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 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
438 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
440 VM_BUG_ON(!is_vm_hugetlb_page(vma));
441 if (!(vma->vm_flags & VM_MAYSHARE))
442 vma->vm_private_data = (void *)0;
445 /* Returns true if the VMA has associated reserve pages */
446 static int vma_has_reserves(struct vm_area_struct *vma)
448 if (vma->vm_flags & VM_NORESERVE)
451 /* Shared mappings always use reserves */
452 if (vma->vm_flags & VM_MAYSHARE)
456 * Only the process that called mmap() has reserves for
459 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
465 static void copy_gigantic_page(struct page *dst, struct page *src)
468 struct hstate *h = page_hstate(src);
469 struct page *dst_base = dst;
470 struct page *src_base = src;
472 for (i = 0; i < pages_per_huge_page(h); ) {
474 copy_highpage(dst, src);
477 dst = mem_map_next(dst, dst_base, i);
478 src = mem_map_next(src, src_base, i);
482 void copy_huge_page(struct page *dst, struct page *src)
485 struct hstate *h = page_hstate(src);
487 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
488 copy_gigantic_page(dst, src);
493 for (i = 0; i < pages_per_huge_page(h); i++) {
495 copy_highpage(dst + i, src + i);
499 static void enqueue_huge_page(struct hstate *h, struct page *page)
501 int nid = page_to_nid(page);
502 list_move(&page->lru, &h->hugepage_freelists[nid]);
503 h->free_huge_pages++;
504 h->free_huge_pages_node[nid]++;
507 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
511 if (list_empty(&h->hugepage_freelists[nid]))
513 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
514 list_move(&page->lru, &h->hugepage_activelist);
515 set_page_refcounted(page);
516 h->free_huge_pages--;
517 h->free_huge_pages_node[nid]--;
521 static struct page *dequeue_huge_page_vma(struct hstate *h,
522 struct vm_area_struct *vma,
523 unsigned long address, int avoid_reserve)
525 struct page *page = NULL;
526 struct mempolicy *mpol;
527 nodemask_t *nodemask;
528 struct zonelist *zonelist;
531 unsigned int cpuset_mems_cookie;
534 * A child process with MAP_PRIVATE mappings created by their parent
535 * have no page reserves. This check ensures that reservations are
536 * not "stolen". The child may still get SIGKILLed
538 if (!vma_has_reserves(vma) &&
539 h->free_huge_pages - h->resv_huge_pages == 0)
542 /* If reserves cannot be used, ensure enough pages are in the pool */
543 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
547 cpuset_mems_cookie = get_mems_allowed();
548 zonelist = huge_zonelist(vma, address,
549 htlb_alloc_mask, &mpol, &nodemask);
551 for_each_zone_zonelist_nodemask(zone, z, zonelist,
552 MAX_NR_ZONES - 1, nodemask) {
553 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
554 page = dequeue_huge_page_node(h, zone_to_nid(zone));
556 if (!avoid_reserve && vma_has_reserves(vma))
557 h->resv_huge_pages--;
564 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
572 static void update_and_free_page(struct hstate *h, struct page *page)
576 VM_BUG_ON(h->order >= MAX_ORDER);
579 h->nr_huge_pages_node[page_to_nid(page)]--;
580 for (i = 0; i < pages_per_huge_page(h); i++) {
581 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
582 1 << PG_referenced | 1 << PG_dirty |
583 1 << PG_active | 1 << PG_reserved |
584 1 << PG_private | 1 << PG_writeback);
586 VM_BUG_ON(hugetlb_cgroup_from_page(page));
587 set_compound_page_dtor(page, NULL);
588 set_page_refcounted(page);
589 arch_release_hugepage(page);
590 __free_pages(page, huge_page_order(h));
593 struct hstate *size_to_hstate(unsigned long size)
598 if (huge_page_size(h) == size)
604 static void free_huge_page(struct page *page)
607 * Can't pass hstate in here because it is called from the
608 * compound page destructor.
610 struct hstate *h = page_hstate(page);
611 int nid = page_to_nid(page);
612 struct hugepage_subpool *spool =
613 (struct hugepage_subpool *)page_private(page);
615 set_page_private(page, 0);
616 page->mapping = NULL;
617 BUG_ON(page_count(page));
618 BUG_ON(page_mapcount(page));
620 spin_lock(&hugetlb_lock);
621 hugetlb_cgroup_uncharge_page(hstate_index(h),
622 pages_per_huge_page(h), page);
623 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
624 /* remove the page from active list */
625 list_del(&page->lru);
626 update_and_free_page(h, page);
627 h->surplus_huge_pages--;
628 h->surplus_huge_pages_node[nid]--;
630 arch_clear_hugepage_flags(page);
631 enqueue_huge_page(h, page);
633 spin_unlock(&hugetlb_lock);
634 hugepage_subpool_put_pages(spool, 1);
637 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
639 INIT_LIST_HEAD(&page->lru);
640 set_compound_page_dtor(page, free_huge_page);
641 spin_lock(&hugetlb_lock);
642 set_hugetlb_cgroup(page, NULL);
644 h->nr_huge_pages_node[nid]++;
645 spin_unlock(&hugetlb_lock);
646 put_page(page); /* free it into the hugepage allocator */
649 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
652 int nr_pages = 1 << order;
653 struct page *p = page + 1;
655 /* we rely on prep_new_huge_page to set the destructor */
656 set_compound_order(page, order);
658 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
660 set_page_count(p, 0);
661 p->first_page = page;
666 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
667 * transparent huge pages. See the PageTransHuge() documentation for more
670 int PageHuge(struct page *page)
672 compound_page_dtor *dtor;
674 if (!PageCompound(page))
677 page = compound_head(page);
678 dtor = get_compound_page_dtor(page);
680 return dtor == free_huge_page;
682 EXPORT_SYMBOL_GPL(PageHuge);
684 pgoff_t __basepage_index(struct page *page)
686 struct page *page_head = compound_head(page);
687 pgoff_t index = page_index(page_head);
688 unsigned long compound_idx;
690 if (!PageHuge(page_head))
691 return page_index(page);
693 if (compound_order(page_head) >= MAX_ORDER)
694 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
696 compound_idx = page - page_head;
698 return (index << compound_order(page_head)) + compound_idx;
701 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
705 if (h->order >= MAX_ORDER)
708 page = alloc_pages_exact_node(nid,
709 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
710 __GFP_REPEAT|__GFP_NOWARN,
713 if (arch_prepare_hugepage(page)) {
714 __free_pages(page, huge_page_order(h));
717 prep_new_huge_page(h, page, nid);
724 * common helper functions for hstate_next_node_to_{alloc|free}.
725 * We may have allocated or freed a huge page based on a different
726 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
727 * be outside of *nodes_allowed. Ensure that we use an allowed
728 * node for alloc or free.
730 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
732 nid = next_node(nid, *nodes_allowed);
733 if (nid == MAX_NUMNODES)
734 nid = first_node(*nodes_allowed);
735 VM_BUG_ON(nid >= MAX_NUMNODES);
740 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
742 if (!node_isset(nid, *nodes_allowed))
743 nid = next_node_allowed(nid, nodes_allowed);
748 * returns the previously saved node ["this node"] from which to
749 * allocate a persistent huge page for the pool and advance the
750 * next node from which to allocate, handling wrap at end of node
753 static int hstate_next_node_to_alloc(struct hstate *h,
754 nodemask_t *nodes_allowed)
758 VM_BUG_ON(!nodes_allowed);
760 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
761 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
767 * helper for free_pool_huge_page() - return the previously saved
768 * node ["this node"] from which to free a huge page. Advance the
769 * next node id whether or not we find a free huge page to free so
770 * that the next attempt to free addresses the next node.
772 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
776 VM_BUG_ON(!nodes_allowed);
778 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
779 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
784 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
785 for (nr_nodes = nodes_weight(*mask); \
787 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
790 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
791 for (nr_nodes = nodes_weight(*mask); \
793 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
796 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
802 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
803 page = alloc_fresh_huge_page_node(h, node);
811 count_vm_event(HTLB_BUDDY_PGALLOC);
813 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
819 * Free huge page from pool from next node to free.
820 * Attempt to keep persistent huge pages more or less
821 * balanced over allowed nodes.
822 * Called with hugetlb_lock locked.
824 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
830 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
832 * If we're returning unused surplus pages, only examine
833 * nodes with surplus pages.
835 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
836 !list_empty(&h->hugepage_freelists[node])) {
838 list_entry(h->hugepage_freelists[node].next,
840 list_del(&page->lru);
841 h->free_huge_pages--;
842 h->free_huge_pages_node[node]--;
844 h->surplus_huge_pages--;
845 h->surplus_huge_pages_node[node]--;
847 update_and_free_page(h, page);
856 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
861 if (h->order >= MAX_ORDER)
865 * Assume we will successfully allocate the surplus page to
866 * prevent racing processes from causing the surplus to exceed
869 * This however introduces a different race, where a process B
870 * tries to grow the static hugepage pool while alloc_pages() is
871 * called by process A. B will only examine the per-node
872 * counters in determining if surplus huge pages can be
873 * converted to normal huge pages in adjust_pool_surplus(). A
874 * won't be able to increment the per-node counter, until the
875 * lock is dropped by B, but B doesn't drop hugetlb_lock until
876 * no more huge pages can be converted from surplus to normal
877 * state (and doesn't try to convert again). Thus, we have a
878 * case where a surplus huge page exists, the pool is grown, and
879 * the surplus huge page still exists after, even though it
880 * should just have been converted to a normal huge page. This
881 * does not leak memory, though, as the hugepage will be freed
882 * once it is out of use. It also does not allow the counters to
883 * go out of whack in adjust_pool_surplus() as we don't modify
884 * the node values until we've gotten the hugepage and only the
885 * per-node value is checked there.
887 spin_lock(&hugetlb_lock);
888 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
889 spin_unlock(&hugetlb_lock);
893 h->surplus_huge_pages++;
895 spin_unlock(&hugetlb_lock);
897 if (nid == NUMA_NO_NODE)
898 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
899 __GFP_REPEAT|__GFP_NOWARN,
902 page = alloc_pages_exact_node(nid,
903 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
904 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
906 if (page && arch_prepare_hugepage(page)) {
907 __free_pages(page, huge_page_order(h));
911 spin_lock(&hugetlb_lock);
913 INIT_LIST_HEAD(&page->lru);
914 r_nid = page_to_nid(page);
915 set_compound_page_dtor(page, free_huge_page);
916 set_hugetlb_cgroup(page, NULL);
918 * We incremented the global counters already
920 h->nr_huge_pages_node[r_nid]++;
921 h->surplus_huge_pages_node[r_nid]++;
922 __count_vm_event(HTLB_BUDDY_PGALLOC);
925 h->surplus_huge_pages--;
926 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
928 spin_unlock(&hugetlb_lock);
934 * This allocation function is useful in the context where vma is irrelevant.
935 * E.g. soft-offlining uses this function because it only cares physical
936 * address of error page.
938 struct page *alloc_huge_page_node(struct hstate *h, int nid)
942 spin_lock(&hugetlb_lock);
943 page = dequeue_huge_page_node(h, nid);
944 spin_unlock(&hugetlb_lock);
947 page = alloc_buddy_huge_page(h, nid);
953 * Increase the hugetlb pool such that it can accommodate a reservation
956 static int gather_surplus_pages(struct hstate *h, int delta)
958 struct list_head surplus_list;
959 struct page *page, *tmp;
961 int needed, allocated;
962 bool alloc_ok = true;
964 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
966 h->resv_huge_pages += delta;
971 INIT_LIST_HEAD(&surplus_list);
975 spin_unlock(&hugetlb_lock);
976 for (i = 0; i < needed; i++) {
977 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
982 list_add(&page->lru, &surplus_list);
987 * After retaking hugetlb_lock, we need to recalculate 'needed'
988 * because either resv_huge_pages or free_huge_pages may have changed.
990 spin_lock(&hugetlb_lock);
991 needed = (h->resv_huge_pages + delta) -
992 (h->free_huge_pages + allocated);
997 * We were not able to allocate enough pages to
998 * satisfy the entire reservation so we free what
999 * we've allocated so far.
1004 * The surplus_list now contains _at_least_ the number of extra pages
1005 * needed to accommodate the reservation. Add the appropriate number
1006 * of pages to the hugetlb pool and free the extras back to the buddy
1007 * allocator. Commit the entire reservation here to prevent another
1008 * process from stealing the pages as they are added to the pool but
1009 * before they are reserved.
1011 needed += allocated;
1012 h->resv_huge_pages += delta;
1015 /* Free the needed pages to the hugetlb pool */
1016 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1020 * This page is now managed by the hugetlb allocator and has
1021 * no users -- drop the buddy allocator's reference.
1023 put_page_testzero(page);
1024 VM_BUG_ON(page_count(page));
1025 enqueue_huge_page(h, page);
1028 spin_unlock(&hugetlb_lock);
1030 /* Free unnecessary surplus pages to the buddy allocator */
1031 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1033 spin_lock(&hugetlb_lock);
1039 * When releasing a hugetlb pool reservation, any surplus pages that were
1040 * allocated to satisfy the reservation must be explicitly freed if they were
1042 * Called with hugetlb_lock held.
1044 static void return_unused_surplus_pages(struct hstate *h,
1045 unsigned long unused_resv_pages)
1047 unsigned long nr_pages;
1049 /* Uncommit the reservation */
1050 h->resv_huge_pages -= unused_resv_pages;
1052 /* Cannot return gigantic pages currently */
1053 if (h->order >= MAX_ORDER)
1056 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1059 * We want to release as many surplus pages as possible, spread
1060 * evenly across all nodes with memory. Iterate across these nodes
1061 * until we can no longer free unreserved surplus pages. This occurs
1062 * when the nodes with surplus pages have no free pages.
1063 * free_pool_huge_page() will balance the the freed pages across the
1064 * on-line nodes with memory and will handle the hstate accounting.
1066 while (nr_pages--) {
1067 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1073 * Determine if the huge page at addr within the vma has an associated
1074 * reservation. Where it does not we will need to logically increase
1075 * reservation and actually increase subpool usage before an allocation
1076 * can occur. Where any new reservation would be required the
1077 * reservation change is prepared, but not committed. Once the page
1078 * has been allocated from the subpool and instantiated the change should
1079 * be committed via vma_commit_reservation. No action is required on
1082 static long vma_needs_reservation(struct hstate *h,
1083 struct vm_area_struct *vma, unsigned long addr)
1085 struct address_space *mapping = vma->vm_file->f_mapping;
1086 struct inode *inode = mapping->host;
1088 if (vma->vm_flags & VM_MAYSHARE) {
1089 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1090 return region_chg(&inode->i_mapping->private_list,
1093 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1098 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1099 struct resv_map *reservations = vma_resv_map(vma);
1101 err = region_chg(&reservations->regions, idx, idx + 1);
1107 static void vma_commit_reservation(struct hstate *h,
1108 struct vm_area_struct *vma, unsigned long addr)
1110 struct address_space *mapping = vma->vm_file->f_mapping;
1111 struct inode *inode = mapping->host;
1113 if (vma->vm_flags & VM_MAYSHARE) {
1114 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1115 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1117 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1118 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1119 struct resv_map *reservations = vma_resv_map(vma);
1121 /* Mark this page used in the map. */
1122 region_add(&reservations->regions, idx, idx + 1);
1126 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1127 unsigned long addr, int avoid_reserve)
1129 struct hugepage_subpool *spool = subpool_vma(vma);
1130 struct hstate *h = hstate_vma(vma);
1134 struct hugetlb_cgroup *h_cg;
1136 idx = hstate_index(h);
1138 * Processes that did not create the mapping will have no
1139 * reserves and will not have accounted against subpool
1140 * limit. Check that the subpool limit can be made before
1141 * satisfying the allocation MAP_NORESERVE mappings may also
1142 * need pages and subpool limit allocated allocated if no reserve
1145 chg = vma_needs_reservation(h, vma, addr);
1147 return ERR_PTR(-ENOMEM);
1149 if (hugepage_subpool_get_pages(spool, chg))
1150 return ERR_PTR(-ENOSPC);
1152 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1154 hugepage_subpool_put_pages(spool, chg);
1155 return ERR_PTR(-ENOSPC);
1157 spin_lock(&hugetlb_lock);
1158 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1160 spin_unlock(&hugetlb_lock);
1161 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1163 hugetlb_cgroup_uncharge_cgroup(idx,
1164 pages_per_huge_page(h),
1166 hugepage_subpool_put_pages(spool, chg);
1167 return ERR_PTR(-ENOSPC);
1169 spin_lock(&hugetlb_lock);
1170 list_move(&page->lru, &h->hugepage_activelist);
1173 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1174 spin_unlock(&hugetlb_lock);
1176 set_page_private(page, (unsigned long)spool);
1178 vma_commit_reservation(h, vma, addr);
1182 int __weak alloc_bootmem_huge_page(struct hstate *h)
1184 struct huge_bootmem_page *m;
1187 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1190 addr = __alloc_bootmem_node_nopanic(NODE_DATA(node),
1191 huge_page_size(h), huge_page_size(h), 0);
1195 * Use the beginning of the huge page to store the
1196 * huge_bootmem_page struct (until gather_bootmem
1197 * puts them into the mem_map).
1206 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1207 /* Put them into a private list first because mem_map is not up yet */
1208 list_add(&m->list, &huge_boot_pages);
1213 static void prep_compound_huge_page(struct page *page, int order)
1215 if (unlikely(order > (MAX_ORDER - 1)))
1216 prep_compound_gigantic_page(page, order);
1218 prep_compound_page(page, order);
1221 /* Put bootmem huge pages into the standard lists after mem_map is up */
1222 static void __init gather_bootmem_prealloc(void)
1224 struct huge_bootmem_page *m;
1226 list_for_each_entry(m, &huge_boot_pages, list) {
1227 struct hstate *h = m->hstate;
1230 #ifdef CONFIG_HIGHMEM
1231 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1232 free_bootmem_late((unsigned long)m,
1233 sizeof(struct huge_bootmem_page));
1235 page = virt_to_page(m);
1237 __ClearPageReserved(page);
1238 WARN_ON(page_count(page) != 1);
1239 prep_compound_huge_page(page, h->order);
1240 prep_new_huge_page(h, page, page_to_nid(page));
1242 * If we had gigantic hugepages allocated at boot time, we need
1243 * to restore the 'stolen' pages to totalram_pages in order to
1244 * fix confusing memory reports from free(1) and another
1245 * side-effects, like CommitLimit going negative.
1247 if (h->order > (MAX_ORDER - 1))
1248 adjust_managed_page_count(page, 1 << h->order);
1252 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1256 for (i = 0; i < h->max_huge_pages; ++i) {
1257 if (h->order >= MAX_ORDER) {
1258 if (!alloc_bootmem_huge_page(h))
1260 } else if (!alloc_fresh_huge_page(h,
1261 &node_states[N_MEMORY]))
1264 h->max_huge_pages = i;
1267 static void __init hugetlb_init_hstates(void)
1271 for_each_hstate(h) {
1272 /* oversize hugepages were init'ed in early boot */
1273 if (h->order < MAX_ORDER)
1274 hugetlb_hstate_alloc_pages(h);
1278 static char * __init memfmt(char *buf, unsigned long n)
1280 if (n >= (1UL << 30))
1281 sprintf(buf, "%lu GB", n >> 30);
1282 else if (n >= (1UL << 20))
1283 sprintf(buf, "%lu MB", n >> 20);
1285 sprintf(buf, "%lu KB", n >> 10);
1289 static void __init report_hugepages(void)
1293 for_each_hstate(h) {
1295 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1296 memfmt(buf, huge_page_size(h)),
1297 h->free_huge_pages);
1301 #ifdef CONFIG_HIGHMEM
1302 static void try_to_free_low(struct hstate *h, unsigned long count,
1303 nodemask_t *nodes_allowed)
1307 if (h->order >= MAX_ORDER)
1310 for_each_node_mask(i, *nodes_allowed) {
1311 struct page *page, *next;
1312 struct list_head *freel = &h->hugepage_freelists[i];
1313 list_for_each_entry_safe(page, next, freel, lru) {
1314 if (count >= h->nr_huge_pages)
1316 if (PageHighMem(page))
1318 list_del(&page->lru);
1319 update_and_free_page(h, page);
1320 h->free_huge_pages--;
1321 h->free_huge_pages_node[page_to_nid(page)]--;
1326 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1327 nodemask_t *nodes_allowed)
1333 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1334 * balanced by operating on them in a round-robin fashion.
1335 * Returns 1 if an adjustment was made.
1337 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1342 VM_BUG_ON(delta != -1 && delta != 1);
1345 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1346 if (h->surplus_huge_pages_node[node])
1350 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1351 if (h->surplus_huge_pages_node[node] <
1352 h->nr_huge_pages_node[node])
1359 h->surplus_huge_pages += delta;
1360 h->surplus_huge_pages_node[node] += delta;
1364 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1365 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1366 nodemask_t *nodes_allowed)
1368 unsigned long min_count, ret;
1370 if (h->order >= MAX_ORDER)
1371 return h->max_huge_pages;
1374 * Increase the pool size
1375 * First take pages out of surplus state. Then make up the
1376 * remaining difference by allocating fresh huge pages.
1378 * We might race with alloc_buddy_huge_page() here and be unable
1379 * to convert a surplus huge page to a normal huge page. That is
1380 * not critical, though, it just means the overall size of the
1381 * pool might be one hugepage larger than it needs to be, but
1382 * within all the constraints specified by the sysctls.
1384 spin_lock(&hugetlb_lock);
1385 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1386 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1390 while (count > persistent_huge_pages(h)) {
1392 * If this allocation races such that we no longer need the
1393 * page, free_huge_page will handle it by freeing the page
1394 * and reducing the surplus.
1396 spin_unlock(&hugetlb_lock);
1397 ret = alloc_fresh_huge_page(h, nodes_allowed);
1398 spin_lock(&hugetlb_lock);
1402 /* Bail for signals. Probably ctrl-c from user */
1403 if (signal_pending(current))
1408 * Decrease the pool size
1409 * First return free pages to the buddy allocator (being careful
1410 * to keep enough around to satisfy reservations). Then place
1411 * pages into surplus state as needed so the pool will shrink
1412 * to the desired size as pages become free.
1414 * By placing pages into the surplus state independent of the
1415 * overcommit value, we are allowing the surplus pool size to
1416 * exceed overcommit. There are few sane options here. Since
1417 * alloc_buddy_huge_page() is checking the global counter,
1418 * though, we'll note that we're not allowed to exceed surplus
1419 * and won't grow the pool anywhere else. Not until one of the
1420 * sysctls are changed, or the surplus pages go out of use.
1422 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1423 min_count = max(count, min_count);
1424 try_to_free_low(h, min_count, nodes_allowed);
1425 while (min_count < persistent_huge_pages(h)) {
1426 if (!free_pool_huge_page(h, nodes_allowed, 0))
1429 while (count < persistent_huge_pages(h)) {
1430 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1434 ret = persistent_huge_pages(h);
1435 spin_unlock(&hugetlb_lock);
1439 #define HSTATE_ATTR_RO(_name) \
1440 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1442 #define HSTATE_ATTR(_name) \
1443 static struct kobj_attribute _name##_attr = \
1444 __ATTR(_name, 0644, _name##_show, _name##_store)
1446 static struct kobject *hugepages_kobj;
1447 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1449 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1451 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1455 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1456 if (hstate_kobjs[i] == kobj) {
1458 *nidp = NUMA_NO_NODE;
1462 return kobj_to_node_hstate(kobj, nidp);
1465 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1466 struct kobj_attribute *attr, char *buf)
1469 unsigned long nr_huge_pages;
1472 h = kobj_to_hstate(kobj, &nid);
1473 if (nid == NUMA_NO_NODE)
1474 nr_huge_pages = h->nr_huge_pages;
1476 nr_huge_pages = h->nr_huge_pages_node[nid];
1478 return sprintf(buf, "%lu\n", nr_huge_pages);
1481 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1482 struct kobject *kobj, struct kobj_attribute *attr,
1483 const char *buf, size_t len)
1487 unsigned long count;
1489 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1491 err = kstrtoul(buf, 10, &count);
1495 h = kobj_to_hstate(kobj, &nid);
1496 if (h->order >= MAX_ORDER) {
1501 if (nid == NUMA_NO_NODE) {
1503 * global hstate attribute
1505 if (!(obey_mempolicy &&
1506 init_nodemask_of_mempolicy(nodes_allowed))) {
1507 NODEMASK_FREE(nodes_allowed);
1508 nodes_allowed = &node_states[N_MEMORY];
1510 } else if (nodes_allowed) {
1512 * per node hstate attribute: adjust count to global,
1513 * but restrict alloc/free to the specified node.
1515 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1516 init_nodemask_of_node(nodes_allowed, nid);
1518 nodes_allowed = &node_states[N_MEMORY];
1520 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1522 if (nodes_allowed != &node_states[N_MEMORY])
1523 NODEMASK_FREE(nodes_allowed);
1527 NODEMASK_FREE(nodes_allowed);
1531 static ssize_t nr_hugepages_show(struct kobject *kobj,
1532 struct kobj_attribute *attr, char *buf)
1534 return nr_hugepages_show_common(kobj, attr, buf);
1537 static ssize_t nr_hugepages_store(struct kobject *kobj,
1538 struct kobj_attribute *attr, const char *buf, size_t len)
1540 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1542 HSTATE_ATTR(nr_hugepages);
1547 * hstate attribute for optionally mempolicy-based constraint on persistent
1548 * huge page alloc/free.
1550 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1551 struct kobj_attribute *attr, char *buf)
1553 return nr_hugepages_show_common(kobj, attr, buf);
1556 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1557 struct kobj_attribute *attr, const char *buf, size_t len)
1559 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1561 HSTATE_ATTR(nr_hugepages_mempolicy);
1565 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1566 struct kobj_attribute *attr, char *buf)
1568 struct hstate *h = kobj_to_hstate(kobj, NULL);
1569 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1572 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1573 struct kobj_attribute *attr, const char *buf, size_t count)
1576 unsigned long input;
1577 struct hstate *h = kobj_to_hstate(kobj, NULL);
1579 if (h->order >= MAX_ORDER)
1582 err = kstrtoul(buf, 10, &input);
1586 spin_lock(&hugetlb_lock);
1587 h->nr_overcommit_huge_pages = input;
1588 spin_unlock(&hugetlb_lock);
1592 HSTATE_ATTR(nr_overcommit_hugepages);
1594 static ssize_t free_hugepages_show(struct kobject *kobj,
1595 struct kobj_attribute *attr, char *buf)
1598 unsigned long free_huge_pages;
1601 h = kobj_to_hstate(kobj, &nid);
1602 if (nid == NUMA_NO_NODE)
1603 free_huge_pages = h->free_huge_pages;
1605 free_huge_pages = h->free_huge_pages_node[nid];
1607 return sprintf(buf, "%lu\n", free_huge_pages);
1609 HSTATE_ATTR_RO(free_hugepages);
1611 static ssize_t resv_hugepages_show(struct kobject *kobj,
1612 struct kobj_attribute *attr, char *buf)
1614 struct hstate *h = kobj_to_hstate(kobj, NULL);
1615 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1617 HSTATE_ATTR_RO(resv_hugepages);
1619 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1620 struct kobj_attribute *attr, char *buf)
1623 unsigned long surplus_huge_pages;
1626 h = kobj_to_hstate(kobj, &nid);
1627 if (nid == NUMA_NO_NODE)
1628 surplus_huge_pages = h->surplus_huge_pages;
1630 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1632 return sprintf(buf, "%lu\n", surplus_huge_pages);
1634 HSTATE_ATTR_RO(surplus_hugepages);
1636 static struct attribute *hstate_attrs[] = {
1637 &nr_hugepages_attr.attr,
1638 &nr_overcommit_hugepages_attr.attr,
1639 &free_hugepages_attr.attr,
1640 &resv_hugepages_attr.attr,
1641 &surplus_hugepages_attr.attr,
1643 &nr_hugepages_mempolicy_attr.attr,
1648 static struct attribute_group hstate_attr_group = {
1649 .attrs = hstate_attrs,
1652 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1653 struct kobject **hstate_kobjs,
1654 struct attribute_group *hstate_attr_group)
1657 int hi = hstate_index(h);
1659 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1660 if (!hstate_kobjs[hi])
1663 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1665 kobject_put(hstate_kobjs[hi]);
1670 static void __init hugetlb_sysfs_init(void)
1675 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1676 if (!hugepages_kobj)
1679 for_each_hstate(h) {
1680 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1681 hstate_kobjs, &hstate_attr_group);
1683 pr_err("Hugetlb: Unable to add hstate %s", h->name);
1690 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1691 * with node devices in node_devices[] using a parallel array. The array
1692 * index of a node device or _hstate == node id.
1693 * This is here to avoid any static dependency of the node device driver, in
1694 * the base kernel, on the hugetlb module.
1696 struct node_hstate {
1697 struct kobject *hugepages_kobj;
1698 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1700 struct node_hstate node_hstates[MAX_NUMNODES];
1703 * A subset of global hstate attributes for node devices
1705 static struct attribute *per_node_hstate_attrs[] = {
1706 &nr_hugepages_attr.attr,
1707 &free_hugepages_attr.attr,
1708 &surplus_hugepages_attr.attr,
1712 static struct attribute_group per_node_hstate_attr_group = {
1713 .attrs = per_node_hstate_attrs,
1717 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1718 * Returns node id via non-NULL nidp.
1720 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1724 for (nid = 0; nid < nr_node_ids; nid++) {
1725 struct node_hstate *nhs = &node_hstates[nid];
1727 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1728 if (nhs->hstate_kobjs[i] == kobj) {
1740 * Unregister hstate attributes from a single node device.
1741 * No-op if no hstate attributes attached.
1743 static void hugetlb_unregister_node(struct node *node)
1746 struct node_hstate *nhs = &node_hstates[node->dev.id];
1748 if (!nhs->hugepages_kobj)
1749 return; /* no hstate attributes */
1751 for_each_hstate(h) {
1752 int idx = hstate_index(h);
1753 if (nhs->hstate_kobjs[idx]) {
1754 kobject_put(nhs->hstate_kobjs[idx]);
1755 nhs->hstate_kobjs[idx] = NULL;
1759 kobject_put(nhs->hugepages_kobj);
1760 nhs->hugepages_kobj = NULL;
1764 * hugetlb module exit: unregister hstate attributes from node devices
1767 static void hugetlb_unregister_all_nodes(void)
1772 * disable node device registrations.
1774 register_hugetlbfs_with_node(NULL, NULL);
1777 * remove hstate attributes from any nodes that have them.
1779 for (nid = 0; nid < nr_node_ids; nid++)
1780 hugetlb_unregister_node(node_devices[nid]);
1784 * Register hstate attributes for a single node device.
1785 * No-op if attributes already registered.
1787 static void hugetlb_register_node(struct node *node)
1790 struct node_hstate *nhs = &node_hstates[node->dev.id];
1793 if (nhs->hugepages_kobj)
1794 return; /* already allocated */
1796 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1798 if (!nhs->hugepages_kobj)
1801 for_each_hstate(h) {
1802 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1804 &per_node_hstate_attr_group);
1806 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1807 h->name, node->dev.id);
1808 hugetlb_unregister_node(node);
1815 * hugetlb init time: register hstate attributes for all registered node
1816 * devices of nodes that have memory. All on-line nodes should have
1817 * registered their associated device by this time.
1819 static void hugetlb_register_all_nodes(void)
1823 for_each_node_state(nid, N_MEMORY) {
1824 struct node *node = node_devices[nid];
1825 if (node->dev.id == nid)
1826 hugetlb_register_node(node);
1830 * Let the node device driver know we're here so it can
1831 * [un]register hstate attributes on node hotplug.
1833 register_hugetlbfs_with_node(hugetlb_register_node,
1834 hugetlb_unregister_node);
1836 #else /* !CONFIG_NUMA */
1838 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1846 static void hugetlb_unregister_all_nodes(void) { }
1848 static void hugetlb_register_all_nodes(void) { }
1852 static void __exit hugetlb_exit(void)
1856 hugetlb_unregister_all_nodes();
1858 for_each_hstate(h) {
1859 kobject_put(hstate_kobjs[hstate_index(h)]);
1862 kobject_put(hugepages_kobj);
1864 module_exit(hugetlb_exit);
1866 static int __init hugetlb_init(void)
1868 /* Some platform decide whether they support huge pages at boot
1869 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1870 * there is no such support
1872 if (HPAGE_SHIFT == 0)
1875 if (!size_to_hstate(default_hstate_size)) {
1876 default_hstate_size = HPAGE_SIZE;
1877 if (!size_to_hstate(default_hstate_size))
1878 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1880 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1881 if (default_hstate_max_huge_pages)
1882 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1884 hugetlb_init_hstates();
1885 gather_bootmem_prealloc();
1888 hugetlb_sysfs_init();
1889 hugetlb_register_all_nodes();
1890 hugetlb_cgroup_file_init();
1894 module_init(hugetlb_init);
1896 /* Should be called on processing a hugepagesz=... option */
1897 void __init hugetlb_add_hstate(unsigned order)
1902 if (size_to_hstate(PAGE_SIZE << order)) {
1903 pr_warning("hugepagesz= specified twice, ignoring\n");
1906 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1908 h = &hstates[hugetlb_max_hstate++];
1910 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1911 h->nr_huge_pages = 0;
1912 h->free_huge_pages = 0;
1913 for (i = 0; i < MAX_NUMNODES; ++i)
1914 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1915 INIT_LIST_HEAD(&h->hugepage_activelist);
1916 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
1917 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
1918 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1919 huge_page_size(h)/1024);
1924 static int __init hugetlb_nrpages_setup(char *s)
1927 static unsigned long *last_mhp;
1930 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1931 * so this hugepages= parameter goes to the "default hstate".
1933 if (!hugetlb_max_hstate)
1934 mhp = &default_hstate_max_huge_pages;
1936 mhp = &parsed_hstate->max_huge_pages;
1938 if (mhp == last_mhp) {
1939 pr_warning("hugepages= specified twice without "
1940 "interleaving hugepagesz=, ignoring\n");
1944 if (sscanf(s, "%lu", mhp) <= 0)
1948 * Global state is always initialized later in hugetlb_init.
1949 * But we need to allocate >= MAX_ORDER hstates here early to still
1950 * use the bootmem allocator.
1952 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
1953 hugetlb_hstate_alloc_pages(parsed_hstate);
1959 __setup("hugepages=", hugetlb_nrpages_setup);
1961 static int __init hugetlb_default_setup(char *s)
1963 default_hstate_size = memparse(s, &s);
1966 __setup("default_hugepagesz=", hugetlb_default_setup);
1968 static unsigned int cpuset_mems_nr(unsigned int *array)
1971 unsigned int nr = 0;
1973 for_each_node_mask(node, cpuset_current_mems_allowed)
1979 #ifdef CONFIG_SYSCTL
1980 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1981 struct ctl_table *table, int write,
1982 void __user *buffer, size_t *length, loff_t *ppos)
1984 struct hstate *h = &default_hstate;
1988 tmp = h->max_huge_pages;
1990 if (write && h->order >= MAX_ORDER)
1994 table->maxlen = sizeof(unsigned long);
1995 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2000 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2001 GFP_KERNEL | __GFP_NORETRY);
2002 if (!(obey_mempolicy &&
2003 init_nodemask_of_mempolicy(nodes_allowed))) {
2004 NODEMASK_FREE(nodes_allowed);
2005 nodes_allowed = &node_states[N_MEMORY];
2007 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2009 if (nodes_allowed != &node_states[N_MEMORY])
2010 NODEMASK_FREE(nodes_allowed);
2016 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2017 void __user *buffer, size_t *length, loff_t *ppos)
2020 return hugetlb_sysctl_handler_common(false, table, write,
2021 buffer, length, ppos);
2025 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2026 void __user *buffer, size_t *length, loff_t *ppos)
2028 return hugetlb_sysctl_handler_common(true, table, write,
2029 buffer, length, ppos);
2031 #endif /* CONFIG_NUMA */
2033 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2034 void __user *buffer,
2035 size_t *length, loff_t *ppos)
2037 proc_dointvec(table, write, buffer, length, ppos);
2038 if (hugepages_treat_as_movable)
2039 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2041 htlb_alloc_mask = GFP_HIGHUSER;
2045 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2046 void __user *buffer,
2047 size_t *length, loff_t *ppos)
2049 struct hstate *h = &default_hstate;
2053 tmp = h->nr_overcommit_huge_pages;
2055 if (write && h->order >= MAX_ORDER)
2059 table->maxlen = sizeof(unsigned long);
2060 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2065 spin_lock(&hugetlb_lock);
2066 h->nr_overcommit_huge_pages = tmp;
2067 spin_unlock(&hugetlb_lock);
2073 #endif /* CONFIG_SYSCTL */
2075 void hugetlb_report_meminfo(struct seq_file *m)
2077 struct hstate *h = &default_hstate;
2079 "HugePages_Total: %5lu\n"
2080 "HugePages_Free: %5lu\n"
2081 "HugePages_Rsvd: %5lu\n"
2082 "HugePages_Surp: %5lu\n"
2083 "Hugepagesize: %8lu kB\n",
2087 h->surplus_huge_pages,
2088 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2091 int hugetlb_report_node_meminfo(int nid, char *buf)
2093 struct hstate *h = &default_hstate;
2095 "Node %d HugePages_Total: %5u\n"
2096 "Node %d HugePages_Free: %5u\n"
2097 "Node %d HugePages_Surp: %5u\n",
2098 nid, h->nr_huge_pages_node[nid],
2099 nid, h->free_huge_pages_node[nid],
2100 nid, h->surplus_huge_pages_node[nid]);
2103 void hugetlb_show_meminfo(void)
2108 for_each_node_state(nid, N_MEMORY)
2110 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2112 h->nr_huge_pages_node[nid],
2113 h->free_huge_pages_node[nid],
2114 h->surplus_huge_pages_node[nid],
2115 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2118 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2119 unsigned long hugetlb_total_pages(void)
2122 unsigned long nr_total_pages = 0;
2125 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2126 return nr_total_pages;
2129 static int hugetlb_acct_memory(struct hstate *h, long delta)
2133 spin_lock(&hugetlb_lock);
2135 * When cpuset is configured, it breaks the strict hugetlb page
2136 * reservation as the accounting is done on a global variable. Such
2137 * reservation is completely rubbish in the presence of cpuset because
2138 * the reservation is not checked against page availability for the
2139 * current cpuset. Application can still potentially OOM'ed by kernel
2140 * with lack of free htlb page in cpuset that the task is in.
2141 * Attempt to enforce strict accounting with cpuset is almost
2142 * impossible (or too ugly) because cpuset is too fluid that
2143 * task or memory node can be dynamically moved between cpusets.
2145 * The change of semantics for shared hugetlb mapping with cpuset is
2146 * undesirable. However, in order to preserve some of the semantics,
2147 * we fall back to check against current free page availability as
2148 * a best attempt and hopefully to minimize the impact of changing
2149 * semantics that cpuset has.
2152 if (gather_surplus_pages(h, delta) < 0)
2155 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2156 return_unused_surplus_pages(h, delta);
2163 return_unused_surplus_pages(h, (unsigned long) -delta);
2166 spin_unlock(&hugetlb_lock);
2170 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2172 struct resv_map *reservations = vma_resv_map(vma);
2175 * This new VMA should share its siblings reservation map if present.
2176 * The VMA will only ever have a valid reservation map pointer where
2177 * it is being copied for another still existing VMA. As that VMA
2178 * has a reference to the reservation map it cannot disappear until
2179 * after this open call completes. It is therefore safe to take a
2180 * new reference here without additional locking.
2183 kref_get(&reservations->refs);
2186 static void resv_map_put(struct vm_area_struct *vma)
2188 struct resv_map *reservations = vma_resv_map(vma);
2192 kref_put(&reservations->refs, resv_map_release);
2195 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2197 struct hstate *h = hstate_vma(vma);
2198 struct resv_map *reservations = vma_resv_map(vma);
2199 struct hugepage_subpool *spool = subpool_vma(vma);
2200 unsigned long reserve;
2201 unsigned long start;
2205 start = vma_hugecache_offset(h, vma, vma->vm_start);
2206 end = vma_hugecache_offset(h, vma, vma->vm_end);
2208 reserve = (end - start) -
2209 region_count(&reservations->regions, start, end);
2214 hugetlb_acct_memory(h, -reserve);
2215 hugepage_subpool_put_pages(spool, reserve);
2221 * We cannot handle pagefaults against hugetlb pages at all. They cause
2222 * handle_mm_fault() to try to instantiate regular-sized pages in the
2223 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2226 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2232 const struct vm_operations_struct hugetlb_vm_ops = {
2233 .fault = hugetlb_vm_op_fault,
2234 .open = hugetlb_vm_op_open,
2235 .close = hugetlb_vm_op_close,
2238 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2244 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2245 vma->vm_page_prot)));
2247 entry = huge_pte_wrprotect(mk_huge_pte(page,
2248 vma->vm_page_prot));
2250 entry = pte_mkyoung(entry);
2251 entry = pte_mkhuge(entry);
2252 entry = arch_make_huge_pte(entry, vma, page, writable);
2257 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2258 unsigned long address, pte_t *ptep)
2262 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2263 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2264 update_mmu_cache(vma, address, ptep);
2268 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2269 struct vm_area_struct *vma)
2271 pte_t *src_pte, *dst_pte, entry;
2272 struct page *ptepage;
2275 struct hstate *h = hstate_vma(vma);
2276 unsigned long sz = huge_page_size(h);
2278 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2280 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2281 src_pte = huge_pte_offset(src, addr);
2284 dst_pte = huge_pte_alloc(dst, addr, sz);
2288 /* If the pagetables are shared don't copy or take references */
2289 if (dst_pte == src_pte)
2292 spin_lock(&dst->page_table_lock);
2293 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2294 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2296 huge_ptep_set_wrprotect(src, addr, src_pte);
2297 entry = huge_ptep_get(src_pte);
2298 ptepage = pte_page(entry);
2300 page_dup_rmap(ptepage);
2301 set_huge_pte_at(dst, addr, dst_pte, entry);
2303 spin_unlock(&src->page_table_lock);
2304 spin_unlock(&dst->page_table_lock);
2312 static int is_hugetlb_entry_migration(pte_t pte)
2316 if (huge_pte_none(pte) || pte_present(pte))
2318 swp = pte_to_swp_entry(pte);
2319 if (non_swap_entry(swp) && is_migration_entry(swp))
2325 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2329 if (huge_pte_none(pte) || pte_present(pte))
2331 swp = pte_to_swp_entry(pte);
2332 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2338 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2339 unsigned long start, unsigned long end,
2340 struct page *ref_page)
2342 int force_flush = 0;
2343 struct mm_struct *mm = vma->vm_mm;
2344 unsigned long address;
2348 struct hstate *h = hstate_vma(vma);
2349 unsigned long sz = huge_page_size(h);
2350 const unsigned long mmun_start = start; /* For mmu_notifiers */
2351 const unsigned long mmun_end = end; /* For mmu_notifiers */
2353 WARN_ON(!is_vm_hugetlb_page(vma));
2354 BUG_ON(start & ~huge_page_mask(h));
2355 BUG_ON(end & ~huge_page_mask(h));
2357 tlb_start_vma(tlb, vma);
2358 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2360 spin_lock(&mm->page_table_lock);
2361 for (address = start; address < end; address += sz) {
2362 ptep = huge_pte_offset(mm, address);
2366 if (huge_pmd_unshare(mm, &address, ptep))
2369 pte = huge_ptep_get(ptep);
2370 if (huge_pte_none(pte))
2374 * HWPoisoned hugepage is already unmapped and dropped reference
2376 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
2377 huge_pte_clear(mm, address, ptep);
2381 page = pte_page(pte);
2383 * If a reference page is supplied, it is because a specific
2384 * page is being unmapped, not a range. Ensure the page we
2385 * are about to unmap is the actual page of interest.
2388 if (page != ref_page)
2392 * Mark the VMA as having unmapped its page so that
2393 * future faults in this VMA will fail rather than
2394 * looking like data was lost
2396 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2399 pte = huge_ptep_get_and_clear(mm, address, ptep);
2400 tlb_remove_tlb_entry(tlb, ptep, address);
2401 if (huge_pte_dirty(pte))
2402 set_page_dirty(page);
2404 page_remove_rmap(page);
2405 force_flush = !__tlb_remove_page(tlb, page);
2408 /* Bail out after unmapping reference page if supplied */
2412 spin_unlock(&mm->page_table_lock);
2414 * mmu_gather ran out of room to batch pages, we break out of
2415 * the PTE lock to avoid doing the potential expensive TLB invalidate
2416 * and page-free while holding it.
2421 if (address < end && !ref_page)
2424 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2425 tlb_end_vma(tlb, vma);
2428 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2429 struct vm_area_struct *vma, unsigned long start,
2430 unsigned long end, struct page *ref_page)
2432 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2435 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2436 * test will fail on a vma being torn down, and not grab a page table
2437 * on its way out. We're lucky that the flag has such an appropriate
2438 * name, and can in fact be safely cleared here. We could clear it
2439 * before the __unmap_hugepage_range above, but all that's necessary
2440 * is to clear it before releasing the i_mmap_mutex. This works
2441 * because in the context this is called, the VMA is about to be
2442 * destroyed and the i_mmap_mutex is held.
2444 vma->vm_flags &= ~VM_MAYSHARE;
2447 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2448 unsigned long end, struct page *ref_page)
2450 struct mm_struct *mm;
2451 struct mmu_gather tlb;
2455 tlb_gather_mmu(&tlb, mm, start, end);
2456 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2457 tlb_finish_mmu(&tlb, start, end);
2461 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2462 * mappping it owns the reserve page for. The intention is to unmap the page
2463 * from other VMAs and let the children be SIGKILLed if they are faulting the
2466 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2467 struct page *page, unsigned long address)
2469 struct hstate *h = hstate_vma(vma);
2470 struct vm_area_struct *iter_vma;
2471 struct address_space *mapping;
2475 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2476 * from page cache lookup which is in HPAGE_SIZE units.
2478 address = address & huge_page_mask(h);
2479 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2481 mapping = file_inode(vma->vm_file)->i_mapping;
2484 * Take the mapping lock for the duration of the table walk. As
2485 * this mapping should be shared between all the VMAs,
2486 * __unmap_hugepage_range() is called as the lock is already held
2488 mutex_lock(&mapping->i_mmap_mutex);
2489 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2490 /* Do not unmap the current VMA */
2491 if (iter_vma == vma)
2495 * Unmap the page from other VMAs without their own reserves.
2496 * They get marked to be SIGKILLed if they fault in these
2497 * areas. This is because a future no-page fault on this VMA
2498 * could insert a zeroed page instead of the data existing
2499 * from the time of fork. This would look like data corruption
2501 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2502 unmap_hugepage_range(iter_vma, address,
2503 address + huge_page_size(h), page);
2505 mutex_unlock(&mapping->i_mmap_mutex);
2511 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2512 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2513 * cannot race with other handlers or page migration.
2514 * Keep the pte_same checks anyway to make transition from the mutex easier.
2516 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2517 unsigned long address, pte_t *ptep, pte_t pte,
2518 struct page *pagecache_page)
2520 struct hstate *h = hstate_vma(vma);
2521 struct page *old_page, *new_page;
2522 int outside_reserve = 0;
2523 unsigned long mmun_start; /* For mmu_notifiers */
2524 unsigned long mmun_end; /* For mmu_notifiers */
2526 old_page = pte_page(pte);
2529 /* If no-one else is actually using this page, avoid the copy
2530 * and just make the page writable */
2531 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2532 page_move_anon_rmap(old_page, vma, address);
2533 set_huge_ptep_writable(vma, address, ptep);
2538 * If the process that created a MAP_PRIVATE mapping is about to
2539 * perform a COW due to a shared page count, attempt to satisfy
2540 * the allocation without using the existing reserves. The pagecache
2541 * page is used to determine if the reserve at this address was
2542 * consumed or not. If reserves were used, a partial faulted mapping
2543 * at the time of fork() could consume its reserves on COW instead
2544 * of the full address range.
2546 if (!(vma->vm_flags & VM_MAYSHARE) &&
2547 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2548 old_page != pagecache_page)
2549 outside_reserve = 1;
2551 page_cache_get(old_page);
2553 /* Drop page_table_lock as buddy allocator may be called */
2554 spin_unlock(&mm->page_table_lock);
2555 new_page = alloc_huge_page(vma, address, outside_reserve);
2557 if (IS_ERR(new_page)) {
2558 long err = PTR_ERR(new_page);
2559 page_cache_release(old_page);
2562 * If a process owning a MAP_PRIVATE mapping fails to COW,
2563 * it is due to references held by a child and an insufficient
2564 * huge page pool. To guarantee the original mappers
2565 * reliability, unmap the page from child processes. The child
2566 * may get SIGKILLed if it later faults.
2568 if (outside_reserve) {
2569 BUG_ON(huge_pte_none(pte));
2570 if (unmap_ref_private(mm, vma, old_page, address)) {
2571 BUG_ON(huge_pte_none(pte));
2572 spin_lock(&mm->page_table_lock);
2573 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2574 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2575 goto retry_avoidcopy;
2577 * race occurs while re-acquiring page_table_lock, and
2585 /* Caller expects lock to be held */
2586 spin_lock(&mm->page_table_lock);
2588 return VM_FAULT_OOM;
2590 return VM_FAULT_SIGBUS;
2594 * When the original hugepage is shared one, it does not have
2595 * anon_vma prepared.
2597 if (unlikely(anon_vma_prepare(vma))) {
2598 page_cache_release(new_page);
2599 page_cache_release(old_page);
2600 /* Caller expects lock to be held */
2601 spin_lock(&mm->page_table_lock);
2602 return VM_FAULT_OOM;
2605 copy_user_huge_page(new_page, old_page, address, vma,
2606 pages_per_huge_page(h));
2607 __SetPageUptodate(new_page);
2609 mmun_start = address & huge_page_mask(h);
2610 mmun_end = mmun_start + huge_page_size(h);
2611 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2613 * Retake the page_table_lock to check for racing updates
2614 * before the page tables are altered
2616 spin_lock(&mm->page_table_lock);
2617 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2618 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2620 huge_ptep_clear_flush(vma, address, ptep);
2621 set_huge_pte_at(mm, address, ptep,
2622 make_huge_pte(vma, new_page, 1));
2623 page_remove_rmap(old_page);
2624 hugepage_add_new_anon_rmap(new_page, vma, address);
2625 /* Make the old page be freed below */
2626 new_page = old_page;
2628 spin_unlock(&mm->page_table_lock);
2629 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2630 /* Caller expects lock to be held */
2631 spin_lock(&mm->page_table_lock);
2632 page_cache_release(new_page);
2633 page_cache_release(old_page);
2637 /* Return the pagecache page at a given address within a VMA */
2638 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2639 struct vm_area_struct *vma, unsigned long address)
2641 struct address_space *mapping;
2644 mapping = vma->vm_file->f_mapping;
2645 idx = vma_hugecache_offset(h, vma, address);
2647 return find_lock_page(mapping, idx);
2651 * Return whether there is a pagecache page to back given address within VMA.
2652 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2654 static bool hugetlbfs_pagecache_present(struct hstate *h,
2655 struct vm_area_struct *vma, unsigned long address)
2657 struct address_space *mapping;
2661 mapping = vma->vm_file->f_mapping;
2662 idx = vma_hugecache_offset(h, vma, address);
2664 page = find_get_page(mapping, idx);
2667 return page != NULL;
2670 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2671 unsigned long address, pte_t *ptep, unsigned int flags)
2673 struct hstate *h = hstate_vma(vma);
2674 int ret = VM_FAULT_SIGBUS;
2679 struct address_space *mapping;
2683 * Currently, we are forced to kill the process in the event the
2684 * original mapper has unmapped pages from the child due to a failed
2685 * COW. Warn that such a situation has occurred as it may not be obvious
2687 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2688 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2693 mapping = vma->vm_file->f_mapping;
2694 idx = vma_hugecache_offset(h, vma, address);
2697 * Use page lock to guard against racing truncation
2698 * before we get page_table_lock.
2701 page = find_lock_page(mapping, idx);
2703 size = i_size_read(mapping->host) >> huge_page_shift(h);
2706 page = alloc_huge_page(vma, address, 0);
2708 ret = PTR_ERR(page);
2712 ret = VM_FAULT_SIGBUS;
2715 clear_huge_page(page, address, pages_per_huge_page(h));
2716 __SetPageUptodate(page);
2718 if (vma->vm_flags & VM_MAYSHARE) {
2720 struct inode *inode = mapping->host;
2722 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2730 spin_lock(&inode->i_lock);
2731 inode->i_blocks += blocks_per_huge_page(h);
2732 spin_unlock(&inode->i_lock);
2735 if (unlikely(anon_vma_prepare(vma))) {
2737 goto backout_unlocked;
2743 * If memory error occurs between mmap() and fault, some process
2744 * don't have hwpoisoned swap entry for errored virtual address.
2745 * So we need to block hugepage fault by PG_hwpoison bit check.
2747 if (unlikely(PageHWPoison(page))) {
2748 ret = VM_FAULT_HWPOISON |
2749 VM_FAULT_SET_HINDEX(hstate_index(h));
2750 goto backout_unlocked;
2755 * If we are going to COW a private mapping later, we examine the
2756 * pending reservations for this page now. This will ensure that
2757 * any allocations necessary to record that reservation occur outside
2760 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2761 if (vma_needs_reservation(h, vma, address) < 0) {
2763 goto backout_unlocked;
2766 spin_lock(&mm->page_table_lock);
2767 size = i_size_read(mapping->host) >> huge_page_shift(h);
2772 if (!huge_pte_none(huge_ptep_get(ptep)))
2776 hugepage_add_new_anon_rmap(page, vma, address);
2778 page_dup_rmap(page);
2779 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2780 && (vma->vm_flags & VM_SHARED)));
2781 set_huge_pte_at(mm, address, ptep, new_pte);
2783 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2784 /* Optimization, do the COW without a second fault */
2785 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2788 spin_unlock(&mm->page_table_lock);
2794 spin_unlock(&mm->page_table_lock);
2801 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2802 unsigned long address, unsigned int flags)
2807 struct page *page = NULL;
2808 struct page *pagecache_page = NULL;
2809 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2810 struct hstate *h = hstate_vma(vma);
2812 address &= huge_page_mask(h);
2814 ptep = huge_pte_offset(mm, address);
2816 entry = huge_ptep_get(ptep);
2817 if (unlikely(is_hugetlb_entry_migration(entry))) {
2818 migration_entry_wait_huge(mm, ptep);
2820 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2821 return VM_FAULT_HWPOISON_LARGE |
2822 VM_FAULT_SET_HINDEX(hstate_index(h));
2825 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2827 return VM_FAULT_OOM;
2830 * Serialize hugepage allocation and instantiation, so that we don't
2831 * get spurious allocation failures if two CPUs race to instantiate
2832 * the same page in the page cache.
2834 mutex_lock(&hugetlb_instantiation_mutex);
2835 entry = huge_ptep_get(ptep);
2836 if (huge_pte_none(entry)) {
2837 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2844 * If we are going to COW the mapping later, we examine the pending
2845 * reservations for this page now. This will ensure that any
2846 * allocations necessary to record that reservation occur outside the
2847 * spinlock. For private mappings, we also lookup the pagecache
2848 * page now as it is used to determine if a reservation has been
2851 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
2852 if (vma_needs_reservation(h, vma, address) < 0) {
2857 if (!(vma->vm_flags & VM_MAYSHARE))
2858 pagecache_page = hugetlbfs_pagecache_page(h,
2863 * hugetlb_cow() requires page locks of pte_page(entry) and
2864 * pagecache_page, so here we need take the former one
2865 * when page != pagecache_page or !pagecache_page.
2866 * Note that locking order is always pagecache_page -> page,
2867 * so no worry about deadlock.
2869 page = pte_page(entry);
2871 if (page != pagecache_page)
2874 spin_lock(&mm->page_table_lock);
2875 /* Check for a racing update before calling hugetlb_cow */
2876 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2877 goto out_page_table_lock;
2880 if (flags & FAULT_FLAG_WRITE) {
2881 if (!huge_pte_write(entry)) {
2882 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2884 goto out_page_table_lock;
2886 entry = huge_pte_mkdirty(entry);
2888 entry = pte_mkyoung(entry);
2889 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2890 flags & FAULT_FLAG_WRITE))
2891 update_mmu_cache(vma, address, ptep);
2893 out_page_table_lock:
2894 spin_unlock(&mm->page_table_lock);
2896 if (pagecache_page) {
2897 unlock_page(pagecache_page);
2898 put_page(pagecache_page);
2900 if (page != pagecache_page)
2905 mutex_unlock(&hugetlb_instantiation_mutex);
2910 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2911 struct page **pages, struct vm_area_struct **vmas,
2912 unsigned long *position, unsigned long *nr_pages,
2913 long i, unsigned int flags)
2915 unsigned long pfn_offset;
2916 unsigned long vaddr = *position;
2917 unsigned long remainder = *nr_pages;
2918 struct hstate *h = hstate_vma(vma);
2920 spin_lock(&mm->page_table_lock);
2921 while (vaddr < vma->vm_end && remainder) {
2927 * Some archs (sparc64, sh*) have multiple pte_ts to
2928 * each hugepage. We have to make sure we get the
2929 * first, for the page indexing below to work.
2931 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2932 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2935 * When coredumping, it suits get_dump_page if we just return
2936 * an error where there's an empty slot with no huge pagecache
2937 * to back it. This way, we avoid allocating a hugepage, and
2938 * the sparse dumpfile avoids allocating disk blocks, but its
2939 * huge holes still show up with zeroes where they need to be.
2941 if (absent && (flags & FOLL_DUMP) &&
2942 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2948 * We need call hugetlb_fault for both hugepages under migration
2949 * (in which case hugetlb_fault waits for the migration,) and
2950 * hwpoisoned hugepages (in which case we need to prevent the
2951 * caller from accessing to them.) In order to do this, we use
2952 * here is_swap_pte instead of is_hugetlb_entry_migration and
2953 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
2954 * both cases, and because we can't follow correct pages
2955 * directly from any kind of swap entries.
2957 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
2958 ((flags & FOLL_WRITE) &&
2959 !huge_pte_write(huge_ptep_get(pte)))) {
2962 spin_unlock(&mm->page_table_lock);
2963 ret = hugetlb_fault(mm, vma, vaddr,
2964 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2965 spin_lock(&mm->page_table_lock);
2966 if (!(ret & VM_FAULT_ERROR))
2973 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2974 page = pte_page(huge_ptep_get(pte));
2977 pages[i] = mem_map_offset(page, pfn_offset);
2988 if (vaddr < vma->vm_end && remainder &&
2989 pfn_offset < pages_per_huge_page(h)) {
2991 * We use pfn_offset to avoid touching the pageframes
2992 * of this compound page.
2997 spin_unlock(&mm->page_table_lock);
2998 *nr_pages = remainder;
3001 return i ? i : -EFAULT;
3004 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3005 unsigned long address, unsigned long end, pgprot_t newprot)
3007 struct mm_struct *mm = vma->vm_mm;
3008 unsigned long start = address;
3011 struct hstate *h = hstate_vma(vma);
3012 unsigned long pages = 0;
3014 BUG_ON(address >= end);
3015 flush_cache_range(vma, address, end);
3017 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3018 spin_lock(&mm->page_table_lock);
3019 for (; address < end; address += huge_page_size(h)) {
3020 ptep = huge_pte_offset(mm, address);
3023 if (huge_pmd_unshare(mm, &address, ptep)) {
3027 if (!huge_pte_none(huge_ptep_get(ptep))) {
3028 pte = huge_ptep_get_and_clear(mm, address, ptep);
3029 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3030 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3031 set_huge_pte_at(mm, address, ptep, pte);
3035 spin_unlock(&mm->page_table_lock);
3037 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3038 * may have cleared our pud entry and done put_page on the page table:
3039 * once we release i_mmap_mutex, another task can do the final put_page
3040 * and that page table be reused and filled with junk.
3042 flush_tlb_range(vma, start, end);
3043 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3045 return pages << h->order;
3048 int hugetlb_reserve_pages(struct inode *inode,
3050 struct vm_area_struct *vma,
3051 vm_flags_t vm_flags)
3054 struct hstate *h = hstate_inode(inode);
3055 struct hugepage_subpool *spool = subpool_inode(inode);
3058 * Only apply hugepage reservation if asked. At fault time, an
3059 * attempt will be made for VM_NORESERVE to allocate a page
3060 * without using reserves
3062 if (vm_flags & VM_NORESERVE)
3066 * Shared mappings base their reservation on the number of pages that
3067 * are already allocated on behalf of the file. Private mappings need
3068 * to reserve the full area even if read-only as mprotect() may be
3069 * called to make the mapping read-write. Assume !vma is a shm mapping
3071 if (!vma || vma->vm_flags & VM_MAYSHARE)
3072 chg = region_chg(&inode->i_mapping->private_list, from, to);
3074 struct resv_map *resv_map = resv_map_alloc();
3080 set_vma_resv_map(vma, resv_map);
3081 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3089 /* There must be enough pages in the subpool for the mapping */
3090 if (hugepage_subpool_get_pages(spool, chg)) {
3096 * Check enough hugepages are available for the reservation.
3097 * Hand the pages back to the subpool if there are not
3099 ret = hugetlb_acct_memory(h, chg);
3101 hugepage_subpool_put_pages(spool, chg);
3106 * Account for the reservations made. Shared mappings record regions
3107 * that have reservations as they are shared by multiple VMAs.
3108 * When the last VMA disappears, the region map says how much
3109 * the reservation was and the page cache tells how much of
3110 * the reservation was consumed. Private mappings are per-VMA and
3111 * only the consumed reservations are tracked. When the VMA
3112 * disappears, the original reservation is the VMA size and the
3113 * consumed reservations are stored in the map. Hence, nothing
3114 * else has to be done for private mappings here
3116 if (!vma || vma->vm_flags & VM_MAYSHARE)
3117 region_add(&inode->i_mapping->private_list, from, to);
3125 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3127 struct hstate *h = hstate_inode(inode);
3128 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3129 struct hugepage_subpool *spool = subpool_inode(inode);
3131 spin_lock(&inode->i_lock);
3132 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3133 spin_unlock(&inode->i_lock);
3135 hugepage_subpool_put_pages(spool, (chg - freed));
3136 hugetlb_acct_memory(h, -(chg - freed));
3139 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3140 static unsigned long page_table_shareable(struct vm_area_struct *svma,
3141 struct vm_area_struct *vma,
3142 unsigned long addr, pgoff_t idx)
3144 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3146 unsigned long sbase = saddr & PUD_MASK;
3147 unsigned long s_end = sbase + PUD_SIZE;
3149 /* Allow segments to share if only one is marked locked */
3150 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3151 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3154 * match the virtual addresses, permission and the alignment of the
3157 if (pmd_index(addr) != pmd_index(saddr) ||
3158 vm_flags != svm_flags ||
3159 sbase < svma->vm_start || svma->vm_end < s_end)
3165 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3167 unsigned long base = addr & PUD_MASK;
3168 unsigned long end = base + PUD_SIZE;
3171 * check on proper vm_flags and page table alignment
3173 if (vma->vm_flags & VM_MAYSHARE &&
3174 vma->vm_start <= base && end <= vma->vm_end)
3180 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3181 * and returns the corresponding pte. While this is not necessary for the
3182 * !shared pmd case because we can allocate the pmd later as well, it makes the
3183 * code much cleaner. pmd allocation is essential for the shared case because
3184 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3185 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3186 * bad pmd for sharing.
3188 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3190 struct vm_area_struct *vma = find_vma(mm, addr);
3191 struct address_space *mapping = vma->vm_file->f_mapping;
3192 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3194 struct vm_area_struct *svma;
3195 unsigned long saddr;
3199 if (!vma_shareable(vma, addr))
3200 return (pte_t *)pmd_alloc(mm, pud, addr);
3202 mutex_lock(&mapping->i_mmap_mutex);
3203 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3207 saddr = page_table_shareable(svma, vma, addr, idx);
3209 spte = huge_pte_offset(svma->vm_mm, saddr);
3211 get_page(virt_to_page(spte));
3220 spin_lock(&mm->page_table_lock);
3222 pud_populate(mm, pud,
3223 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3225 put_page(virt_to_page(spte));
3226 spin_unlock(&mm->page_table_lock);
3228 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3229 mutex_unlock(&mapping->i_mmap_mutex);
3234 * unmap huge page backed by shared pte.
3236 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3237 * indicated by page_count > 1, unmap is achieved by clearing pud and
3238 * decrementing the ref count. If count == 1, the pte page is not shared.
3240 * called with vma->vm_mm->page_table_lock held.
3242 * returns: 1 successfully unmapped a shared pte page
3243 * 0 the underlying pte page is not shared, or it is the last user
3245 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3247 pgd_t *pgd = pgd_offset(mm, *addr);
3248 pud_t *pud = pud_offset(pgd, *addr);
3250 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3251 if (page_count(virt_to_page(ptep)) == 1)
3255 put_page(virt_to_page(ptep));
3256 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3259 #define want_pmd_share() (1)
3260 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3261 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3265 #define want_pmd_share() (0)
3266 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3268 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3269 pte_t *huge_pte_alloc(struct mm_struct *mm,
3270 unsigned long addr, unsigned long sz)
3276 pgd = pgd_offset(mm, addr);
3277 pud = pud_alloc(mm, pgd, addr);
3279 if (sz == PUD_SIZE) {
3282 BUG_ON(sz != PMD_SIZE);
3283 if (want_pmd_share() && pud_none(*pud))
3284 pte = huge_pmd_share(mm, addr, pud);
3286 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3289 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3294 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3300 pgd = pgd_offset(mm, addr);
3301 if (pgd_present(*pgd)) {
3302 pud = pud_offset(pgd, addr);
3303 if (pud_present(*pud)) {
3305 return (pte_t *)pud;
3306 pmd = pmd_offset(pud, addr);
3309 return (pte_t *) pmd;
3313 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3314 pmd_t *pmd, int write)
3318 page = pte_page(*(pte_t *)pmd);
3320 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
3325 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3326 pud_t *pud, int write)
3330 page = pte_page(*(pte_t *)pud);
3332 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
3336 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3338 /* Can be overriden by architectures */
3339 __attribute__((weak)) struct page *
3340 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3341 pud_t *pud, int write)
3347 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3349 #ifdef CONFIG_MEMORY_FAILURE
3351 /* Should be called in hugetlb_lock */
3352 static int is_hugepage_on_freelist(struct page *hpage)
3356 struct hstate *h = page_hstate(hpage);
3357 int nid = page_to_nid(hpage);
3359 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3366 * This function is called from memory failure code.
3367 * Assume the caller holds page lock of the head page.
3369 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3371 struct hstate *h = page_hstate(hpage);
3372 int nid = page_to_nid(hpage);
3375 spin_lock(&hugetlb_lock);
3376 if (is_hugepage_on_freelist(hpage)) {
3378 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3379 * but dangling hpage->lru can trigger list-debug warnings
3380 * (this happens when we call unpoison_memory() on it),
3381 * so let it point to itself with list_del_init().
3383 list_del_init(&hpage->lru);
3384 set_page_refcounted(hpage);
3385 h->free_huge_pages--;
3386 h->free_huge_pages_node[nid]--;
3389 spin_unlock(&hugetlb_lock);