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, long chg)
448 if (vma->vm_flags & VM_NORESERVE) {
450 * This address is already reserved by other process(chg == 0),
451 * so, we should decrement reserved count. Without decrementing,
452 * reserve count remains after releasing inode, because this
453 * allocated page will go into page cache and is regarded as
454 * coming from reserved pool in releasing step. Currently, we
455 * don't have any other solution to deal with this situation
456 * properly, so add work-around here.
458 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
464 /* Shared mappings always use reserves */
465 if (vma->vm_flags & VM_MAYSHARE)
469 * Only the process that called mmap() has reserves for
472 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
478 static void copy_gigantic_page(struct page *dst, struct page *src)
481 struct hstate *h = page_hstate(src);
482 struct page *dst_base = dst;
483 struct page *src_base = src;
485 for (i = 0; i < pages_per_huge_page(h); ) {
487 copy_highpage(dst, src);
490 dst = mem_map_next(dst, dst_base, i);
491 src = mem_map_next(src, src_base, i);
495 void copy_huge_page(struct page *dst, struct page *src)
498 struct hstate *h = page_hstate(src);
500 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
501 copy_gigantic_page(dst, src);
506 for (i = 0; i < pages_per_huge_page(h); i++) {
508 copy_highpage(dst + i, src + i);
512 static void enqueue_huge_page(struct hstate *h, struct page *page)
514 int nid = page_to_nid(page);
515 list_move(&page->lru, &h->hugepage_freelists[nid]);
516 h->free_huge_pages++;
517 h->free_huge_pages_node[nid]++;
520 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
524 if (list_empty(&h->hugepage_freelists[nid]))
526 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
527 list_move(&page->lru, &h->hugepage_activelist);
528 set_page_refcounted(page);
529 h->free_huge_pages--;
530 h->free_huge_pages_node[nid]--;
534 static struct page *dequeue_huge_page_vma(struct hstate *h,
535 struct vm_area_struct *vma,
536 unsigned long address, int avoid_reserve,
539 struct page *page = NULL;
540 struct mempolicy *mpol;
541 nodemask_t *nodemask;
542 struct zonelist *zonelist;
545 unsigned int cpuset_mems_cookie;
548 * A child process with MAP_PRIVATE mappings created by their parent
549 * have no page reserves. This check ensures that reservations are
550 * not "stolen". The child may still get SIGKILLed
552 if (!vma_has_reserves(vma, chg) &&
553 h->free_huge_pages - h->resv_huge_pages == 0)
556 /* If reserves cannot be used, ensure enough pages are in the pool */
557 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
561 cpuset_mems_cookie = get_mems_allowed();
562 zonelist = huge_zonelist(vma, address,
563 htlb_alloc_mask, &mpol, &nodemask);
565 for_each_zone_zonelist_nodemask(zone, z, zonelist,
566 MAX_NR_ZONES - 1, nodemask) {
567 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
568 page = dequeue_huge_page_node(h, zone_to_nid(zone));
572 if (!vma_has_reserves(vma, chg))
575 h->resv_huge_pages--;
582 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
590 static void update_and_free_page(struct hstate *h, struct page *page)
594 VM_BUG_ON(h->order >= MAX_ORDER);
597 h->nr_huge_pages_node[page_to_nid(page)]--;
598 for (i = 0; i < pages_per_huge_page(h); i++) {
599 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
600 1 << PG_referenced | 1 << PG_dirty |
601 1 << PG_active | 1 << PG_reserved |
602 1 << PG_private | 1 << PG_writeback);
604 VM_BUG_ON(hugetlb_cgroup_from_page(page));
605 set_compound_page_dtor(page, NULL);
606 set_page_refcounted(page);
607 arch_release_hugepage(page);
608 __free_pages(page, huge_page_order(h));
611 struct hstate *size_to_hstate(unsigned long size)
616 if (huge_page_size(h) == size)
622 static void free_huge_page(struct page *page)
625 * Can't pass hstate in here because it is called from the
626 * compound page destructor.
628 struct hstate *h = page_hstate(page);
629 int nid = page_to_nid(page);
630 struct hugepage_subpool *spool =
631 (struct hugepage_subpool *)page_private(page);
633 set_page_private(page, 0);
634 page->mapping = NULL;
635 BUG_ON(page_count(page));
636 BUG_ON(page_mapcount(page));
638 spin_lock(&hugetlb_lock);
639 hugetlb_cgroup_uncharge_page(hstate_index(h),
640 pages_per_huge_page(h), page);
641 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
642 /* remove the page from active list */
643 list_del(&page->lru);
644 update_and_free_page(h, page);
645 h->surplus_huge_pages--;
646 h->surplus_huge_pages_node[nid]--;
648 arch_clear_hugepage_flags(page);
649 enqueue_huge_page(h, page);
651 spin_unlock(&hugetlb_lock);
652 hugepage_subpool_put_pages(spool, 1);
655 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
657 INIT_LIST_HEAD(&page->lru);
658 set_compound_page_dtor(page, free_huge_page);
659 spin_lock(&hugetlb_lock);
660 set_hugetlb_cgroup(page, NULL);
662 h->nr_huge_pages_node[nid]++;
663 spin_unlock(&hugetlb_lock);
664 put_page(page); /* free it into the hugepage allocator */
667 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
670 int nr_pages = 1 << order;
671 struct page *p = page + 1;
673 /* we rely on prep_new_huge_page to set the destructor */
674 set_compound_order(page, order);
676 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
678 set_page_count(p, 0);
679 p->first_page = page;
684 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
685 * transparent huge pages. See the PageTransHuge() documentation for more
688 int PageHuge(struct page *page)
690 compound_page_dtor *dtor;
692 if (!PageCompound(page))
695 page = compound_head(page);
696 dtor = get_compound_page_dtor(page);
698 return dtor == free_huge_page;
700 EXPORT_SYMBOL_GPL(PageHuge);
702 pgoff_t __basepage_index(struct page *page)
704 struct page *page_head = compound_head(page);
705 pgoff_t index = page_index(page_head);
706 unsigned long compound_idx;
708 if (!PageHuge(page_head))
709 return page_index(page);
711 if (compound_order(page_head) >= MAX_ORDER)
712 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
714 compound_idx = page - page_head;
716 return (index << compound_order(page_head)) + compound_idx;
719 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
723 if (h->order >= MAX_ORDER)
726 page = alloc_pages_exact_node(nid,
727 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
728 __GFP_REPEAT|__GFP_NOWARN,
731 if (arch_prepare_hugepage(page)) {
732 __free_pages(page, huge_page_order(h));
735 prep_new_huge_page(h, page, nid);
742 * common helper functions for hstate_next_node_to_{alloc|free}.
743 * We may have allocated or freed a huge page based on a different
744 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
745 * be outside of *nodes_allowed. Ensure that we use an allowed
746 * node for alloc or free.
748 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
750 nid = next_node(nid, *nodes_allowed);
751 if (nid == MAX_NUMNODES)
752 nid = first_node(*nodes_allowed);
753 VM_BUG_ON(nid >= MAX_NUMNODES);
758 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
760 if (!node_isset(nid, *nodes_allowed))
761 nid = next_node_allowed(nid, nodes_allowed);
766 * returns the previously saved node ["this node"] from which to
767 * allocate a persistent huge page for the pool and advance the
768 * next node from which to allocate, handling wrap at end of node
771 static int hstate_next_node_to_alloc(struct hstate *h,
772 nodemask_t *nodes_allowed)
776 VM_BUG_ON(!nodes_allowed);
778 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
779 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
785 * helper for free_pool_huge_page() - return the previously saved
786 * node ["this node"] from which to free a huge page. Advance the
787 * next node id whether or not we find a free huge page to free so
788 * that the next attempt to free addresses the next node.
790 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
794 VM_BUG_ON(!nodes_allowed);
796 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
797 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
802 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
803 for (nr_nodes = nodes_weight(*mask); \
805 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
808 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
809 for (nr_nodes = nodes_weight(*mask); \
811 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
814 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
820 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
821 page = alloc_fresh_huge_page_node(h, node);
829 count_vm_event(HTLB_BUDDY_PGALLOC);
831 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
837 * Free huge page from pool from next node to free.
838 * Attempt to keep persistent huge pages more or less
839 * balanced over allowed nodes.
840 * Called with hugetlb_lock locked.
842 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
848 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
850 * If we're returning unused surplus pages, only examine
851 * nodes with surplus pages.
853 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
854 !list_empty(&h->hugepage_freelists[node])) {
856 list_entry(h->hugepage_freelists[node].next,
858 list_del(&page->lru);
859 h->free_huge_pages--;
860 h->free_huge_pages_node[node]--;
862 h->surplus_huge_pages--;
863 h->surplus_huge_pages_node[node]--;
865 update_and_free_page(h, page);
874 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
879 if (h->order >= MAX_ORDER)
883 * Assume we will successfully allocate the surplus page to
884 * prevent racing processes from causing the surplus to exceed
887 * This however introduces a different race, where a process B
888 * tries to grow the static hugepage pool while alloc_pages() is
889 * called by process A. B will only examine the per-node
890 * counters in determining if surplus huge pages can be
891 * converted to normal huge pages in adjust_pool_surplus(). A
892 * won't be able to increment the per-node counter, until the
893 * lock is dropped by B, but B doesn't drop hugetlb_lock until
894 * no more huge pages can be converted from surplus to normal
895 * state (and doesn't try to convert again). Thus, we have a
896 * case where a surplus huge page exists, the pool is grown, and
897 * the surplus huge page still exists after, even though it
898 * should just have been converted to a normal huge page. This
899 * does not leak memory, though, as the hugepage will be freed
900 * once it is out of use. It also does not allow the counters to
901 * go out of whack in adjust_pool_surplus() as we don't modify
902 * the node values until we've gotten the hugepage and only the
903 * per-node value is checked there.
905 spin_lock(&hugetlb_lock);
906 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
907 spin_unlock(&hugetlb_lock);
911 h->surplus_huge_pages++;
913 spin_unlock(&hugetlb_lock);
915 if (nid == NUMA_NO_NODE)
916 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
917 __GFP_REPEAT|__GFP_NOWARN,
920 page = alloc_pages_exact_node(nid,
921 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
922 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
924 if (page && arch_prepare_hugepage(page)) {
925 __free_pages(page, huge_page_order(h));
929 spin_lock(&hugetlb_lock);
931 INIT_LIST_HEAD(&page->lru);
932 r_nid = page_to_nid(page);
933 set_compound_page_dtor(page, free_huge_page);
934 set_hugetlb_cgroup(page, NULL);
936 * We incremented the global counters already
938 h->nr_huge_pages_node[r_nid]++;
939 h->surplus_huge_pages_node[r_nid]++;
940 __count_vm_event(HTLB_BUDDY_PGALLOC);
943 h->surplus_huge_pages--;
944 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
946 spin_unlock(&hugetlb_lock);
952 * This allocation function is useful in the context where vma is irrelevant.
953 * E.g. soft-offlining uses this function because it only cares physical
954 * address of error page.
956 struct page *alloc_huge_page_node(struct hstate *h, int nid)
958 struct page *page = NULL;
960 spin_lock(&hugetlb_lock);
961 if (h->free_huge_pages - h->resv_huge_pages > 0)
962 page = dequeue_huge_page_node(h, nid);
963 spin_unlock(&hugetlb_lock);
966 page = alloc_buddy_huge_page(h, nid);
972 * Increase the hugetlb pool such that it can accommodate a reservation
975 static int gather_surplus_pages(struct hstate *h, int delta)
977 struct list_head surplus_list;
978 struct page *page, *tmp;
980 int needed, allocated;
981 bool alloc_ok = true;
983 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
985 h->resv_huge_pages += delta;
990 INIT_LIST_HEAD(&surplus_list);
994 spin_unlock(&hugetlb_lock);
995 for (i = 0; i < needed; i++) {
996 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1001 list_add(&page->lru, &surplus_list);
1006 * After retaking hugetlb_lock, we need to recalculate 'needed'
1007 * because either resv_huge_pages or free_huge_pages may have changed.
1009 spin_lock(&hugetlb_lock);
1010 needed = (h->resv_huge_pages + delta) -
1011 (h->free_huge_pages + allocated);
1016 * We were not able to allocate enough pages to
1017 * satisfy the entire reservation so we free what
1018 * we've allocated so far.
1023 * The surplus_list now contains _at_least_ the number of extra pages
1024 * needed to accommodate the reservation. Add the appropriate number
1025 * of pages to the hugetlb pool and free the extras back to the buddy
1026 * allocator. Commit the entire reservation here to prevent another
1027 * process from stealing the pages as they are added to the pool but
1028 * before they are reserved.
1030 needed += allocated;
1031 h->resv_huge_pages += delta;
1034 /* Free the needed pages to the hugetlb pool */
1035 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1039 * This page is now managed by the hugetlb allocator and has
1040 * no users -- drop the buddy allocator's reference.
1042 put_page_testzero(page);
1043 VM_BUG_ON(page_count(page));
1044 enqueue_huge_page(h, page);
1047 spin_unlock(&hugetlb_lock);
1049 /* Free unnecessary surplus pages to the buddy allocator */
1050 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1052 spin_lock(&hugetlb_lock);
1058 * When releasing a hugetlb pool reservation, any surplus pages that were
1059 * allocated to satisfy the reservation must be explicitly freed if they were
1061 * Called with hugetlb_lock held.
1063 static void return_unused_surplus_pages(struct hstate *h,
1064 unsigned long unused_resv_pages)
1066 unsigned long nr_pages;
1068 /* Uncommit the reservation */
1069 h->resv_huge_pages -= unused_resv_pages;
1071 /* Cannot return gigantic pages currently */
1072 if (h->order >= MAX_ORDER)
1075 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1078 * We want to release as many surplus pages as possible, spread
1079 * evenly across all nodes with memory. Iterate across these nodes
1080 * until we can no longer free unreserved surplus pages. This occurs
1081 * when the nodes with surplus pages have no free pages.
1082 * free_pool_huge_page() will balance the the freed pages across the
1083 * on-line nodes with memory and will handle the hstate accounting.
1085 while (nr_pages--) {
1086 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1092 * Determine if the huge page at addr within the vma has an associated
1093 * reservation. Where it does not we will need to logically increase
1094 * reservation and actually increase subpool usage before an allocation
1095 * can occur. Where any new reservation would be required the
1096 * reservation change is prepared, but not committed. Once the page
1097 * has been allocated from the subpool and instantiated the change should
1098 * be committed via vma_commit_reservation. No action is required on
1101 static long vma_needs_reservation(struct hstate *h,
1102 struct vm_area_struct *vma, unsigned long addr)
1104 struct address_space *mapping = vma->vm_file->f_mapping;
1105 struct inode *inode = mapping->host;
1107 if (vma->vm_flags & VM_MAYSHARE) {
1108 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1109 return region_chg(&inode->i_mapping->private_list,
1112 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1117 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1118 struct resv_map *reservations = vma_resv_map(vma);
1120 err = region_chg(&reservations->regions, idx, idx + 1);
1126 static void vma_commit_reservation(struct hstate *h,
1127 struct vm_area_struct *vma, unsigned long addr)
1129 struct address_space *mapping = vma->vm_file->f_mapping;
1130 struct inode *inode = mapping->host;
1132 if (vma->vm_flags & VM_MAYSHARE) {
1133 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1134 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1136 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1137 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1138 struct resv_map *reservations = vma_resv_map(vma);
1140 /* Mark this page used in the map. */
1141 region_add(&reservations->regions, idx, idx + 1);
1145 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1146 unsigned long addr, int avoid_reserve)
1148 struct hugepage_subpool *spool = subpool_vma(vma);
1149 struct hstate *h = hstate_vma(vma);
1153 struct hugetlb_cgroup *h_cg;
1155 idx = hstate_index(h);
1157 * Processes that did not create the mapping will have no
1158 * reserves and will not have accounted against subpool
1159 * limit. Check that the subpool limit can be made before
1160 * satisfying the allocation MAP_NORESERVE mappings may also
1161 * need pages and subpool limit allocated allocated if no reserve
1164 chg = vma_needs_reservation(h, vma, addr);
1166 return ERR_PTR(-ENOMEM);
1168 if (hugepage_subpool_get_pages(spool, chg))
1169 return ERR_PTR(-ENOSPC);
1171 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1173 hugepage_subpool_put_pages(spool, chg);
1174 return ERR_PTR(-ENOSPC);
1176 spin_lock(&hugetlb_lock);
1177 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
1179 spin_unlock(&hugetlb_lock);
1180 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1182 hugetlb_cgroup_uncharge_cgroup(idx,
1183 pages_per_huge_page(h),
1185 hugepage_subpool_put_pages(spool, chg);
1186 return ERR_PTR(-ENOSPC);
1188 spin_lock(&hugetlb_lock);
1189 list_move(&page->lru, &h->hugepage_activelist);
1192 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1193 spin_unlock(&hugetlb_lock);
1195 set_page_private(page, (unsigned long)spool);
1197 vma_commit_reservation(h, vma, addr);
1201 int __weak alloc_bootmem_huge_page(struct hstate *h)
1203 struct huge_bootmem_page *m;
1206 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1209 addr = __alloc_bootmem_node_nopanic(NODE_DATA(node),
1210 huge_page_size(h), huge_page_size(h), 0);
1214 * Use the beginning of the huge page to store the
1215 * huge_bootmem_page struct (until gather_bootmem
1216 * puts them into the mem_map).
1225 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1226 /* Put them into a private list first because mem_map is not up yet */
1227 list_add(&m->list, &huge_boot_pages);
1232 static void prep_compound_huge_page(struct page *page, int order)
1234 if (unlikely(order > (MAX_ORDER - 1)))
1235 prep_compound_gigantic_page(page, order);
1237 prep_compound_page(page, order);
1240 /* Put bootmem huge pages into the standard lists after mem_map is up */
1241 static void __init gather_bootmem_prealloc(void)
1243 struct huge_bootmem_page *m;
1245 list_for_each_entry(m, &huge_boot_pages, list) {
1246 struct hstate *h = m->hstate;
1249 #ifdef CONFIG_HIGHMEM
1250 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1251 free_bootmem_late((unsigned long)m,
1252 sizeof(struct huge_bootmem_page));
1254 page = virt_to_page(m);
1256 __ClearPageReserved(page);
1257 WARN_ON(page_count(page) != 1);
1258 prep_compound_huge_page(page, h->order);
1259 prep_new_huge_page(h, page, page_to_nid(page));
1261 * If we had gigantic hugepages allocated at boot time, we need
1262 * to restore the 'stolen' pages to totalram_pages in order to
1263 * fix confusing memory reports from free(1) and another
1264 * side-effects, like CommitLimit going negative.
1266 if (h->order > (MAX_ORDER - 1))
1267 adjust_managed_page_count(page, 1 << h->order);
1271 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1275 for (i = 0; i < h->max_huge_pages; ++i) {
1276 if (h->order >= MAX_ORDER) {
1277 if (!alloc_bootmem_huge_page(h))
1279 } else if (!alloc_fresh_huge_page(h,
1280 &node_states[N_MEMORY]))
1283 h->max_huge_pages = i;
1286 static void __init hugetlb_init_hstates(void)
1290 for_each_hstate(h) {
1291 /* oversize hugepages were init'ed in early boot */
1292 if (h->order < MAX_ORDER)
1293 hugetlb_hstate_alloc_pages(h);
1297 static char * __init memfmt(char *buf, unsigned long n)
1299 if (n >= (1UL << 30))
1300 sprintf(buf, "%lu GB", n >> 30);
1301 else if (n >= (1UL << 20))
1302 sprintf(buf, "%lu MB", n >> 20);
1304 sprintf(buf, "%lu KB", n >> 10);
1308 static void __init report_hugepages(void)
1312 for_each_hstate(h) {
1314 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1315 memfmt(buf, huge_page_size(h)),
1316 h->free_huge_pages);
1320 #ifdef CONFIG_HIGHMEM
1321 static void try_to_free_low(struct hstate *h, unsigned long count,
1322 nodemask_t *nodes_allowed)
1326 if (h->order >= MAX_ORDER)
1329 for_each_node_mask(i, *nodes_allowed) {
1330 struct page *page, *next;
1331 struct list_head *freel = &h->hugepage_freelists[i];
1332 list_for_each_entry_safe(page, next, freel, lru) {
1333 if (count >= h->nr_huge_pages)
1335 if (PageHighMem(page))
1337 list_del(&page->lru);
1338 update_and_free_page(h, page);
1339 h->free_huge_pages--;
1340 h->free_huge_pages_node[page_to_nid(page)]--;
1345 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1346 nodemask_t *nodes_allowed)
1352 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1353 * balanced by operating on them in a round-robin fashion.
1354 * Returns 1 if an adjustment was made.
1356 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1361 VM_BUG_ON(delta != -1 && delta != 1);
1364 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1365 if (h->surplus_huge_pages_node[node])
1369 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1370 if (h->surplus_huge_pages_node[node] <
1371 h->nr_huge_pages_node[node])
1378 h->surplus_huge_pages += delta;
1379 h->surplus_huge_pages_node[node] += delta;
1383 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1384 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1385 nodemask_t *nodes_allowed)
1387 unsigned long min_count, ret;
1389 if (h->order >= MAX_ORDER)
1390 return h->max_huge_pages;
1393 * Increase the pool size
1394 * First take pages out of surplus state. Then make up the
1395 * remaining difference by allocating fresh huge pages.
1397 * We might race with alloc_buddy_huge_page() here and be unable
1398 * to convert a surplus huge page to a normal huge page. That is
1399 * not critical, though, it just means the overall size of the
1400 * pool might be one hugepage larger than it needs to be, but
1401 * within all the constraints specified by the sysctls.
1403 spin_lock(&hugetlb_lock);
1404 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1405 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1409 while (count > persistent_huge_pages(h)) {
1411 * If this allocation races such that we no longer need the
1412 * page, free_huge_page will handle it by freeing the page
1413 * and reducing the surplus.
1415 spin_unlock(&hugetlb_lock);
1416 ret = alloc_fresh_huge_page(h, nodes_allowed);
1417 spin_lock(&hugetlb_lock);
1421 /* Bail for signals. Probably ctrl-c from user */
1422 if (signal_pending(current))
1427 * Decrease the pool size
1428 * First return free pages to the buddy allocator (being careful
1429 * to keep enough around to satisfy reservations). Then place
1430 * pages into surplus state as needed so the pool will shrink
1431 * to the desired size as pages become free.
1433 * By placing pages into the surplus state independent of the
1434 * overcommit value, we are allowing the surplus pool size to
1435 * exceed overcommit. There are few sane options here. Since
1436 * alloc_buddy_huge_page() is checking the global counter,
1437 * though, we'll note that we're not allowed to exceed surplus
1438 * and won't grow the pool anywhere else. Not until one of the
1439 * sysctls are changed, or the surplus pages go out of use.
1441 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1442 min_count = max(count, min_count);
1443 try_to_free_low(h, min_count, nodes_allowed);
1444 while (min_count < persistent_huge_pages(h)) {
1445 if (!free_pool_huge_page(h, nodes_allowed, 0))
1448 while (count < persistent_huge_pages(h)) {
1449 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1453 ret = persistent_huge_pages(h);
1454 spin_unlock(&hugetlb_lock);
1458 #define HSTATE_ATTR_RO(_name) \
1459 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1461 #define HSTATE_ATTR(_name) \
1462 static struct kobj_attribute _name##_attr = \
1463 __ATTR(_name, 0644, _name##_show, _name##_store)
1465 static struct kobject *hugepages_kobj;
1466 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1468 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1470 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1474 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1475 if (hstate_kobjs[i] == kobj) {
1477 *nidp = NUMA_NO_NODE;
1481 return kobj_to_node_hstate(kobj, nidp);
1484 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1485 struct kobj_attribute *attr, char *buf)
1488 unsigned long nr_huge_pages;
1491 h = kobj_to_hstate(kobj, &nid);
1492 if (nid == NUMA_NO_NODE)
1493 nr_huge_pages = h->nr_huge_pages;
1495 nr_huge_pages = h->nr_huge_pages_node[nid];
1497 return sprintf(buf, "%lu\n", nr_huge_pages);
1500 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1501 struct kobject *kobj, struct kobj_attribute *attr,
1502 const char *buf, size_t len)
1506 unsigned long count;
1508 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1510 err = kstrtoul(buf, 10, &count);
1514 h = kobj_to_hstate(kobj, &nid);
1515 if (h->order >= MAX_ORDER) {
1520 if (nid == NUMA_NO_NODE) {
1522 * global hstate attribute
1524 if (!(obey_mempolicy &&
1525 init_nodemask_of_mempolicy(nodes_allowed))) {
1526 NODEMASK_FREE(nodes_allowed);
1527 nodes_allowed = &node_states[N_MEMORY];
1529 } else if (nodes_allowed) {
1531 * per node hstate attribute: adjust count to global,
1532 * but restrict alloc/free to the specified node.
1534 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1535 init_nodemask_of_node(nodes_allowed, nid);
1537 nodes_allowed = &node_states[N_MEMORY];
1539 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1541 if (nodes_allowed != &node_states[N_MEMORY])
1542 NODEMASK_FREE(nodes_allowed);
1546 NODEMASK_FREE(nodes_allowed);
1550 static ssize_t nr_hugepages_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_store(struct kobject *kobj,
1557 struct kobj_attribute *attr, const char *buf, size_t len)
1559 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1561 HSTATE_ATTR(nr_hugepages);
1566 * hstate attribute for optionally mempolicy-based constraint on persistent
1567 * huge page alloc/free.
1569 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1570 struct kobj_attribute *attr, char *buf)
1572 return nr_hugepages_show_common(kobj, attr, buf);
1575 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1576 struct kobj_attribute *attr, const char *buf, size_t len)
1578 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1580 HSTATE_ATTR(nr_hugepages_mempolicy);
1584 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1585 struct kobj_attribute *attr, char *buf)
1587 struct hstate *h = kobj_to_hstate(kobj, NULL);
1588 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1591 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1592 struct kobj_attribute *attr, const char *buf, size_t count)
1595 unsigned long input;
1596 struct hstate *h = kobj_to_hstate(kobj, NULL);
1598 if (h->order >= MAX_ORDER)
1601 err = kstrtoul(buf, 10, &input);
1605 spin_lock(&hugetlb_lock);
1606 h->nr_overcommit_huge_pages = input;
1607 spin_unlock(&hugetlb_lock);
1611 HSTATE_ATTR(nr_overcommit_hugepages);
1613 static ssize_t free_hugepages_show(struct kobject *kobj,
1614 struct kobj_attribute *attr, char *buf)
1617 unsigned long free_huge_pages;
1620 h = kobj_to_hstate(kobj, &nid);
1621 if (nid == NUMA_NO_NODE)
1622 free_huge_pages = h->free_huge_pages;
1624 free_huge_pages = h->free_huge_pages_node[nid];
1626 return sprintf(buf, "%lu\n", free_huge_pages);
1628 HSTATE_ATTR_RO(free_hugepages);
1630 static ssize_t resv_hugepages_show(struct kobject *kobj,
1631 struct kobj_attribute *attr, char *buf)
1633 struct hstate *h = kobj_to_hstate(kobj, NULL);
1634 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1636 HSTATE_ATTR_RO(resv_hugepages);
1638 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1639 struct kobj_attribute *attr, char *buf)
1642 unsigned long surplus_huge_pages;
1645 h = kobj_to_hstate(kobj, &nid);
1646 if (nid == NUMA_NO_NODE)
1647 surplus_huge_pages = h->surplus_huge_pages;
1649 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1651 return sprintf(buf, "%lu\n", surplus_huge_pages);
1653 HSTATE_ATTR_RO(surplus_hugepages);
1655 static struct attribute *hstate_attrs[] = {
1656 &nr_hugepages_attr.attr,
1657 &nr_overcommit_hugepages_attr.attr,
1658 &free_hugepages_attr.attr,
1659 &resv_hugepages_attr.attr,
1660 &surplus_hugepages_attr.attr,
1662 &nr_hugepages_mempolicy_attr.attr,
1667 static struct attribute_group hstate_attr_group = {
1668 .attrs = hstate_attrs,
1671 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1672 struct kobject **hstate_kobjs,
1673 struct attribute_group *hstate_attr_group)
1676 int hi = hstate_index(h);
1678 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1679 if (!hstate_kobjs[hi])
1682 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1684 kobject_put(hstate_kobjs[hi]);
1689 static void __init hugetlb_sysfs_init(void)
1694 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1695 if (!hugepages_kobj)
1698 for_each_hstate(h) {
1699 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1700 hstate_kobjs, &hstate_attr_group);
1702 pr_err("Hugetlb: Unable to add hstate %s", h->name);
1709 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1710 * with node devices in node_devices[] using a parallel array. The array
1711 * index of a node device or _hstate == node id.
1712 * This is here to avoid any static dependency of the node device driver, in
1713 * the base kernel, on the hugetlb module.
1715 struct node_hstate {
1716 struct kobject *hugepages_kobj;
1717 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1719 struct node_hstate node_hstates[MAX_NUMNODES];
1722 * A subset of global hstate attributes for node devices
1724 static struct attribute *per_node_hstate_attrs[] = {
1725 &nr_hugepages_attr.attr,
1726 &free_hugepages_attr.attr,
1727 &surplus_hugepages_attr.attr,
1731 static struct attribute_group per_node_hstate_attr_group = {
1732 .attrs = per_node_hstate_attrs,
1736 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1737 * Returns node id via non-NULL nidp.
1739 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1743 for (nid = 0; nid < nr_node_ids; nid++) {
1744 struct node_hstate *nhs = &node_hstates[nid];
1746 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1747 if (nhs->hstate_kobjs[i] == kobj) {
1759 * Unregister hstate attributes from a single node device.
1760 * No-op if no hstate attributes attached.
1762 static void hugetlb_unregister_node(struct node *node)
1765 struct node_hstate *nhs = &node_hstates[node->dev.id];
1767 if (!nhs->hugepages_kobj)
1768 return; /* no hstate attributes */
1770 for_each_hstate(h) {
1771 int idx = hstate_index(h);
1772 if (nhs->hstate_kobjs[idx]) {
1773 kobject_put(nhs->hstate_kobjs[idx]);
1774 nhs->hstate_kobjs[idx] = NULL;
1778 kobject_put(nhs->hugepages_kobj);
1779 nhs->hugepages_kobj = NULL;
1783 * hugetlb module exit: unregister hstate attributes from node devices
1786 static void hugetlb_unregister_all_nodes(void)
1791 * disable node device registrations.
1793 register_hugetlbfs_with_node(NULL, NULL);
1796 * remove hstate attributes from any nodes that have them.
1798 for (nid = 0; nid < nr_node_ids; nid++)
1799 hugetlb_unregister_node(node_devices[nid]);
1803 * Register hstate attributes for a single node device.
1804 * No-op if attributes already registered.
1806 static void hugetlb_register_node(struct node *node)
1809 struct node_hstate *nhs = &node_hstates[node->dev.id];
1812 if (nhs->hugepages_kobj)
1813 return; /* already allocated */
1815 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1817 if (!nhs->hugepages_kobj)
1820 for_each_hstate(h) {
1821 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1823 &per_node_hstate_attr_group);
1825 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1826 h->name, node->dev.id);
1827 hugetlb_unregister_node(node);
1834 * hugetlb init time: register hstate attributes for all registered node
1835 * devices of nodes that have memory. All on-line nodes should have
1836 * registered their associated device by this time.
1838 static void hugetlb_register_all_nodes(void)
1842 for_each_node_state(nid, N_MEMORY) {
1843 struct node *node = node_devices[nid];
1844 if (node->dev.id == nid)
1845 hugetlb_register_node(node);
1849 * Let the node device driver know we're here so it can
1850 * [un]register hstate attributes on node hotplug.
1852 register_hugetlbfs_with_node(hugetlb_register_node,
1853 hugetlb_unregister_node);
1855 #else /* !CONFIG_NUMA */
1857 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1865 static void hugetlb_unregister_all_nodes(void) { }
1867 static void hugetlb_register_all_nodes(void) { }
1871 static void __exit hugetlb_exit(void)
1875 hugetlb_unregister_all_nodes();
1877 for_each_hstate(h) {
1878 kobject_put(hstate_kobjs[hstate_index(h)]);
1881 kobject_put(hugepages_kobj);
1883 module_exit(hugetlb_exit);
1885 static int __init hugetlb_init(void)
1887 /* Some platform decide whether they support huge pages at boot
1888 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1889 * there is no such support
1891 if (HPAGE_SHIFT == 0)
1894 if (!size_to_hstate(default_hstate_size)) {
1895 default_hstate_size = HPAGE_SIZE;
1896 if (!size_to_hstate(default_hstate_size))
1897 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1899 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1900 if (default_hstate_max_huge_pages)
1901 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1903 hugetlb_init_hstates();
1904 gather_bootmem_prealloc();
1907 hugetlb_sysfs_init();
1908 hugetlb_register_all_nodes();
1909 hugetlb_cgroup_file_init();
1913 module_init(hugetlb_init);
1915 /* Should be called on processing a hugepagesz=... option */
1916 void __init hugetlb_add_hstate(unsigned order)
1921 if (size_to_hstate(PAGE_SIZE << order)) {
1922 pr_warning("hugepagesz= specified twice, ignoring\n");
1925 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1927 h = &hstates[hugetlb_max_hstate++];
1929 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1930 h->nr_huge_pages = 0;
1931 h->free_huge_pages = 0;
1932 for (i = 0; i < MAX_NUMNODES; ++i)
1933 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1934 INIT_LIST_HEAD(&h->hugepage_activelist);
1935 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
1936 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
1937 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1938 huge_page_size(h)/1024);
1943 static int __init hugetlb_nrpages_setup(char *s)
1946 static unsigned long *last_mhp;
1949 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1950 * so this hugepages= parameter goes to the "default hstate".
1952 if (!hugetlb_max_hstate)
1953 mhp = &default_hstate_max_huge_pages;
1955 mhp = &parsed_hstate->max_huge_pages;
1957 if (mhp == last_mhp) {
1958 pr_warning("hugepages= specified twice without "
1959 "interleaving hugepagesz=, ignoring\n");
1963 if (sscanf(s, "%lu", mhp) <= 0)
1967 * Global state is always initialized later in hugetlb_init.
1968 * But we need to allocate >= MAX_ORDER hstates here early to still
1969 * use the bootmem allocator.
1971 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
1972 hugetlb_hstate_alloc_pages(parsed_hstate);
1978 __setup("hugepages=", hugetlb_nrpages_setup);
1980 static int __init hugetlb_default_setup(char *s)
1982 default_hstate_size = memparse(s, &s);
1985 __setup("default_hugepagesz=", hugetlb_default_setup);
1987 static unsigned int cpuset_mems_nr(unsigned int *array)
1990 unsigned int nr = 0;
1992 for_each_node_mask(node, cpuset_current_mems_allowed)
1998 #ifdef CONFIG_SYSCTL
1999 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2000 struct ctl_table *table, int write,
2001 void __user *buffer, size_t *length, loff_t *ppos)
2003 struct hstate *h = &default_hstate;
2007 tmp = h->max_huge_pages;
2009 if (write && h->order >= MAX_ORDER)
2013 table->maxlen = sizeof(unsigned long);
2014 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2019 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2020 GFP_KERNEL | __GFP_NORETRY);
2021 if (!(obey_mempolicy &&
2022 init_nodemask_of_mempolicy(nodes_allowed))) {
2023 NODEMASK_FREE(nodes_allowed);
2024 nodes_allowed = &node_states[N_MEMORY];
2026 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2028 if (nodes_allowed != &node_states[N_MEMORY])
2029 NODEMASK_FREE(nodes_allowed);
2035 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2036 void __user *buffer, size_t *length, loff_t *ppos)
2039 return hugetlb_sysctl_handler_common(false, table, write,
2040 buffer, length, ppos);
2044 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2045 void __user *buffer, size_t *length, loff_t *ppos)
2047 return hugetlb_sysctl_handler_common(true, table, write,
2048 buffer, length, ppos);
2050 #endif /* CONFIG_NUMA */
2052 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2053 void __user *buffer,
2054 size_t *length, loff_t *ppos)
2056 proc_dointvec(table, write, buffer, length, ppos);
2057 if (hugepages_treat_as_movable)
2058 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2060 htlb_alloc_mask = GFP_HIGHUSER;
2064 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2065 void __user *buffer,
2066 size_t *length, loff_t *ppos)
2068 struct hstate *h = &default_hstate;
2072 tmp = h->nr_overcommit_huge_pages;
2074 if (write && h->order >= MAX_ORDER)
2078 table->maxlen = sizeof(unsigned long);
2079 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2084 spin_lock(&hugetlb_lock);
2085 h->nr_overcommit_huge_pages = tmp;
2086 spin_unlock(&hugetlb_lock);
2092 #endif /* CONFIG_SYSCTL */
2094 void hugetlb_report_meminfo(struct seq_file *m)
2096 struct hstate *h = &default_hstate;
2098 "HugePages_Total: %5lu\n"
2099 "HugePages_Free: %5lu\n"
2100 "HugePages_Rsvd: %5lu\n"
2101 "HugePages_Surp: %5lu\n"
2102 "Hugepagesize: %8lu kB\n",
2106 h->surplus_huge_pages,
2107 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2110 int hugetlb_report_node_meminfo(int nid, char *buf)
2112 struct hstate *h = &default_hstate;
2114 "Node %d HugePages_Total: %5u\n"
2115 "Node %d HugePages_Free: %5u\n"
2116 "Node %d HugePages_Surp: %5u\n",
2117 nid, h->nr_huge_pages_node[nid],
2118 nid, h->free_huge_pages_node[nid],
2119 nid, h->surplus_huge_pages_node[nid]);
2122 void hugetlb_show_meminfo(void)
2127 for_each_node_state(nid, N_MEMORY)
2129 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2131 h->nr_huge_pages_node[nid],
2132 h->free_huge_pages_node[nid],
2133 h->surplus_huge_pages_node[nid],
2134 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2137 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2138 unsigned long hugetlb_total_pages(void)
2141 unsigned long nr_total_pages = 0;
2144 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2145 return nr_total_pages;
2148 static int hugetlb_acct_memory(struct hstate *h, long delta)
2152 spin_lock(&hugetlb_lock);
2154 * When cpuset is configured, it breaks the strict hugetlb page
2155 * reservation as the accounting is done on a global variable. Such
2156 * reservation is completely rubbish in the presence of cpuset because
2157 * the reservation is not checked against page availability for the
2158 * current cpuset. Application can still potentially OOM'ed by kernel
2159 * with lack of free htlb page in cpuset that the task is in.
2160 * Attempt to enforce strict accounting with cpuset is almost
2161 * impossible (or too ugly) because cpuset is too fluid that
2162 * task or memory node can be dynamically moved between cpusets.
2164 * The change of semantics for shared hugetlb mapping with cpuset is
2165 * undesirable. However, in order to preserve some of the semantics,
2166 * we fall back to check against current free page availability as
2167 * a best attempt and hopefully to minimize the impact of changing
2168 * semantics that cpuset has.
2171 if (gather_surplus_pages(h, delta) < 0)
2174 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2175 return_unused_surplus_pages(h, delta);
2182 return_unused_surplus_pages(h, (unsigned long) -delta);
2185 spin_unlock(&hugetlb_lock);
2189 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2191 struct resv_map *reservations = vma_resv_map(vma);
2194 * This new VMA should share its siblings reservation map if present.
2195 * The VMA will only ever have a valid reservation map pointer where
2196 * it is being copied for another still existing VMA. As that VMA
2197 * has a reference to the reservation map it cannot disappear until
2198 * after this open call completes. It is therefore safe to take a
2199 * new reference here without additional locking.
2202 kref_get(&reservations->refs);
2205 static void resv_map_put(struct vm_area_struct *vma)
2207 struct resv_map *reservations = vma_resv_map(vma);
2211 kref_put(&reservations->refs, resv_map_release);
2214 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2216 struct hstate *h = hstate_vma(vma);
2217 struct resv_map *reservations = vma_resv_map(vma);
2218 struct hugepage_subpool *spool = subpool_vma(vma);
2219 unsigned long reserve;
2220 unsigned long start;
2224 start = vma_hugecache_offset(h, vma, vma->vm_start);
2225 end = vma_hugecache_offset(h, vma, vma->vm_end);
2227 reserve = (end - start) -
2228 region_count(&reservations->regions, start, end);
2233 hugetlb_acct_memory(h, -reserve);
2234 hugepage_subpool_put_pages(spool, reserve);
2240 * We cannot handle pagefaults against hugetlb pages at all. They cause
2241 * handle_mm_fault() to try to instantiate regular-sized pages in the
2242 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2245 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2251 const struct vm_operations_struct hugetlb_vm_ops = {
2252 .fault = hugetlb_vm_op_fault,
2253 .open = hugetlb_vm_op_open,
2254 .close = hugetlb_vm_op_close,
2257 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2263 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2264 vma->vm_page_prot)));
2266 entry = huge_pte_wrprotect(mk_huge_pte(page,
2267 vma->vm_page_prot));
2269 entry = pte_mkyoung(entry);
2270 entry = pte_mkhuge(entry);
2271 entry = arch_make_huge_pte(entry, vma, page, writable);
2276 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2277 unsigned long address, pte_t *ptep)
2281 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2282 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2283 update_mmu_cache(vma, address, ptep);
2287 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2288 struct vm_area_struct *vma)
2290 pte_t *src_pte, *dst_pte, entry;
2291 struct page *ptepage;
2294 struct hstate *h = hstate_vma(vma);
2295 unsigned long sz = huge_page_size(h);
2297 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2299 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2300 src_pte = huge_pte_offset(src, addr);
2303 dst_pte = huge_pte_alloc(dst, addr, sz);
2307 /* If the pagetables are shared don't copy or take references */
2308 if (dst_pte == src_pte)
2311 spin_lock(&dst->page_table_lock);
2312 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2313 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2315 huge_ptep_set_wrprotect(src, addr, src_pte);
2316 entry = huge_ptep_get(src_pte);
2317 ptepage = pte_page(entry);
2319 page_dup_rmap(ptepage);
2320 set_huge_pte_at(dst, addr, dst_pte, entry);
2322 spin_unlock(&src->page_table_lock);
2323 spin_unlock(&dst->page_table_lock);
2331 static int is_hugetlb_entry_migration(pte_t pte)
2335 if (huge_pte_none(pte) || pte_present(pte))
2337 swp = pte_to_swp_entry(pte);
2338 if (non_swap_entry(swp) && is_migration_entry(swp))
2344 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2348 if (huge_pte_none(pte) || pte_present(pte))
2350 swp = pte_to_swp_entry(pte);
2351 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2357 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2358 unsigned long start, unsigned long end,
2359 struct page *ref_page)
2361 int force_flush = 0;
2362 struct mm_struct *mm = vma->vm_mm;
2363 unsigned long address;
2367 struct hstate *h = hstate_vma(vma);
2368 unsigned long sz = huge_page_size(h);
2369 const unsigned long mmun_start = start; /* For mmu_notifiers */
2370 const unsigned long mmun_end = end; /* For mmu_notifiers */
2372 WARN_ON(!is_vm_hugetlb_page(vma));
2373 BUG_ON(start & ~huge_page_mask(h));
2374 BUG_ON(end & ~huge_page_mask(h));
2376 tlb_start_vma(tlb, vma);
2377 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2379 spin_lock(&mm->page_table_lock);
2380 for (address = start; address < end; address += sz) {
2381 ptep = huge_pte_offset(mm, address);
2385 if (huge_pmd_unshare(mm, &address, ptep))
2388 pte = huge_ptep_get(ptep);
2389 if (huge_pte_none(pte))
2393 * HWPoisoned hugepage is already unmapped and dropped reference
2395 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
2396 huge_pte_clear(mm, address, ptep);
2400 page = pte_page(pte);
2402 * If a reference page is supplied, it is because a specific
2403 * page is being unmapped, not a range. Ensure the page we
2404 * are about to unmap is the actual page of interest.
2407 if (page != ref_page)
2411 * Mark the VMA as having unmapped its page so that
2412 * future faults in this VMA will fail rather than
2413 * looking like data was lost
2415 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2418 pte = huge_ptep_get_and_clear(mm, address, ptep);
2419 tlb_remove_tlb_entry(tlb, ptep, address);
2420 if (huge_pte_dirty(pte))
2421 set_page_dirty(page);
2423 page_remove_rmap(page);
2424 force_flush = !__tlb_remove_page(tlb, page);
2427 /* Bail out after unmapping reference page if supplied */
2431 spin_unlock(&mm->page_table_lock);
2433 * mmu_gather ran out of room to batch pages, we break out of
2434 * the PTE lock to avoid doing the potential expensive TLB invalidate
2435 * and page-free while holding it.
2440 if (address < end && !ref_page)
2443 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2444 tlb_end_vma(tlb, vma);
2447 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2448 struct vm_area_struct *vma, unsigned long start,
2449 unsigned long end, struct page *ref_page)
2451 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2454 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2455 * test will fail on a vma being torn down, and not grab a page table
2456 * on its way out. We're lucky that the flag has such an appropriate
2457 * name, and can in fact be safely cleared here. We could clear it
2458 * before the __unmap_hugepage_range above, but all that's necessary
2459 * is to clear it before releasing the i_mmap_mutex. This works
2460 * because in the context this is called, the VMA is about to be
2461 * destroyed and the i_mmap_mutex is held.
2463 vma->vm_flags &= ~VM_MAYSHARE;
2466 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2467 unsigned long end, struct page *ref_page)
2469 struct mm_struct *mm;
2470 struct mmu_gather tlb;
2474 tlb_gather_mmu(&tlb, mm, start, end);
2475 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2476 tlb_finish_mmu(&tlb, start, end);
2480 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2481 * mappping it owns the reserve page for. The intention is to unmap the page
2482 * from other VMAs and let the children be SIGKILLed if they are faulting the
2485 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2486 struct page *page, unsigned long address)
2488 struct hstate *h = hstate_vma(vma);
2489 struct vm_area_struct *iter_vma;
2490 struct address_space *mapping;
2494 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2495 * from page cache lookup which is in HPAGE_SIZE units.
2497 address = address & huge_page_mask(h);
2498 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2500 mapping = file_inode(vma->vm_file)->i_mapping;
2503 * Take the mapping lock for the duration of the table walk. As
2504 * this mapping should be shared between all the VMAs,
2505 * __unmap_hugepage_range() is called as the lock is already held
2507 mutex_lock(&mapping->i_mmap_mutex);
2508 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2509 /* Do not unmap the current VMA */
2510 if (iter_vma == vma)
2514 * Unmap the page from other VMAs without their own reserves.
2515 * They get marked to be SIGKILLed if they fault in these
2516 * areas. This is because a future no-page fault on this VMA
2517 * could insert a zeroed page instead of the data existing
2518 * from the time of fork. This would look like data corruption
2520 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2521 unmap_hugepage_range(iter_vma, address,
2522 address + huge_page_size(h), page);
2524 mutex_unlock(&mapping->i_mmap_mutex);
2530 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2531 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2532 * cannot race with other handlers or page migration.
2533 * Keep the pte_same checks anyway to make transition from the mutex easier.
2535 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2536 unsigned long address, pte_t *ptep, pte_t pte,
2537 struct page *pagecache_page)
2539 struct hstate *h = hstate_vma(vma);
2540 struct page *old_page, *new_page;
2541 int outside_reserve = 0;
2542 unsigned long mmun_start; /* For mmu_notifiers */
2543 unsigned long mmun_end; /* For mmu_notifiers */
2545 old_page = pte_page(pte);
2548 /* If no-one else is actually using this page, avoid the copy
2549 * and just make the page writable */
2550 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2551 page_move_anon_rmap(old_page, vma, address);
2552 set_huge_ptep_writable(vma, address, ptep);
2557 * If the process that created a MAP_PRIVATE mapping is about to
2558 * perform a COW due to a shared page count, attempt to satisfy
2559 * the allocation without using the existing reserves. The pagecache
2560 * page is used to determine if the reserve at this address was
2561 * consumed or not. If reserves were used, a partial faulted mapping
2562 * at the time of fork() could consume its reserves on COW instead
2563 * of the full address range.
2565 if (!(vma->vm_flags & VM_MAYSHARE) &&
2566 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2567 old_page != pagecache_page)
2568 outside_reserve = 1;
2570 page_cache_get(old_page);
2572 /* Drop page_table_lock as buddy allocator may be called */
2573 spin_unlock(&mm->page_table_lock);
2574 new_page = alloc_huge_page(vma, address, outside_reserve);
2576 if (IS_ERR(new_page)) {
2577 long err = PTR_ERR(new_page);
2578 page_cache_release(old_page);
2581 * If a process owning a MAP_PRIVATE mapping fails to COW,
2582 * it is due to references held by a child and an insufficient
2583 * huge page pool. To guarantee the original mappers
2584 * reliability, unmap the page from child processes. The child
2585 * may get SIGKILLed if it later faults.
2587 if (outside_reserve) {
2588 BUG_ON(huge_pte_none(pte));
2589 if (unmap_ref_private(mm, vma, old_page, address)) {
2590 BUG_ON(huge_pte_none(pte));
2591 spin_lock(&mm->page_table_lock);
2592 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2593 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2594 goto retry_avoidcopy;
2596 * race occurs while re-acquiring page_table_lock, and
2604 /* Caller expects lock to be held */
2605 spin_lock(&mm->page_table_lock);
2607 return VM_FAULT_OOM;
2609 return VM_FAULT_SIGBUS;
2613 * When the original hugepage is shared one, it does not have
2614 * anon_vma prepared.
2616 if (unlikely(anon_vma_prepare(vma))) {
2617 page_cache_release(new_page);
2618 page_cache_release(old_page);
2619 /* Caller expects lock to be held */
2620 spin_lock(&mm->page_table_lock);
2621 return VM_FAULT_OOM;
2624 copy_user_huge_page(new_page, old_page, address, vma,
2625 pages_per_huge_page(h));
2626 __SetPageUptodate(new_page);
2628 mmun_start = address & huge_page_mask(h);
2629 mmun_end = mmun_start + huge_page_size(h);
2630 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2632 * Retake the page_table_lock to check for racing updates
2633 * before the page tables are altered
2635 spin_lock(&mm->page_table_lock);
2636 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2637 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2639 huge_ptep_clear_flush(vma, address, ptep);
2640 set_huge_pte_at(mm, address, ptep,
2641 make_huge_pte(vma, new_page, 1));
2642 page_remove_rmap(old_page);
2643 hugepage_add_new_anon_rmap(new_page, vma, address);
2644 /* Make the old page be freed below */
2645 new_page = old_page;
2647 spin_unlock(&mm->page_table_lock);
2648 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2649 /* Caller expects lock to be held */
2650 spin_lock(&mm->page_table_lock);
2651 page_cache_release(new_page);
2652 page_cache_release(old_page);
2656 /* Return the pagecache page at a given address within a VMA */
2657 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2658 struct vm_area_struct *vma, unsigned long address)
2660 struct address_space *mapping;
2663 mapping = vma->vm_file->f_mapping;
2664 idx = vma_hugecache_offset(h, vma, address);
2666 return find_lock_page(mapping, idx);
2670 * Return whether there is a pagecache page to back given address within VMA.
2671 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2673 static bool hugetlbfs_pagecache_present(struct hstate *h,
2674 struct vm_area_struct *vma, unsigned long address)
2676 struct address_space *mapping;
2680 mapping = vma->vm_file->f_mapping;
2681 idx = vma_hugecache_offset(h, vma, address);
2683 page = find_get_page(mapping, idx);
2686 return page != NULL;
2689 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2690 unsigned long address, pte_t *ptep, unsigned int flags)
2692 struct hstate *h = hstate_vma(vma);
2693 int ret = VM_FAULT_SIGBUS;
2698 struct address_space *mapping;
2702 * Currently, we are forced to kill the process in the event the
2703 * original mapper has unmapped pages from the child due to a failed
2704 * COW. Warn that such a situation has occurred as it may not be obvious
2706 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2707 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2712 mapping = vma->vm_file->f_mapping;
2713 idx = vma_hugecache_offset(h, vma, address);
2716 * Use page lock to guard against racing truncation
2717 * before we get page_table_lock.
2720 page = find_lock_page(mapping, idx);
2722 size = i_size_read(mapping->host) >> huge_page_shift(h);
2725 page = alloc_huge_page(vma, address, 0);
2727 ret = PTR_ERR(page);
2731 ret = VM_FAULT_SIGBUS;
2734 clear_huge_page(page, address, pages_per_huge_page(h));
2735 __SetPageUptodate(page);
2737 if (vma->vm_flags & VM_MAYSHARE) {
2739 struct inode *inode = mapping->host;
2741 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2749 spin_lock(&inode->i_lock);
2750 inode->i_blocks += blocks_per_huge_page(h);
2751 spin_unlock(&inode->i_lock);
2754 if (unlikely(anon_vma_prepare(vma))) {
2756 goto backout_unlocked;
2762 * If memory error occurs between mmap() and fault, some process
2763 * don't have hwpoisoned swap entry for errored virtual address.
2764 * So we need to block hugepage fault by PG_hwpoison bit check.
2766 if (unlikely(PageHWPoison(page))) {
2767 ret = VM_FAULT_HWPOISON |
2768 VM_FAULT_SET_HINDEX(hstate_index(h));
2769 goto backout_unlocked;
2774 * If we are going to COW a private mapping later, we examine the
2775 * pending reservations for this page now. This will ensure that
2776 * any allocations necessary to record that reservation occur outside
2779 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2780 if (vma_needs_reservation(h, vma, address) < 0) {
2782 goto backout_unlocked;
2785 spin_lock(&mm->page_table_lock);
2786 size = i_size_read(mapping->host) >> huge_page_shift(h);
2791 if (!huge_pte_none(huge_ptep_get(ptep)))
2795 hugepage_add_new_anon_rmap(page, vma, address);
2797 page_dup_rmap(page);
2798 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2799 && (vma->vm_flags & VM_SHARED)));
2800 set_huge_pte_at(mm, address, ptep, new_pte);
2802 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2803 /* Optimization, do the COW without a second fault */
2804 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2807 spin_unlock(&mm->page_table_lock);
2813 spin_unlock(&mm->page_table_lock);
2820 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2821 unsigned long address, unsigned int flags)
2826 struct page *page = NULL;
2827 struct page *pagecache_page = NULL;
2828 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2829 struct hstate *h = hstate_vma(vma);
2831 address &= huge_page_mask(h);
2833 ptep = huge_pte_offset(mm, address);
2835 entry = huge_ptep_get(ptep);
2836 if (unlikely(is_hugetlb_entry_migration(entry))) {
2837 migration_entry_wait_huge(mm, ptep);
2839 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2840 return VM_FAULT_HWPOISON_LARGE |
2841 VM_FAULT_SET_HINDEX(hstate_index(h));
2844 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2846 return VM_FAULT_OOM;
2849 * Serialize hugepage allocation and instantiation, so that we don't
2850 * get spurious allocation failures if two CPUs race to instantiate
2851 * the same page in the page cache.
2853 mutex_lock(&hugetlb_instantiation_mutex);
2854 entry = huge_ptep_get(ptep);
2855 if (huge_pte_none(entry)) {
2856 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2863 * If we are going to COW the mapping later, we examine the pending
2864 * reservations for this page now. This will ensure that any
2865 * allocations necessary to record that reservation occur outside the
2866 * spinlock. For private mappings, we also lookup the pagecache
2867 * page now as it is used to determine if a reservation has been
2870 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
2871 if (vma_needs_reservation(h, vma, address) < 0) {
2876 if (!(vma->vm_flags & VM_MAYSHARE))
2877 pagecache_page = hugetlbfs_pagecache_page(h,
2882 * hugetlb_cow() requires page locks of pte_page(entry) and
2883 * pagecache_page, so here we need take the former one
2884 * when page != pagecache_page or !pagecache_page.
2885 * Note that locking order is always pagecache_page -> page,
2886 * so no worry about deadlock.
2888 page = pte_page(entry);
2890 if (page != pagecache_page)
2893 spin_lock(&mm->page_table_lock);
2894 /* Check for a racing update before calling hugetlb_cow */
2895 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2896 goto out_page_table_lock;
2899 if (flags & FAULT_FLAG_WRITE) {
2900 if (!huge_pte_write(entry)) {
2901 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2903 goto out_page_table_lock;
2905 entry = huge_pte_mkdirty(entry);
2907 entry = pte_mkyoung(entry);
2908 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2909 flags & FAULT_FLAG_WRITE))
2910 update_mmu_cache(vma, address, ptep);
2912 out_page_table_lock:
2913 spin_unlock(&mm->page_table_lock);
2915 if (pagecache_page) {
2916 unlock_page(pagecache_page);
2917 put_page(pagecache_page);
2919 if (page != pagecache_page)
2924 mutex_unlock(&hugetlb_instantiation_mutex);
2929 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2930 struct page **pages, struct vm_area_struct **vmas,
2931 unsigned long *position, unsigned long *nr_pages,
2932 long i, unsigned int flags)
2934 unsigned long pfn_offset;
2935 unsigned long vaddr = *position;
2936 unsigned long remainder = *nr_pages;
2937 struct hstate *h = hstate_vma(vma);
2939 spin_lock(&mm->page_table_lock);
2940 while (vaddr < vma->vm_end && remainder) {
2946 * Some archs (sparc64, sh*) have multiple pte_ts to
2947 * each hugepage. We have to make sure we get the
2948 * first, for the page indexing below to work.
2950 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2951 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2954 * When coredumping, it suits get_dump_page if we just return
2955 * an error where there's an empty slot with no huge pagecache
2956 * to back it. This way, we avoid allocating a hugepage, and
2957 * the sparse dumpfile avoids allocating disk blocks, but its
2958 * huge holes still show up with zeroes where they need to be.
2960 if (absent && (flags & FOLL_DUMP) &&
2961 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2967 * We need call hugetlb_fault for both hugepages under migration
2968 * (in which case hugetlb_fault waits for the migration,) and
2969 * hwpoisoned hugepages (in which case we need to prevent the
2970 * caller from accessing to them.) In order to do this, we use
2971 * here is_swap_pte instead of is_hugetlb_entry_migration and
2972 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
2973 * both cases, and because we can't follow correct pages
2974 * directly from any kind of swap entries.
2976 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
2977 ((flags & FOLL_WRITE) &&
2978 !huge_pte_write(huge_ptep_get(pte)))) {
2981 spin_unlock(&mm->page_table_lock);
2982 ret = hugetlb_fault(mm, vma, vaddr,
2983 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2984 spin_lock(&mm->page_table_lock);
2985 if (!(ret & VM_FAULT_ERROR))
2992 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2993 page = pte_page(huge_ptep_get(pte));
2996 pages[i] = mem_map_offset(page, pfn_offset);
3007 if (vaddr < vma->vm_end && remainder &&
3008 pfn_offset < pages_per_huge_page(h)) {
3010 * We use pfn_offset to avoid touching the pageframes
3011 * of this compound page.
3016 spin_unlock(&mm->page_table_lock);
3017 *nr_pages = remainder;
3020 return i ? i : -EFAULT;
3023 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3024 unsigned long address, unsigned long end, pgprot_t newprot)
3026 struct mm_struct *mm = vma->vm_mm;
3027 unsigned long start = address;
3030 struct hstate *h = hstate_vma(vma);
3031 unsigned long pages = 0;
3033 BUG_ON(address >= end);
3034 flush_cache_range(vma, address, end);
3036 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3037 spin_lock(&mm->page_table_lock);
3038 for (; address < end; address += huge_page_size(h)) {
3039 ptep = huge_pte_offset(mm, address);
3042 if (huge_pmd_unshare(mm, &address, ptep)) {
3046 if (!huge_pte_none(huge_ptep_get(ptep))) {
3047 pte = huge_ptep_get_and_clear(mm, address, ptep);
3048 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3049 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3050 set_huge_pte_at(mm, address, ptep, pte);
3054 spin_unlock(&mm->page_table_lock);
3056 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3057 * may have cleared our pud entry and done put_page on the page table:
3058 * once we release i_mmap_mutex, another task can do the final put_page
3059 * and that page table be reused and filled with junk.
3061 flush_tlb_range(vma, start, end);
3062 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3064 return pages << h->order;
3067 int hugetlb_reserve_pages(struct inode *inode,
3069 struct vm_area_struct *vma,
3070 vm_flags_t vm_flags)
3073 struct hstate *h = hstate_inode(inode);
3074 struct hugepage_subpool *spool = subpool_inode(inode);
3077 * Only apply hugepage reservation if asked. At fault time, an
3078 * attempt will be made for VM_NORESERVE to allocate a page
3079 * without using reserves
3081 if (vm_flags & VM_NORESERVE)
3085 * Shared mappings base their reservation on the number of pages that
3086 * are already allocated on behalf of the file. Private mappings need
3087 * to reserve the full area even if read-only as mprotect() may be
3088 * called to make the mapping read-write. Assume !vma is a shm mapping
3090 if (!vma || vma->vm_flags & VM_MAYSHARE)
3091 chg = region_chg(&inode->i_mapping->private_list, from, to);
3093 struct resv_map *resv_map = resv_map_alloc();
3099 set_vma_resv_map(vma, resv_map);
3100 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3108 /* There must be enough pages in the subpool for the mapping */
3109 if (hugepage_subpool_get_pages(spool, chg)) {
3115 * Check enough hugepages are available for the reservation.
3116 * Hand the pages back to the subpool if there are not
3118 ret = hugetlb_acct_memory(h, chg);
3120 hugepage_subpool_put_pages(spool, chg);
3125 * Account for the reservations made. Shared mappings record regions
3126 * that have reservations as they are shared by multiple VMAs.
3127 * When the last VMA disappears, the region map says how much
3128 * the reservation was and the page cache tells how much of
3129 * the reservation was consumed. Private mappings are per-VMA and
3130 * only the consumed reservations are tracked. When the VMA
3131 * disappears, the original reservation is the VMA size and the
3132 * consumed reservations are stored in the map. Hence, nothing
3133 * else has to be done for private mappings here
3135 if (!vma || vma->vm_flags & VM_MAYSHARE)
3136 region_add(&inode->i_mapping->private_list, from, to);
3144 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3146 struct hstate *h = hstate_inode(inode);
3147 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3148 struct hugepage_subpool *spool = subpool_inode(inode);
3150 spin_lock(&inode->i_lock);
3151 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3152 spin_unlock(&inode->i_lock);
3154 hugepage_subpool_put_pages(spool, (chg - freed));
3155 hugetlb_acct_memory(h, -(chg - freed));
3158 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3159 static unsigned long page_table_shareable(struct vm_area_struct *svma,
3160 struct vm_area_struct *vma,
3161 unsigned long addr, pgoff_t idx)
3163 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3165 unsigned long sbase = saddr & PUD_MASK;
3166 unsigned long s_end = sbase + PUD_SIZE;
3168 /* Allow segments to share if only one is marked locked */
3169 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3170 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3173 * match the virtual addresses, permission and the alignment of the
3176 if (pmd_index(addr) != pmd_index(saddr) ||
3177 vm_flags != svm_flags ||
3178 sbase < svma->vm_start || svma->vm_end < s_end)
3184 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3186 unsigned long base = addr & PUD_MASK;
3187 unsigned long end = base + PUD_SIZE;
3190 * check on proper vm_flags and page table alignment
3192 if (vma->vm_flags & VM_MAYSHARE &&
3193 vma->vm_start <= base && end <= vma->vm_end)
3199 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3200 * and returns the corresponding pte. While this is not necessary for the
3201 * !shared pmd case because we can allocate the pmd later as well, it makes the
3202 * code much cleaner. pmd allocation is essential for the shared case because
3203 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3204 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3205 * bad pmd for sharing.
3207 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3209 struct vm_area_struct *vma = find_vma(mm, addr);
3210 struct address_space *mapping = vma->vm_file->f_mapping;
3211 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3213 struct vm_area_struct *svma;
3214 unsigned long saddr;
3218 if (!vma_shareable(vma, addr))
3219 return (pte_t *)pmd_alloc(mm, pud, addr);
3221 mutex_lock(&mapping->i_mmap_mutex);
3222 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3226 saddr = page_table_shareable(svma, vma, addr, idx);
3228 spte = huge_pte_offset(svma->vm_mm, saddr);
3230 get_page(virt_to_page(spte));
3239 spin_lock(&mm->page_table_lock);
3241 pud_populate(mm, pud,
3242 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3244 put_page(virt_to_page(spte));
3245 spin_unlock(&mm->page_table_lock);
3247 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3248 mutex_unlock(&mapping->i_mmap_mutex);
3253 * unmap huge page backed by shared pte.
3255 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3256 * indicated by page_count > 1, unmap is achieved by clearing pud and
3257 * decrementing the ref count. If count == 1, the pte page is not shared.
3259 * called with vma->vm_mm->page_table_lock held.
3261 * returns: 1 successfully unmapped a shared pte page
3262 * 0 the underlying pte page is not shared, or it is the last user
3264 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3266 pgd_t *pgd = pgd_offset(mm, *addr);
3267 pud_t *pud = pud_offset(pgd, *addr);
3269 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3270 if (page_count(virt_to_page(ptep)) == 1)
3274 put_page(virt_to_page(ptep));
3275 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3278 #define want_pmd_share() (1)
3279 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3280 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3284 #define want_pmd_share() (0)
3285 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3287 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3288 pte_t *huge_pte_alloc(struct mm_struct *mm,
3289 unsigned long addr, unsigned long sz)
3295 pgd = pgd_offset(mm, addr);
3296 pud = pud_alloc(mm, pgd, addr);
3298 if (sz == PUD_SIZE) {
3301 BUG_ON(sz != PMD_SIZE);
3302 if (want_pmd_share() && pud_none(*pud))
3303 pte = huge_pmd_share(mm, addr, pud);
3305 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3308 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3313 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3319 pgd = pgd_offset(mm, addr);
3320 if (pgd_present(*pgd)) {
3321 pud = pud_offset(pgd, addr);
3322 if (pud_present(*pud)) {
3324 return (pte_t *)pud;
3325 pmd = pmd_offset(pud, addr);
3328 return (pte_t *) pmd;
3332 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3333 pmd_t *pmd, int write)
3337 page = pte_page(*(pte_t *)pmd);
3339 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
3344 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3345 pud_t *pud, int write)
3349 page = pte_page(*(pte_t *)pud);
3351 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
3355 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3357 /* Can be overriden by architectures */
3358 __attribute__((weak)) struct page *
3359 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3360 pud_t *pud, int write)
3366 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3368 #ifdef CONFIG_MEMORY_FAILURE
3370 /* Should be called in hugetlb_lock */
3371 static int is_hugepage_on_freelist(struct page *hpage)
3375 struct hstate *h = page_hstate(hpage);
3376 int nid = page_to_nid(hpage);
3378 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3385 * This function is called from memory failure code.
3386 * Assume the caller holds page lock of the head page.
3388 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3390 struct hstate *h = page_hstate(hpage);
3391 int nid = page_to_nid(hpage);
3394 spin_lock(&hugetlb_lock);
3395 if (is_hugepage_on_freelist(hpage)) {
3397 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3398 * but dangling hpage->lru can trigger list-debug warnings
3399 * (this happens when we call unpoison_memory() on it),
3400 * so let it point to itself with list_del_init().
3402 list_del_init(&hpage->lru);
3403 set_page_refcounted(hpage);
3404 h->free_huge_pages--;
3405 h->free_huge_pages_node[nid]--;
3408 spin_unlock(&hugetlb_lock);