2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/module.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/rmap.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
26 #include <asm/pgtable.h>
30 #include <linux/hugetlb.h>
31 #include <linux/hugetlb_cgroup.h>
32 #include <linux/node.h>
35 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
36 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
37 unsigned long hugepages_treat_as_movable;
39 int hugetlb_max_hstate __read_mostly;
40 unsigned int default_hstate_idx;
41 struct hstate hstates[HUGE_MAX_HSTATE];
43 __initdata LIST_HEAD(huge_boot_pages);
45 /* for command line parsing */
46 static struct hstate * __initdata parsed_hstate;
47 static unsigned long __initdata default_hstate_max_huge_pages;
48 static unsigned long __initdata default_hstate_size;
51 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
53 DEFINE_SPINLOCK(hugetlb_lock);
55 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
57 bool free = (spool->count == 0) && (spool->used_hpages == 0);
59 spin_unlock(&spool->lock);
61 /* If no pages are used, and no other handles to the subpool
62 * remain, free the subpool the subpool remain */
67 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
69 struct hugepage_subpool *spool;
71 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
75 spin_lock_init(&spool->lock);
77 spool->max_hpages = nr_blocks;
78 spool->used_hpages = 0;
83 void hugepage_put_subpool(struct hugepage_subpool *spool)
85 spin_lock(&spool->lock);
86 BUG_ON(!spool->count);
88 unlock_or_release_subpool(spool);
91 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
99 spin_lock(&spool->lock);
100 if ((spool->used_hpages + delta) <= spool->max_hpages) {
101 spool->used_hpages += delta;
105 spin_unlock(&spool->lock);
110 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
116 spin_lock(&spool->lock);
117 spool->used_hpages -= delta;
118 /* If hugetlbfs_put_super couldn't free spool due to
119 * an outstanding quota reference, free it now. */
120 unlock_or_release_subpool(spool);
123 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
125 return HUGETLBFS_SB(inode->i_sb)->spool;
128 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
130 return subpool_inode(file_inode(vma->vm_file));
134 * Region tracking -- allows tracking of reservations and instantiated pages
135 * across the pages in a mapping.
137 * The region data structures are protected by a combination of the mmap_sem
138 * and the hugetlb_instantiation_mutex. To access or modify a region the caller
139 * must either hold the mmap_sem for write, or the mmap_sem for read and
140 * the hugetlb_instantiation_mutex:
142 * down_write(&mm->mmap_sem);
144 * down_read(&mm->mmap_sem);
145 * mutex_lock(&hugetlb_instantiation_mutex);
148 struct list_head link;
153 static long region_add(struct list_head *head, long f, long t)
155 struct file_region *rg, *nrg, *trg;
157 /* Locate the region we are either in or before. */
158 list_for_each_entry(rg, head, link)
162 /* Round our left edge to the current segment if it encloses us. */
166 /* Check for and consume any regions we now overlap with. */
168 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
169 if (&rg->link == head)
174 /* If this area reaches higher then extend our area to
175 * include it completely. If this is not the first area
176 * which we intend to reuse, free it. */
189 static long region_chg(struct list_head *head, long f, long t)
191 struct file_region *rg, *nrg;
194 /* Locate the region we are before or in. */
195 list_for_each_entry(rg, head, link)
199 /* If we are below the current region then a new region is required.
200 * Subtle, allocate a new region at the position but make it zero
201 * size such that we can guarantee to record the reservation. */
202 if (&rg->link == head || t < rg->from) {
203 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
208 INIT_LIST_HEAD(&nrg->link);
209 list_add(&nrg->link, rg->link.prev);
214 /* Round our left edge to the current segment if it encloses us. */
219 /* Check for and consume any regions we now overlap with. */
220 list_for_each_entry(rg, rg->link.prev, link) {
221 if (&rg->link == head)
226 /* We overlap with this area, if it extends further than
227 * us then we must extend ourselves. Account for its
228 * existing reservation. */
233 chg -= rg->to - rg->from;
238 static long region_truncate(struct list_head *head, long end)
240 struct file_region *rg, *trg;
243 /* Locate the region we are either in or before. */
244 list_for_each_entry(rg, head, link)
247 if (&rg->link == head)
250 /* If we are in the middle of a region then adjust it. */
251 if (end > rg->from) {
254 rg = list_entry(rg->link.next, typeof(*rg), link);
257 /* Drop any remaining regions. */
258 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
259 if (&rg->link == head)
261 chg += rg->to - rg->from;
268 static long region_count(struct list_head *head, long f, long t)
270 struct file_region *rg;
273 /* Locate each segment we overlap with, and count that overlap. */
274 list_for_each_entry(rg, head, link) {
283 seg_from = max(rg->from, f);
284 seg_to = min(rg->to, t);
286 chg += seg_to - seg_from;
293 * Convert the address within this vma to the page offset within
294 * the mapping, in pagecache page units; huge pages here.
296 static pgoff_t vma_hugecache_offset(struct hstate *h,
297 struct vm_area_struct *vma, unsigned long address)
299 return ((address - vma->vm_start) >> huge_page_shift(h)) +
300 (vma->vm_pgoff >> huge_page_order(h));
303 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
304 unsigned long address)
306 return vma_hugecache_offset(hstate_vma(vma), vma, address);
310 * Return the size of the pages allocated when backing a VMA. In the majority
311 * cases this will be same size as used by the page table entries.
313 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
315 struct hstate *hstate;
317 if (!is_vm_hugetlb_page(vma))
320 hstate = hstate_vma(vma);
322 return 1UL << huge_page_shift(hstate);
324 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
327 * Return the page size being used by the MMU to back a VMA. In the majority
328 * of cases, the page size used by the kernel matches the MMU size. On
329 * architectures where it differs, an architecture-specific version of this
330 * function is required.
332 #ifndef vma_mmu_pagesize
333 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
335 return vma_kernel_pagesize(vma);
340 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
341 * bits of the reservation map pointer, which are always clear due to
344 #define HPAGE_RESV_OWNER (1UL << 0)
345 #define HPAGE_RESV_UNMAPPED (1UL << 1)
346 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
349 * These helpers are used to track how many pages are reserved for
350 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
351 * is guaranteed to have their future faults succeed.
353 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
354 * the reserve counters are updated with the hugetlb_lock held. It is safe
355 * to reset the VMA at fork() time as it is not in use yet and there is no
356 * chance of the global counters getting corrupted as a result of the values.
358 * The private mapping reservation is represented in a subtly different
359 * manner to a shared mapping. A shared mapping has a region map associated
360 * with the underlying file, this region map represents the backing file
361 * pages which have ever had a reservation assigned which this persists even
362 * after the page is instantiated. A private mapping has a region map
363 * associated with the original mmap which is attached to all VMAs which
364 * reference it, this region map represents those offsets which have consumed
365 * reservation ie. where pages have been instantiated.
367 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
369 return (unsigned long)vma->vm_private_data;
372 static void set_vma_private_data(struct vm_area_struct *vma,
375 vma->vm_private_data = (void *)value;
380 struct list_head regions;
383 static struct resv_map *resv_map_alloc(void)
385 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
389 kref_init(&resv_map->refs);
390 INIT_LIST_HEAD(&resv_map->regions);
395 static void resv_map_release(struct kref *ref)
397 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
399 /* Clear out any active regions before we release the map. */
400 region_truncate(&resv_map->regions, 0);
404 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
406 VM_BUG_ON(!is_vm_hugetlb_page(vma));
407 if (!(vma->vm_flags & VM_MAYSHARE))
408 return (struct resv_map *)(get_vma_private_data(vma) &
413 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
415 VM_BUG_ON(!is_vm_hugetlb_page(vma));
416 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
418 set_vma_private_data(vma, (get_vma_private_data(vma) &
419 HPAGE_RESV_MASK) | (unsigned long)map);
422 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
424 VM_BUG_ON(!is_vm_hugetlb_page(vma));
425 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
427 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
430 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
432 VM_BUG_ON(!is_vm_hugetlb_page(vma));
434 return (get_vma_private_data(vma) & flag) != 0;
437 /* Decrement the reserved pages in the hugepage pool by one */
438 static void decrement_hugepage_resv_vma(struct hstate *h,
439 struct vm_area_struct *vma)
441 if (vma->vm_flags & VM_NORESERVE)
444 if (vma->vm_flags & VM_MAYSHARE) {
445 /* Shared mappings always use reserves */
446 h->resv_huge_pages--;
447 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
449 * Only the process that called mmap() has reserves for
452 h->resv_huge_pages--;
456 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
457 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
459 VM_BUG_ON(!is_vm_hugetlb_page(vma));
460 if (!(vma->vm_flags & VM_MAYSHARE))
461 vma->vm_private_data = (void *)0;
464 /* Returns true if the VMA has associated reserve pages */
465 static int vma_has_reserves(struct vm_area_struct *vma)
467 if (vma->vm_flags & VM_MAYSHARE)
469 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
474 static void copy_gigantic_page(struct page *dst, struct page *src)
477 struct hstate *h = page_hstate(src);
478 struct page *dst_base = dst;
479 struct page *src_base = src;
481 for (i = 0; i < pages_per_huge_page(h); ) {
483 copy_highpage(dst, src);
486 dst = mem_map_next(dst, dst_base, i);
487 src = mem_map_next(src, src_base, i);
491 void copy_huge_page(struct page *dst, struct page *src)
494 struct hstate *h = page_hstate(src);
496 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
497 copy_gigantic_page(dst, src);
502 for (i = 0; i < pages_per_huge_page(h); i++) {
504 copy_highpage(dst + i, src + i);
508 static void enqueue_huge_page(struct hstate *h, struct page *page)
510 int nid = page_to_nid(page);
511 list_move(&page->lru, &h->hugepage_freelists[nid]);
512 h->free_huge_pages++;
513 h->free_huge_pages_node[nid]++;
516 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
520 if (list_empty(&h->hugepage_freelists[nid]))
522 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
523 list_move(&page->lru, &h->hugepage_activelist);
524 set_page_refcounted(page);
525 h->free_huge_pages--;
526 h->free_huge_pages_node[nid]--;
530 static struct page *dequeue_huge_page_vma(struct hstate *h,
531 struct vm_area_struct *vma,
532 unsigned long address, int avoid_reserve)
534 struct page *page = NULL;
535 struct mempolicy *mpol;
536 nodemask_t *nodemask;
537 struct zonelist *zonelist;
540 unsigned int cpuset_mems_cookie;
543 * A child process with MAP_PRIVATE mappings created by their parent
544 * have no page reserves. This check ensures that reservations are
545 * not "stolen". The child may still get SIGKILLed
547 if (!vma_has_reserves(vma) &&
548 h->free_huge_pages - h->resv_huge_pages == 0)
551 /* If reserves cannot be used, ensure enough pages are in the pool */
552 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
556 cpuset_mems_cookie = get_mems_allowed();
557 zonelist = huge_zonelist(vma, address,
558 htlb_alloc_mask, &mpol, &nodemask);
560 for_each_zone_zonelist_nodemask(zone, z, zonelist,
561 MAX_NR_ZONES - 1, nodemask) {
562 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
563 page = dequeue_huge_page_node(h, zone_to_nid(zone));
566 decrement_hugepage_resv_vma(h, vma);
573 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
581 static void update_and_free_page(struct hstate *h, struct page *page)
585 VM_BUG_ON(h->order >= MAX_ORDER);
588 h->nr_huge_pages_node[page_to_nid(page)]--;
589 for (i = 0; i < pages_per_huge_page(h); i++) {
590 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
591 1 << PG_referenced | 1 << PG_dirty |
592 1 << PG_active | 1 << PG_reserved |
593 1 << PG_private | 1 << PG_writeback);
595 VM_BUG_ON(hugetlb_cgroup_from_page(page));
596 set_compound_page_dtor(page, NULL);
597 set_page_refcounted(page);
598 arch_release_hugepage(page);
599 __free_pages(page, huge_page_order(h));
602 struct hstate *size_to_hstate(unsigned long size)
607 if (huge_page_size(h) == size)
613 static void free_huge_page(struct page *page)
616 * Can't pass hstate in here because it is called from the
617 * compound page destructor.
619 struct hstate *h = page_hstate(page);
620 int nid = page_to_nid(page);
621 struct hugepage_subpool *spool =
622 (struct hugepage_subpool *)page_private(page);
624 set_page_private(page, 0);
625 page->mapping = NULL;
626 BUG_ON(page_count(page));
627 BUG_ON(page_mapcount(page));
629 spin_lock(&hugetlb_lock);
630 hugetlb_cgroup_uncharge_page(hstate_index(h),
631 pages_per_huge_page(h), page);
632 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
633 /* remove the page from active list */
634 list_del(&page->lru);
635 update_and_free_page(h, page);
636 h->surplus_huge_pages--;
637 h->surplus_huge_pages_node[nid]--;
639 arch_clear_hugepage_flags(page);
640 enqueue_huge_page(h, page);
642 spin_unlock(&hugetlb_lock);
643 hugepage_subpool_put_pages(spool, 1);
646 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
648 INIT_LIST_HEAD(&page->lru);
649 set_compound_page_dtor(page, free_huge_page);
650 spin_lock(&hugetlb_lock);
651 set_hugetlb_cgroup(page, NULL);
653 h->nr_huge_pages_node[nid]++;
654 spin_unlock(&hugetlb_lock);
655 put_page(page); /* free it into the hugepage allocator */
658 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
661 int nr_pages = 1 << order;
662 struct page *p = page + 1;
664 /* we rely on prep_new_huge_page to set the destructor */
665 set_compound_order(page, order);
667 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
669 set_page_count(p, 0);
670 p->first_page = page;
675 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
676 * transparent huge pages. See the PageTransHuge() documentation for more
679 int PageHuge(struct page *page)
681 compound_page_dtor *dtor;
683 if (!PageCompound(page))
686 page = compound_head(page);
687 dtor = get_compound_page_dtor(page);
689 return dtor == free_huge_page;
691 EXPORT_SYMBOL_GPL(PageHuge);
693 pgoff_t __basepage_index(struct page *page)
695 struct page *page_head = compound_head(page);
696 pgoff_t index = page_index(page_head);
697 unsigned long compound_idx;
699 if (!PageHuge(page_head))
700 return page_index(page);
702 if (compound_order(page_head) >= MAX_ORDER)
703 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
705 compound_idx = page - page_head;
707 return (index << compound_order(page_head)) + compound_idx;
710 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
714 if (h->order >= MAX_ORDER)
717 page = alloc_pages_exact_node(nid,
718 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
719 __GFP_REPEAT|__GFP_NOWARN,
722 if (arch_prepare_hugepage(page)) {
723 __free_pages(page, huge_page_order(h));
726 prep_new_huge_page(h, page, nid);
733 * common helper functions for hstate_next_node_to_{alloc|free}.
734 * We may have allocated or freed a huge page based on a different
735 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
736 * be outside of *nodes_allowed. Ensure that we use an allowed
737 * node for alloc or free.
739 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
741 nid = next_node(nid, *nodes_allowed);
742 if (nid == MAX_NUMNODES)
743 nid = first_node(*nodes_allowed);
744 VM_BUG_ON(nid >= MAX_NUMNODES);
749 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
751 if (!node_isset(nid, *nodes_allowed))
752 nid = next_node_allowed(nid, nodes_allowed);
757 * returns the previously saved node ["this node"] from which to
758 * allocate a persistent huge page for the pool and advance the
759 * next node from which to allocate, handling wrap at end of node
762 static int hstate_next_node_to_alloc(struct hstate *h,
763 nodemask_t *nodes_allowed)
767 VM_BUG_ON(!nodes_allowed);
769 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
770 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
776 * helper for free_pool_huge_page() - return the previously saved
777 * node ["this node"] from which to free a huge page. Advance the
778 * next node id whether or not we find a free huge page to free so
779 * that the next attempt to free addresses the next node.
781 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
785 VM_BUG_ON(!nodes_allowed);
787 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
788 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
793 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
794 for (nr_nodes = nodes_weight(*mask); \
796 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
799 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
800 for (nr_nodes = nodes_weight(*mask); \
802 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
805 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
811 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
812 page = alloc_fresh_huge_page_node(h, node);
820 count_vm_event(HTLB_BUDDY_PGALLOC);
822 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
828 * Free huge page from pool from next node to free.
829 * Attempt to keep persistent huge pages more or less
830 * balanced over allowed nodes.
831 * Called with hugetlb_lock locked.
833 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
839 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
841 * If we're returning unused surplus pages, only examine
842 * nodes with surplus pages.
844 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
845 !list_empty(&h->hugepage_freelists[node])) {
847 list_entry(h->hugepage_freelists[node].next,
849 list_del(&page->lru);
850 h->free_huge_pages--;
851 h->free_huge_pages_node[node]--;
853 h->surplus_huge_pages--;
854 h->surplus_huge_pages_node[node]--;
856 update_and_free_page(h, page);
865 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
870 if (h->order >= MAX_ORDER)
874 * Assume we will successfully allocate the surplus page to
875 * prevent racing processes from causing the surplus to exceed
878 * This however introduces a different race, where a process B
879 * tries to grow the static hugepage pool while alloc_pages() is
880 * called by process A. B will only examine the per-node
881 * counters in determining if surplus huge pages can be
882 * converted to normal huge pages in adjust_pool_surplus(). A
883 * won't be able to increment the per-node counter, until the
884 * lock is dropped by B, but B doesn't drop hugetlb_lock until
885 * no more huge pages can be converted from surplus to normal
886 * state (and doesn't try to convert again). Thus, we have a
887 * case where a surplus huge page exists, the pool is grown, and
888 * the surplus huge page still exists after, even though it
889 * should just have been converted to a normal huge page. This
890 * does not leak memory, though, as the hugepage will be freed
891 * once it is out of use. It also does not allow the counters to
892 * go out of whack in adjust_pool_surplus() as we don't modify
893 * the node values until we've gotten the hugepage and only the
894 * per-node value is checked there.
896 spin_lock(&hugetlb_lock);
897 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
898 spin_unlock(&hugetlb_lock);
902 h->surplus_huge_pages++;
904 spin_unlock(&hugetlb_lock);
906 if (nid == NUMA_NO_NODE)
907 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
908 __GFP_REPEAT|__GFP_NOWARN,
911 page = alloc_pages_exact_node(nid,
912 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
913 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
915 if (page && arch_prepare_hugepage(page)) {
916 __free_pages(page, huge_page_order(h));
920 spin_lock(&hugetlb_lock);
922 INIT_LIST_HEAD(&page->lru);
923 r_nid = page_to_nid(page);
924 set_compound_page_dtor(page, free_huge_page);
925 set_hugetlb_cgroup(page, NULL);
927 * We incremented the global counters already
929 h->nr_huge_pages_node[r_nid]++;
930 h->surplus_huge_pages_node[r_nid]++;
931 __count_vm_event(HTLB_BUDDY_PGALLOC);
934 h->surplus_huge_pages--;
935 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
937 spin_unlock(&hugetlb_lock);
943 * This allocation function is useful in the context where vma is irrelevant.
944 * E.g. soft-offlining uses this function because it only cares physical
945 * address of error page.
947 struct page *alloc_huge_page_node(struct hstate *h, int nid)
951 spin_lock(&hugetlb_lock);
952 page = dequeue_huge_page_node(h, nid);
953 spin_unlock(&hugetlb_lock);
956 page = alloc_buddy_huge_page(h, nid);
962 * Increase the hugetlb pool such that it can accommodate a reservation
965 static int gather_surplus_pages(struct hstate *h, int delta)
967 struct list_head surplus_list;
968 struct page *page, *tmp;
970 int needed, allocated;
971 bool alloc_ok = true;
973 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
975 h->resv_huge_pages += delta;
980 INIT_LIST_HEAD(&surplus_list);
984 spin_unlock(&hugetlb_lock);
985 for (i = 0; i < needed; i++) {
986 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
991 list_add(&page->lru, &surplus_list);
996 * After retaking hugetlb_lock, we need to recalculate 'needed'
997 * because either resv_huge_pages or free_huge_pages may have changed.
999 spin_lock(&hugetlb_lock);
1000 needed = (h->resv_huge_pages + delta) -
1001 (h->free_huge_pages + allocated);
1006 * We were not able to allocate enough pages to
1007 * satisfy the entire reservation so we free what
1008 * we've allocated so far.
1013 * The surplus_list now contains _at_least_ the number of extra pages
1014 * needed to accommodate the reservation. Add the appropriate number
1015 * of pages to the hugetlb pool and free the extras back to the buddy
1016 * allocator. Commit the entire reservation here to prevent another
1017 * process from stealing the pages as they are added to the pool but
1018 * before they are reserved.
1020 needed += allocated;
1021 h->resv_huge_pages += delta;
1024 /* Free the needed pages to the hugetlb pool */
1025 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1029 * This page is now managed by the hugetlb allocator and has
1030 * no users -- drop the buddy allocator's reference.
1032 put_page_testzero(page);
1033 VM_BUG_ON(page_count(page));
1034 enqueue_huge_page(h, page);
1037 spin_unlock(&hugetlb_lock);
1039 /* Free unnecessary surplus pages to the buddy allocator */
1040 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1042 spin_lock(&hugetlb_lock);
1048 * When releasing a hugetlb pool reservation, any surplus pages that were
1049 * allocated to satisfy the reservation must be explicitly freed if they were
1051 * Called with hugetlb_lock held.
1053 static void return_unused_surplus_pages(struct hstate *h,
1054 unsigned long unused_resv_pages)
1056 unsigned long nr_pages;
1058 /* Uncommit the reservation */
1059 h->resv_huge_pages -= unused_resv_pages;
1061 /* Cannot return gigantic pages currently */
1062 if (h->order >= MAX_ORDER)
1065 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1068 * We want to release as many surplus pages as possible, spread
1069 * evenly across all nodes with memory. Iterate across these nodes
1070 * until we can no longer free unreserved surplus pages. This occurs
1071 * when the nodes with surplus pages have no free pages.
1072 * free_pool_huge_page() will balance the the freed pages across the
1073 * on-line nodes with memory and will handle the hstate accounting.
1075 while (nr_pages--) {
1076 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1082 * Determine if the huge page at addr within the vma has an associated
1083 * reservation. Where it does not we will need to logically increase
1084 * reservation and actually increase subpool usage before an allocation
1085 * can occur. Where any new reservation would be required the
1086 * reservation change is prepared, but not committed. Once the page
1087 * has been allocated from the subpool and instantiated the change should
1088 * be committed via vma_commit_reservation. No action is required on
1091 static long vma_needs_reservation(struct hstate *h,
1092 struct vm_area_struct *vma, unsigned long addr)
1094 struct address_space *mapping = vma->vm_file->f_mapping;
1095 struct inode *inode = mapping->host;
1097 if (vma->vm_flags & VM_MAYSHARE) {
1098 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1099 return region_chg(&inode->i_mapping->private_list,
1102 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1107 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1108 struct resv_map *reservations = vma_resv_map(vma);
1110 err = region_chg(&reservations->regions, idx, idx + 1);
1116 static void vma_commit_reservation(struct hstate *h,
1117 struct vm_area_struct *vma, unsigned long addr)
1119 struct address_space *mapping = vma->vm_file->f_mapping;
1120 struct inode *inode = mapping->host;
1122 if (vma->vm_flags & VM_MAYSHARE) {
1123 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1124 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1126 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1127 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1128 struct resv_map *reservations = vma_resv_map(vma);
1130 /* Mark this page used in the map. */
1131 region_add(&reservations->regions, idx, idx + 1);
1135 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1136 unsigned long addr, int avoid_reserve)
1138 struct hugepage_subpool *spool = subpool_vma(vma);
1139 struct hstate *h = hstate_vma(vma);
1143 struct hugetlb_cgroup *h_cg;
1145 idx = hstate_index(h);
1147 * Processes that did not create the mapping will have no
1148 * reserves and will not have accounted against subpool
1149 * limit. Check that the subpool limit can be made before
1150 * satisfying the allocation MAP_NORESERVE mappings may also
1151 * need pages and subpool limit allocated allocated if no reserve
1154 chg = vma_needs_reservation(h, vma, addr);
1156 return ERR_PTR(-ENOMEM);
1158 if (hugepage_subpool_get_pages(spool, chg))
1159 return ERR_PTR(-ENOSPC);
1161 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1163 hugepage_subpool_put_pages(spool, chg);
1164 return ERR_PTR(-ENOSPC);
1166 spin_lock(&hugetlb_lock);
1167 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1169 spin_unlock(&hugetlb_lock);
1170 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1172 hugetlb_cgroup_uncharge_cgroup(idx,
1173 pages_per_huge_page(h),
1175 hugepage_subpool_put_pages(spool, chg);
1176 return ERR_PTR(-ENOSPC);
1178 spin_lock(&hugetlb_lock);
1179 list_move(&page->lru, &h->hugepage_activelist);
1182 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1183 spin_unlock(&hugetlb_lock);
1185 set_page_private(page, (unsigned long)spool);
1187 vma_commit_reservation(h, vma, addr);
1191 int __weak alloc_bootmem_huge_page(struct hstate *h)
1193 struct huge_bootmem_page *m;
1196 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1199 addr = __alloc_bootmem_node_nopanic(NODE_DATA(node),
1200 huge_page_size(h), huge_page_size(h), 0);
1204 * Use the beginning of the huge page to store the
1205 * huge_bootmem_page struct (until gather_bootmem
1206 * puts them into the mem_map).
1215 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1216 /* Put them into a private list first because mem_map is not up yet */
1217 list_add(&m->list, &huge_boot_pages);
1222 static void prep_compound_huge_page(struct page *page, int order)
1224 if (unlikely(order > (MAX_ORDER - 1)))
1225 prep_compound_gigantic_page(page, order);
1227 prep_compound_page(page, order);
1230 /* Put bootmem huge pages into the standard lists after mem_map is up */
1231 static void __init gather_bootmem_prealloc(void)
1233 struct huge_bootmem_page *m;
1235 list_for_each_entry(m, &huge_boot_pages, list) {
1236 struct hstate *h = m->hstate;
1239 #ifdef CONFIG_HIGHMEM
1240 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1241 free_bootmem_late((unsigned long)m,
1242 sizeof(struct huge_bootmem_page));
1244 page = virt_to_page(m);
1246 __ClearPageReserved(page);
1247 WARN_ON(page_count(page) != 1);
1248 prep_compound_huge_page(page, h->order);
1249 prep_new_huge_page(h, page, page_to_nid(page));
1251 * If we had gigantic hugepages allocated at boot time, we need
1252 * to restore the 'stolen' pages to totalram_pages in order to
1253 * fix confusing memory reports from free(1) and another
1254 * side-effects, like CommitLimit going negative.
1256 if (h->order > (MAX_ORDER - 1))
1257 adjust_managed_page_count(page, 1 << h->order);
1261 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1265 for (i = 0; i < h->max_huge_pages; ++i) {
1266 if (h->order >= MAX_ORDER) {
1267 if (!alloc_bootmem_huge_page(h))
1269 } else if (!alloc_fresh_huge_page(h,
1270 &node_states[N_MEMORY]))
1273 h->max_huge_pages = i;
1276 static void __init hugetlb_init_hstates(void)
1280 for_each_hstate(h) {
1281 /* oversize hugepages were init'ed in early boot */
1282 if (h->order < MAX_ORDER)
1283 hugetlb_hstate_alloc_pages(h);
1287 static char * __init memfmt(char *buf, unsigned long n)
1289 if (n >= (1UL << 30))
1290 sprintf(buf, "%lu GB", n >> 30);
1291 else if (n >= (1UL << 20))
1292 sprintf(buf, "%lu MB", n >> 20);
1294 sprintf(buf, "%lu KB", n >> 10);
1298 static void __init report_hugepages(void)
1302 for_each_hstate(h) {
1304 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1305 memfmt(buf, huge_page_size(h)),
1306 h->free_huge_pages);
1310 #ifdef CONFIG_HIGHMEM
1311 static void try_to_free_low(struct hstate *h, unsigned long count,
1312 nodemask_t *nodes_allowed)
1316 if (h->order >= MAX_ORDER)
1319 for_each_node_mask(i, *nodes_allowed) {
1320 struct page *page, *next;
1321 struct list_head *freel = &h->hugepage_freelists[i];
1322 list_for_each_entry_safe(page, next, freel, lru) {
1323 if (count >= h->nr_huge_pages)
1325 if (PageHighMem(page))
1327 list_del(&page->lru);
1328 update_and_free_page(h, page);
1329 h->free_huge_pages--;
1330 h->free_huge_pages_node[page_to_nid(page)]--;
1335 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1336 nodemask_t *nodes_allowed)
1342 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1343 * balanced by operating on them in a round-robin fashion.
1344 * Returns 1 if an adjustment was made.
1346 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1351 VM_BUG_ON(delta != -1 && delta != 1);
1354 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1355 if (h->surplus_huge_pages_node[node])
1359 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1360 if (h->surplus_huge_pages_node[node] <
1361 h->nr_huge_pages_node[node])
1368 h->surplus_huge_pages += delta;
1369 h->surplus_huge_pages_node[node] += delta;
1373 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1374 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1375 nodemask_t *nodes_allowed)
1377 unsigned long min_count, ret;
1379 if (h->order >= MAX_ORDER)
1380 return h->max_huge_pages;
1383 * Increase the pool size
1384 * First take pages out of surplus state. Then make up the
1385 * remaining difference by allocating fresh huge pages.
1387 * We might race with alloc_buddy_huge_page() here and be unable
1388 * to convert a surplus huge page to a normal huge page. That is
1389 * not critical, though, it just means the overall size of the
1390 * pool might be one hugepage larger than it needs to be, but
1391 * within all the constraints specified by the sysctls.
1393 spin_lock(&hugetlb_lock);
1394 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1395 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1399 while (count > persistent_huge_pages(h)) {
1401 * If this allocation races such that we no longer need the
1402 * page, free_huge_page will handle it by freeing the page
1403 * and reducing the surplus.
1405 spin_unlock(&hugetlb_lock);
1406 ret = alloc_fresh_huge_page(h, nodes_allowed);
1407 spin_lock(&hugetlb_lock);
1411 /* Bail for signals. Probably ctrl-c from user */
1412 if (signal_pending(current))
1417 * Decrease the pool size
1418 * First return free pages to the buddy allocator (being careful
1419 * to keep enough around to satisfy reservations). Then place
1420 * pages into surplus state as needed so the pool will shrink
1421 * to the desired size as pages become free.
1423 * By placing pages into the surplus state independent of the
1424 * overcommit value, we are allowing the surplus pool size to
1425 * exceed overcommit. There are few sane options here. Since
1426 * alloc_buddy_huge_page() is checking the global counter,
1427 * though, we'll note that we're not allowed to exceed surplus
1428 * and won't grow the pool anywhere else. Not until one of the
1429 * sysctls are changed, or the surplus pages go out of use.
1431 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1432 min_count = max(count, min_count);
1433 try_to_free_low(h, min_count, nodes_allowed);
1434 while (min_count < persistent_huge_pages(h)) {
1435 if (!free_pool_huge_page(h, nodes_allowed, 0))
1438 while (count < persistent_huge_pages(h)) {
1439 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1443 ret = persistent_huge_pages(h);
1444 spin_unlock(&hugetlb_lock);
1448 #define HSTATE_ATTR_RO(_name) \
1449 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1451 #define HSTATE_ATTR(_name) \
1452 static struct kobj_attribute _name##_attr = \
1453 __ATTR(_name, 0644, _name##_show, _name##_store)
1455 static struct kobject *hugepages_kobj;
1456 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1458 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1460 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1464 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1465 if (hstate_kobjs[i] == kobj) {
1467 *nidp = NUMA_NO_NODE;
1471 return kobj_to_node_hstate(kobj, nidp);
1474 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1475 struct kobj_attribute *attr, char *buf)
1478 unsigned long nr_huge_pages;
1481 h = kobj_to_hstate(kobj, &nid);
1482 if (nid == NUMA_NO_NODE)
1483 nr_huge_pages = h->nr_huge_pages;
1485 nr_huge_pages = h->nr_huge_pages_node[nid];
1487 return sprintf(buf, "%lu\n", nr_huge_pages);
1490 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1491 struct kobject *kobj, struct kobj_attribute *attr,
1492 const char *buf, size_t len)
1496 unsigned long count;
1498 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1500 err = kstrtoul(buf, 10, &count);
1504 h = kobj_to_hstate(kobj, &nid);
1505 if (h->order >= MAX_ORDER) {
1510 if (nid == NUMA_NO_NODE) {
1512 * global hstate attribute
1514 if (!(obey_mempolicy &&
1515 init_nodemask_of_mempolicy(nodes_allowed))) {
1516 NODEMASK_FREE(nodes_allowed);
1517 nodes_allowed = &node_states[N_MEMORY];
1519 } else if (nodes_allowed) {
1521 * per node hstate attribute: adjust count to global,
1522 * but restrict alloc/free to the specified node.
1524 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1525 init_nodemask_of_node(nodes_allowed, nid);
1527 nodes_allowed = &node_states[N_MEMORY];
1529 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1531 if (nodes_allowed != &node_states[N_MEMORY])
1532 NODEMASK_FREE(nodes_allowed);
1536 NODEMASK_FREE(nodes_allowed);
1540 static ssize_t nr_hugepages_show(struct kobject *kobj,
1541 struct kobj_attribute *attr, char *buf)
1543 return nr_hugepages_show_common(kobj, attr, buf);
1546 static ssize_t nr_hugepages_store(struct kobject *kobj,
1547 struct kobj_attribute *attr, const char *buf, size_t len)
1549 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1551 HSTATE_ATTR(nr_hugepages);
1556 * hstate attribute for optionally mempolicy-based constraint on persistent
1557 * huge page alloc/free.
1559 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1560 struct kobj_attribute *attr, char *buf)
1562 return nr_hugepages_show_common(kobj, attr, buf);
1565 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1566 struct kobj_attribute *attr, const char *buf, size_t len)
1568 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1570 HSTATE_ATTR(nr_hugepages_mempolicy);
1574 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1575 struct kobj_attribute *attr, char *buf)
1577 struct hstate *h = kobj_to_hstate(kobj, NULL);
1578 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1581 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1582 struct kobj_attribute *attr, const char *buf, size_t count)
1585 unsigned long input;
1586 struct hstate *h = kobj_to_hstate(kobj, NULL);
1588 if (h->order >= MAX_ORDER)
1591 err = kstrtoul(buf, 10, &input);
1595 spin_lock(&hugetlb_lock);
1596 h->nr_overcommit_huge_pages = input;
1597 spin_unlock(&hugetlb_lock);
1601 HSTATE_ATTR(nr_overcommit_hugepages);
1603 static ssize_t free_hugepages_show(struct kobject *kobj,
1604 struct kobj_attribute *attr, char *buf)
1607 unsigned long free_huge_pages;
1610 h = kobj_to_hstate(kobj, &nid);
1611 if (nid == NUMA_NO_NODE)
1612 free_huge_pages = h->free_huge_pages;
1614 free_huge_pages = h->free_huge_pages_node[nid];
1616 return sprintf(buf, "%lu\n", free_huge_pages);
1618 HSTATE_ATTR_RO(free_hugepages);
1620 static ssize_t resv_hugepages_show(struct kobject *kobj,
1621 struct kobj_attribute *attr, char *buf)
1623 struct hstate *h = kobj_to_hstate(kobj, NULL);
1624 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1626 HSTATE_ATTR_RO(resv_hugepages);
1628 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1629 struct kobj_attribute *attr, char *buf)
1632 unsigned long surplus_huge_pages;
1635 h = kobj_to_hstate(kobj, &nid);
1636 if (nid == NUMA_NO_NODE)
1637 surplus_huge_pages = h->surplus_huge_pages;
1639 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1641 return sprintf(buf, "%lu\n", surplus_huge_pages);
1643 HSTATE_ATTR_RO(surplus_hugepages);
1645 static struct attribute *hstate_attrs[] = {
1646 &nr_hugepages_attr.attr,
1647 &nr_overcommit_hugepages_attr.attr,
1648 &free_hugepages_attr.attr,
1649 &resv_hugepages_attr.attr,
1650 &surplus_hugepages_attr.attr,
1652 &nr_hugepages_mempolicy_attr.attr,
1657 static struct attribute_group hstate_attr_group = {
1658 .attrs = hstate_attrs,
1661 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1662 struct kobject **hstate_kobjs,
1663 struct attribute_group *hstate_attr_group)
1666 int hi = hstate_index(h);
1668 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1669 if (!hstate_kobjs[hi])
1672 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1674 kobject_put(hstate_kobjs[hi]);
1679 static void __init hugetlb_sysfs_init(void)
1684 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1685 if (!hugepages_kobj)
1688 for_each_hstate(h) {
1689 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1690 hstate_kobjs, &hstate_attr_group);
1692 pr_err("Hugetlb: Unable to add hstate %s", h->name);
1699 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1700 * with node devices in node_devices[] using a parallel array. The array
1701 * index of a node device or _hstate == node id.
1702 * This is here to avoid any static dependency of the node device driver, in
1703 * the base kernel, on the hugetlb module.
1705 struct node_hstate {
1706 struct kobject *hugepages_kobj;
1707 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1709 struct node_hstate node_hstates[MAX_NUMNODES];
1712 * A subset of global hstate attributes for node devices
1714 static struct attribute *per_node_hstate_attrs[] = {
1715 &nr_hugepages_attr.attr,
1716 &free_hugepages_attr.attr,
1717 &surplus_hugepages_attr.attr,
1721 static struct attribute_group per_node_hstate_attr_group = {
1722 .attrs = per_node_hstate_attrs,
1726 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1727 * Returns node id via non-NULL nidp.
1729 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1733 for (nid = 0; nid < nr_node_ids; nid++) {
1734 struct node_hstate *nhs = &node_hstates[nid];
1736 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1737 if (nhs->hstate_kobjs[i] == kobj) {
1749 * Unregister hstate attributes from a single node device.
1750 * No-op if no hstate attributes attached.
1752 static void hugetlb_unregister_node(struct node *node)
1755 struct node_hstate *nhs = &node_hstates[node->dev.id];
1757 if (!nhs->hugepages_kobj)
1758 return; /* no hstate attributes */
1760 for_each_hstate(h) {
1761 int idx = hstate_index(h);
1762 if (nhs->hstate_kobjs[idx]) {
1763 kobject_put(nhs->hstate_kobjs[idx]);
1764 nhs->hstate_kobjs[idx] = NULL;
1768 kobject_put(nhs->hugepages_kobj);
1769 nhs->hugepages_kobj = NULL;
1773 * hugetlb module exit: unregister hstate attributes from node devices
1776 static void hugetlb_unregister_all_nodes(void)
1781 * disable node device registrations.
1783 register_hugetlbfs_with_node(NULL, NULL);
1786 * remove hstate attributes from any nodes that have them.
1788 for (nid = 0; nid < nr_node_ids; nid++)
1789 hugetlb_unregister_node(node_devices[nid]);
1793 * Register hstate attributes for a single node device.
1794 * No-op if attributes already registered.
1796 static void hugetlb_register_node(struct node *node)
1799 struct node_hstate *nhs = &node_hstates[node->dev.id];
1802 if (nhs->hugepages_kobj)
1803 return; /* already allocated */
1805 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1807 if (!nhs->hugepages_kobj)
1810 for_each_hstate(h) {
1811 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1813 &per_node_hstate_attr_group);
1815 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1816 h->name, node->dev.id);
1817 hugetlb_unregister_node(node);
1824 * hugetlb init time: register hstate attributes for all registered node
1825 * devices of nodes that have memory. All on-line nodes should have
1826 * registered their associated device by this time.
1828 static void hugetlb_register_all_nodes(void)
1832 for_each_node_state(nid, N_MEMORY) {
1833 struct node *node = node_devices[nid];
1834 if (node->dev.id == nid)
1835 hugetlb_register_node(node);
1839 * Let the node device driver know we're here so it can
1840 * [un]register hstate attributes on node hotplug.
1842 register_hugetlbfs_with_node(hugetlb_register_node,
1843 hugetlb_unregister_node);
1845 #else /* !CONFIG_NUMA */
1847 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1855 static void hugetlb_unregister_all_nodes(void) { }
1857 static void hugetlb_register_all_nodes(void) { }
1861 static void __exit hugetlb_exit(void)
1865 hugetlb_unregister_all_nodes();
1867 for_each_hstate(h) {
1868 kobject_put(hstate_kobjs[hstate_index(h)]);
1871 kobject_put(hugepages_kobj);
1873 module_exit(hugetlb_exit);
1875 static int __init hugetlb_init(void)
1877 /* Some platform decide whether they support huge pages at boot
1878 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1879 * there is no such support
1881 if (HPAGE_SHIFT == 0)
1884 if (!size_to_hstate(default_hstate_size)) {
1885 default_hstate_size = HPAGE_SIZE;
1886 if (!size_to_hstate(default_hstate_size))
1887 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1889 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1890 if (default_hstate_max_huge_pages)
1891 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1893 hugetlb_init_hstates();
1894 gather_bootmem_prealloc();
1897 hugetlb_sysfs_init();
1898 hugetlb_register_all_nodes();
1899 hugetlb_cgroup_file_init();
1903 module_init(hugetlb_init);
1905 /* Should be called on processing a hugepagesz=... option */
1906 void __init hugetlb_add_hstate(unsigned order)
1911 if (size_to_hstate(PAGE_SIZE << order)) {
1912 pr_warning("hugepagesz= specified twice, ignoring\n");
1915 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1917 h = &hstates[hugetlb_max_hstate++];
1919 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1920 h->nr_huge_pages = 0;
1921 h->free_huge_pages = 0;
1922 for (i = 0; i < MAX_NUMNODES; ++i)
1923 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1924 INIT_LIST_HEAD(&h->hugepage_activelist);
1925 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
1926 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
1927 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1928 huge_page_size(h)/1024);
1933 static int __init hugetlb_nrpages_setup(char *s)
1936 static unsigned long *last_mhp;
1939 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1940 * so this hugepages= parameter goes to the "default hstate".
1942 if (!hugetlb_max_hstate)
1943 mhp = &default_hstate_max_huge_pages;
1945 mhp = &parsed_hstate->max_huge_pages;
1947 if (mhp == last_mhp) {
1948 pr_warning("hugepages= specified twice without "
1949 "interleaving hugepagesz=, ignoring\n");
1953 if (sscanf(s, "%lu", mhp) <= 0)
1957 * Global state is always initialized later in hugetlb_init.
1958 * But we need to allocate >= MAX_ORDER hstates here early to still
1959 * use the bootmem allocator.
1961 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
1962 hugetlb_hstate_alloc_pages(parsed_hstate);
1968 __setup("hugepages=", hugetlb_nrpages_setup);
1970 static int __init hugetlb_default_setup(char *s)
1972 default_hstate_size = memparse(s, &s);
1975 __setup("default_hugepagesz=", hugetlb_default_setup);
1977 static unsigned int cpuset_mems_nr(unsigned int *array)
1980 unsigned int nr = 0;
1982 for_each_node_mask(node, cpuset_current_mems_allowed)
1988 #ifdef CONFIG_SYSCTL
1989 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1990 struct ctl_table *table, int write,
1991 void __user *buffer, size_t *length, loff_t *ppos)
1993 struct hstate *h = &default_hstate;
1997 tmp = h->max_huge_pages;
1999 if (write && h->order >= MAX_ORDER)
2003 table->maxlen = sizeof(unsigned long);
2004 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2009 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2010 GFP_KERNEL | __GFP_NORETRY);
2011 if (!(obey_mempolicy &&
2012 init_nodemask_of_mempolicy(nodes_allowed))) {
2013 NODEMASK_FREE(nodes_allowed);
2014 nodes_allowed = &node_states[N_MEMORY];
2016 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2018 if (nodes_allowed != &node_states[N_MEMORY])
2019 NODEMASK_FREE(nodes_allowed);
2025 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2026 void __user *buffer, size_t *length, loff_t *ppos)
2029 return hugetlb_sysctl_handler_common(false, table, write,
2030 buffer, length, ppos);
2034 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2035 void __user *buffer, size_t *length, loff_t *ppos)
2037 return hugetlb_sysctl_handler_common(true, table, write,
2038 buffer, length, ppos);
2040 #endif /* CONFIG_NUMA */
2042 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2043 void __user *buffer,
2044 size_t *length, loff_t *ppos)
2046 proc_dointvec(table, write, buffer, length, ppos);
2047 if (hugepages_treat_as_movable)
2048 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2050 htlb_alloc_mask = GFP_HIGHUSER;
2054 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2055 void __user *buffer,
2056 size_t *length, loff_t *ppos)
2058 struct hstate *h = &default_hstate;
2062 tmp = h->nr_overcommit_huge_pages;
2064 if (write && h->order >= MAX_ORDER)
2068 table->maxlen = sizeof(unsigned long);
2069 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2074 spin_lock(&hugetlb_lock);
2075 h->nr_overcommit_huge_pages = tmp;
2076 spin_unlock(&hugetlb_lock);
2082 #endif /* CONFIG_SYSCTL */
2084 void hugetlb_report_meminfo(struct seq_file *m)
2086 struct hstate *h = &default_hstate;
2088 "HugePages_Total: %5lu\n"
2089 "HugePages_Free: %5lu\n"
2090 "HugePages_Rsvd: %5lu\n"
2091 "HugePages_Surp: %5lu\n"
2092 "Hugepagesize: %8lu kB\n",
2096 h->surplus_huge_pages,
2097 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2100 int hugetlb_report_node_meminfo(int nid, char *buf)
2102 struct hstate *h = &default_hstate;
2104 "Node %d HugePages_Total: %5u\n"
2105 "Node %d HugePages_Free: %5u\n"
2106 "Node %d HugePages_Surp: %5u\n",
2107 nid, h->nr_huge_pages_node[nid],
2108 nid, h->free_huge_pages_node[nid],
2109 nid, h->surplus_huge_pages_node[nid]);
2112 void hugetlb_show_meminfo(void)
2117 for_each_node_state(nid, N_MEMORY)
2119 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2121 h->nr_huge_pages_node[nid],
2122 h->free_huge_pages_node[nid],
2123 h->surplus_huge_pages_node[nid],
2124 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2127 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2128 unsigned long hugetlb_total_pages(void)
2131 unsigned long nr_total_pages = 0;
2134 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2135 return nr_total_pages;
2138 static int hugetlb_acct_memory(struct hstate *h, long delta)
2142 spin_lock(&hugetlb_lock);
2144 * When cpuset is configured, it breaks the strict hugetlb page
2145 * reservation as the accounting is done on a global variable. Such
2146 * reservation is completely rubbish in the presence of cpuset because
2147 * the reservation is not checked against page availability for the
2148 * current cpuset. Application can still potentially OOM'ed by kernel
2149 * with lack of free htlb page in cpuset that the task is in.
2150 * Attempt to enforce strict accounting with cpuset is almost
2151 * impossible (or too ugly) because cpuset is too fluid that
2152 * task or memory node can be dynamically moved between cpusets.
2154 * The change of semantics for shared hugetlb mapping with cpuset is
2155 * undesirable. However, in order to preserve some of the semantics,
2156 * we fall back to check against current free page availability as
2157 * a best attempt and hopefully to minimize the impact of changing
2158 * semantics that cpuset has.
2161 if (gather_surplus_pages(h, delta) < 0)
2164 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2165 return_unused_surplus_pages(h, delta);
2172 return_unused_surplus_pages(h, (unsigned long) -delta);
2175 spin_unlock(&hugetlb_lock);
2179 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2181 struct resv_map *reservations = vma_resv_map(vma);
2184 * This new VMA should share its siblings reservation map if present.
2185 * The VMA will only ever have a valid reservation map pointer where
2186 * it is being copied for another still existing VMA. As that VMA
2187 * has a reference to the reservation map it cannot disappear until
2188 * after this open call completes. It is therefore safe to take a
2189 * new reference here without additional locking.
2192 kref_get(&reservations->refs);
2195 static void resv_map_put(struct vm_area_struct *vma)
2197 struct resv_map *reservations = vma_resv_map(vma);
2201 kref_put(&reservations->refs, resv_map_release);
2204 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2206 struct hstate *h = hstate_vma(vma);
2207 struct resv_map *reservations = vma_resv_map(vma);
2208 struct hugepage_subpool *spool = subpool_vma(vma);
2209 unsigned long reserve;
2210 unsigned long start;
2214 start = vma_hugecache_offset(h, vma, vma->vm_start);
2215 end = vma_hugecache_offset(h, vma, vma->vm_end);
2217 reserve = (end - start) -
2218 region_count(&reservations->regions, start, end);
2223 hugetlb_acct_memory(h, -reserve);
2224 hugepage_subpool_put_pages(spool, reserve);
2230 * We cannot handle pagefaults against hugetlb pages at all. They cause
2231 * handle_mm_fault() to try to instantiate regular-sized pages in the
2232 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2235 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2241 const struct vm_operations_struct hugetlb_vm_ops = {
2242 .fault = hugetlb_vm_op_fault,
2243 .open = hugetlb_vm_op_open,
2244 .close = hugetlb_vm_op_close,
2247 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2253 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2254 vma->vm_page_prot)));
2256 entry = huge_pte_wrprotect(mk_huge_pte(page,
2257 vma->vm_page_prot));
2259 entry = pte_mkyoung(entry);
2260 entry = pte_mkhuge(entry);
2261 entry = arch_make_huge_pte(entry, vma, page, writable);
2266 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2267 unsigned long address, pte_t *ptep)
2271 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2272 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2273 update_mmu_cache(vma, address, ptep);
2277 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2278 struct vm_area_struct *vma)
2280 pte_t *src_pte, *dst_pte, entry;
2281 struct page *ptepage;
2284 struct hstate *h = hstate_vma(vma);
2285 unsigned long sz = huge_page_size(h);
2287 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2289 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2290 src_pte = huge_pte_offset(src, addr);
2293 dst_pte = huge_pte_alloc(dst, addr, sz);
2297 /* If the pagetables are shared don't copy or take references */
2298 if (dst_pte == src_pte)
2301 spin_lock(&dst->page_table_lock);
2302 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2303 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2305 huge_ptep_set_wrprotect(src, addr, src_pte);
2306 entry = huge_ptep_get(src_pte);
2307 ptepage = pte_page(entry);
2309 page_dup_rmap(ptepage);
2310 set_huge_pte_at(dst, addr, dst_pte, entry);
2312 spin_unlock(&src->page_table_lock);
2313 spin_unlock(&dst->page_table_lock);
2321 static int is_hugetlb_entry_migration(pte_t pte)
2325 if (huge_pte_none(pte) || pte_present(pte))
2327 swp = pte_to_swp_entry(pte);
2328 if (non_swap_entry(swp) && is_migration_entry(swp))
2334 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2338 if (huge_pte_none(pte) || pte_present(pte))
2340 swp = pte_to_swp_entry(pte);
2341 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2347 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2348 unsigned long start, unsigned long end,
2349 struct page *ref_page)
2351 int force_flush = 0;
2352 struct mm_struct *mm = vma->vm_mm;
2353 unsigned long address;
2357 struct hstate *h = hstate_vma(vma);
2358 unsigned long sz = huge_page_size(h);
2359 const unsigned long mmun_start = start; /* For mmu_notifiers */
2360 const unsigned long mmun_end = end; /* For mmu_notifiers */
2362 WARN_ON(!is_vm_hugetlb_page(vma));
2363 BUG_ON(start & ~huge_page_mask(h));
2364 BUG_ON(end & ~huge_page_mask(h));
2366 tlb_start_vma(tlb, vma);
2367 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2369 spin_lock(&mm->page_table_lock);
2370 for (address = start; address < end; address += sz) {
2371 ptep = huge_pte_offset(mm, address);
2375 if (huge_pmd_unshare(mm, &address, ptep))
2378 pte = huge_ptep_get(ptep);
2379 if (huge_pte_none(pte))
2383 * HWPoisoned hugepage is already unmapped and dropped reference
2385 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
2386 huge_pte_clear(mm, address, ptep);
2390 page = pte_page(pte);
2392 * If a reference page is supplied, it is because a specific
2393 * page is being unmapped, not a range. Ensure the page we
2394 * are about to unmap is the actual page of interest.
2397 if (page != ref_page)
2401 * Mark the VMA as having unmapped its page so that
2402 * future faults in this VMA will fail rather than
2403 * looking like data was lost
2405 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2408 pte = huge_ptep_get_and_clear(mm, address, ptep);
2409 tlb_remove_tlb_entry(tlb, ptep, address);
2410 if (huge_pte_dirty(pte))
2411 set_page_dirty(page);
2413 page_remove_rmap(page);
2414 force_flush = !__tlb_remove_page(tlb, page);
2417 /* Bail out after unmapping reference page if supplied */
2421 spin_unlock(&mm->page_table_lock);
2423 * mmu_gather ran out of room to batch pages, we break out of
2424 * the PTE lock to avoid doing the potential expensive TLB invalidate
2425 * and page-free while holding it.
2430 if (address < end && !ref_page)
2433 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2434 tlb_end_vma(tlb, vma);
2437 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2438 struct vm_area_struct *vma, unsigned long start,
2439 unsigned long end, struct page *ref_page)
2441 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2444 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2445 * test will fail on a vma being torn down, and not grab a page table
2446 * on its way out. We're lucky that the flag has such an appropriate
2447 * name, and can in fact be safely cleared here. We could clear it
2448 * before the __unmap_hugepage_range above, but all that's necessary
2449 * is to clear it before releasing the i_mmap_mutex. This works
2450 * because in the context this is called, the VMA is about to be
2451 * destroyed and the i_mmap_mutex is held.
2453 vma->vm_flags &= ~VM_MAYSHARE;
2456 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2457 unsigned long end, struct page *ref_page)
2459 struct mm_struct *mm;
2460 struct mmu_gather tlb;
2464 tlb_gather_mmu(&tlb, mm, start, end);
2465 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2466 tlb_finish_mmu(&tlb, start, end);
2470 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2471 * mappping it owns the reserve page for. The intention is to unmap the page
2472 * from other VMAs and let the children be SIGKILLed if they are faulting the
2475 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2476 struct page *page, unsigned long address)
2478 struct hstate *h = hstate_vma(vma);
2479 struct vm_area_struct *iter_vma;
2480 struct address_space *mapping;
2484 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2485 * from page cache lookup which is in HPAGE_SIZE units.
2487 address = address & huge_page_mask(h);
2488 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2490 mapping = file_inode(vma->vm_file)->i_mapping;
2493 * Take the mapping lock for the duration of the table walk. As
2494 * this mapping should be shared between all the VMAs,
2495 * __unmap_hugepage_range() is called as the lock is already held
2497 mutex_lock(&mapping->i_mmap_mutex);
2498 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2499 /* Do not unmap the current VMA */
2500 if (iter_vma == vma)
2504 * Unmap the page from other VMAs without their own reserves.
2505 * They get marked to be SIGKILLed if they fault in these
2506 * areas. This is because a future no-page fault on this VMA
2507 * could insert a zeroed page instead of the data existing
2508 * from the time of fork. This would look like data corruption
2510 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2511 unmap_hugepage_range(iter_vma, address,
2512 address + huge_page_size(h), page);
2514 mutex_unlock(&mapping->i_mmap_mutex);
2520 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2521 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2522 * cannot race with other handlers or page migration.
2523 * Keep the pte_same checks anyway to make transition from the mutex easier.
2525 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2526 unsigned long address, pte_t *ptep, pte_t pte,
2527 struct page *pagecache_page)
2529 struct hstate *h = hstate_vma(vma);
2530 struct page *old_page, *new_page;
2531 int outside_reserve = 0;
2532 unsigned long mmun_start; /* For mmu_notifiers */
2533 unsigned long mmun_end; /* For mmu_notifiers */
2535 old_page = pte_page(pte);
2538 /* If no-one else is actually using this page, avoid the copy
2539 * and just make the page writable */
2540 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2541 page_move_anon_rmap(old_page, vma, address);
2542 set_huge_ptep_writable(vma, address, ptep);
2547 * If the process that created a MAP_PRIVATE mapping is about to
2548 * perform a COW due to a shared page count, attempt to satisfy
2549 * the allocation without using the existing reserves. The pagecache
2550 * page is used to determine if the reserve at this address was
2551 * consumed or not. If reserves were used, a partial faulted mapping
2552 * at the time of fork() could consume its reserves on COW instead
2553 * of the full address range.
2555 if (!(vma->vm_flags & VM_MAYSHARE) &&
2556 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2557 old_page != pagecache_page)
2558 outside_reserve = 1;
2560 page_cache_get(old_page);
2562 /* Drop page_table_lock as buddy allocator may be called */
2563 spin_unlock(&mm->page_table_lock);
2564 new_page = alloc_huge_page(vma, address, outside_reserve);
2566 if (IS_ERR(new_page)) {
2567 long err = PTR_ERR(new_page);
2568 page_cache_release(old_page);
2571 * If a process owning a MAP_PRIVATE mapping fails to COW,
2572 * it is due to references held by a child and an insufficient
2573 * huge page pool. To guarantee the original mappers
2574 * reliability, unmap the page from child processes. The child
2575 * may get SIGKILLed if it later faults.
2577 if (outside_reserve) {
2578 BUG_ON(huge_pte_none(pte));
2579 if (unmap_ref_private(mm, vma, old_page, address)) {
2580 BUG_ON(huge_pte_none(pte));
2581 spin_lock(&mm->page_table_lock);
2582 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2583 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2584 goto retry_avoidcopy;
2586 * race occurs while re-acquiring page_table_lock, and
2594 /* Caller expects lock to be held */
2595 spin_lock(&mm->page_table_lock);
2597 return VM_FAULT_OOM;
2599 return VM_FAULT_SIGBUS;
2603 * When the original hugepage is shared one, it does not have
2604 * anon_vma prepared.
2606 if (unlikely(anon_vma_prepare(vma))) {
2607 page_cache_release(new_page);
2608 page_cache_release(old_page);
2609 /* Caller expects lock to be held */
2610 spin_lock(&mm->page_table_lock);
2611 return VM_FAULT_OOM;
2614 copy_user_huge_page(new_page, old_page, address, vma,
2615 pages_per_huge_page(h));
2616 __SetPageUptodate(new_page);
2618 mmun_start = address & huge_page_mask(h);
2619 mmun_end = mmun_start + huge_page_size(h);
2620 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2622 * Retake the page_table_lock to check for racing updates
2623 * before the page tables are altered
2625 spin_lock(&mm->page_table_lock);
2626 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2627 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2629 huge_ptep_clear_flush(vma, address, ptep);
2630 set_huge_pte_at(mm, address, ptep,
2631 make_huge_pte(vma, new_page, 1));
2632 page_remove_rmap(old_page);
2633 hugepage_add_new_anon_rmap(new_page, vma, address);
2634 /* Make the old page be freed below */
2635 new_page = old_page;
2637 spin_unlock(&mm->page_table_lock);
2638 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2639 /* Caller expects lock to be held */
2640 spin_lock(&mm->page_table_lock);
2641 page_cache_release(new_page);
2642 page_cache_release(old_page);
2646 /* Return the pagecache page at a given address within a VMA */
2647 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2648 struct vm_area_struct *vma, unsigned long address)
2650 struct address_space *mapping;
2653 mapping = vma->vm_file->f_mapping;
2654 idx = vma_hugecache_offset(h, vma, address);
2656 return find_lock_page(mapping, idx);
2660 * Return whether there is a pagecache page to back given address within VMA.
2661 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2663 static bool hugetlbfs_pagecache_present(struct hstate *h,
2664 struct vm_area_struct *vma, unsigned long address)
2666 struct address_space *mapping;
2670 mapping = vma->vm_file->f_mapping;
2671 idx = vma_hugecache_offset(h, vma, address);
2673 page = find_get_page(mapping, idx);
2676 return page != NULL;
2679 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2680 unsigned long address, pte_t *ptep, unsigned int flags)
2682 struct hstate *h = hstate_vma(vma);
2683 int ret = VM_FAULT_SIGBUS;
2688 struct address_space *mapping;
2692 * Currently, we are forced to kill the process in the event the
2693 * original mapper has unmapped pages from the child due to a failed
2694 * COW. Warn that such a situation has occurred as it may not be obvious
2696 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2697 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2702 mapping = vma->vm_file->f_mapping;
2703 idx = vma_hugecache_offset(h, vma, address);
2706 * Use page lock to guard against racing truncation
2707 * before we get page_table_lock.
2710 page = find_lock_page(mapping, idx);
2712 size = i_size_read(mapping->host) >> huge_page_shift(h);
2715 page = alloc_huge_page(vma, address, 0);
2717 ret = PTR_ERR(page);
2721 ret = VM_FAULT_SIGBUS;
2724 clear_huge_page(page, address, pages_per_huge_page(h));
2725 __SetPageUptodate(page);
2727 if (vma->vm_flags & VM_MAYSHARE) {
2729 struct inode *inode = mapping->host;
2731 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2739 spin_lock(&inode->i_lock);
2740 inode->i_blocks += blocks_per_huge_page(h);
2741 spin_unlock(&inode->i_lock);
2744 if (unlikely(anon_vma_prepare(vma))) {
2746 goto backout_unlocked;
2752 * If memory error occurs between mmap() and fault, some process
2753 * don't have hwpoisoned swap entry for errored virtual address.
2754 * So we need to block hugepage fault by PG_hwpoison bit check.
2756 if (unlikely(PageHWPoison(page))) {
2757 ret = VM_FAULT_HWPOISON |
2758 VM_FAULT_SET_HINDEX(hstate_index(h));
2759 goto backout_unlocked;
2764 * If we are going to COW a private mapping later, we examine the
2765 * pending reservations for this page now. This will ensure that
2766 * any allocations necessary to record that reservation occur outside
2769 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2770 if (vma_needs_reservation(h, vma, address) < 0) {
2772 goto backout_unlocked;
2775 spin_lock(&mm->page_table_lock);
2776 size = i_size_read(mapping->host) >> huge_page_shift(h);
2781 if (!huge_pte_none(huge_ptep_get(ptep)))
2785 hugepage_add_new_anon_rmap(page, vma, address);
2787 page_dup_rmap(page);
2788 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2789 && (vma->vm_flags & VM_SHARED)));
2790 set_huge_pte_at(mm, address, ptep, new_pte);
2792 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2793 /* Optimization, do the COW without a second fault */
2794 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2797 spin_unlock(&mm->page_table_lock);
2803 spin_unlock(&mm->page_table_lock);
2810 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2811 unsigned long address, unsigned int flags)
2816 struct page *page = NULL;
2817 struct page *pagecache_page = NULL;
2818 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2819 struct hstate *h = hstate_vma(vma);
2821 address &= huge_page_mask(h);
2823 ptep = huge_pte_offset(mm, address);
2825 entry = huge_ptep_get(ptep);
2826 if (unlikely(is_hugetlb_entry_migration(entry))) {
2827 migration_entry_wait_huge(mm, ptep);
2829 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2830 return VM_FAULT_HWPOISON_LARGE |
2831 VM_FAULT_SET_HINDEX(hstate_index(h));
2834 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2836 return VM_FAULT_OOM;
2839 * Serialize hugepage allocation and instantiation, so that we don't
2840 * get spurious allocation failures if two CPUs race to instantiate
2841 * the same page in the page cache.
2843 mutex_lock(&hugetlb_instantiation_mutex);
2844 entry = huge_ptep_get(ptep);
2845 if (huge_pte_none(entry)) {
2846 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2853 * If we are going to COW the mapping later, we examine the pending
2854 * reservations for this page now. This will ensure that any
2855 * allocations necessary to record that reservation occur outside the
2856 * spinlock. For private mappings, we also lookup the pagecache
2857 * page now as it is used to determine if a reservation has been
2860 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
2861 if (vma_needs_reservation(h, vma, address) < 0) {
2866 if (!(vma->vm_flags & VM_MAYSHARE))
2867 pagecache_page = hugetlbfs_pagecache_page(h,
2872 * hugetlb_cow() requires page locks of pte_page(entry) and
2873 * pagecache_page, so here we need take the former one
2874 * when page != pagecache_page or !pagecache_page.
2875 * Note that locking order is always pagecache_page -> page,
2876 * so no worry about deadlock.
2878 page = pte_page(entry);
2880 if (page != pagecache_page)
2883 spin_lock(&mm->page_table_lock);
2884 /* Check for a racing update before calling hugetlb_cow */
2885 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2886 goto out_page_table_lock;
2889 if (flags & FAULT_FLAG_WRITE) {
2890 if (!huge_pte_write(entry)) {
2891 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2893 goto out_page_table_lock;
2895 entry = huge_pte_mkdirty(entry);
2897 entry = pte_mkyoung(entry);
2898 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2899 flags & FAULT_FLAG_WRITE))
2900 update_mmu_cache(vma, address, ptep);
2902 out_page_table_lock:
2903 spin_unlock(&mm->page_table_lock);
2905 if (pagecache_page) {
2906 unlock_page(pagecache_page);
2907 put_page(pagecache_page);
2909 if (page != pagecache_page)
2914 mutex_unlock(&hugetlb_instantiation_mutex);
2919 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2920 struct page **pages, struct vm_area_struct **vmas,
2921 unsigned long *position, unsigned long *nr_pages,
2922 long i, unsigned int flags)
2924 unsigned long pfn_offset;
2925 unsigned long vaddr = *position;
2926 unsigned long remainder = *nr_pages;
2927 struct hstate *h = hstate_vma(vma);
2929 spin_lock(&mm->page_table_lock);
2930 while (vaddr < vma->vm_end && remainder) {
2936 * Some archs (sparc64, sh*) have multiple pte_ts to
2937 * each hugepage. We have to make sure we get the
2938 * first, for the page indexing below to work.
2940 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2941 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2944 * When coredumping, it suits get_dump_page if we just return
2945 * an error where there's an empty slot with no huge pagecache
2946 * to back it. This way, we avoid allocating a hugepage, and
2947 * the sparse dumpfile avoids allocating disk blocks, but its
2948 * huge holes still show up with zeroes where they need to be.
2950 if (absent && (flags & FOLL_DUMP) &&
2951 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2957 * We need call hugetlb_fault for both hugepages under migration
2958 * (in which case hugetlb_fault waits for the migration,) and
2959 * hwpoisoned hugepages (in which case we need to prevent the
2960 * caller from accessing to them.) In order to do this, we use
2961 * here is_swap_pte instead of is_hugetlb_entry_migration and
2962 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
2963 * both cases, and because we can't follow correct pages
2964 * directly from any kind of swap entries.
2966 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
2967 ((flags & FOLL_WRITE) &&
2968 !huge_pte_write(huge_ptep_get(pte)))) {
2971 spin_unlock(&mm->page_table_lock);
2972 ret = hugetlb_fault(mm, vma, vaddr,
2973 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2974 spin_lock(&mm->page_table_lock);
2975 if (!(ret & VM_FAULT_ERROR))
2982 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2983 page = pte_page(huge_ptep_get(pte));
2986 pages[i] = mem_map_offset(page, pfn_offset);
2997 if (vaddr < vma->vm_end && remainder &&
2998 pfn_offset < pages_per_huge_page(h)) {
3000 * We use pfn_offset to avoid touching the pageframes
3001 * of this compound page.
3006 spin_unlock(&mm->page_table_lock);
3007 *nr_pages = remainder;
3010 return i ? i : -EFAULT;
3013 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3014 unsigned long address, unsigned long end, pgprot_t newprot)
3016 struct mm_struct *mm = vma->vm_mm;
3017 unsigned long start = address;
3020 struct hstate *h = hstate_vma(vma);
3021 unsigned long pages = 0;
3023 BUG_ON(address >= end);
3024 flush_cache_range(vma, address, end);
3026 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3027 spin_lock(&mm->page_table_lock);
3028 for (; address < end; address += huge_page_size(h)) {
3029 ptep = huge_pte_offset(mm, address);
3032 if (huge_pmd_unshare(mm, &address, ptep)) {
3036 if (!huge_pte_none(huge_ptep_get(ptep))) {
3037 pte = huge_ptep_get_and_clear(mm, address, ptep);
3038 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3039 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3040 set_huge_pte_at(mm, address, ptep, pte);
3044 spin_unlock(&mm->page_table_lock);
3046 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3047 * may have cleared our pud entry and done put_page on the page table:
3048 * once we release i_mmap_mutex, another task can do the final put_page
3049 * and that page table be reused and filled with junk.
3051 flush_tlb_range(vma, start, end);
3052 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3054 return pages << h->order;
3057 int hugetlb_reserve_pages(struct inode *inode,
3059 struct vm_area_struct *vma,
3060 vm_flags_t vm_flags)
3063 struct hstate *h = hstate_inode(inode);
3064 struct hugepage_subpool *spool = subpool_inode(inode);
3067 * Only apply hugepage reservation if asked. At fault time, an
3068 * attempt will be made for VM_NORESERVE to allocate a page
3069 * without using reserves
3071 if (vm_flags & VM_NORESERVE)
3075 * Shared mappings base their reservation on the number of pages that
3076 * are already allocated on behalf of the file. Private mappings need
3077 * to reserve the full area even if read-only as mprotect() may be
3078 * called to make the mapping read-write. Assume !vma is a shm mapping
3080 if (!vma || vma->vm_flags & VM_MAYSHARE)
3081 chg = region_chg(&inode->i_mapping->private_list, from, to);
3083 struct resv_map *resv_map = resv_map_alloc();
3089 set_vma_resv_map(vma, resv_map);
3090 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3098 /* There must be enough pages in the subpool for the mapping */
3099 if (hugepage_subpool_get_pages(spool, chg)) {
3105 * Check enough hugepages are available for the reservation.
3106 * Hand the pages back to the subpool if there are not
3108 ret = hugetlb_acct_memory(h, chg);
3110 hugepage_subpool_put_pages(spool, chg);
3115 * Account for the reservations made. Shared mappings record regions
3116 * that have reservations as they are shared by multiple VMAs.
3117 * When the last VMA disappears, the region map says how much
3118 * the reservation was and the page cache tells how much of
3119 * the reservation was consumed. Private mappings are per-VMA and
3120 * only the consumed reservations are tracked. When the VMA
3121 * disappears, the original reservation is the VMA size and the
3122 * consumed reservations are stored in the map. Hence, nothing
3123 * else has to be done for private mappings here
3125 if (!vma || vma->vm_flags & VM_MAYSHARE)
3126 region_add(&inode->i_mapping->private_list, from, to);
3134 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3136 struct hstate *h = hstate_inode(inode);
3137 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3138 struct hugepage_subpool *spool = subpool_inode(inode);
3140 spin_lock(&inode->i_lock);
3141 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3142 spin_unlock(&inode->i_lock);
3144 hugepage_subpool_put_pages(spool, (chg - freed));
3145 hugetlb_acct_memory(h, -(chg - freed));
3148 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3149 static unsigned long page_table_shareable(struct vm_area_struct *svma,
3150 struct vm_area_struct *vma,
3151 unsigned long addr, pgoff_t idx)
3153 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3155 unsigned long sbase = saddr & PUD_MASK;
3156 unsigned long s_end = sbase + PUD_SIZE;
3158 /* Allow segments to share if only one is marked locked */
3159 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3160 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3163 * match the virtual addresses, permission and the alignment of the
3166 if (pmd_index(addr) != pmd_index(saddr) ||
3167 vm_flags != svm_flags ||
3168 sbase < svma->vm_start || svma->vm_end < s_end)
3174 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3176 unsigned long base = addr & PUD_MASK;
3177 unsigned long end = base + PUD_SIZE;
3180 * check on proper vm_flags and page table alignment
3182 if (vma->vm_flags & VM_MAYSHARE &&
3183 vma->vm_start <= base && end <= vma->vm_end)
3189 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3190 * and returns the corresponding pte. While this is not necessary for the
3191 * !shared pmd case because we can allocate the pmd later as well, it makes the
3192 * code much cleaner. pmd allocation is essential for the shared case because
3193 * pud has to be populated inside the same i_mmap_mutex section - otherwise
3194 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3195 * bad pmd for sharing.
3197 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3199 struct vm_area_struct *vma = find_vma(mm, addr);
3200 struct address_space *mapping = vma->vm_file->f_mapping;
3201 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3203 struct vm_area_struct *svma;
3204 unsigned long saddr;
3208 if (!vma_shareable(vma, addr))
3209 return (pte_t *)pmd_alloc(mm, pud, addr);
3211 mutex_lock(&mapping->i_mmap_mutex);
3212 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3216 saddr = page_table_shareable(svma, vma, addr, idx);
3218 spte = huge_pte_offset(svma->vm_mm, saddr);
3220 get_page(virt_to_page(spte));
3229 spin_lock(&mm->page_table_lock);
3231 pud_populate(mm, pud,
3232 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3234 put_page(virt_to_page(spte));
3235 spin_unlock(&mm->page_table_lock);
3237 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3238 mutex_unlock(&mapping->i_mmap_mutex);
3243 * unmap huge page backed by shared pte.
3245 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3246 * indicated by page_count > 1, unmap is achieved by clearing pud and
3247 * decrementing the ref count. If count == 1, the pte page is not shared.
3249 * called with vma->vm_mm->page_table_lock held.
3251 * returns: 1 successfully unmapped a shared pte page
3252 * 0 the underlying pte page is not shared, or it is the last user
3254 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3256 pgd_t *pgd = pgd_offset(mm, *addr);
3257 pud_t *pud = pud_offset(pgd, *addr);
3259 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3260 if (page_count(virt_to_page(ptep)) == 1)
3264 put_page(virt_to_page(ptep));
3265 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3268 #define want_pmd_share() (1)
3269 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3270 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3274 #define want_pmd_share() (0)
3275 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3277 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3278 pte_t *huge_pte_alloc(struct mm_struct *mm,
3279 unsigned long addr, unsigned long sz)
3285 pgd = pgd_offset(mm, addr);
3286 pud = pud_alloc(mm, pgd, addr);
3288 if (sz == PUD_SIZE) {
3291 BUG_ON(sz != PMD_SIZE);
3292 if (want_pmd_share() && pud_none(*pud))
3293 pte = huge_pmd_share(mm, addr, pud);
3295 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3298 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3303 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3309 pgd = pgd_offset(mm, addr);
3310 if (pgd_present(*pgd)) {
3311 pud = pud_offset(pgd, addr);
3312 if (pud_present(*pud)) {
3314 return (pte_t *)pud;
3315 pmd = pmd_offset(pud, addr);
3318 return (pte_t *) pmd;
3322 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3323 pmd_t *pmd, int write)
3327 page = pte_page(*(pte_t *)pmd);
3329 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
3334 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3335 pud_t *pud, int write)
3339 page = pte_page(*(pte_t *)pud);
3341 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
3345 #else /* !CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3347 /* Can be overriden by architectures */
3348 __attribute__((weak)) struct page *
3349 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3350 pud_t *pud, int write)
3356 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3358 #ifdef CONFIG_MEMORY_FAILURE
3360 /* Should be called in hugetlb_lock */
3361 static int is_hugepage_on_freelist(struct page *hpage)
3365 struct hstate *h = page_hstate(hpage);
3366 int nid = page_to_nid(hpage);
3368 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3375 * This function is called from memory failure code.
3376 * Assume the caller holds page lock of the head page.
3378 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3380 struct hstate *h = page_hstate(hpage);
3381 int nid = page_to_nid(hpage);
3384 spin_lock(&hugetlb_lock);
3385 if (is_hugepage_on_freelist(hpage)) {
3387 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3388 * but dangling hpage->lru can trigger list-debug warnings
3389 * (this happens when we call unpoison_memory() on it),
3390 * so let it point to itself with list_del_init().
3392 list_del_init(&hpage->lru);
3393 set_page_refcounted(hpage);
3394 h->free_huge_pages--;
3395 h->free_huge_pages_node[nid]--;
3398 spin_unlock(&hugetlb_lock);