepared.
2633 if (unlikely(anon_vma_prepare(vma))) {
2634 page_cache_release(new_page);
2635 page_cache_release(old_page);
2636 /* Caller expects lock to be held */
2637 spin_lock(&mm->page_table_lock);
2638 return VM_FAULT_OOM;
2641 copy_user_huge_page(new_page, old_page, address, vma,
2642 pages_per_huge_page(h));
2643 __SetPageUptodate(new_page);
2645 mmun_start = address & huge_page_mask(h);
2646 mmun_end = mmun_start + huge_page_size(h);
2647 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2649 * Retake the page_table_lock to check for racing updates
2650 * before the page tables are altered
2652 spin_lock(&mm->page_table_lock);
2653 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2654 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2656 huge_ptep_clear_flush(vma, address, ptep);
2657 set_huge_pte_at(mm, address, ptep,
2658 make_huge_pte(vma, new_page, 1));
2659 page_remove_rmap(old_page);
2660 hugepage_add_new_anon_rmap(new_page, vma, address);
2661 /* Make the old page be freed below */
2662 new_page = old_page;
2664 spin_unlock(&mm->page_table_lock);
2665 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2666 /* Caller expects lock to be held */
2667 spin_lock(&mm->page_table_lock);
2668 page_cache_release(new_page);
2669 page_cache_release(old_page);
2673 /* Return the pagecache page at a given address within a VMA */
2674 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2675 struct vm_area_struct *vma, unsigned long address)
2677 struct address_space *mapping;
2680 mapping = vma->vm_file->f_mapping;
2681 idx = vma_hugecache_offset(h, vma, address);
2683 return find_lock_page(mapping, idx);
2687 * Return whether there is a pagecache page to back given address within VMA.
2688 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2690 static bool hugetlbfs_pagecache_present(struct hstate *h,
2691 struct vm_area_struct *vma, unsigned long address)
2693 struct address_space *mapping;
2697 mapping = vma->vm_file->f_mapping;
2698 idx = vma_hugecache_offset(h, vma, address);
2700 page = find_get_page(mapping, idx);
2703 return page != NULL;
2706 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2707 unsigned long address, pte_t *ptep, unsigned int flags)
2709 struct hstate *h = hstate_vma(vma);
2710 int ret = VM_FAULT_SIGBUS;
2715 struct address_space *mapping;
2719 * Currently, we are forced to kill the process in the event the
2720 * original mapper has unmapped pages from the child due to a failed
2721 * COW. Warn that such a situation has occurred as it may not be obvious
2723 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2724 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2729 mapping = vma->vm_file->f_mapping;
2730 idx = vma_hugecache_offset(h, vma, address);
2733 * Use page lock to guard against racing truncation
2734 * before we get page_table_lock.
2737 page = find_lock_page(mapping, idx);
2739 size = i_size_read(mapping->host) >> huge_page_shift(h);
2742 page = alloc_huge_page(vma, address, 0);
2744 ret = PTR_ERR(page);
2748 ret = VM_FAULT_SIGBUS;
2751 clear_huge_page(page, address, pages_per_huge_page(h));
2752 __SetPageUptodate(page);
2754 if (vma->vm_flags & VM_MAYSHARE) {
2756 struct inode *inode = mapping->host;
2758 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2766 spin_lock(&inode->i_lock);
2767 inode->i_blocks += blocks_per_huge_page(h);
2768 spin_unlock(&inode->i_lock);
2771 if (unlikely(anon_vma_prepare(vma))) {
2773 goto backout_unlocked;
2779 * If memory error occurs between mmap() and fault, some process
2780 * don't have hwpoisoned swap entry for errored virtual address.
2781 * So we need to block hugepage fault by PG_hwpoison bit check.
2783 if (unlikely(PageHWPoison(page))) {
2784 ret = VM_FAULT_HWPOISON |
2785 VM_FAULT_SET_HINDEX(hstate_index(h));
2786 goto backout_unlocked;
2791 * If we are going to COW a private mapping later, we examine the
2792 * pending reservations for this page now. This will ensure that
2793 * any allocations necessary to record that reservation occur outside
2796 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2797 if (vma_needs_reservation(h, vma, address) < 0) {
2799 goto backout_unlocked;
2802 spin_lock(&mm->page_table_lock);
2803 size = i_size_read(mapping->host) >> huge_page_shift(h);
2808 if (!huge_pte_none(huge_ptep_get(ptep)))
2812 hugepage_add_new_anon_rmap(page, vma, address);
2814 page_dup_rmap(page);
2815 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2816 && (vma->vm_flags & VM_SHARED)));
2817 set_huge_pte_at(mm, address, ptep, new_pte);
2819 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2820 /* Optimization, do the COW without a second fault */
2821 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2824 spin_unlock(&mm->page_table_lock);
2830 spin_unlock(&mm->page_table_lock);
2837 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2838 unsigned long address, unsigned int flags)
2843 struct page *page = NULL;
2844 struct page *pagecache_page = NULL;
2845 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2846 struct hstate *h = hstate_vma(vma);
2848 address &= huge_page_mask(h);
2850 ptep = huge_pte_offset(mm, address);
2852 entry = huge_ptep_get(ptep);
2853 if (unlikely(is_hugetlb_entry_migration(entry))) {
2854 migration_entry_wait_huge(mm, ptep);
2856 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2857 return VM_FAULT_HWPOISON_LARGE |
2858 VM_FAULT_SET_HINDEX(hstate_index(h));
2861 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2863 return VM_FAULT_OOM;
2866 * Serialize hugepage allocation and instantiation, so that we don't
2867 * get spurious allocation failures if two CPUs race to instantiate
2868 * the same page in the page cache.
2870 mutex_lock(&hugetlb_instantiation_mutex);
2871 entry = huge_ptep_get(ptep);
2872 if (huge_pte_none(entry)) {
2873 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2880 * If we are going to COW the mapping later, we examine the pending
2881 * reservations for this page now. This will ensure that any
2882 * allocations necessary to record that reservation occur outside the
2883 * spinlock. For private mappings, we also lookup the pagecache
2884 * page now as it is used to determine if a reservation has been
2887 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
2888 if (vma_needs_reservation(h, vma, address) < 0) {
2893 if (!(vma->vm_flags & VM_MAYSHARE))
2894 pagecache_page = hugetlbfs_pagecache_page(h,
2899 * hugetlb_cow() requires page locks of pte_page(entry) and
2900 * pagecache_page, so here we need take the former one
2901 * when page != pagecache_page or !pagecache_page.
2902 * Note that locking order is always pagecache_page -> page,
2903 * so no worry about deadlock.
2905 page = pte_page(entry);
2907 if (page != pagecache_page)
2910 spin_lock(&mm->page_table_lock);
2911 /* Check for a racing update before calling hugetlb_cow */
2912 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2913 goto out_page_table_lock;
2916 if (flags & FAULT_FLAG_WRITE) {
2917 if (!huge_pte_write(entry)) {
2918 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2920 goto out_page_table_lock;
2922 entry = huge_pte_mkdirty(entry);
2924 entry = pte_mkyoung(entry);
2925 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2926 flags & FAULT_FLAG_WRITE))
2927 update_mmu_cache(vma, address, ptep);
2929 out_page_table_lock:
2930 spin_unlock(&mm->page_table_lock);
2932 if (pagecache_page) {
2933 unlock_page(pagecache_page);
2934 put_page(pagecache_page);
2936 if (page != pagecache_page)
2941 mutex_unlock(&hugetlb_instantiation_mutex);
2946 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2947 struct page **pages, struct vm_area_struct **vmas,
2948 unsigned long *position, unsigned long *nr_pages,
2949 long i, unsigned int flags)
2951 unsigned long pfn_offset;
2952 unsigned long vaddr = *position;
2953 unsigned long remainder = *nr_pages;
2954 struct hstate *h = hstate_vma(vma);
2956 spin_lock(&mm->page_table_lock);
2957 while (vaddr < vma->vm_end && remainder) {
2963 * Some archs (sparc64, sh*) have multiple pte_ts to
2964 * each hugepage. We have to make sure we get the
2965 * first, for the page indexing below to work.
2967 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2968 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2971 * When coredumping, it suits get_dump_page if we just return
2972 * an error where there's an empty slot with no huge pagecache
2973 * to back it. This way, we avoid allocating a hugepage, and
2974 * the sparse dumpfile avoids allocating disk blocks, but its
2975 * huge holes still show up with zeroes where they need to be.
2977 if (absent && (flags & FOLL_DUMP) &&
2978 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2984 * We need call hugetlb_fault for both hugepages under migration
2985 * (in which case hugetlb_fault waits for the migration,) and
2986 * hwpoisoned hugepages (in which case we need to prevent the
2987 * caller from accessing to them.) In order to do this, we use
2988 * here is_swap_pte instead of is_hugetlb_entry_migration and
2989 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
2990 * both cases, and because we can't follow correct pages
2991 * directly from any kind of swap entries.
2993 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
2994 ((flags & FOLL_WRITE) &&
2995 !huge_pte_write(huge_ptep_get(pte)))) {
2998 spin_unlock(&mm->page_table_lock);
2999 ret = hugetlb_fault(mm, vma, vaddr,
3000 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3001 spin_lock(&mm->page_table_lock);
3002 if (!(ret & VM_FAULT_ERROR))
3009 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3010 page = pte_page(huge_ptep_get(pte));
3013 pages[i] = mem_map_offset(page, pfn_offset);
3024 if (vaddr < vma->vm_end && remainder &&
3025 pfn_offset < pages_per_huge_page(h)) {
3027 * We use pfn_offset to avoid touching the pageframes
3028 * of this compound page.
3033 spin_unlock(&mm->page_table_lock);
3034 *nr_pages = remainder;
3037 return i ? i : -EFAULT;
3040 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3041 unsigned long address, unsigned long end, pgprot_t newprot)
3043 struct mm_struct *mm = vma->vm_mm;
3044 unsigned long start = address;
3047 struct hstate *h = hstate_vma(vma);
3048 unsigned long pages = 0;
3050 BUG_ON(address >= end);
3051 flush_cache_range(vma, address, end);
3053 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3054 spin_lock(&mm->page_table_lock);
3055 for (; address < end; address += huge_page_size(h)) {
3056 ptep = huge_pte_offset(mm, address);
3059 if (huge_pmd_unshare(mm, &address, ptep)) {
3063 if (!huge_pte_none(huge_ptep_get(ptep))) {
3064 pte = huge_ptep_get_and_clear(mm, address, ptep);
3065 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3066 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3067 set_huge_pte_at(mm, address, ptep, pte);
3071 spin_unlock(&mm->page_table_lock);
3073 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3074 * may have cleared our pud entry and done put_page on the page table:
3075 * once we release i_mmap_mutex, another task can do the final put_page
3076 * and that page table be reused and filled with junk.
3078 flush_tlb_range(vma, start, end);
3079 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3081 return pages << h->order;
3084 int hugetlb_reserve_pages(struct inode *inode,
3086 struct vm_area_struct *vma,
3087 vm_flags_t vm_flags)
3090 struct hstate *h = hstate_inode(inode);
3091 struct hugepage_subpool *spool = subpool_inode(inode);
3094 * Only apply hugepage reservation if asked. At fault time, an
3095 * attempt will be made for VM_NORESERVE to allocate a page
3096 * without using reserves
3098 if (vm_flags & VM_NORESERVE)
3102 * Shared mappings base their reservation on the number of pages that
3103 * are already allocated on behalf of the file. Private mappings need
3104 * to reserve the full area even if read-only as mprotect() may be
3105 * called to make the mapping read-write. Assume !vma is a shm mapping
3107 if (!vma || vma->vm_flags & VM_MAYSHARE)
3108 chg = region_chg(&inode->i_mapping->private_list, from, to);
3110 struct resv_map *resv_map = resv_map_alloc();
3116 set_vma_resv_map(vma, resv_map);
3117 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3125 /* There must be enough pages in the subpool for the mapping */
3126 if (hugepage_subpool_get_pages(spool, chg)) {
3132 * Check enough hugepages are available for the reservation.
3133 * Hand the pages back to the subpool if there are not
3135 ret = hugetlb_acct_memory(h, chg);
3137 hugepage_subpool_put_pages(spool, chg);
3142 * Account for the reservations made. Shared mappings record regions
3143 * that have reservations as they are shared by multiple VMAs.
3144 * When the last VMA disappears, the region map says how much
3145 * the reservation was and the page cache tells how much of
3146 * the reservation was consumed. Private mappings are per-VMA and
3147 * only the consumed reservations are tracked. When the VMA
3148 * disappears, the original reservation is the VMA size and the
3149 * consumed reservations are stored in the map. Hence, nothing
3150 * else has to be done for private mappings here
3152 if (!vma || vma->vm_flags & VM_MAYSHARE)
3153 region_add(&inode->i_mapping->private_list, from, to);
3161 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3163 struct hstate *h = hstate_inode(inode);
3164 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3165 struct hugepage_subpool *spool = subpool_inode(inode);
3167 spin_lock(&inode->i_lock);
3168 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3169 spin_unlock(&inode->i_lock);
3171 hugepage_subpool_put_pages(spool, (chg - freed));
3172 hugetlb_acct_memory(h, -(chg - freed));
3175 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3176 static unsigned long page_table_shareable(struct vm_area_struct *svma,
3177 struct vm_area_struct *vma,
3178 unsigned long addr, pgoff_t idx)
3180 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3182 unsigned long sbase = saddr & PUD_MASK;
3183 unsigned long s_end = sbase + PUD_SIZE;
3185 /* Allow segments to share if only one is marked locked */
3186 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3187 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3190 * match the virtual addresses, permission and the alignment of the
3193 if (pmd_index(addr) != pmd_index(saddr) ||
3194 vm_flags != svm_flags ||
3195 sbase < svma->vm_start || svma->vm_end < s_end)
3201 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3203 unsigned long base = addr & PUD_MASK;
3204 unsigned long end = base + PUD_SIZE;
3207 * check on proper vm_flags and page table alignment
3209 if (vma->vm_flags & VM_MAYSHARE &&
3210 vma->vm_start <= base && end <= vma->vm_end)
3216 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3217 * and returns the corresponding pte. While this is not necessary for the
3218 * !shared pmd case because we can allocate the pmd later as well, it makes the
3219 * code much cleaner. pmd allocation is essential for the shared case because
3220 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3221 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3222 * bad pmd for sharing.
3224 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3226 struct vm_area_struct *vma = find_vma(mm, addr);
3227 struct address_space *mapping = vma->vm_file->f_mapping;
3228 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3230 struct vm_area_struct *svma;
3231 unsigned long saddr;
3235 if (!vma_shareable(vma, addr))
3236 return (pte_t *)pmd_alloc(mm, pud, addr);
3238 mutex_lock(&mapping->i_mmap_mutex);
3239 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3243 saddr = page_table_shareable(svma, vma, addr, idx);
3245 spte = huge_pte_offset(svma->vm_mm, saddr);
3247 get_page(virt_to_page(spte));
3256 spin_lock(&mm->page_table_lock);
3258 pud_populate(mm, pud,
3259 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3261 put_page(virt_to_page(spte));
3262 spin_unlock(&mm->page_table_lock);
3264 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3265 mutex_unlock(&mapping->i_mmap_mutex);
3270 * unmap huge page backed by shared pte.
3272 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3273 * indicated by page_count > 1, unmap is achieved by clearing pud and
3274 * decrementing the ref count. If count == 1, the pte page is not shared.
3276 * called with vma->vm_mm->page_table_lock held.
3278 * returns: 1 successfully unmapped a shared pte page
3279 * 0 the underlying pte page is not shared, or it is the last user
3281 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3283 pgd_t *pgd = pgd_offset(mm, *addr);
3284 pud_t *pud = pud_offset(pgd, *addr);
3286 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3287 if (page_count(virt_to_page(ptep)) == 1)
3291 put_page(virt_to_page(ptep));
3292 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3295 #define want_pmd_share() (1)
3296 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3297 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3301 #define want_pmd_share() (0)
3302 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3304 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3305 pte_t *huge_pte_alloc(struct mm_struct *mm,
3306 unsigned long addr, unsigned long sz)
3312 pgd = pgd_offset(mm, addr);
3313 pud = pud_alloc(mm, pgd, addr);
3315 if (sz == PUD_SIZE) {
3318 BUG_ON(sz != PMD_SIZE);
3319 if (want_pmd_share() && pud_none(*pud))
3320 pte = huge_pmd_share(mm, addr, pud);
3322 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3325 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3330 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3336 pgd = pgd_offset(mm, addr);
3337 if (pgd_present(*pgd)) {
3338 pud = pud_offset(pgd, addr);
3339 if (pud_present(*pud)) {
3341 return (pte_t *)pud;
3342 pmd = pmd_offset(pud, addr);
3345 return (pte_t *) pmd;
3349 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3350 pmd_t *pmd, int write)
3354 page = pte_page(*(pte_t *)pmd);
3356 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
3361 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3362 pud_t *pud, int write)
3366 page = pte_page(*(pte_t *)pud);
3368 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
3372 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3374 /* Can be overriden by architectures */
3375 __attribute__((weak)) struct page *
3376 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3377 pud_t *pud, int write)
3383 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3385 #ifdef CONFIG_MEMORY_FAILURE
3387 /* Should be called in hugetlb_lock */
3388 static int is_hugepage_on_freelist(struct page *hpage)
3392 struct hstate *h = page_hstate(hpage);
3393 int nid = page_to_nid(hpage);
3395 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3402 * This function is called from memory failure code.
3403 * Assume the caller holds page lock of the head page.
3405 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3407 struct hstate *h = page_hstate(hpage);
3408 int nid = page_to_nid(hpage);
3411 spin_lock(&hugetlb_lock);
3412 if (is_hugepage_on_freelist(hpage)) {
3414 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3415 * but dangling hpage->lru can trigger list-debug warnings
3416 * (this happens when we call unpoison_memory() on it),
3417 * so let it point to itself with list_del_init().
3419 list_del_init(&hpage->lru);
3420 set_page_refcounted(hpage);
3421 h->free_huge_pages--;
3422 h->free_huge_pages_node[nid]--;
3425 spin_unlock(&hugetlb_lock);