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_NORESERVE)
469 if (vma->vm_flags & VM_MAYSHARE)
471 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
476 static void copy_gigantic_page(struct page *dst, struct page *src)
479 struct hstate *h = page_hstate(src);
480 struct page *dst_base = dst;
481 struct page *src_base = src;
483 for (i = 0; i < pages_per_huge_page(h); ) {
485 copy_highpage(dst, src);
488 dst = mem_map_next(dst, dst_base, i);
489 src = mem_map_next(src, src_base, i);
493 void copy_huge_page(struct page *dst, struct page *src)
496 struct hstate *h = page_hstate(src);
498 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
499 copy_gigantic_page(dst, src);
504 for (i = 0; i < pages_per_huge_page(h); i++) {
506 copy_highpage(dst + i, src + i);
510 static void enqueue_huge_page(struct hstate *h, struct page *page)
512 int nid = page_to_nid(page);
513 list_move(&page->lru, &h->hugepage_freelists[nid]);
514 h->free_huge_pages++;
515 h->free_huge_pages_node[nid]++;
518 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
522 if (list_empty(&h->hugepage_freelists[nid]))
524 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
525 list_move(&page->lru, &h->hugepage_activelist);
526 set_page_refcounted(page);
527 h->free_huge_pages--;
528 h->free_huge_pages_node[nid]--;
532 static struct page *dequeue_huge_page_vma(struct hstate *h,
533 struct vm_area_struct *vma,
534 unsigned long address, int avoid_reserve)
536 struct page *page = NULL;
537 struct mempolicy *mpol;
538 nodemask_t *nodemask;
539 struct zonelist *zonelist;
542 unsigned int cpuset_mems_cookie;
545 * A child process with MAP_PRIVATE mappings created by their parent
546 * have no page reserves. This check ensures that reservations are
547 * not "stolen". The child may still get SIGKILLed
549 if (!vma_has_reserves(vma) &&
550 h->free_huge_pages - h->resv_huge_pages == 0)
553 /* If reserves cannot be used, ensure enough pages are in the pool */
554 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
558 cpuset_mems_cookie = get_mems_allowed();
559 zonelist = huge_zonelist(vma, address,
560 htlb_alloc_mask, &mpol, &nodemask);
562 for_each_zone_zonelist_nodemask(zone, z, zonelist,
563 MAX_NR_ZONES - 1, nodemask) {
564 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
565 page = dequeue_huge_page_node(h, zone_to_nid(zone));
568 decrement_hugepage_resv_vma(h, vma);
575 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
583 static void update_and_free_page(struct hstate *h, struct page *page)
587 VM_BUG_ON(h->order >= MAX_ORDER);
590 h->nr_huge_pages_node[page_to_nid(page)]--;
591 for (i = 0; i < pages_per_huge_page(h); i++) {
592 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
593 1 << PG_referenced | 1 << PG_dirty |
594 1 << PG_active | 1 << PG_reserved |
595 1 << PG_private | 1 << PG_writeback);
597 VM_BUG_ON(hugetlb_cgroup_from_page(page));
598 set_compound_page_dtor(page, NULL);
599 set_page_refcounted(page);
600 arch_release_hugepage(page);
601 __free_pages(page, huge_page_order(h));
604 struct hstate *size_to_hstate(unsigned long size)
609 if (huge_page_size(h) == size)
615 static void free_huge_page(struct page *page)
618 * Can't pass hstate in here because it is called from the
619 * compound page destructor.
621 struct hstate *h = page_hstate(page);
622 int nid = page_to_nid(page);
623 struct hugepage_subpool *spool =
624 (struct hugepage_subpool *)page_private(page);
626 set_page_private(page, 0);
627 page->mapping = NULL;
628 BUG_ON(page_count(page));
629 BUG_ON(page_mapcount(page));
631 spin_lock(&hugetlb_lock);
632 hugetlb_cgroup_uncharge_page(hstate_index(h),
633 pages_per_huge_page(h), page);
634 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
635 /* remove the page from active list */
636 list_del(&page->lru);
637 update_and_free_page(h, page);
638 h->surplus_huge_pages--;
639 h->surplus_huge_pages_node[nid]--;
641 arch_clear_hugepage_flags(page);
642 enqueue_huge_page(h, page);
644 spin_unlock(&hugetlb_lock);
645 hugepage_subpool_put_pages(spool, 1);
648 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
650 INIT_LIST_HEAD(&page->lru);
651 set_compound_page_dtor(page, free_huge_page);
652 spin_lock(&hugetlb_lock);
653 set_hugetlb_cgroup(page, NULL);
655 h->nr_huge_pages_node[nid]++;
656 spin_unlock(&hugetlb_lock);
657 put_page(page); /* free it into the hugepage allocator */
660 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
663 int nr_pages = 1 << order;
664 struct page *p = page + 1;
666 /* we rely on prep_new_huge_page to set the destructor */
667 set_compound_order(page, order);
669 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
671 set_page_count(p, 0);
672 p->first_page = page;
677 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
678 * transparent huge pages. See the PageTransHuge() documentation for more
681 int PageHuge(struct page *page)
683 compound_page_dtor *dtor;
685 if (!PageCompound(page))
688 page = compound_head(page);
689 dtor = get_compound_page_dtor(page);
691 return dtor == free_huge_page;
693 EXPORT_SYMBOL_GPL(PageHuge);
695 pgoff_t __basepage_index(struct page *page)
697 struct page *page_head = compound_head(page);
698 pgoff_t index = page_index(page_head);
699 unsigned long compound_idx;
701 if (!PageHuge(page_head))
702 return page_index(page);
704 if (compound_order(page_head) >= MAX_ORDER)
705 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
707 compound_idx = page - page_head;
709 return (index << compound_order(page_head)) + compound_idx;
712 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
716 if (h->order >= MAX_ORDER)
719 page = alloc_pages_exact_node(nid,
720 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
721 __GFP_REPEAT|__GFP_NOWARN,
724 if (arch_prepare_hugepage(page)) {
725 __free_pages(page, huge_page_order(h));
728 prep_new_huge_page(h, page, nid);
735 * common helper functions for hstate_next_node_to_{alloc|free}.
736 * We may have allocated or freed a huge page based on a different
737 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
738 * be outside of *nodes_allowed. Ensure that we use an allowed
739 * node for alloc or free.
741 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
743 nid = next_node(nid, *nodes_allowed);
744 if (nid == MAX_NUMNODES)
745 nid = first_node(*nodes_allowed);
746 VM_BUG_ON(nid >= MAX_NUMNODES);
751 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
753 if (!node_isset(nid, *nodes_allowed))
754 nid = next_node_allowed(nid, nodes_allowed);
759 * returns the previously saved node ["this node"] from which to
760 * allocate a persistent huge page for the pool and advance the
761 * next node from which to allocate, handling wrap at end of node
764 static int hstate_next_node_to_alloc(struct hstate *h,
765 nodemask_t *nodes_allowed)
769 VM_BUG_ON(!nodes_allowed);
771 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
772 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
778 * helper for free_pool_huge_page() - return the previously saved
779 * node ["this node"] from which to free a huge page. Advance the
780 * next node id whether or not we find a free huge page to free so
781 * that the next attempt to free addresses the next node.
783 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
787 VM_BUG_ON(!nodes_allowed);
789 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
790 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
795 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
796 for (nr_nodes = nodes_weight(*mask); \
798 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
801 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
802 for (nr_nodes = nodes_weight(*mask); \
804 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
807 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
813 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
814 page = alloc_fresh_huge_page_node(h, node);
822 count_vm_event(HTLB_BUDDY_PGALLOC);
824 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
830 * Free huge page from pool from next node to free.
831 * Attempt to keep persistent huge pages more or less
832 * balanced over allowed nodes.
833 * Called with hugetlb_lock locked.
835 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
841 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
843 * If we're returning unused surplus pages, only examine
844 * nodes with surplus pages.
846 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
847 !list_empty(&h->hugepage_freelists[node])) {
849 list_entry(h->hugepage_freelists[node].next,
851 list_del(&page->lru);
852 h->free_huge_pages--;
853 h->free_huge_pages_node[node]--;
855 h->surplus_huge_pages--;
856 h->surplus_huge_pages_node[node]--;
858 update_and_free_page(h, page);
867 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
872 if (h->order >= MAX_ORDER)
876 * Assume we will successfully allocate the surplus page to
877 * prevent racing processes from causing the surplus to exceed
880 * This however introduces a different race, where a process B
881 * tries to grow the static hugepage pool while alloc_pages() is
882 * called by process A. B will only examine the per-node
883 * counters in determining if surplus huge pages can be
884 * converted to normal huge pages in adjust_pool_surplus(). A
885 * won't be able to increment the per-node counter, until the
886 * lock is dropped by B, but B doesn't drop hugetlb_lock until
887 * no more huge pages can be converted from surplus to normal
888 * state (and doesn't try to convert again). Thus, we have a
889 * case where a surplus huge page exists, the pool is grown, and
890 * the surplus huge page still exists after, even though it
891 * should just have been converted to a normal huge page. This
892 * does not leak memory, though, as the hugepage will be freed
893 * once it is out of use. It also does not allow the counters to
894 * go out of whack in adjust_pool_surplus() as we don't modify
895 * the node values until we've gotten the hugepage and only the
896 * per-node value is checked there.
898 spin_lock(&hugetlb_lock);
899 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
900 spin_unlock(&hugetlb_lock);
904 h->surplus_huge_pages++;
906 spin_unlock(&hugetlb_lock);
908 if (nid == NUMA_NO_NODE)
909 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
910 __GFP_REPEAT|__GFP_NOWARN,
913 page = alloc_pages_exact_node(nid,
914 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
915 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
917 if (page && arch_prepare_hugepage(page)) {
918 __free_pages(page, huge_page_order(h));
922 spin_lock(&hugetlb_lock);
924 INIT_LIST_HEAD(&page->lru);
925 r_nid = page_to_nid(page);
926 set_compound_page_dtor(page, free_huge_page);
927 set_hugetlb_cgroup(page, NULL);
929 * We incremented the global counters already
931 h->nr_huge_pages_node[r_nid]++;
932 h->surplus_huge_pages_node[r_nid]++;
933 __count_vm_event(HTLB_BUDDY_PGALLOC);
936 h->surplus_huge_pages--;
937 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
939 spin_unlock(&hugetlb_lock);
945 * This allocation function is useful in the context where vma is irrelevant.
946 * E.g. soft-offlining uses this function because it only cares physical
947 * address of error page.
949 struct page *alloc_huge_page_node(struct hstate *h, int nid)
953 spin_lock(&hugetlb_lock);
954 page = dequeue_huge_page_node(h, nid);
955 spin_unlock(&hugetlb_lock);
958 page = alloc_buddy_huge_page(h, nid);
964 * Increase the hugetlb pool such that it can accommodate a reservation
967 static int gather_surplus_pages(struct hstate *h, int delta)
969 struct list_head surplus_list;
970 struct page *page, *tmp;
972 int needed, allocated;
973 bool alloc_ok = true;
975 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
977 h->resv_huge_pages += delta;
982 INIT_LIST_HEAD(&surplus_list);
986 spin_unlock(&hugetlb_lock);
987 for (i = 0; i < needed; i++) {
988 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
993 list_add(&page->lru, &surplus_list);
998 * After retaking hugetlb_lock, we need to recalculate 'needed'
999 * because either resv_huge_pages or free_huge_pages may have changed.
1001 spin_lock(&hugetlb_lock);
1002 needed = (h->resv_huge_pages + delta) -
1003 (h->free_huge_pages + allocated);
1008 * We were not able to allocate enough pages to
1009 * satisfy the entire reservation so we free what
1010 * we've allocated so far.
1015 * The surplus_list now contains _at_least_ the number of extra pages
1016 * needed to accommodate the reservation. Add the appropriate number
1017 * of pages to the hugetlb pool and free the extras back to the buddy
1018 * allocator. Commit the entire reservation here to prevent another
1019 * process from stealing the pages as they are added to the pool but
1020 * before they are reserved.
1022 needed += allocated;
1023 h->resv_huge_pages += delta;
1026 /* Free the needed pages to the hugetlb pool */
1027 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1031 * This page is now managed by the hugetlb allocator and has
1032 * no users -- drop the buddy allocator's reference.
1034 put_page_testzero(page);
1035 VM_BUG_ON(page_count(page));
1036 enqueue_huge_page(h, page);
1039 spin_unlock(&hugetlb_lock);
1041 /* Free unnecessary surplus pages to the buddy allocator */
1042 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1044 spin_lock(&hugetlb_lock);
1050 * When releasing a hugetlb pool reservation, any surplus pages that were
1051 * allocated to satisfy the reservation must be explicitly freed if they were
1053 * Called with hugetlb_lock held.
1055 static void return_unused_surplus_pages(struct hstate *h,
1056 unsigned long unused_resv_pages)
1058 unsigned long nr_pages;
1060 /* Uncommit the reservation */
1061 h->resv_huge_pages -= unused_resv_pages;
1063 /* Cannot return gigantic pages currently */
1064 if (h->order >= MAX_ORDER)
1067 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1070 * We want to release as many surplus pages as possible, spread
1071 * evenly across all nodes with memory. Iterate across these nodes
1072 * until we can no longer free unreserved surplus pages. This occurs
1073 * when the nodes with surplus pages have no free pages.
1074 * free_pool_huge_page() will balance the the freed pages across the
1075 * on-line nodes with memory and will handle the hstate accounting.
1077 while (nr_pages--) {
1078 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1084 * Determine if the huge page at addr within the vma has an associated
1085 * reservation. Where it does not we will need to logically increase
1086 * reservation and actually increase subpool usage before an allocation
1087 * can occur. Where any new reservation would be required the
1088 * reservation change is prepared, but not committed. Once the page
1089 * has been allocated from the subpool and instantiated the change should
1090 * be committed via vma_commit_reservation. No action is required on
1093 static long vma_needs_reservation(struct hstate *h,
1094 struct vm_area_struct *vma, unsigned long addr)
1096 struct address_space *mapping = vma->vm_file->f_mapping;
1097 struct inode *inode = mapping->host;
1099 if (vma->vm_flags & VM_MAYSHARE) {
1100 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1101 return region_chg(&inode->i_mapping->private_list,
1104 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1109 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1110 struct resv_map *reservations = vma_resv_map(vma);
1112 err = region_chg(&reservations->regions, idx, idx + 1);
1118 static void vma_commit_reservation(struct hstate *h,
1119 struct vm_area_struct *vma, unsigned long addr)
1121 struct address_space *mapping = vma->vm_file->f_mapping;
1122 struct inode *inode = mapping->host;
1124 if (vma->vm_flags & VM_MAYSHARE) {
1125 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1126 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1128 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1129 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1130 struct resv_map *reservations = vma_resv_map(vma);
1132 /* Mark this page used in the map. */
1133 region_add(&reservations->regions, idx, idx + 1);
1137 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1138 unsigned long addr, int avoid_reserve)
1140 struct hugepage_subpool *spool = subpool_vma(vma);
1141 struct hstate *h = hstate_vma(vma);
1145 struct hugetlb_cgroup *h_cg;
1147 idx = hstate_index(h);
1149 * Processes that did not create the mapping will have no
1150 * reserves and will not have accounted against subpool
1151 * limit. Check that the subpool limit can be made before
1152 * satisfying the allocation MAP_NORESERVE mappings may also
1153 * need pages and subpool limit allocated allocated if no reserve
1156 chg = vma_needs_reservation(h, vma, addr);
1158 return ERR_PTR(-ENOMEM);
1160 if (hugepage_subpool_get_pages(spool, chg))
1161 return ERR_PTR(-ENOSPC);
1163 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1165 hugepage_subpool_put_pages(spool, chg);
1166 return ERR_PTR(-ENOSPC);
1168 spin_lock(&hugetlb_lock);
1169 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1171 spin_unlock(&hugetlb_lock);
1172 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1174 hugetlb_cgroup_uncharge_cgroup(idx,
1175 pages_per_huge_page(h),
1177 hugepage_subpool_put_pages(spool, chg);
1178 return ERR_PTR(-ENOSPC);
1180 spin_lock(&hugetlb_lock);
1181 list_move(&page->lru, &h->hugepage_activelist);
1184 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1185 spin_unlock(&hugetlb_lock);
1187 set_page_private(page, (unsigned long)spool);
1189 vma_commit_reservation(h, vma, addr);
1193 int __weak alloc_bootmem_huge_page(struct hstate *h)
1195 struct huge_bootmem_page *m;
1198 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1201 addr = __alloc_bootmem_node_nopanic(NODE_DATA(node),
1202 huge_page_size(h), huge_page_size(h), 0);
1206 * Use the beginning of the huge page to store the
1207 * huge_bootmem_page struct (until gather_bootmem
1208 * puts them into the mem_map).
1217 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1218 /* Put them into a private list first because mem_map is not up yet */
1219 list_add(&m->list, &huge_boot_pages);
1224 static void prep_compound_huge_page(struct page *page, int order)
1226 if (unlikely(order > (MAX_ORDER - 1)))
1227 prep_compound_gigantic_page(page, order);
1229 prep_compound_page(page, order);
1232 /* Put bootmem huge pages into the standard lists after mem_map is up */
1233 static void __init gather_bootmem_prealloc(void)
1235 struct huge_bootmem_page *m;
1237 list_for_each_entry(m, &huge_boot_pages, list) {
1238 struct hstate *h = m->hstate;
1241 #ifdef CONFIG_HIGHMEM
1242 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1243 free_bootmem_late((unsigned long)m,
1244 sizeof(struct huge_bootmem_page));
1246 page = virt_to_page(m);
1248 __ClearPageReserved(page);
1249 WARN_ON(page_count(page) != 1);
1250 prep_compound_huge_page(page, h->order);
1251 prep_new_huge_page(h, page, page_to_nid(page));
1253 * If we had gigantic hugepages allocated at boot time, we need
1254 * to restore the 'stolen' pages to totalram_pages in order to
1255 * fix confusing memory reports from free(1) and another
1256 * side-effects, like CommitLimit going negative.
1258 if (h->order > (MAX_ORDER - 1))
1259 adjust_managed_page_count(page, 1 << h->order);
1263 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1267 for (i = 0; i < h->max_huge_pages; ++i) {
1268 if (h->order >= MAX_ORDER) {
1269 if (!alloc_bootmem_huge_page(h))
1271 } else if (!alloc_fresh_huge_page(h,
1272 &node_states[N_MEMORY]))
1275 h->max_huge_pages = i;
1278 static void __init hugetlb_init_hstates(void)
1282 for_each_hstate(h) {
1283 /* oversize hugepages were init'ed in early boot */
1284 if (h->order < MAX_ORDER)
1285 hugetlb_hstate_alloc_pages(h);
1289 static char * __init memfmt(char *buf, unsigned long n)
1291 if (n >= (1UL << 30))
1292 sprintf(buf, "%lu GB", n >> 30);
1293 else if (n >= (1UL << 20))
1294 sprintf(buf, "%lu MB", n >> 20);
1296 sprintf(buf, "%lu KB", n >> 10);
1300 static void __init report_hugepages(void)
1304 for_each_hstate(h) {
1306 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1307 memfmt(buf, huge_page_size(h)),
1308 h->free_huge_pages);
1312 #ifdef CONFIG_HIGHMEM
1313 static void try_to_free_low(struct hstate *h, unsigned long count,
1314 nodemask_t *nodes_allowed)
1318 if (h->order >= MAX_ORDER)
1321 for_each_node_mask(i, *nodes_allowed) {
1322 struct page *page, *next;
1323 struct list_head *freel = &h->hugepage_freelists[i];
1324 list_for_each_entry_safe(page, next, freel, lru) {
1325 if (count >= h->nr_huge_pages)
1327 if (PageHighMem(page))
1329 list_del(&page->lru);
1330 update_and_free_page(h, page);
1331 h->free_huge_pages--;
1332 h->free_huge_pages_node[page_to_nid(page)]--;
1337 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1338 nodemask_t *nodes_allowed)
1344 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1345 * balanced by operating on them in a round-robin fashion.
1346 * Returns 1 if an adjustment was made.
1348 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1353 VM_BUG_ON(delta != -1 && delta != 1);
1356 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1357 if (h->surplus_huge_pages_node[node])
1361 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1362 if (h->surplus_huge_pages_node[node] <
1363 h->nr_huge_pages_node[node])
1370 h->surplus_huge_pages += delta;
1371 h->surplus_huge_pages_node[node] += delta;
1375 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1376 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1377 nodemask_t *nodes_allowed)
1379 unsigned long min_count, ret;
1381 if (h->order >= MAX_ORDER)
1382 return h->max_huge_pages;
1385 * Increase the pool size
1386 * First take pages out of surplus state. Then make up the
1387 * remaining difference by allocating fresh huge pages.
1389 * We might race with alloc_buddy_huge_page() here and be unable
1390 * to convert a surplus huge page to a normal huge page. That is
1391 * not critical, though, it just means the overall size of the
1392 * pool might be one hugepage larger than it needs to be, but
1393 * within all the constraints specified by the sysctls.
1395 spin_lock(&hugetlb_lock);
1396 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1397 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1401 while (count > persistent_huge_pages(h)) {
1403 * If this allocation races such that we no longer need the
1404 * page, free_huge_page will handle it by freeing the page
1405 * and reducing the surplus.
1407 spin_unlock(&hugetlb_lock);
1408 ret = alloc_fresh_huge_page(h, nodes_allowed);
1409 spin_lock(&hugetlb_lock);
1413 /* Bail for signals. Probably ctrl-c from user */
1414 if (signal_pending(current))
1419 * Decrease the pool size
1420 * First return free pages to the buddy allocator (being careful
1421 * to keep enough around to satisfy reservations). Then place
1422 * pages into surplus state as needed so the pool will shrink
1423 * to the desired size as pages become free.
1425 * By placing pages into the surplus state independent of the
1426 * overcommit value, we are allowing the surplus pool size to
1427 * exceed overcommit. There are few sane options here. Since
1428 * alloc_buddy_huge_page() is checking the global counter,
1429 * though, we'll note that we're not allowed to exceed surplus
1430 * and won't grow the pool anywhere else. Not until one of the
1431 * sysctls are changed, or the surplus pages go out of use.
1433 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1434 min_count = max(count, min_count);
1435 try_to_free_low(h, min_count, nodes_allowed);
1436 while (min_count < persistent_huge_pages(h)) {
1437 if (!free_pool_huge_page(h, nodes_allowed, 0))
1440 while (count < persistent_huge_pages(h)) {
1441 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1445 ret = persistent_huge_pages(h);
1446 spin_unlock(&hugetlb_lock);
1450 #define HSTATE_ATTR_RO(_name) \
1451 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1453 #define HSTATE_ATTR(_name) \
1454 static struct kobj_attribute _name##_attr = \
1455 __ATTR(_name, 0644, _name##_show, _name##_store)
1457 static struct kobject *hugepages_kobj;
1458 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1460 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1462 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1466 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1467 if (hstate_kobjs[i] == kobj) {
1469 *nidp = NUMA_NO_NODE;
1473 return kobj_to_node_hstate(kobj, nidp);
1476 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1477 struct kobj_attribute *attr, char *buf)
1480 unsigned long nr_huge_pages;
1483 h = kobj_to_hstate(kobj, &nid);
1484 if (nid == NUMA_NO_NODE)
1485 nr_huge_pages = h->nr_huge_pages;
1487 nr_huge_pages = h->nr_huge_pages_node[nid];
1489 return sprintf(buf, "%lu\n", nr_huge_pages);
1492 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1493 struct kobject *kobj, struct kobj_attribute *attr,
1494 const char *buf, size_t len)
1498 unsigned long count;
1500 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1502 err = kstrtoul(buf, 10, &count);
1506 h = kobj_to_hstate(kobj, &nid);
1507 if (h->order >= MAX_ORDER) {
1512 if (nid == NUMA_NO_NODE) {
1514 * global hstate attribute
1516 if (!(obey_mempolicy &&
1517 init_nodemask_of_mempolicy(nodes_allowed))) {
1518 NODEMASK_FREE(nodes_allowed);
1519 nodes_allowed = &node_states[N_MEMORY];
1521 } else if (nodes_allowed) {
1523 * per node hstate attribute: adjust count to global,
1524 * but restrict alloc/free to the specified node.
1526 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1527 init_nodemask_of_node(nodes_allowed, nid);
1529 nodes_allowed = &node_states[N_MEMORY];
1531 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1533 if (nodes_allowed != &node_states[N_MEMORY])
1534 NODEMASK_FREE(nodes_allowed);
1538 NODEMASK_FREE(nodes_allowed);
1542 static ssize_t nr_hugepages_show(struct kobject *kobj,
1543 struct kobj_attribute *attr, char *buf)
1545 return nr_hugepages_show_common(kobj, attr, buf);
1548 static ssize_t nr_hugepages_store(struct kobject *kobj,
1549 struct kobj_attribute *attr, const char *buf, size_t len)
1551 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1553 HSTATE_ATTR(nr_hugepages);
1558 * hstate attribute for optionally mempolicy-based constraint on persistent
1559 * huge page alloc/free.
1561 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1562 struct kobj_attribute *attr, char *buf)
1564 return nr_hugepages_show_common(kobj, attr, buf);
1567 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1568 struct kobj_attribute *attr, const char *buf, size_t len)
1570 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1572 HSTATE_ATTR(nr_hugepages_mempolicy);
1576 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1577 struct kobj_attribute *attr, char *buf)
1579 struct hstate *h = kobj_to_hstate(kobj, NULL);
1580 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1583 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1584 struct kobj_attribute *attr, const char *buf, size_t count)
1587 unsigned long input;
1588 struct hstate *h = kobj_to_hstate(kobj, NULL);
1590 if (h->order >= MAX_ORDER)
1593 err = kstrtoul(buf, 10, &input);
1597 spin_lock(&hugetlb_lock);
1598 h->nr_overcommit_huge_pages = input;
1599 spin_unlock(&hugetlb_lock);
1603 HSTATE_ATTR(nr_overcommit_hugepages);
1605 static ssize_t free_hugepages_show(struct kobject *kobj,
1606 struct kobj_attribute *attr, char *buf)
1609 unsigned long free_huge_pages;
1612 h = kobj_to_hstate(kobj, &nid);
1613 if (nid == NUMA_NO_NODE)
1614 free_huge_pages = h->free_huge_pages;
1616 free_huge_pages = h->free_huge_pages_node[nid];
1618 return sprintf(buf, "%lu\n", free_huge_pages);
1620 HSTATE_ATTR_RO(free_hugepages);
1622 static ssize_t resv_hugepages_show(struct kobject *kobj,
1623 struct kobj_attribute *attr, char *buf)
1625 struct hstate *h = kobj_to_hstate(kobj, NULL);
1626 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1628 HSTATE_ATTR_RO(resv_hugepages);
1630 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1631 struct kobj_attribute *attr, char *buf)
1634 unsigned long surplus_huge_pages;
1637 h = kobj_to_hstate(kobj, &nid);
1638 if (nid == NUMA_NO_NODE)
1639 surplus_huge_pages = h->surplus_huge_pages;
1641 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1643 return sprintf(buf, "%lu\n", surplus_huge_pages);
1645 HSTATE_ATTR_RO(surplus_hugepages);
1647 static struct attribute *hstate_attrs[] = {
1648 &nr_hugepages_attr.attr,
1649 &nr_overcommit_hugepages_attr.attr,
1650 &free_hugepages_attr.attr,
1651 &resv_hugepages_attr.attr,
1652 &surplus_hugepages_attr.attr,
1654 &nr_hugepages_mempolicy_attr.attr,
1659 static struct attribute_group hstate_attr_group = {
1660 .attrs = hstate_attrs,
1663 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1664 struct kobject **hstate_kobjs,
1665 struct attribute_group *hstate_attr_group)
1668 int hi = hstate_index(h);
1670 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1671 if (!hstate_kobjs[hi])
1674 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1676 kobject_put(hstate_kobjs[hi]);
1681 static void __init hugetlb_sysfs_init(void)
1686 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1687 if (!hugepages_kobj)
1690 for_each_hstate(h) {
1691 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1692 hstate_kobjs, &hstate_attr_group);
1694 pr_err("Hugetlb: Unable to add hstate %s", h->name);
1701 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1702 * with node devices in node_devices[] using a parallel array. The array
1703 * index of a node device or _hstate == node id.
1704 * This is here to avoid any static dependency of the node device driver, in
1705 * the base kernel, on the hugetlb module.
1707 struct node_hstate {
1708 struct kobject *hugepages_kobj;
1709 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1711 struct node_hstate node_hstates[MAX_NUMNODES];
1714 * A subset of global hstate attributes for node devices
1716 static struct attribute *per_node_hstate_attrs[] = {
1717 &nr_hugepages_attr.attr,
1718 &free_hugepages_attr.attr,
1719 &surplus_hugepages_attr.attr,
1723 static struct attribute_group per_node_hstate_attr_group = {
1724 .attrs = per_node_hstate_attrs,
1728 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1729 * Returns node id via non-NULL nidp.
1731 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1735 for (nid = 0; nid < nr_node_ids; nid++) {
1736 struct node_hstate *nhs = &node_hstates[nid];
1738 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1739 if (nhs->hstate_kobjs[i] == kobj) {
1751 * Unregister hstate attributes from a single node device.
1752 * No-op if no hstate attributes attached.
1754 static void hugetlb_unregister_node(struct node *node)
1757 struct node_hstate *nhs = &node_hstates[node->dev.id];
1759 if (!nhs->hugepages_kobj)
1760 return; /* no hstate attributes */
1762 for_each_hstate(h) {
1763 int idx = hstate_index(h);
1764 if (nhs->hstate_kobjs[idx]) {
1765 kobject_put(nhs->hstate_kobjs[idx]);
1766 nhs->hstate_kobjs[idx] = NULL;
1770 kobject_put(nhs->hugepages_kobj);
1771 nhs->hugepages_kobj = NULL;
1775 * hugetlb module exit: unregister hstate attributes from node devices
1778 static void hugetlb_unregister_all_nodes(void)
1783 * disable node device registrations.
1785 register_hugetlbfs_with_node(NULL, NULL);
1788 * remove hstate attributes from any nodes that have them.
1790 for (nid = 0; nid < nr_node_ids; nid++)
1791 hugetlb_unregister_node(node_devices[nid]);
1795 * Register hstate attributes for a single node device.
1796 * No-op if attributes already registered.
1798 static void hugetlb_register_node(struct node *node)
1801 struct node_hstate *nhs = &node_hstates[node->dev.id];
1804 if (nhs->hugepages_kobj)
1805 return; /* already allocated */
1807 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1809 if (!nhs->hugepages_kobj)
1812 for_each_hstate(h) {
1813 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1815 &per_node_hstate_attr_group);
1817 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1818 h->name, node->dev.id);
1819 hugetlb_unregister_node(node);
1826 * hugetlb init time: register hstate attributes for all registered node
1827 * devices of nodes that have memory. All on-line nodes should have
1828 * registered their associated device by this time.
1830 static void hugetlb_register_all_nodes(void)
1834 for_each_node_state(nid, N_MEMORY) {
1835 struct node *node = node_devices[nid];
1836 if (node->dev.id == nid)
1837 hugetlb_register_node(node);
1841 * Let the node device driver know we're here so it can
1842 * [un]register hstate attributes on node hotplug.
1844 register_hugetlbfs_with_node(hugetlb_register_node,
1845 hugetlb_unregister_node);
1847 #else /* !CONFIG_NUMA */
1849 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1857 static void hugetlb_unregister_all_nodes(void) { }
1859 static void hugetlb_register_all_nodes(void) { }
1863 static void __exit hugetlb_exit(void)
1867 hugetlb_unregister_all_nodes();
1869 for_each_hstate(h) {
1870 kobject_put(hstate_kobjs[hstate_index(h)]);
1873 kobject_put(hugepages_kobj);
1875 module_exit(hugetlb_exit);
1877 static int __init hugetlb_init(void)
1879 /* Some platform decide whether they support huge pages at boot
1880 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1881 * there is no such support
1883 if (HPAGE_SHIFT == 0)
1886 if (!size_to_hstate(default_hstate_size)) {
1887 default_hstate_size = HPAGE_SIZE;
1888 if (!size_to_hstate(default_hstate_size))
1889 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1891 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1892 if (default_hstate_max_huge_pages)
1893 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1895 hugetlb_init_hstates();
1896 gather_bootmem_prealloc();
1899 hugetlb_sysfs_init();
1900 hugetlb_register_all_nodes();
1901 hugetlb_cgroup_file_init();
1905 module_init(hugetlb_init);
1907 /* Should be called on processing a hugepagesz=... option */
1908 void __init hugetlb_add_hstate(unsigned order)
1913 if (size_to_hstate(PAGE_SIZE << order)) {
1914 pr_warning("hugepagesz= specified twice, ignoring\n");
1917 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1919 h = &hstates[hugetlb_max_hstate++];
1921 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1922 h->nr_huge_pages = 0;
1923 h->free_huge_pages = 0;
1924 for (i = 0; i < MAX_NUMNODES; ++i)
1925 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1926 INIT_LIST_HEAD(&h->hugepage_activelist);
1927 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
1928 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
1929 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1930 huge_page_size(h)/1024);
1935 static int __init hugetlb_nrpages_setup(char *s)
1938 static unsigned long *last_mhp;
1941 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1942 * so this hugepages= parameter goes to the "default hstate".
1944 if (!hugetlb_max_hstate)
1945 mhp = &default_hstate_max_huge_pages;
1947 mhp = &parsed_hstate->max_huge_pages;
1949 if (mhp == last_mhp) {
1950 pr_warning("hugepages= specified twice without "
1951 "interleaving hugepagesz=, ignoring\n");
1955 if (sscanf(s, "%lu", mhp) <= 0)
1959 * Global state is always initialized later in hugetlb_init.
1960 * But we need to allocate >= MAX_ORDER hstates here early to still
1961 * use the bootmem allocator.
1963 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
1964 hugetlb_hstate_alloc_pages(parsed_hstate);
1970 __setup("hugepages=", hugetlb_nrpages_setup);
1972 static int __init hugetlb_default_setup(char *s)
1974 default_hstate_size = memparse(s, &s);
1977 __setup("default_hugepagesz=", hugetlb_default_setup);
1979 static unsigned int cpuset_mems_nr(unsigned int *array)
1982 unsigned int nr = 0;
1984 for_each_node_mask(node, cpuset_current_mems_allowed)
1990 #ifdef CONFIG_SYSCTL
1991 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1992 struct ctl_table *table, int write,
1993 void __user *buffer, size_t *length, loff_t *ppos)
1995 struct hstate *h = &default_hstate;
1999 tmp = h->max_huge_pages;
2001 if (write && h->order >= MAX_ORDER)
2005 table->maxlen = sizeof(unsigned long);
2006 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2011 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2012 GFP_KERNEL | __GFP_NORETRY);
2013 if (!(obey_mempolicy &&
2014 init_nodemask_of_mempolicy(nodes_allowed))) {
2015 NODEMASK_FREE(nodes_allowed);
2016 nodes_allowed = &node_states[N_MEMORY];
2018 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2020 if (nodes_allowed != &node_states[N_MEMORY])
2021 NODEMASK_FREE(nodes_allowed);
2027 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2028 void __user *buffer, size_t *length, loff_t *ppos)
2031 return hugetlb_sysctl_handler_common(false, table, write,
2032 buffer, length, ppos);
2036 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2037 void __user *buffer, size_t *length, loff_t *ppos)
2039 return hugetlb_sysctl_handler_common(true, table, write,
2040 buffer, length, ppos);
2042 #endif /* CONFIG_NUMA */
2044 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2045 void __user *buffer,
2046 size_t *length, loff_t *ppos)
2048 proc_dointvec(table, write, buffer, length, ppos);
2049 if (hugepages_treat_as_movable)
2050 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2052 htlb_alloc_mask = GFP_HIGHUSER;
2056 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2057 void __user *buffer,
2058 size_t *length, loff_t *ppos)
2060 struct hstate *h = &default_hstate;
2064 tmp = h->nr_overcommit_huge_pages;
2066 if (write && h->order >= MAX_ORDER)
2070 table->maxlen = sizeof(unsigned long);
2071 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2076 spin_lock(&hugetlb_lock);
2077 h->nr_overcommit_huge_pages = tmp;
2078 spin_unlock(&hugetlb_lock);
2084 #endif /* CONFIG_SYSCTL */
2086 void hugetlb_report_meminfo(struct seq_file *m)
2088 struct hstate *h = &default_hstate;
2090 "HugePages_Total: %5lu\n"
2091 "HugePages_Free: %5lu\n"
2092 "HugePages_Rsvd: %5lu\n"
2093 "HugePages_Surp: %5lu\n"
2094 "Hugepagesize: %8lu kB\n",
2098 h->surplus_huge_pages,
2099 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2102 int hugetlb_report_node_meminfo(int nid, char *buf)
2104 struct hstate *h = &default_hstate;
2106 "Node %d HugePages_Total: %5u\n"
2107 "Node %d HugePages_Free: %5u\n"
2108 "Node %d HugePages_Surp: %5u\n",
2109 nid, h->nr_huge_pages_node[nid],
2110 nid, h->free_huge_pages_node[nid],
2111 nid, h->surplus_huge_pages_node[nid]);
2114 void hugetlb_show_meminfo(void)
2119 for_each_node_state(nid, N_MEMORY)
2121 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2123 h->nr_huge_pages_node[nid],
2124 h->free_huge_pages_node[nid],
2125 h->surplus_huge_pages_node[nid],
2126 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2129 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2130 unsigned long hugetlb_total_pages(void)
2133 unsigned long nr_total_pages = 0;
2136 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2137 return nr_total_pages;
2140 static int hugetlb_acct_memory(struct hstate *h, long delta)
2144 spin_lock(&hugetlb_lock);
2146 * When cpuset is configured, it breaks the strict hugetlb page
2147 * reservation as the accounting is done on a global variable. Such
2148 * reservation is completely rubbish in the presence of cpuset because
2149 * the reservation is not checked against page availability for the
2150 * current cpuset. Application can still potentially OOM'ed by kernel
2151 * with lack of free htlb page in cpuset that the task is in.
2152 * Attempt to enforce strict accounting with cpuset is almost
2153 * impossible (or too ugly) because cpuset is too fluid that
2154 * task or memory node can be dynamically moved between cpusets.
2156 * The change of semantics for shared hugetlb mapping with cpuset is
2157 * undesirable. However, in order to preserve some of the semantics,
2158 * we fall back to check against current free page availability as
2159 * a best attempt and hopefully to minimize the impact of changing
2160 * semantics that cpuset has.
2163 if (gather_surplus_pages(h, delta) < 0)
2166 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2167 return_unused_surplus_pages(h, delta);
2174 return_unused_surplus_pages(h, (unsigned long) -delta);
2177 spin_unlock(&hugetlb_lock);
2181 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2183 struct resv_map *reservations = vma_resv_map(vma);
2186 * This new VMA should share its siblings reservation map if present.
2187 * The VMA will only ever have a valid reservation map pointer where
2188 * it is being copied for another still existing VMA. As that VMA
2189 * has a reference to the reservation map it cannot disappear until
2190 * after this open call completes. It is therefore safe to take a
2191 * new reference here without additional locking.
2194 kref_get(&reservations->refs);
2197 static void resv_map_put(struct vm_area_struct *vma)
2199 struct resv_map *reservations = vma_resv_map(vma);
2203 kref_put(&reservations->refs, resv_map_release);
2206 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2208 struct hstate *h = hstate_vma(vma);
2209 struct resv_map *reservations = vma_resv_map(vma);
2210 struct hugepage_subpool *spool = subpool_vma(vma);
2211 unsigned long reserve;
2212 unsigned long start;
2216 start = vma_hugecache_offset(h, vma, vma->vm_start);
2217 end = vma_hugecache_offset(h, vma, vma->vm_end);
2219 reserve = (end - start) -
2220 region_count(&reservations->regions, start, end);
2225 hugetlb_acct_memory(h, -reserve);
2226 hugepage_subpool_put_pages(spool, reserve);
2232 * We cannot handle pagefaults against hugetlb pages at all. They cause
2233 * handle_mm_fault() to try to instantiate regular-sized pages in the
2234 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2237 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2243 const struct vm_operations_struct hugetlb_vm_ops = {
2244 .fault = hugetlb_vm_op_fault,
2245 .open = hugetlb_vm_op_open,
2246 .close = hugetlb_vm_op_close,
2249 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2255 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2256 vma->vm_page_prot)));
2258 entry = huge_pte_wrprotect(mk_huge_pte(page,
2259 vma->vm_page_prot));
2261 entry = pte_mkyoung(entry);
2262 entry = pte_mkhuge(entry);
2263 entry = arch_make_huge_pte(entry, vma, page, writable);
2268 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2269 unsigned long address, pte_t *ptep)
2273 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2274 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2275 update_mmu_cache(vma, address, ptep);
2279 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2280 struct vm_area_struct *vma)
2282 pte_t *src_pte, *dst_pte, entry;
2283 struct page *ptepage;
2286 struct hstate *h = hstate_vma(vma);
2287 unsigned long sz = huge_page_size(h);
2289 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2291 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2292 src_pte = huge_pte_offset(src, addr);
2295 dst_pte = huge_pte_alloc(dst, addr, sz);
2299 /* If the pagetables are shared don't copy or take references */
2300 if (dst_pte == src_pte)
2303 spin_lock(&dst->page_table_lock);
2304 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2305 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2307 huge_ptep_set_wrprotect(src, addr, src_pte);
2308 entry = huge_ptep_get(src_pte);
2309 ptepage = pte_page(entry);
2311 page_dup_rmap(ptepage);
2312 set_huge_pte_at(dst, addr, dst_pte, entry);
2314 spin_unlock(&src->page_table_lock);
2315 spin_unlock(&dst->page_table_lock);
2323 static int is_hugetlb_entry_migration(pte_t pte)
2327 if (huge_pte_none(pte) || pte_present(pte))
2329 swp = pte_to_swp_entry(pte);
2330 if (non_swap_entry(swp) && is_migration_entry(swp))
2336 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2340 if (huge_pte_none(pte) || pte_present(pte))
2342 swp = pte_to_swp_entry(pte);
2343 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2349 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2350 unsigned long start, unsigned long end,
2351 struct page *ref_page)
2353 int force_flush = 0;
2354 struct mm_struct *mm = vma->vm_mm;
2355 unsigned long address;
2359 struct hstate *h = hstate_vma(vma);
2360 unsigned long sz = huge_page_size(h);
2361 const unsigned long mmun_start = start; /* For mmu_notifiers */
2362 const unsigned long mmun_end = end; /* For mmu_notifiers */
2364 WARN_ON(!is_vm_hugetlb_page(vma));
2365 BUG_ON(start & ~huge_page_mask(h));
2366 BUG_ON(end & ~huge_page_mask(h));
2368 tlb_start_vma(tlb, vma);
2369 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2371 spin_lock(&mm->page_table_lock);
2372 for (address = start; address < end; address += sz) {
2373 ptep = huge_pte_offset(mm, address);
2377 if (huge_pmd_unshare(mm, &address, ptep))
2380 pte = huge_ptep_get(ptep);
2381 if (huge_pte_none(pte))
2385 * HWPoisoned hugepage is already unmapped and dropped reference
2387 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
2388 huge_pte_clear(mm, address, ptep);
2392 page = pte_page(pte);
2394 * If a reference page is supplied, it is because a specific
2395 * page is being unmapped, not a range. Ensure the page we
2396 * are about to unmap is the actual page of interest.
2399 if (page != ref_page)
2403 * Mark the VMA as having unmapped its page so that
2404 * future faults in this VMA will fail rather than
2405 * looking like data was lost
2407 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2410 pte = huge_ptep_get_and_clear(mm, address, ptep);
2411 tlb_remove_tlb_entry(tlb, ptep, address);
2412 if (huge_pte_dirty(pte))
2413 set_page_dirty(page);
2415 page_remove_rmap(page);
2416 force_flush = !__tlb_remove_page(tlb, page);
2419 /* Bail out after unmapping reference page if supplied */
2423 spin_unlock(&mm->page_table_lock);
2425 * mmu_gather ran out of room to batch pages, we break out of
2426 * the PTE lock to avoid doing the potential expensive TLB invalidate
2427 * and page-free while holding it.
2432 if (address < end && !ref_page)
2435 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2436 tlb_end_vma(tlb, vma);
2439 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2440 struct vm_area_struct *vma, unsigned long start,
2441 unsigned long end, struct page *ref_page)
2443 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2446 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2447 * test will fail on a vma being torn down, and not grab a page table
2448 * on its way out. We're lucky that the flag has such an appropriate
2449 * name, and can in fact be safely cleared here. We could clear it
2450 * before the __unmap_hugepage_range above, but all that's necessary
2451 * is to clear it before releasing the i_mmap_mutex. This works
2452 * because in the context this is called, the VMA is about to be
2453 * destroyed and the i_mmap_mutex is held.
2455 vma->vm_flags &= ~VM_MAYSHARE;
2458 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2459 unsigned long end, struct page *ref_page)
2461 struct mm_struct *mm;
2462 struct mmu_gather tlb;
2466 tlb_gather_mmu(&tlb, mm, start, end);
2467 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2468 tlb_finish_mmu(&tlb, start, end);
2472 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2473 * mappping it owns the reserve page for. The intention is to unmap the page
2474 * from other VMAs and let the children be SIGKILLed if they are faulting the
2477 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2478 struct page *page, unsigned long address)
2480 struct hstate *h = hstate_vma(vma);
2481 struct vm_area_struct *iter_vma;
2482 struct address_space *mapping;
2486 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2487 * from page cache lookup which is in HPAGE_SIZE units.
2489 address = address & huge_page_mask(h);
2490 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2492 mapping = file_inode(vma->vm_file)->i_mapping;
2495 * Take the mapping lock for the duration of the table walk. As
2496 * this mapping should be shared between all the VMAs,
2497 * __unmap_hugepage_range() is called as the lock is already held
2499 mutex_lock(&mapping->i_mmap_mutex);
2500 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2501 /* Do not unmap the current VMA */
2502 if (iter_vma == vma)
2506 * Unmap the page from other VMAs without their own reserves.
2507 * They get marked to be SIGKILLed if they fault in these
2508 * areas. This is because a future no-page fault on this VMA
2509 * could insert a zeroed page instead of the data existing
2510 * from the time of fork. This would look like data corruption
2512 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2513 unmap_hugepage_range(iter_vma, address,
2514 address + huge_page_size(h), page);
2516 mutex_unlock(&mapping->i_mmap_mutex);
2522 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2523 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2524 * cannot race with other handlers or page migration.
2525 * Keep the pte_same checks anyway to make transition from the mutex easier.
2527 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2528 unsigned long address, pte_t *ptep, pte_t pte,
2529 struct page *pagecache_page)
2531 struct hstate *h = hstate_vma(vma);
2532 struct page *old_page, *new_page;
2533 int outside_reserve = 0;
2534 unsigned long mmun_start; /* For mmu_notifiers */
2535 unsigned long mmun_end; /* For mmu_notifiers */
2537 old_page = pte_page(pte);
2540 /* If no-one else is actually using this page, avoid the copy
2541 * and just make the page writable */
2542 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2543 page_move_anon_rmap(old_page, vma, address);
2544 set_huge_ptep_writable(vma, address, ptep);
2549 * If the process that created a MAP_PRIVATE mapping is about to
2550 * perform a COW due to a shared page count, attempt to satisfy
2551 * the allocation without using the existing reserves. The pagecache
2552 * page is used to determine if the reserve at this address was
2553 * consumed or not. If reserves were used, a partial faulted mapping
2554 * at the time of fork() could consume its reserves on COW instead
2555 * of the full address range.
2557 if (!(vma->vm_flags & VM_MAYSHARE) &&
2558 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2559 old_page != pagecache_page)
2560 outside_reserve = 1;
2562 page_cache_get(old_page);
2564 /* Drop page_table_lock as buddy allocator may be called */
2565 spin_unlock(&mm->page_table_lock);
2566 new_page = alloc_huge_page(vma, address, outside_reserve);
2568 if (IS_ERR(new_page)) {
2569 long err = PTR_ERR(new_page);
2570 page_cache_release(old_page);
2573 * If a process owning a MAP_PRIVATE mapping fails to COW,
2574 * it is due to references held by a child and an insufficient
2575 * huge page pool. To guarantee the original mappers
2576 * reliability, unmap the page from child processes. The child
2577 * may get SIGKILLed if it later faults.
2579 if (outside_reserve) {
2580 BUG_ON(huge_pte_none(pte));
2581 if (unmap_ref_private(mm, vma, old_page, address)) {
2582 BUG_ON(huge_pte_none(pte));
2583 spin_lock(&mm->page_table_lock);
2584 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2585 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2586 goto retry_avoidcopy;
2588 * race occurs while re-acquiring page_table_lock, and
2596 /* Caller expects lock to be held */
2597 spin_lock(&mm->page_table_lock);
2599 return VM_FAULT_OOM;
2601 return VM_FAULT_SIGBUS;
2605 * When the original hugepage is shared one, it does not have
2606 * anon_vma prepared.
2608 if (unlikely(anon_vma_prepare(vma))) {
2609 page_cache_release(new_page);
2610 page_cache_release(old_page);
2611 /* Caller expects lock to be held */
2612 spin_lock(&mm->page_table_lock);
2613 return VM_FAULT_OOM;
2616 copy_user_huge_page(new_page, old_page, address, vma,
2617 pages_per_huge_page(h));
2618 __SetPageUptodate(new_page);
2620 mmun_start = address & huge_page_mask(h);
2621 mmun_end = mmun_start + huge_page_size(h);
2622 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2624 * Retake the page_table_lock to check for racing updates
2625 * before the page tables are altered
2627 spin_lock(&mm->page_table_lock);
2628 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2629 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2631 huge_ptep_clear_flush(vma, address, ptep);
2632 set_huge_pte_at(mm, address, ptep,
2633 make_huge_pte(vma, new_page, 1));
2634 page_remove_rmap(old_page);
2635 hugepage_add_new_anon_rmap(new_page, vma, address);
2636 /* Make the old page be freed below */
2637 new_page = old_page;
2639 spin_unlock(&mm->page_table_lock);
2640 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2641 /* Caller expects lock to be held */
2642 spin_lock(&mm->page_table_lock);
2643 page_cache_release(new_page);
2644 page_cache_release(old_page);
2648 /* Return the pagecache page at a given address within a VMA */
2649 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2650 struct vm_area_struct *vma, unsigned long address)
2652 struct address_space *mapping;
2655 mapping = vma->vm_file->f_mapping;
2656 idx = vma_hugecache_offset(h, vma, address);
2658 return find_lock_page(mapping, idx);
2662 * Return whether there is a pagecache page to back given address within VMA.
2663 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2665 static bool hugetlbfs_pagecache_present(struct hstate *h,
2666 struct vm_area_struct *vma, unsigned long address)
2668 struct address_space *mapping;
2672 mapping = vma->vm_file->f_mapping;
2673 idx = vma_hugecache_offset(h, vma, address);
2675 page = find_get_page(mapping, idx);
2678 return page != NULL;
2681 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2682 unsigned long address, pte_t *ptep, unsigned int flags)
2684 struct hstate *h = hstate_vma(vma);
2685 int ret = VM_FAULT_SIGBUS;
2690 struct address_space *mapping;
2694 * Currently, we are forced to kill the process in the event the
2695 * original mapper has unmapped pages from the child due to a failed
2696 * COW. Warn that such a situation has occurred as it may not be obvious
2698 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2699 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2704 mapping = vma->vm_file->f_mapping;
2705 idx = vma_hugecache_offset(h, vma, address);
2708 * Use page lock to guard against racing truncation
2709 * before we get page_table_lock.
2712 page = find_lock_page(mapping, idx);
2714 size = i_size_read(mapping->host) >> huge_page_shift(h);
2717 page = alloc_huge_page(vma, address, 0);
2719 ret = PTR_ERR(page);
2723 ret = VM_FAULT_SIGBUS;
2726 clear_huge_page(page, address, pages_per_huge_page(h));
2727 __SetPageUptodate(page);
2729 if (vma->vm_flags & VM_MAYSHARE) {
2731 struct inode *inode = mapping->host;
2733 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2741 spin_lock(&inode->i_lock);
2742 inode->i_blocks += blocks_per_huge_page(h);
2743 spin_unlock(&inode->i_lock);
2746 if (unlikely(anon_vma_prepare(vma))) {
2748 goto backout_unlocked;
2754 * If memory error occurs between mmap() and fault, some process
2755 * don't have hwpoisoned swap entry for errored virtual address.
2756 * So we need to block hugepage fault by PG_hwpoison bit check.
2758 if (unlikely(PageHWPoison(page))) {
2759 ret = VM_FAULT_HWPOISON |
2760 VM_FAULT_SET_HINDEX(hstate_index(h));
2761 goto backout_unlocked;
2766 * If we are going to COW a private mapping later, we examine the
2767 * pending reservations for this page now. This will ensure that
2768 * any allocations necessary to record that reservation occur outside
2771 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2772 if (vma_needs_reservation(h, vma, address) < 0) {
2774 goto backout_unlocked;
2777 spin_lock(&mm->page_table_lock);
2778 size = i_size_read(mapping->host) >> huge_page_shift(h);
2783 if (!huge_pte_none(huge_ptep_get(ptep)))
2787 hugepage_add_new_anon_rmap(page, vma, address);
2789 page_dup_rmap(page);
2790 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2791 && (vma->vm_flags & VM_SHARED)));
2792 set_huge_pte_at(mm, address, ptep, new_pte);
2794 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2795 /* Optimization, do the COW without a second fault */
2796 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2799 spin_unlock(&mm->page_table_lock);
2805 spin_unlock(&mm->page_table_lock);
2812 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2813 unsigned long address, unsigned int flags)
2818 struct page *page = NULL;
2819 struct page *pagecache_page = NULL;
2820 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2821 struct hstate *h = hstate_vma(vma);
2823 address &= huge_page_mask(h);
2825 ptep = huge_pte_offset(mm, address);
2827 entry = huge_ptep_get(ptep);
2828 if (unlikely(is_hugetlb_entry_migration(entry))) {
2829 migration_entry_wait_huge(mm, ptep);
2831 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2832 return VM_FAULT_HWPOISON_LARGE |
2833 VM_FAULT_SET_HINDEX(hstate_index(h));
2836 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2838 return VM_FAULT_OOM;
2841 * Serialize hugepage allocation and instantiation, so that we don't
2842 * get spurious allocation failures if two CPUs race to instantiate
2843 * the same page in the page cache.
2845 mutex_lock(&hugetlb_instantiation_mutex);
2846 entry = huge_ptep_get(ptep);
2847 if (huge_pte_none(entry)) {
2848 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2855 * If we are going to COW the mapping later, we examine the pending
2856 * reservations for this page now. This will ensure that any
2857 * allocations necessary to record that reservation occur outside the
2858 * spinlock. For private mappings, we also lookup the pagecache
2859 * page now as it is used to determine if a reservation has been
2862 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
2863 if (vma_needs_reservation(h, vma, address) < 0) {
2868 if (!(vma->vm_flags & VM_MAYSHARE))
2869 pagecache_page = hugetlbfs_pagecache_page(h,
2874 * hugetlb_cow() requires page locks of pte_page(entry) and
2875 * pagecache_page, so here we need take the former one
2876 * when page != pagecache_page or !pagecache_page.
2877 * Note that locking order is always pagecache_page -> page,
2878 * so no worry about deadlock.
2880 page = pte_page(entry);
2882 if (page != pagecache_page)
2885 spin_lock(&mm->page_table_lock);
2886 /* Check for a racing update before calling hugetlb_cow */
2887 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2888 goto out_page_table_lock;
2891 if (flags & FAULT_FLAG_WRITE) {
2892 if (!huge_pte_write(entry)) {
2893 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2895 goto out_page_table_lock;
2897 entry = huge_pte_mkdirty(entry);
2899 entry = pte_mkyoung(entry);
2900 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2901 flags & FAULT_FLAG_WRITE))
2902 update_mmu_cache(vma, address, ptep);
2904 out_page_table_lock:
2905 spin_unlock(&mm->page_table_lock);
2907 if (pagecache_page) {
2908 unlock_page(pagecache_page);
2909 put_page(pagecache_page);
2911 if (page != pagecache_page)
2916 mutex_unlock(&hugetlb_instantiation_mutex);
2921 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2922 struct page **pages, struct vm_area_struct **vmas,
2923 unsigned long *position, unsigned long *nr_pages,
2924 long i, unsigned int flags)
2926 unsigned long pfn_offset;
2927 unsigned long vaddr = *position;
2928 unsigned long remainder = *nr_pages;
2929 struct hstate *h = hstate_vma(vma);
2931 spin_lock(&mm->page_table_lock);
2932 while (vaddr < vma->vm_end && remainder) {
2938 * Some archs (sparc64, sh*) have multiple pte_ts to
2939 * each hugepage. We have to make sure we get the
2940 * first, for the page indexing below to work.
2942 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2943 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2946 * When coredumping, it suits get_dump_page if we just return
2947 * an error where there's an empty slot with no huge pagecache
2948 * to back it. This way, we avoid allocating a hugepage, and
2949 * the sparse dumpfile avoids allocating disk blocks, but its
2950 * huge holes still show up with zeroes where they need to be.
2952 if (absent && (flags & FOLL_DUMP) &&
2953 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2959 * We need call hugetlb_fault for both hugepages under migration
2960 * (in which case hugetlb_fault waits for the migration,) and
2961 * hwpoisoned hugepages (in which case we need to prevent the
2962 * caller from accessing to them.) In order to do this, we use
2963 * here is_swap_pte instead of is_hugetlb_entry_migration and
2964 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
2965 * both cases, and because we can't follow correct pages
2966 * directly from any kind of swap entries.
2968 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
2969 ((flags & FOLL_WRITE) &&
2970 !huge_pte_write(huge_ptep_get(pte)))) {
2973 spin_unlock(&mm->page_table_lock);
2974 ret = hugetlb_fault(mm, vma, vaddr,
2975 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2976 spin_lock(&mm->page_table_lock);
2977 if (!(ret & VM_FAULT_ERROR))
2984 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2985 page = pte_page(huge_ptep_get(pte));
2988 pages[i] = mem_map_offset(page, pfn_offset);
2999 if (vaddr < vma->vm_end && remainder &&
3000 pfn_offset < pages_per_huge_page(h)) {
3002 * We use pfn_offset to avoid touching the pageframes
3003 * of this compound page.
3008 spin_unlock(&mm->page_table_lock);
3009 *nr_pages = remainder;
3012 return i ? i : -EFAULT;
3015 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3016 unsigned long address, unsigned long end, pgprot_t newprot)
3018 struct mm_struct *mm = vma->vm_mm;
3019 unsigned long start = address;
3022 struct hstate *h = hstate_vma(vma);
3023 unsigned long pages = 0;
3025 BUG_ON(address >= end);
3026 flush_cache_range(vma, address, end);
3028 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3029 spin_lock(&mm->page_table_lock);
3030 for (; address < end; address += huge_page_size(h)) {
3031 ptep = huge_pte_offset(mm, address);
3034 if (huge_pmd_unshare(mm, &address, ptep)) {
3038 if (!huge_pte_none(huge_ptep_get(ptep))) {
3039 pte = huge_ptep_get_and_clear(mm, address, ptep);
3040 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3041 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3042 set_huge_pte_at(mm, address, ptep, pte);
3046 spin_unlock(&mm->page_table_lock);
3048 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3049 * may have cleared our pud entry and done put_page on the page table:
3050 * once we release i_mmap_mutex, another task can do the final put_page
3051 * and that page table be reused and filled with junk.
3053 flush_tlb_range(vma, start, end);
3054 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3056 return pages << h->order;
3059 int hugetlb_reserve_pages(struct inode *inode,
3061 struct vm_area_struct *vma,
3062 vm_flags_t vm_flags)
3065 struct hstate *h = hstate_inode(inode);
3066 struct hugepage_subpool *spool = subpool_inode(inode);
3069 * Only apply hugepage reservation if asked. At fault time, an
3070 * attempt will be made for VM_NORESERVE to allocate a page
3071 * without using reserves
3073 if (vm_flags & VM_NORESERVE)
3077 * Shared mappings base their reservation on the number of pages that
3078 * are already allocated on behalf of the file. Private mappings need
3079 * to reserve the full area even if read-only as mprotect() may be
3080 * called to make the mapping read-write. Assume !vma is a shm mapping
3082 if (!vma || vma->vm_flags & VM_MAYSHARE)
3083 chg = region_chg(&inode->i_mapping->private_list, from, to);
3085 struct resv_map *resv_map = resv_map_alloc();
3091 set_vma_resv_map(vma, resv_map);
3092 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3100 /* There must be enough pages in the subpool for the mapping */
3101 if (hugepage_subpool_get_pages(spool, chg)) {
3107 * Check enough hugepages are available for the reservation.
3108 * Hand the pages back to the subpool if there are not
3110 ret = hugetlb_acct_memory(h, chg);
3112 hugepage_subpool_put_pages(spool, chg);
3117 * Account for the reservations made. Shared mappings record regions
3118 * that have reservations as they are shared by multiple VMAs.
3119 * When the last VMA disappears, the region map says how much
3120 * the reservation was and the page cache tells how much of
3121 * the reservation was consumed. Private mappings are per-VMA and
3122 * only the consumed reservations are tracked. When the VMA
3123 * disappears, the original reservation is the VMA size and the
3124 * consumed reservations are stored in the map. Hence, nothing
3125 * else has to be done for private mappings here
3127 if (!vma || vma->vm_flags & VM_MAYSHARE)
3128 region_add(&inode->i_mapping->private_list, from, to);
3136 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3138 struct hstate *h = hstate_inode(inode);
3139 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3140 struct hugepage_subpool *spool = subpool_inode(inode);
3142 spin_lock(&inode->i_lock);
3143 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3144 spin_unlock(&inode->i_lock);
3146 hugepage_subpool_put_pages(spool, (chg - freed));
3147 hugetlb_acct_memory(h, -(chg - freed));
3150 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3151 static unsigned long page_table_shareable(struct vm_area_struct *svma,
3152 struct vm_area_struct *vma,
3153 unsigned long addr, pgoff_t idx)
3155 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3157 unsigned long sbase = saddr & PUD_MASK;
3158 unsigned long s_end = sbase + PUD_SIZE;
3160 /* Allow segments to share if only one is marked locked */
3161 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3162 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3165 * match the virtual addresses, permission and the alignment of the
3168 if (pmd_index(addr) != pmd_index(saddr) ||
3169 vm_flags != svm_flags ||
3170 sbase < svma->vm_start || svma->vm_end < s_end)
3176 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3178 unsigned long base = addr & PUD_MASK;
3179 unsigned long end = base + PUD_SIZE;
3182 * check on proper vm_flags and page table alignment
3184 if (vma->vm_flags & VM_MAYSHARE &&
3185 vma->vm_start <= base && end <= vma->vm_end)
3191 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3192 * and returns the corresponding pte. While this is not necessary for the
3193 * !shared pmd case because we can allocate the pmd later as well, it makes the
3194 * code much cleaner. pmd allocation is essential for the shared case because
3195 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3196 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3197 * bad pmd for sharing.
3199 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3201 struct vm_area_struct *vma = find_vma(mm, addr);
3202 struct address_space *mapping = vma->vm_file->f_mapping;
3203 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3205 struct vm_area_struct *svma;
3206 unsigned long saddr;
3210 if (!vma_shareable(vma, addr))
3211 return (pte_t *)pmd_alloc(mm, pud, addr);
3213 mutex_lock(&mapping->i_mmap_mutex);
3214 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3218 saddr = page_table_shareable(svma, vma, addr, idx);
3220 spte = huge_pte_offset(svma->vm_mm, saddr);
3222 get_page(virt_to_page(spte));
3231 spin_lock(&mm->page_table_lock);
3233 pud_populate(mm, pud,
3234 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3236 put_page(virt_to_page(spte));
3237 spin_unlock(&mm->page_table_lock);
3239 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3240 mutex_unlock(&mapping->i_mmap_mutex);
3245 * unmap huge page backed by shared pte.
3247 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3248 * indicated by page_count > 1, unmap is achieved by clearing pud and
3249 * decrementing the ref count. If count == 1, the pte page is not shared.
3251 * called with vma->vm_mm->page_table_lock held.
3253 * returns: 1 successfully unmapped a shared pte page
3254 * 0 the underlying pte page is not shared, or it is the last user
3256 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3258 pgd_t *pgd = pgd_offset(mm, *addr);
3259 pud_t *pud = pud_offset(pgd, *addr);
3261 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3262 if (page_count(virt_to_page(ptep)) == 1)
3266 put_page(virt_to_page(ptep));
3267 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3270 #define want_pmd_share() (1)
3271 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3272 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3276 #define want_pmd_share() (0)
3277 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3279 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3280 pte_t *huge_pte_alloc(struct mm_struct *mm,
3281 unsigned long addr, unsigned long sz)
3287 pgd = pgd_offset(mm, addr);
3288 pud = pud_alloc(mm, pgd, addr);
3290 if (sz == PUD_SIZE) {
3293 BUG_ON(sz != PMD_SIZE);
3294 if (want_pmd_share() && pud_none(*pud))
3295 pte = huge_pmd_share(mm, addr, pud);
3297 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3300 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3305 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3311 pgd = pgd_offset(mm, addr);
3312 if (pgd_present(*pgd)) {
3313 pud = pud_offset(pgd, addr);
3314 if (pud_present(*pud)) {
3316 return (pte_t *)pud;
3317 pmd = pmd_offset(pud, addr);
3320 return (pte_t *) pmd;
3324 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3325 pmd_t *pmd, int write)
3329 page = pte_page(*(pte_t *)pmd);
3331 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
3336 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3337 pud_t *pud, int write)
3341 page = pte_page(*(pte_t *)pud);
3343 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
3347 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3349 /* Can be overriden by architectures */
3350 __attribute__((weak)) struct page *
3351 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3352 pud_t *pud, int write)
3358 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3360 #ifdef CONFIG_MEMORY_FAILURE
3362 /* Should be called in hugetlb_lock */
3363 static int is_hugepage_on_freelist(struct page *hpage)
3367 struct hstate *h = page_hstate(hpage);
3368 int nid = page_to_nid(hpage);
3370 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3377 * This function is called from memory failure code.
3378 * Assume the caller holds page lock of the head page.
3380 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3382 struct hstate *h = page_hstate(hpage);
3383 int nid = page_to_nid(hpage);
3386 spin_lock(&hugetlb_lock);
3387 if (is_hugepage_on_freelist(hpage)) {
3389 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3390 * but dangling hpage->lru can trigger list-debug warnings
3391 * (this happens when we call unpoison_memory() on it),
3392 * so let it point to itself with list_del_init().
3394 list_del_init(&hpage->lru);
3395 set_page_refcounted(hpage);
3396 h->free_huge_pages--;
3397 h->free_huge_pages_node[nid]--;
3400 spin_unlock(&hugetlb_lock);