2 * Generic hugetlb support.
3 * (C) William Irwin, 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>
29 #include <linux/hugetlb.h>
30 #include <linux/node.h>
33 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
34 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
35 unsigned long hugepages_treat_as_movable;
37 static int max_hstate;
38 unsigned int default_hstate_idx;
39 struct hstate hstates[HUGE_MAX_HSTATE];
41 __initdata LIST_HEAD(huge_boot_pages);
43 /* for command line parsing */
44 static struct hstate * __initdata parsed_hstate;
45 static unsigned long __initdata default_hstate_max_huge_pages;
46 static unsigned long __initdata default_hstate_size;
48 #define for_each_hstate(h) \
49 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
52 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
54 static DEFINE_SPINLOCK(hugetlb_lock);
56 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
58 bool free = (spool->count == 0) && (spool->used_hpages == 0);
60 spin_unlock(&spool->lock);
62 /* If no pages are used, and no other handles to the subpool
63 * remain, free the subpool the subpool remain */
68 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
70 struct hugepage_subpool *spool;
72 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
76 spin_lock_init(&spool->lock);
78 spool->max_hpages = nr_blocks;
79 spool->used_hpages = 0;
84 void hugepage_put_subpool(struct hugepage_subpool *spool)
86 spin_lock(&spool->lock);
87 BUG_ON(!spool->count);
89 unlock_or_release_subpool(spool);
92 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
100 spin_lock(&spool->lock);
101 if ((spool->used_hpages + delta) <= spool->max_hpages) {
102 spool->used_hpages += delta;
106 spin_unlock(&spool->lock);
111 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
117 spin_lock(&spool->lock);
118 spool->used_hpages -= delta;
119 /* If hugetlbfs_put_super couldn't free spool due to
120 * an outstanding quota reference, free it now. */
121 unlock_or_release_subpool(spool);
124 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
126 return HUGETLBFS_SB(inode->i_sb)->spool;
129 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
131 return subpool_inode(vma->vm_file->f_dentry->d_inode);
135 * Region tracking -- allows tracking of reservations and instantiated pages
136 * across the pages in a mapping.
138 * The region data structures are protected by a combination of the mmap_sem
139 * and the hugetlb_instantion_mutex. To access or modify a region the caller
140 * must either hold the mmap_sem for write, or the mmap_sem for read and
141 * the hugetlb_instantiation mutex:
143 * down_write(&mm->mmap_sem);
145 * down_read(&mm->mmap_sem);
146 * mutex_lock(&hugetlb_instantiation_mutex);
149 struct list_head link;
154 static long region_add(struct list_head *head, long f, long t)
156 struct file_region *rg, *nrg, *trg;
158 /* Locate the region we are either in or before. */
159 list_for_each_entry(rg, head, link)
163 /* Round our left edge to the current segment if it encloses us. */
167 /* Check for and consume any regions we now overlap with. */
169 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
170 if (&rg->link == head)
175 /* If this area reaches higher then extend our area to
176 * include it completely. If this is not the first area
177 * which we intend to reuse, free it. */
190 static long region_chg(struct list_head *head, long f, long t)
192 struct file_region *rg, *nrg;
195 /* Locate the region we are before or in. */
196 list_for_each_entry(rg, head, link)
200 /* If we are below the current region then a new region is required.
201 * Subtle, allocate a new region at the position but make it zero
202 * size such that we can guarantee to record the reservation. */
203 if (&rg->link == head || t < rg->from) {
204 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
209 INIT_LIST_HEAD(&nrg->link);
210 list_add(&nrg->link, rg->link.prev);
215 /* Round our left edge to the current segment if it encloses us. */
220 /* Check for and consume any regions we now overlap with. */
221 list_for_each_entry(rg, rg->link.prev, link) {
222 if (&rg->link == head)
227 /* We overlap with this area, if it extends further than
228 * us then we must extend ourselves. Account for its
229 * existing reservation. */
234 chg -= rg->to - rg->from;
239 static long region_truncate(struct list_head *head, long end)
241 struct file_region *rg, *trg;
244 /* Locate the region we are either in or before. */
245 list_for_each_entry(rg, head, link)
248 if (&rg->link == head)
251 /* If we are in the middle of a region then adjust it. */
252 if (end > rg->from) {
255 rg = list_entry(rg->link.next, typeof(*rg), link);
258 /* Drop any remaining regions. */
259 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
260 if (&rg->link == head)
262 chg += rg->to - rg->from;
269 static long region_count(struct list_head *head, long f, long t)
271 struct file_region *rg;
274 /* Locate each segment we overlap with, and count that overlap. */
275 list_for_each_entry(rg, head, link) {
284 seg_from = max(rg->from, f);
285 seg_to = min(rg->to, t);
287 chg += seg_to - seg_from;
294 * Convert the address within this vma to the page offset within
295 * the mapping, in pagecache page units; huge pages here.
297 static pgoff_t vma_hugecache_offset(struct hstate *h,
298 struct vm_area_struct *vma, unsigned long address)
300 return ((address - vma->vm_start) >> huge_page_shift(h)) +
301 (vma->vm_pgoff >> huge_page_order(h));
304 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
305 unsigned long address)
307 return vma_hugecache_offset(hstate_vma(vma), vma, address);
311 * Return the size of the pages allocated when backing a VMA. In the majority
312 * cases this will be same size as used by the page table entries.
314 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
316 struct hstate *hstate;
318 if (!is_vm_hugetlb_page(vma))
321 hstate = hstate_vma(vma);
323 return 1UL << (hstate->order + PAGE_SHIFT);
325 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
328 * Return the page size being used by the MMU to back a VMA. In the majority
329 * of cases, the page size used by the kernel matches the MMU size. On
330 * architectures where it differs, an architecture-specific version of this
331 * function is required.
333 #ifndef vma_mmu_pagesize
334 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
336 return vma_kernel_pagesize(vma);
341 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
342 * bits of the reservation map pointer, which are always clear due to
345 #define HPAGE_RESV_OWNER (1UL << 0)
346 #define HPAGE_RESV_UNMAPPED (1UL << 1)
347 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
350 * These helpers are used to track how many pages are reserved for
351 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
352 * is guaranteed to have their future faults succeed.
354 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
355 * the reserve counters are updated with the hugetlb_lock held. It is safe
356 * to reset the VMA at fork() time as it is not in use yet and there is no
357 * chance of the global counters getting corrupted as a result of the values.
359 * The private mapping reservation is represented in a subtly different
360 * manner to a shared mapping. A shared mapping has a region map associated
361 * with the underlying file, this region map represents the backing file
362 * pages which have ever had a reservation assigned which this persists even
363 * after the page is instantiated. A private mapping has a region map
364 * associated with the original mmap which is attached to all VMAs which
365 * reference it, this region map represents those offsets which have consumed
366 * reservation ie. where pages have been instantiated.
368 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
370 return (unsigned long)vma->vm_private_data;
373 static void set_vma_private_data(struct vm_area_struct *vma,
376 vma->vm_private_data = (void *)value;
381 struct list_head regions;
384 static struct resv_map *resv_map_alloc(void)
386 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
390 kref_init(&resv_map->refs);
391 INIT_LIST_HEAD(&resv_map->regions);
396 static void resv_map_release(struct kref *ref)
398 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
400 /* Clear out any active regions before we release the map. */
401 region_truncate(&resv_map->regions, 0);
405 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
407 VM_BUG_ON(!is_vm_hugetlb_page(vma));
408 if (!(vma->vm_flags & VM_MAYSHARE))
409 return (struct resv_map *)(get_vma_private_data(vma) &
414 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
416 VM_BUG_ON(!is_vm_hugetlb_page(vma));
417 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
419 set_vma_private_data(vma, (get_vma_private_data(vma) &
420 HPAGE_RESV_MASK) | (unsigned long)map);
423 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
425 VM_BUG_ON(!is_vm_hugetlb_page(vma));
426 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
428 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
431 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
433 VM_BUG_ON(!is_vm_hugetlb_page(vma));
435 return (get_vma_private_data(vma) & flag) != 0;
438 /* Decrement the reserved pages in the hugepage pool by one */
439 static void decrement_hugepage_resv_vma(struct hstate *h,
440 struct vm_area_struct *vma)
442 if (vma->vm_flags & VM_NORESERVE)
445 if (vma->vm_flags & VM_MAYSHARE) {
446 /* Shared mappings always use reserves */
447 h->resv_huge_pages--;
448 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
450 * Only the process that called mmap() has reserves for
453 h->resv_huge_pages--;
457 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
458 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
460 VM_BUG_ON(!is_vm_hugetlb_page(vma));
461 if (!(vma->vm_flags & VM_MAYSHARE))
462 vma->vm_private_data = (void *)0;
465 /* Returns true if the VMA has associated reserve pages */
466 static int vma_has_reserves(struct vm_area_struct *vma)
468 if (vma->vm_flags & VM_MAYSHARE)
470 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
475 static void copy_gigantic_page(struct page *dst, struct page *src)
478 struct hstate *h = page_hstate(src);
479 struct page *dst_base = dst;
480 struct page *src_base = src;
482 for (i = 0; i < pages_per_huge_page(h); ) {
484 copy_highpage(dst, src);
487 dst = mem_map_next(dst, dst_base, i);
488 src = mem_map_next(src, src_base, i);
492 void copy_huge_page(struct page *dst, struct page *src)
495 struct hstate *h = page_hstate(src);
497 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
498 copy_gigantic_page(dst, src);
503 for (i = 0; i < pages_per_huge_page(h); i++) {
505 copy_highpage(dst + i, src + i);
509 static void enqueue_huge_page(struct hstate *h, struct page *page)
511 int nid = page_to_nid(page);
512 list_add(&page->lru, &h->hugepage_freelists[nid]);
513 h->free_huge_pages++;
514 h->free_huge_pages_node[nid]++;
517 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
521 if (list_empty(&h->hugepage_freelists[nid]))
523 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
524 list_del(&page->lru);
525 set_page_refcounted(page);
526 h->free_huge_pages--;
527 h->free_huge_pages_node[nid]--;
531 static struct page *dequeue_huge_page_vma(struct hstate *h,
532 struct vm_area_struct *vma,
533 unsigned long address, int avoid_reserve)
535 struct page *page = NULL;
536 struct mempolicy *mpol;
537 nodemask_t *nodemask;
538 struct zonelist *zonelist;
541 unsigned int cpuset_mems_cookie;
544 cpuset_mems_cookie = get_mems_allowed();
545 zonelist = huge_zonelist(vma, address,
546 htlb_alloc_mask, &mpol, &nodemask);
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) &&
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)
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))
582 static void update_and_free_page(struct hstate *h, struct page *page)
586 VM_BUG_ON(h->order >= MAX_ORDER);
589 h->nr_huge_pages_node[page_to_nid(page)]--;
590 for (i = 0; i < pages_per_huge_page(h); i++) {
591 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
592 1 << PG_referenced | 1 << PG_dirty |
593 1 << PG_active | 1 << PG_reserved |
594 1 << PG_private | 1 << PG_writeback);
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));
628 INIT_LIST_HEAD(&page->lru);
630 spin_lock(&hugetlb_lock);
631 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
632 update_and_free_page(h, page);
633 h->surplus_huge_pages--;
634 h->surplus_huge_pages_node[nid]--;
636 arch_clear_hugepage_flags(page);
637 enqueue_huge_page(h, page);
639 spin_unlock(&hugetlb_lock);
640 hugepage_subpool_put_pages(spool, 1);
643 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
645 set_compound_page_dtor(page, free_huge_page);
646 spin_lock(&hugetlb_lock);
648 h->nr_huge_pages_node[nid]++;
649 spin_unlock(&hugetlb_lock);
650 put_page(page); /* free it into the hugepage allocator */
653 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
656 int nr_pages = 1 << order;
657 struct page *p = page + 1;
659 /* we rely on prep_new_huge_page to set the destructor */
660 set_compound_order(page, order);
662 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
664 set_page_count(p, 0);
665 p->first_page = page;
669 int PageHuge(struct page *page)
671 compound_page_dtor *dtor;
673 if (!PageCompound(page))
676 page = compound_head(page);
677 dtor = get_compound_page_dtor(page);
679 return dtor == free_huge_page;
681 EXPORT_SYMBOL_GPL(PageHuge);
683 pgoff_t __basepage_index(struct page *page)
685 struct page *page_head = compound_head(page);
686 pgoff_t index = page_index(page_head);
687 unsigned long compound_idx;
689 if (!PageHuge(page_head))
690 return page_index(page);
692 if (compound_order(page_head) >= MAX_ORDER)
693 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
695 compound_idx = page - page_head;
697 return (index << compound_order(page_head)) + compound_idx;
700 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
704 if (h->order >= MAX_ORDER)
707 page = alloc_pages_exact_node(nid,
708 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
709 __GFP_REPEAT|__GFP_NOWARN,
712 if (arch_prepare_hugepage(page)) {
713 __free_pages(page, huge_page_order(h));
716 prep_new_huge_page(h, page, nid);
723 * common helper functions for hstate_next_node_to_{alloc|free}.
724 * We may have allocated or freed a huge page based on a different
725 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
726 * be outside of *nodes_allowed. Ensure that we use an allowed
727 * node for alloc or free.
729 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
731 nid = next_node(nid, *nodes_allowed);
732 if (nid == MAX_NUMNODES)
733 nid = first_node(*nodes_allowed);
734 VM_BUG_ON(nid >= MAX_NUMNODES);
739 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
741 if (!node_isset(nid, *nodes_allowed))
742 nid = next_node_allowed(nid, nodes_allowed);
747 * returns the previously saved node ["this node"] from which to
748 * allocate a persistent huge page for the pool and advance the
749 * next node from which to allocate, handling wrap at end of node
752 static int hstate_next_node_to_alloc(struct hstate *h,
753 nodemask_t *nodes_allowed)
757 VM_BUG_ON(!nodes_allowed);
759 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
760 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
765 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
772 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
773 next_nid = start_nid;
776 page = alloc_fresh_huge_page_node(h, next_nid);
781 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
782 } while (next_nid != start_nid);
785 count_vm_event(HTLB_BUDDY_PGALLOC);
787 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
793 * helper for free_pool_huge_page() - return the previously saved
794 * node ["this node"] from which to free a huge page. Advance the
795 * next node id whether or not we find a free huge page to free so
796 * that the next attempt to free addresses the next node.
798 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
802 VM_BUG_ON(!nodes_allowed);
804 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
805 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
811 * Free huge page from pool from next node to free.
812 * Attempt to keep persistent huge pages more or less
813 * balanced over allowed nodes.
814 * Called with hugetlb_lock locked.
816 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
823 start_nid = hstate_next_node_to_free(h, nodes_allowed);
824 next_nid = start_nid;
828 * If we're returning unused surplus pages, only examine
829 * nodes with surplus pages.
831 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
832 !list_empty(&h->hugepage_freelists[next_nid])) {
834 list_entry(h->hugepage_freelists[next_nid].next,
836 list_del(&page->lru);
837 h->free_huge_pages--;
838 h->free_huge_pages_node[next_nid]--;
840 h->surplus_huge_pages--;
841 h->surplus_huge_pages_node[next_nid]--;
843 update_and_free_page(h, page);
847 next_nid = hstate_next_node_to_free(h, nodes_allowed);
848 } while (next_nid != start_nid);
853 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
858 if (h->order >= MAX_ORDER)
862 * Assume we will successfully allocate the surplus page to
863 * prevent racing processes from causing the surplus to exceed
866 * This however introduces a different race, where a process B
867 * tries to grow the static hugepage pool while alloc_pages() is
868 * called by process A. B will only examine the per-node
869 * counters in determining if surplus huge pages can be
870 * converted to normal huge pages in adjust_pool_surplus(). A
871 * won't be able to increment the per-node counter, until the
872 * lock is dropped by B, but B doesn't drop hugetlb_lock until
873 * no more huge pages can be converted from surplus to normal
874 * state (and doesn't try to convert again). Thus, we have a
875 * case where a surplus huge page exists, the pool is grown, and
876 * the surplus huge page still exists after, even though it
877 * should just have been converted to a normal huge page. This
878 * does not leak memory, though, as the hugepage will be freed
879 * once it is out of use. It also does not allow the counters to
880 * go out of whack in adjust_pool_surplus() as we don't modify
881 * the node values until we've gotten the hugepage and only the
882 * per-node value is checked there.
884 spin_lock(&hugetlb_lock);
885 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
886 spin_unlock(&hugetlb_lock);
890 h->surplus_huge_pages++;
892 spin_unlock(&hugetlb_lock);
894 if (nid == NUMA_NO_NODE)
895 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
896 __GFP_REPEAT|__GFP_NOWARN,
899 page = alloc_pages_exact_node(nid,
900 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
901 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
903 if (page && arch_prepare_hugepage(page)) {
904 __free_pages(page, huge_page_order(h));
908 spin_lock(&hugetlb_lock);
910 r_nid = page_to_nid(page);
911 set_compound_page_dtor(page, free_huge_page);
913 * We incremented the global counters already
915 h->nr_huge_pages_node[r_nid]++;
916 h->surplus_huge_pages_node[r_nid]++;
917 __count_vm_event(HTLB_BUDDY_PGALLOC);
920 h->surplus_huge_pages--;
921 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
923 spin_unlock(&hugetlb_lock);
929 * This allocation function is useful in the context where vma is irrelevant.
930 * E.g. soft-offlining uses this function because it only cares physical
931 * address of error page.
933 struct page *alloc_huge_page_node(struct hstate *h, int nid)
937 spin_lock(&hugetlb_lock);
938 page = dequeue_huge_page_node(h, nid);
939 spin_unlock(&hugetlb_lock);
942 page = alloc_buddy_huge_page(h, nid);
948 * Increase the hugetlb pool such that it can accommodate a reservation
951 static int gather_surplus_pages(struct hstate *h, int delta)
953 struct list_head surplus_list;
954 struct page *page, *tmp;
956 int needed, allocated;
958 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
960 h->resv_huge_pages += delta;
965 INIT_LIST_HEAD(&surplus_list);
969 spin_unlock(&hugetlb_lock);
970 for (i = 0; i < needed; i++) {
971 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
974 * We were not able to allocate enough pages to
975 * satisfy the entire reservation so we free what
976 * we've allocated so far.
980 list_add(&page->lru, &surplus_list);
985 * After retaking hugetlb_lock, we need to recalculate 'needed'
986 * because either resv_huge_pages or free_huge_pages may have changed.
988 spin_lock(&hugetlb_lock);
989 needed = (h->resv_huge_pages + delta) -
990 (h->free_huge_pages + allocated);
995 * The surplus_list now contains _at_least_ the number of extra pages
996 * needed to accommodate the reservation. Add the appropriate number
997 * of pages to the hugetlb pool and free the extras back to the buddy
998 * allocator. Commit the entire reservation here to prevent another
999 * process from stealing the pages as they are added to the pool but
1000 * before they are reserved.
1002 needed += allocated;
1003 h->resv_huge_pages += delta;
1006 /* Free the needed pages to the hugetlb pool */
1007 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1010 list_del(&page->lru);
1012 * This page is now managed by the hugetlb allocator and has
1013 * no users -- drop the buddy allocator's reference.
1015 put_page_testzero(page);
1016 VM_BUG_ON(page_count(page));
1017 enqueue_huge_page(h, page);
1019 spin_unlock(&hugetlb_lock);
1021 /* Free unnecessary surplus pages to the buddy allocator */
1023 if (!list_empty(&surplus_list)) {
1024 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1025 list_del(&page->lru);
1029 spin_lock(&hugetlb_lock);
1035 * When releasing a hugetlb pool reservation, any surplus pages that were
1036 * allocated to satisfy the reservation must be explicitly freed if they were
1038 * Called with hugetlb_lock held.
1040 static void return_unused_surplus_pages(struct hstate *h,
1041 unsigned long unused_resv_pages)
1043 unsigned long nr_pages;
1045 /* Uncommit the reservation */
1046 h->resv_huge_pages -= unused_resv_pages;
1048 /* Cannot return gigantic pages currently */
1049 if (h->order >= MAX_ORDER)
1052 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1055 * We want to release as many surplus pages as possible, spread
1056 * evenly across all nodes with memory. Iterate across these nodes
1057 * until we can no longer free unreserved surplus pages. This occurs
1058 * when the nodes with surplus pages have no free pages.
1059 * free_pool_huge_page() will balance the the freed pages across the
1060 * on-line nodes with memory and will handle the hstate accounting.
1062 while (nr_pages--) {
1063 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
1069 * Determine if the huge page at addr within the vma has an associated
1070 * reservation. Where it does not we will need to logically increase
1071 * reservation and actually increase subpool usage before an allocation
1072 * can occur. Where any new reservation would be required the
1073 * reservation change is prepared, but not committed. Once the page
1074 * has been allocated from the subpool and instantiated the change should
1075 * be committed via vma_commit_reservation. No action is required on
1078 static long vma_needs_reservation(struct hstate *h,
1079 struct vm_area_struct *vma, unsigned long addr)
1081 struct address_space *mapping = vma->vm_file->f_mapping;
1082 struct inode *inode = mapping->host;
1084 if (vma->vm_flags & VM_MAYSHARE) {
1085 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1086 return region_chg(&inode->i_mapping->private_list,
1089 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1094 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1095 struct resv_map *reservations = vma_resv_map(vma);
1097 err = region_chg(&reservations->regions, idx, idx + 1);
1103 static void vma_commit_reservation(struct hstate *h,
1104 struct vm_area_struct *vma, unsigned long addr)
1106 struct address_space *mapping = vma->vm_file->f_mapping;
1107 struct inode *inode = mapping->host;
1109 if (vma->vm_flags & VM_MAYSHARE) {
1110 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1111 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1113 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1114 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1115 struct resv_map *reservations = vma_resv_map(vma);
1117 /* Mark this page used in the map. */
1118 region_add(&reservations->regions, idx, idx + 1);
1122 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1123 unsigned long addr, int avoid_reserve)
1125 struct hugepage_subpool *spool = subpool_vma(vma);
1126 struct hstate *h = hstate_vma(vma);
1131 * Processes that did not create the mapping will have no
1132 * reserves and will not have accounted against subpool
1133 * limit. Check that the subpool limit can be made before
1134 * satisfying the allocation MAP_NORESERVE mappings may also
1135 * need pages and subpool limit allocated allocated if no reserve
1138 chg = vma_needs_reservation(h, vma, addr);
1140 return ERR_PTR(-VM_FAULT_OOM);
1142 if (hugepage_subpool_get_pages(spool, chg))
1143 return ERR_PTR(-VM_FAULT_SIGBUS);
1145 spin_lock(&hugetlb_lock);
1146 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1147 spin_unlock(&hugetlb_lock);
1150 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1152 hugepage_subpool_put_pages(spool, chg);
1153 return ERR_PTR(-VM_FAULT_SIGBUS);
1157 set_page_private(page, (unsigned long)spool);
1159 vma_commit_reservation(h, vma, addr);
1164 int __weak alloc_bootmem_huge_page(struct hstate *h)
1166 struct huge_bootmem_page *m;
1167 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1172 addr = __alloc_bootmem_node_nopanic(
1173 NODE_DATA(hstate_next_node_to_alloc(h,
1174 &node_states[N_HIGH_MEMORY])),
1175 huge_page_size(h), huge_page_size(h), 0);
1179 * Use the beginning of the huge page to store the
1180 * huge_bootmem_page struct (until gather_bootmem
1181 * puts them into the mem_map).
1191 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1192 /* Put them into a private list first because mem_map is not up yet */
1193 list_add(&m->list, &huge_boot_pages);
1198 static void prep_compound_huge_page(struct page *page, int order)
1200 if (unlikely(order > (MAX_ORDER - 1)))
1201 prep_compound_gigantic_page(page, order);
1203 prep_compound_page(page, order);
1206 /* Put bootmem huge pages into the standard lists after mem_map is up */
1207 static void __init gather_bootmem_prealloc(void)
1209 struct huge_bootmem_page *m;
1211 list_for_each_entry(m, &huge_boot_pages, list) {
1212 struct hstate *h = m->hstate;
1215 #ifdef CONFIG_HIGHMEM
1216 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1217 free_bootmem_late((unsigned long)m,
1218 sizeof(struct huge_bootmem_page));
1220 page = virt_to_page(m);
1222 __ClearPageReserved(page);
1223 WARN_ON(page_count(page) != 1);
1224 prep_compound_huge_page(page, h->order);
1225 prep_new_huge_page(h, page, page_to_nid(page));
1227 * If we had gigantic hugepages allocated at boot time, we need
1228 * to restore the 'stolen' pages to totalram_pages in order to
1229 * fix confusing memory reports from free(1) and another
1230 * side-effects, like CommitLimit going negative.
1232 if (h->order > (MAX_ORDER - 1))
1233 totalram_pages += 1 << h->order;
1237 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1241 for (i = 0; i < h->max_huge_pages; ++i) {
1242 if (h->order >= MAX_ORDER) {
1243 if (!alloc_bootmem_huge_page(h))
1245 } else if (!alloc_fresh_huge_page(h,
1246 &node_states[N_HIGH_MEMORY]))
1249 h->max_huge_pages = i;
1252 static void __init hugetlb_init_hstates(void)
1256 for_each_hstate(h) {
1257 /* oversize hugepages were init'ed in early boot */
1258 if (h->order < MAX_ORDER)
1259 hugetlb_hstate_alloc_pages(h);
1263 static char * __init memfmt(char *buf, unsigned long n)
1265 if (n >= (1UL << 30))
1266 sprintf(buf, "%lu GB", n >> 30);
1267 else if (n >= (1UL << 20))
1268 sprintf(buf, "%lu MB", n >> 20);
1270 sprintf(buf, "%lu KB", n >> 10);
1274 static void __init report_hugepages(void)
1278 for_each_hstate(h) {
1280 printk(KERN_INFO "HugeTLB registered %s page size, "
1281 "pre-allocated %ld pages\n",
1282 memfmt(buf, huge_page_size(h)),
1283 h->free_huge_pages);
1287 #ifdef CONFIG_HIGHMEM
1288 static void try_to_free_low(struct hstate *h, unsigned long count,
1289 nodemask_t *nodes_allowed)
1293 if (h->order >= MAX_ORDER)
1296 for_each_node_mask(i, *nodes_allowed) {
1297 struct page *page, *next;
1298 struct list_head *freel = &h->hugepage_freelists[i];
1299 list_for_each_entry_safe(page, next, freel, lru) {
1300 if (count >= h->nr_huge_pages)
1302 if (PageHighMem(page))
1304 list_del(&page->lru);
1305 update_and_free_page(h, page);
1306 h->free_huge_pages--;
1307 h->free_huge_pages_node[page_to_nid(page)]--;
1312 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1313 nodemask_t *nodes_allowed)
1319 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1320 * balanced by operating on them in a round-robin fashion.
1321 * Returns 1 if an adjustment was made.
1323 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1326 int start_nid, next_nid;
1329 VM_BUG_ON(delta != -1 && delta != 1);
1332 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1334 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1335 next_nid = start_nid;
1341 * To shrink on this node, there must be a surplus page
1343 if (!h->surplus_huge_pages_node[nid]) {
1344 next_nid = hstate_next_node_to_alloc(h,
1351 * Surplus cannot exceed the total number of pages
1353 if (h->surplus_huge_pages_node[nid] >=
1354 h->nr_huge_pages_node[nid]) {
1355 next_nid = hstate_next_node_to_free(h,
1361 h->surplus_huge_pages += delta;
1362 h->surplus_huge_pages_node[nid] += delta;
1365 } while (next_nid != start_nid);
1370 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1371 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1372 nodemask_t *nodes_allowed)
1374 unsigned long min_count, ret;
1376 if (h->order >= MAX_ORDER)
1377 return h->max_huge_pages;
1380 * Increase the pool size
1381 * First take pages out of surplus state. Then make up the
1382 * remaining difference by allocating fresh huge pages.
1384 * We might race with alloc_buddy_huge_page() here and be unable
1385 * to convert a surplus huge page to a normal huge page. That is
1386 * not critical, though, it just means the overall size of the
1387 * pool might be one hugepage larger than it needs to be, but
1388 * within all the constraints specified by the sysctls.
1390 spin_lock(&hugetlb_lock);
1391 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1392 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1396 while (count > persistent_huge_pages(h)) {
1398 * If this allocation races such that we no longer need the
1399 * page, free_huge_page will handle it by freeing the page
1400 * and reducing the surplus.
1402 spin_unlock(&hugetlb_lock);
1403 ret = alloc_fresh_huge_page(h, nodes_allowed);
1404 spin_lock(&hugetlb_lock);
1408 /* Bail for signals. Probably ctrl-c from user */
1409 if (signal_pending(current))
1414 * Decrease the pool size
1415 * First return free pages to the buddy allocator (being careful
1416 * to keep enough around to satisfy reservations). Then place
1417 * pages into surplus state as needed so the pool will shrink
1418 * to the desired size as pages become free.
1420 * By placing pages into the surplus state independent of the
1421 * overcommit value, we are allowing the surplus pool size to
1422 * exceed overcommit. There are few sane options here. Since
1423 * alloc_buddy_huge_page() is checking the global counter,
1424 * though, we'll note that we're not allowed to exceed surplus
1425 * and won't grow the pool anywhere else. Not until one of the
1426 * sysctls are changed, or the surplus pages go out of use.
1428 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1429 min_count = max(count, min_count);
1430 try_to_free_low(h, min_count, nodes_allowed);
1431 while (min_count < persistent_huge_pages(h)) {
1432 if (!free_pool_huge_page(h, nodes_allowed, 0))
1435 while (count < persistent_huge_pages(h)) {
1436 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1440 ret = persistent_huge_pages(h);
1441 spin_unlock(&hugetlb_lock);
1445 #define HSTATE_ATTR_RO(_name) \
1446 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1448 #define HSTATE_ATTR(_name) \
1449 static struct kobj_attribute _name##_attr = \
1450 __ATTR(_name, 0644, _name##_show, _name##_store)
1452 static struct kobject *hugepages_kobj;
1453 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1455 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1457 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1461 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1462 if (hstate_kobjs[i] == kobj) {
1464 *nidp = NUMA_NO_NODE;
1468 return kobj_to_node_hstate(kobj, nidp);
1471 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1472 struct kobj_attribute *attr, char *buf)
1475 unsigned long nr_huge_pages;
1478 h = kobj_to_hstate(kobj, &nid);
1479 if (nid == NUMA_NO_NODE)
1480 nr_huge_pages = h->nr_huge_pages;
1482 nr_huge_pages = h->nr_huge_pages_node[nid];
1484 return sprintf(buf, "%lu\n", nr_huge_pages);
1487 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1488 struct kobject *kobj, struct kobj_attribute *attr,
1489 const char *buf, size_t len)
1493 unsigned long count;
1495 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1497 err = strict_strtoul(buf, 10, &count);
1501 h = kobj_to_hstate(kobj, &nid);
1502 if (h->order >= MAX_ORDER) {
1507 if (nid == NUMA_NO_NODE) {
1509 * global hstate attribute
1511 if (!(obey_mempolicy &&
1512 init_nodemask_of_mempolicy(nodes_allowed))) {
1513 NODEMASK_FREE(nodes_allowed);
1514 nodes_allowed = &node_states[N_HIGH_MEMORY];
1516 } else if (nodes_allowed) {
1518 * per node hstate attribute: adjust count to global,
1519 * but restrict alloc/free to the specified node.
1521 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1522 init_nodemask_of_node(nodes_allowed, nid);
1524 nodes_allowed = &node_states[N_HIGH_MEMORY];
1526 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1528 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1529 NODEMASK_FREE(nodes_allowed);
1533 NODEMASK_FREE(nodes_allowed);
1537 static ssize_t nr_hugepages_show(struct kobject *kobj,
1538 struct kobj_attribute *attr, char *buf)
1540 return nr_hugepages_show_common(kobj, attr, buf);
1543 static ssize_t nr_hugepages_store(struct kobject *kobj,
1544 struct kobj_attribute *attr, const char *buf, size_t len)
1546 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1548 HSTATE_ATTR(nr_hugepages);
1553 * hstate attribute for optionally mempolicy-based constraint on persistent
1554 * huge page alloc/free.
1556 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1557 struct kobj_attribute *attr, char *buf)
1559 return nr_hugepages_show_common(kobj, attr, buf);
1562 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1563 struct kobj_attribute *attr, const char *buf, size_t len)
1565 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1567 HSTATE_ATTR(nr_hugepages_mempolicy);
1571 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1572 struct kobj_attribute *attr, char *buf)
1574 struct hstate *h = kobj_to_hstate(kobj, NULL);
1575 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1578 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1579 struct kobj_attribute *attr, const char *buf, size_t count)
1582 unsigned long input;
1583 struct hstate *h = kobj_to_hstate(kobj, NULL);
1585 if (h->order >= MAX_ORDER)
1588 err = strict_strtoul(buf, 10, &input);
1592 spin_lock(&hugetlb_lock);
1593 h->nr_overcommit_huge_pages = input;
1594 spin_unlock(&hugetlb_lock);
1598 HSTATE_ATTR(nr_overcommit_hugepages);
1600 static ssize_t free_hugepages_show(struct kobject *kobj,
1601 struct kobj_attribute *attr, char *buf)
1604 unsigned long free_huge_pages;
1607 h = kobj_to_hstate(kobj, &nid);
1608 if (nid == NUMA_NO_NODE)
1609 free_huge_pages = h->free_huge_pages;
1611 free_huge_pages = h->free_huge_pages_node[nid];
1613 return sprintf(buf, "%lu\n", free_huge_pages);
1615 HSTATE_ATTR_RO(free_hugepages);
1617 static ssize_t resv_hugepages_show(struct kobject *kobj,
1618 struct kobj_attribute *attr, char *buf)
1620 struct hstate *h = kobj_to_hstate(kobj, NULL);
1621 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1623 HSTATE_ATTR_RO(resv_hugepages);
1625 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1626 struct kobj_attribute *attr, char *buf)
1629 unsigned long surplus_huge_pages;
1632 h = kobj_to_hstate(kobj, &nid);
1633 if (nid == NUMA_NO_NODE)
1634 surplus_huge_pages = h->surplus_huge_pages;
1636 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1638 return sprintf(buf, "%lu\n", surplus_huge_pages);
1640 HSTATE_ATTR_RO(surplus_hugepages);
1642 static struct attribute *hstate_attrs[] = {
1643 &nr_hugepages_attr.attr,
1644 &nr_overcommit_hugepages_attr.attr,
1645 &free_hugepages_attr.attr,
1646 &resv_hugepages_attr.attr,
1647 &surplus_hugepages_attr.attr,
1649 &nr_hugepages_mempolicy_attr.attr,
1654 static struct attribute_group hstate_attr_group = {
1655 .attrs = hstate_attrs,
1658 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1659 struct kobject **hstate_kobjs,
1660 struct attribute_group *hstate_attr_group)
1663 int hi = h - hstates;
1665 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1666 if (!hstate_kobjs[hi])
1669 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1671 kobject_put(hstate_kobjs[hi]);
1676 static void __init hugetlb_sysfs_init(void)
1681 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1682 if (!hugepages_kobj)
1685 for_each_hstate(h) {
1686 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1687 hstate_kobjs, &hstate_attr_group);
1689 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1697 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1698 * with node devices in node_devices[] using a parallel array. The array
1699 * index of a node device or _hstate == node id.
1700 * This is here to avoid any static dependency of the node device driver, in
1701 * the base kernel, on the hugetlb module.
1703 struct node_hstate {
1704 struct kobject *hugepages_kobj;
1705 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1707 struct node_hstate node_hstates[MAX_NUMNODES];
1710 * A subset of global hstate attributes for node devices
1712 static struct attribute *per_node_hstate_attrs[] = {
1713 &nr_hugepages_attr.attr,
1714 &free_hugepages_attr.attr,
1715 &surplus_hugepages_attr.attr,
1719 static struct attribute_group per_node_hstate_attr_group = {
1720 .attrs = per_node_hstate_attrs,
1724 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1725 * Returns node id via non-NULL nidp.
1727 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1731 for (nid = 0; nid < nr_node_ids; nid++) {
1732 struct node_hstate *nhs = &node_hstates[nid];
1734 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1735 if (nhs->hstate_kobjs[i] == kobj) {
1747 * Unregister hstate attributes from a single node device.
1748 * No-op if no hstate attributes attached.
1750 void hugetlb_unregister_node(struct node *node)
1753 struct node_hstate *nhs = &node_hstates[node->dev.id];
1755 if (!nhs->hugepages_kobj)
1756 return; /* no hstate attributes */
1759 if (nhs->hstate_kobjs[h - hstates]) {
1760 kobject_put(nhs->hstate_kobjs[h - hstates]);
1761 nhs->hstate_kobjs[h - hstates] = NULL;
1764 kobject_put(nhs->hugepages_kobj);
1765 nhs->hugepages_kobj = NULL;
1769 * hugetlb module exit: unregister hstate attributes from node devices
1772 static void hugetlb_unregister_all_nodes(void)
1777 * disable node device registrations.
1779 register_hugetlbfs_with_node(NULL, NULL);
1782 * remove hstate attributes from any nodes that have them.
1784 for (nid = 0; nid < nr_node_ids; nid++)
1785 hugetlb_unregister_node(&node_devices[nid]);
1789 * Register hstate attributes for a single node device.
1790 * No-op if attributes already registered.
1792 void hugetlb_register_node(struct node *node)
1795 struct node_hstate *nhs = &node_hstates[node->dev.id];
1798 if (nhs->hugepages_kobj)
1799 return; /* already allocated */
1801 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1803 if (!nhs->hugepages_kobj)
1806 for_each_hstate(h) {
1807 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1809 &per_node_hstate_attr_group);
1811 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1813 h->name, node->dev.id);
1814 hugetlb_unregister_node(node);
1821 * hugetlb init time: register hstate attributes for all registered node
1822 * devices of nodes that have memory. All on-line nodes should have
1823 * registered their associated device by this time.
1825 static void hugetlb_register_all_nodes(void)
1829 for_each_node_state(nid, N_HIGH_MEMORY) {
1830 struct node *node = &node_devices[nid];
1831 if (node->dev.id == nid)
1832 hugetlb_register_node(node);
1836 * Let the node device driver know we're here so it can
1837 * [un]register hstate attributes on node hotplug.
1839 register_hugetlbfs_with_node(hugetlb_register_node,
1840 hugetlb_unregister_node);
1842 #else /* !CONFIG_NUMA */
1844 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1852 static void hugetlb_unregister_all_nodes(void) { }
1854 static void hugetlb_register_all_nodes(void) { }
1858 static void __exit hugetlb_exit(void)
1862 hugetlb_unregister_all_nodes();
1864 for_each_hstate(h) {
1865 kobject_put(hstate_kobjs[h - hstates]);
1868 kobject_put(hugepages_kobj);
1870 module_exit(hugetlb_exit);
1872 static int __init hugetlb_init(void)
1874 /* Some platform decide whether they support huge pages at boot
1875 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1876 * there is no such support
1878 if (HPAGE_SHIFT == 0)
1881 if (!size_to_hstate(default_hstate_size)) {
1882 default_hstate_size = HPAGE_SIZE;
1883 if (!size_to_hstate(default_hstate_size))
1884 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1886 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1887 if (default_hstate_max_huge_pages)
1888 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1890 hugetlb_init_hstates();
1892 gather_bootmem_prealloc();
1896 hugetlb_sysfs_init();
1898 hugetlb_register_all_nodes();
1902 module_init(hugetlb_init);
1904 /* Should be called on processing a hugepagesz=... option */
1905 void __init hugetlb_add_hstate(unsigned order)
1910 if (size_to_hstate(PAGE_SIZE << order)) {
1911 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1914 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1916 h = &hstates[max_hstate++];
1918 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1919 h->nr_huge_pages = 0;
1920 h->free_huge_pages = 0;
1921 for (i = 0; i < MAX_NUMNODES; ++i)
1922 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1923 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1924 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1925 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1926 huge_page_size(h)/1024);
1931 static int __init hugetlb_nrpages_setup(char *s)
1934 static unsigned long *last_mhp;
1937 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1938 * so this hugepages= parameter goes to the "default hstate".
1941 mhp = &default_hstate_max_huge_pages;
1943 mhp = &parsed_hstate->max_huge_pages;
1945 if (mhp == last_mhp) {
1946 printk(KERN_WARNING "hugepages= specified twice without "
1947 "interleaving hugepagesz=, ignoring\n");
1951 if (sscanf(s, "%lu", mhp) <= 0)
1955 * Global state is always initialized later in hugetlb_init.
1956 * But we need to allocate >= MAX_ORDER hstates here early to still
1957 * use the bootmem allocator.
1959 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1960 hugetlb_hstate_alloc_pages(parsed_hstate);
1966 __setup("hugepages=", hugetlb_nrpages_setup);
1968 static int __init hugetlb_default_setup(char *s)
1970 default_hstate_size = memparse(s, &s);
1973 __setup("default_hugepagesz=", hugetlb_default_setup);
1975 static unsigned int cpuset_mems_nr(unsigned int *array)
1978 unsigned int nr = 0;
1980 for_each_node_mask(node, cpuset_current_mems_allowed)
1986 #ifdef CONFIG_SYSCTL
1987 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1988 struct ctl_table *table, int write,
1989 void __user *buffer, size_t *length, loff_t *ppos)
1991 struct hstate *h = &default_hstate;
1995 tmp = h->max_huge_pages;
1997 if (write && h->order >= MAX_ORDER)
2001 table->maxlen = sizeof(unsigned long);
2002 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2007 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2008 GFP_KERNEL | __GFP_NORETRY);
2009 if (!(obey_mempolicy &&
2010 init_nodemask_of_mempolicy(nodes_allowed))) {
2011 NODEMASK_FREE(nodes_allowed);
2012 nodes_allowed = &node_states[N_HIGH_MEMORY];
2014 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2016 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
2017 NODEMASK_FREE(nodes_allowed);
2023 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2024 void __user *buffer, size_t *length, loff_t *ppos)
2027 return hugetlb_sysctl_handler_common(false, table, write,
2028 buffer, length, ppos);
2032 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2033 void __user *buffer, size_t *length, loff_t *ppos)
2035 return hugetlb_sysctl_handler_common(true, table, write,
2036 buffer, length, ppos);
2038 #endif /* CONFIG_NUMA */
2040 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2041 void __user *buffer,
2042 size_t *length, loff_t *ppos)
2044 proc_dointvec(table, write, buffer, length, ppos);
2045 if (hugepages_treat_as_movable)
2046 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2048 htlb_alloc_mask = GFP_HIGHUSER;
2052 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2053 void __user *buffer,
2054 size_t *length, loff_t *ppos)
2056 struct hstate *h = &default_hstate;
2060 tmp = h->nr_overcommit_huge_pages;
2062 if (write && h->order >= MAX_ORDER)
2066 table->maxlen = sizeof(unsigned long);
2067 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2072 spin_lock(&hugetlb_lock);
2073 h->nr_overcommit_huge_pages = tmp;
2074 spin_unlock(&hugetlb_lock);
2080 #endif /* CONFIG_SYSCTL */
2082 void hugetlb_report_meminfo(struct seq_file *m)
2084 struct hstate *h = &default_hstate;
2086 "HugePages_Total: %5lu\n"
2087 "HugePages_Free: %5lu\n"
2088 "HugePages_Rsvd: %5lu\n"
2089 "HugePages_Surp: %5lu\n"
2090 "Hugepagesize: %8lu kB\n",
2094 h->surplus_huge_pages,
2095 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2098 int hugetlb_report_node_meminfo(int nid, char *buf)
2100 struct hstate *h = &default_hstate;
2102 "Node %d HugePages_Total: %5u\n"
2103 "Node %d HugePages_Free: %5u\n"
2104 "Node %d HugePages_Surp: %5u\n",
2105 nid, h->nr_huge_pages_node[nid],
2106 nid, h->free_huge_pages_node[nid],
2107 nid, h->surplus_huge_pages_node[nid]);
2110 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2111 unsigned long hugetlb_total_pages(void)
2114 unsigned long nr_total_pages = 0;
2117 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2118 return nr_total_pages;
2121 static int hugetlb_acct_memory(struct hstate *h, long delta)
2125 spin_lock(&hugetlb_lock);
2127 * When cpuset is configured, it breaks the strict hugetlb page
2128 * reservation as the accounting is done on a global variable. Such
2129 * reservation is completely rubbish in the presence of cpuset because
2130 * the reservation is not checked against page availability for the
2131 * current cpuset. Application can still potentially OOM'ed by kernel
2132 * with lack of free htlb page in cpuset that the task is in.
2133 * Attempt to enforce strict accounting with cpuset is almost
2134 * impossible (or too ugly) because cpuset is too fluid that
2135 * task or memory node can be dynamically moved between cpusets.
2137 * The change of semantics for shared hugetlb mapping with cpuset is
2138 * undesirable. However, in order to preserve some of the semantics,
2139 * we fall back to check against current free page availability as
2140 * a best attempt and hopefully to minimize the impact of changing
2141 * semantics that cpuset has.
2144 if (gather_surplus_pages(h, delta) < 0)
2147 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2148 return_unused_surplus_pages(h, delta);
2155 return_unused_surplus_pages(h, (unsigned long) -delta);
2158 spin_unlock(&hugetlb_lock);
2162 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2164 struct resv_map *reservations = vma_resv_map(vma);
2167 * This new VMA should share its siblings reservation map if present.
2168 * The VMA will only ever have a valid reservation map pointer where
2169 * it is being copied for another still existing VMA. As that VMA
2170 * has a reference to the reservation map it cannot disappear until
2171 * after this open call completes. It is therefore safe to take a
2172 * new reference here without additional locking.
2175 kref_get(&reservations->refs);
2178 static void resv_map_put(struct vm_area_struct *vma)
2180 struct resv_map *reservations = vma_resv_map(vma);
2184 kref_put(&reservations->refs, resv_map_release);
2187 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2189 struct hstate *h = hstate_vma(vma);
2190 struct resv_map *reservations = vma_resv_map(vma);
2191 struct hugepage_subpool *spool = subpool_vma(vma);
2192 unsigned long reserve;
2193 unsigned long start;
2197 start = vma_hugecache_offset(h, vma, vma->vm_start);
2198 end = vma_hugecache_offset(h, vma, vma->vm_end);
2200 reserve = (end - start) -
2201 region_count(&reservations->regions, start, end);
2206 hugetlb_acct_memory(h, -reserve);
2207 hugepage_subpool_put_pages(spool, reserve);
2213 * We cannot handle pagefaults against hugetlb pages at all. They cause
2214 * handle_mm_fault() to try to instantiate regular-sized pages in the
2215 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2218 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2224 const struct vm_operations_struct hugetlb_vm_ops = {
2225 .fault = hugetlb_vm_op_fault,
2226 .open = hugetlb_vm_op_open,
2227 .close = hugetlb_vm_op_close,
2230 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2237 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2239 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2241 entry = pte_mkyoung(entry);
2242 entry = pte_mkhuge(entry);
2247 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2248 unsigned long address, pte_t *ptep)
2252 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2253 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2254 update_mmu_cache(vma, address, ptep);
2258 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2259 struct vm_area_struct *vma)
2261 pte_t *src_pte, *dst_pte, entry;
2262 struct page *ptepage;
2265 struct hstate *h = hstate_vma(vma);
2266 unsigned long sz = huge_page_size(h);
2268 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2270 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2271 src_pte = huge_pte_offset(src, addr);
2274 dst_pte = huge_pte_alloc(dst, addr, sz);
2278 /* If the pagetables are shared don't copy or take references */
2279 if (dst_pte == src_pte)
2282 spin_lock(&dst->page_table_lock);
2283 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2284 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2286 huge_ptep_set_wrprotect(src, addr, src_pte);
2287 entry = huge_ptep_get(src_pte);
2288 ptepage = pte_page(entry);
2290 page_dup_rmap(ptepage);
2291 set_huge_pte_at(dst, addr, dst_pte, entry);
2293 spin_unlock(&src->page_table_lock);
2294 spin_unlock(&dst->page_table_lock);
2302 static int is_hugetlb_entry_migration(pte_t pte)
2306 if (huge_pte_none(pte) || pte_present(pte))
2308 swp = pte_to_swp_entry(pte);
2309 if (non_swap_entry(swp) && is_migration_entry(swp))
2315 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2319 if (huge_pte_none(pte) || pte_present(pte))
2321 swp = pte_to_swp_entry(pte);
2322 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2328 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2329 unsigned long end, struct page *ref_page)
2331 struct mm_struct *mm = vma->vm_mm;
2332 unsigned long address;
2337 struct hstate *h = hstate_vma(vma);
2338 unsigned long sz = huge_page_size(h);
2341 * A page gathering list, protected by per file i_mmap_mutex. The
2342 * lock is used to avoid list corruption from multiple unmapping
2343 * of the same page since we are using page->lru.
2345 LIST_HEAD(page_list);
2347 WARN_ON(!is_vm_hugetlb_page(vma));
2348 BUG_ON(start & ~huge_page_mask(h));
2349 BUG_ON(end & ~huge_page_mask(h));
2351 mmu_notifier_invalidate_range_start(mm, start, end);
2352 spin_lock(&mm->page_table_lock);
2353 for (address = start; address < end; address += sz) {
2354 ptep = huge_pte_offset(mm, address);
2358 if (huge_pmd_unshare(mm, &address, ptep))
2362 * If a reference page is supplied, it is because a specific
2363 * page is being unmapped, not a range. Ensure the page we
2364 * are about to unmap is the actual page of interest.
2367 pte = huge_ptep_get(ptep);
2368 if (huge_pte_none(pte))
2370 page = pte_page(pte);
2371 if (page != ref_page)
2375 * Mark the VMA as having unmapped its page so that
2376 * future faults in this VMA will fail rather than
2377 * looking like data was lost
2379 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2382 pte = huge_ptep_get_and_clear(mm, address, ptep);
2383 if (huge_pte_none(pte))
2387 * HWPoisoned hugepage is already unmapped and dropped reference
2389 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2392 page = pte_page(pte);
2394 set_page_dirty(page);
2395 list_add(&page->lru, &page_list);
2397 spin_unlock(&mm->page_table_lock);
2398 flush_tlb_range(vma, start, end);
2399 mmu_notifier_invalidate_range_end(mm, start, end);
2400 list_for_each_entry_safe(page, tmp, &page_list, lru) {
2401 page_remove_rmap(page);
2402 list_del(&page->lru);
2407 void __unmap_hugepage_range_final(struct vm_area_struct *vma,
2408 unsigned long start, unsigned long end,
2409 struct page *ref_page)
2411 __unmap_hugepage_range(vma, start, end, ref_page);
2414 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2415 * test will fail on a vma being torn down, and not grab a page table
2416 * on its way out. We're lucky that the flag has such an appropriate
2417 * name, and can in fact be safely cleared here. We could clear it
2418 * before the __unmap_hugepage_range above, but all that's necessary
2419 * is to clear it before releasing the i_mmap_mutex. This works
2420 * because in the context this is called, the VMA is about to be
2421 * destroyed and the i_mmap_mutex is held.
2423 vma->vm_flags &= ~VM_MAYSHARE;
2426 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2427 unsigned long end, struct page *ref_page)
2429 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2430 __unmap_hugepage_range(vma, start, end, ref_page);
2431 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2435 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2436 * mappping it owns the reserve page for. The intention is to unmap the page
2437 * from other VMAs and let the children be SIGKILLed if they are faulting the
2440 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2441 struct page *page, unsigned long address)
2443 struct hstate *h = hstate_vma(vma);
2444 struct vm_area_struct *iter_vma;
2445 struct address_space *mapping;
2446 struct prio_tree_iter iter;
2450 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2451 * from page cache lookup which is in HPAGE_SIZE units.
2453 address = address & huge_page_mask(h);
2454 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2456 mapping = vma->vm_file->f_dentry->d_inode->i_mapping;
2459 * Take the mapping lock for the duration of the table walk. As
2460 * this mapping should be shared between all the VMAs,
2461 * __unmap_hugepage_range() is called as the lock is already held
2463 mutex_lock(&mapping->i_mmap_mutex);
2464 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2465 /* Do not unmap the current VMA */
2466 if (iter_vma == vma)
2470 * Unmap the page from other VMAs without their own reserves.
2471 * They get marked to be SIGKILLed if they fault in these
2472 * areas. This is because a future no-page fault on this VMA
2473 * could insert a zeroed page instead of the data existing
2474 * from the time of fork. This would look like data corruption
2476 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2477 __unmap_hugepage_range(iter_vma,
2478 address, address + huge_page_size(h),
2481 mutex_unlock(&mapping->i_mmap_mutex);
2487 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2489 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2490 unsigned long address, pte_t *ptep, pte_t pte,
2491 struct page *pagecache_page)
2493 struct hstate *h = hstate_vma(vma);
2494 struct page *old_page, *new_page;
2496 int outside_reserve = 0;
2498 old_page = pte_page(pte);
2501 /* If no-one else is actually using this page, avoid the copy
2502 * and just make the page writable */
2503 avoidcopy = (page_mapcount(old_page) == 1);
2505 if (PageAnon(old_page))
2506 page_move_anon_rmap(old_page, vma, address);
2507 set_huge_ptep_writable(vma, address, ptep);
2512 * If the process that created a MAP_PRIVATE mapping is about to
2513 * perform a COW due to a shared page count, attempt to satisfy
2514 * the allocation without using the existing reserves. The pagecache
2515 * page is used to determine if the reserve at this address was
2516 * consumed or not. If reserves were used, a partial faulted mapping
2517 * at the time of fork() could consume its reserves on COW instead
2518 * of the full address range.
2520 if (!(vma->vm_flags & VM_MAYSHARE) &&
2521 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2522 old_page != pagecache_page)
2523 outside_reserve = 1;
2525 page_cache_get(old_page);
2527 /* Drop page_table_lock as buddy allocator may be called */
2528 spin_unlock(&mm->page_table_lock);
2529 new_page = alloc_huge_page(vma, address, outside_reserve);
2531 if (IS_ERR(new_page)) {
2532 page_cache_release(old_page);
2535 * If a process owning a MAP_PRIVATE mapping fails to COW,
2536 * it is due to references held by a child and an insufficient
2537 * huge page pool. To guarantee the original mappers
2538 * reliability, unmap the page from child processes. The child
2539 * may get SIGKILLed if it later faults.
2541 if (outside_reserve) {
2542 BUG_ON(huge_pte_none(pte));
2543 if (unmap_ref_private(mm, vma, old_page, address)) {
2544 BUG_ON(huge_pte_none(pte));
2545 spin_lock(&mm->page_table_lock);
2546 goto retry_avoidcopy;
2551 /* Caller expects lock to be held */
2552 spin_lock(&mm->page_table_lock);
2553 return -PTR_ERR(new_page);
2557 * When the original hugepage is shared one, it does not have
2558 * anon_vma prepared.
2560 if (unlikely(anon_vma_prepare(vma))) {
2561 page_cache_release(new_page);
2562 page_cache_release(old_page);
2563 /* Caller expects lock to be held */
2564 spin_lock(&mm->page_table_lock);
2565 return VM_FAULT_OOM;
2568 copy_user_huge_page(new_page, old_page, address, vma,
2569 pages_per_huge_page(h));
2570 __SetPageUptodate(new_page);
2573 * Retake the page_table_lock to check for racing updates
2574 * before the page tables are altered
2576 spin_lock(&mm->page_table_lock);
2577 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2578 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2580 mmu_notifier_invalidate_range_start(mm,
2581 address & huge_page_mask(h),
2582 (address & huge_page_mask(h)) + huge_page_size(h));
2583 huge_ptep_clear_flush(vma, address, ptep);
2584 set_huge_pte_at(mm, address, ptep,
2585 make_huge_pte(vma, new_page, 1));
2586 page_remove_rmap(old_page);
2587 hugepage_add_new_anon_rmap(new_page, vma, address);
2588 /* Make the old page be freed below */
2589 new_page = old_page;
2590 mmu_notifier_invalidate_range_end(mm,
2591 address & huge_page_mask(h),
2592 (address & huge_page_mask(h)) + huge_page_size(h));
2594 page_cache_release(new_page);
2595 page_cache_release(old_page);
2599 /* Return the pagecache page at a given address within a VMA */
2600 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2601 struct vm_area_struct *vma, unsigned long address)
2603 struct address_space *mapping;
2606 mapping = vma->vm_file->f_mapping;
2607 idx = vma_hugecache_offset(h, vma, address);
2609 return find_lock_page(mapping, idx);
2613 * Return whether there is a pagecache page to back given address within VMA.
2614 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2616 static bool hugetlbfs_pagecache_present(struct hstate *h,
2617 struct vm_area_struct *vma, unsigned long address)
2619 struct address_space *mapping;
2623 mapping = vma->vm_file->f_mapping;
2624 idx = vma_hugecache_offset(h, vma, address);
2626 page = find_get_page(mapping, idx);
2629 return page != NULL;
2632 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2633 unsigned long address, pte_t *ptep, unsigned int flags)
2635 struct hstate *h = hstate_vma(vma);
2636 int ret = VM_FAULT_SIGBUS;
2640 struct address_space *mapping;
2644 * Currently, we are forced to kill the process in the event the
2645 * original mapper has unmapped pages from the child due to a failed
2646 * COW. Warn that such a situation has occurred as it may not be obvious
2648 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2650 "PID %d killed due to inadequate hugepage pool\n",
2655 mapping = vma->vm_file->f_mapping;
2656 idx = vma_hugecache_offset(h, vma, address);
2659 * Use page lock to guard against racing truncation
2660 * before we get page_table_lock.
2663 page = find_lock_page(mapping, idx);
2665 size = i_size_read(mapping->host) >> huge_page_shift(h);
2668 page = alloc_huge_page(vma, address, 0);
2670 ret = -PTR_ERR(page);
2673 clear_huge_page(page, address, pages_per_huge_page(h));
2674 __SetPageUptodate(page);
2676 if (vma->vm_flags & VM_MAYSHARE) {
2678 struct inode *inode = mapping->host;
2680 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2688 spin_lock(&inode->i_lock);
2689 inode->i_blocks += blocks_per_huge_page(h);
2690 spin_unlock(&inode->i_lock);
2691 page_dup_rmap(page);
2694 if (unlikely(anon_vma_prepare(vma))) {
2696 goto backout_unlocked;
2698 hugepage_add_new_anon_rmap(page, vma, address);
2702 * If memory error occurs between mmap() and fault, some process
2703 * don't have hwpoisoned swap entry for errored virtual address.
2704 * So we need to block hugepage fault by PG_hwpoison bit check.
2706 if (unlikely(PageHWPoison(page))) {
2707 ret = VM_FAULT_HWPOISON |
2708 VM_FAULT_SET_HINDEX(h - hstates);
2709 goto backout_unlocked;
2711 page_dup_rmap(page);
2715 * If we are going to COW a private mapping later, we examine the
2716 * pending reservations for this page now. This will ensure that
2717 * any allocations necessary to record that reservation occur outside
2720 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2721 if (vma_needs_reservation(h, vma, address) < 0) {
2723 goto backout_unlocked;
2726 spin_lock(&mm->page_table_lock);
2727 size = i_size_read(mapping->host) >> huge_page_shift(h);
2732 if (!huge_pte_none(huge_ptep_get(ptep)))
2735 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2736 && (vma->vm_flags & VM_SHARED)));
2737 set_huge_pte_at(mm, address, ptep, new_pte);
2739 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2740 /* Optimization, do the COW without a second fault */
2741 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2744 spin_unlock(&mm->page_table_lock);
2750 spin_unlock(&mm->page_table_lock);
2757 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2758 unsigned long address, unsigned int flags)
2763 struct page *page = NULL;
2764 struct page *pagecache_page = NULL;
2765 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2766 struct hstate *h = hstate_vma(vma);
2768 ptep = huge_pte_offset(mm, address);
2770 entry = huge_ptep_get(ptep);
2771 if (unlikely(is_hugetlb_entry_migration(entry))) {
2772 migration_entry_wait_huge(mm, ptep);
2774 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2775 return VM_FAULT_HWPOISON_LARGE |
2776 VM_FAULT_SET_HINDEX(h - hstates);
2779 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2781 return VM_FAULT_OOM;
2784 * Serialize hugepage allocation and instantiation, so that we don't
2785 * get spurious allocation failures if two CPUs race to instantiate
2786 * the same page in the page cache.
2788 mutex_lock(&hugetlb_instantiation_mutex);
2789 entry = huge_ptep_get(ptep);
2790 if (huge_pte_none(entry)) {
2791 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2798 * If we are going to COW the mapping later, we examine the pending
2799 * reservations for this page now. This will ensure that any
2800 * allocations necessary to record that reservation occur outside the
2801 * spinlock. For private mappings, we also lookup the pagecache
2802 * page now as it is used to determine if a reservation has been
2805 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2806 if (vma_needs_reservation(h, vma, address) < 0) {
2811 if (!(vma->vm_flags & VM_MAYSHARE))
2812 pagecache_page = hugetlbfs_pagecache_page(h,
2817 * hugetlb_cow() requires page locks of pte_page(entry) and
2818 * pagecache_page, so here we need take the former one
2819 * when page != pagecache_page or !pagecache_page.
2820 * Note that locking order is always pagecache_page -> page,
2821 * so no worry about deadlock.
2823 page = pte_page(entry);
2825 if (page != pagecache_page)
2828 spin_lock(&mm->page_table_lock);
2829 /* Check for a racing update before calling hugetlb_cow */
2830 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2831 goto out_page_table_lock;
2834 if (flags & FAULT_FLAG_WRITE) {
2835 if (!pte_write(entry)) {
2836 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2838 goto out_page_table_lock;
2840 entry = pte_mkdirty(entry);
2842 entry = pte_mkyoung(entry);
2843 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2844 flags & FAULT_FLAG_WRITE))
2845 update_mmu_cache(vma, address, ptep);
2847 out_page_table_lock:
2848 spin_unlock(&mm->page_table_lock);
2850 if (pagecache_page) {
2851 unlock_page(pagecache_page);
2852 put_page(pagecache_page);
2854 if (page != pagecache_page)
2859 mutex_unlock(&hugetlb_instantiation_mutex);
2864 /* Can be overriden by architectures */
2865 __attribute__((weak)) struct page *
2866 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2867 pud_t *pud, int write)
2873 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2874 struct page **pages, struct vm_area_struct **vmas,
2875 unsigned long *position, int *length, int i,
2878 unsigned long pfn_offset;
2879 unsigned long vaddr = *position;
2880 int remainder = *length;
2881 struct hstate *h = hstate_vma(vma);
2883 spin_lock(&mm->page_table_lock);
2884 while (vaddr < vma->vm_end && remainder) {
2890 * Some archs (sparc64, sh*) have multiple pte_ts to
2891 * each hugepage. We have to make sure we get the
2892 * first, for the page indexing below to work.
2894 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2895 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2898 * When coredumping, it suits get_dump_page if we just return
2899 * an error where there's an empty slot with no huge pagecache
2900 * to back it. This way, we avoid allocating a hugepage, and
2901 * the sparse dumpfile avoids allocating disk blocks, but its
2902 * huge holes still show up with zeroes where they need to be.
2904 if (absent && (flags & FOLL_DUMP) &&
2905 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2911 * We need call hugetlb_fault for both hugepages under migration
2912 * (in which case hugetlb_fault waits for the migration,) and
2913 * hwpoisoned hugepages (in which case we need to prevent the
2914 * caller from accessing to them.) In order to do this, we use
2915 * here is_swap_pte instead of is_hugetlb_entry_migration and
2916 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
2917 * both cases, and because we can't follow correct pages
2918 * directly from any kind of swap entries.
2920 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
2921 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2924 spin_unlock(&mm->page_table_lock);
2925 ret = hugetlb_fault(mm, vma, vaddr,
2926 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2927 spin_lock(&mm->page_table_lock);
2928 if (!(ret & VM_FAULT_ERROR))
2935 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2936 page = pte_page(huge_ptep_get(pte));
2939 pages[i] = mem_map_offset(page, pfn_offset);
2950 if (vaddr < vma->vm_end && remainder &&
2951 pfn_offset < pages_per_huge_page(h)) {
2953 * We use pfn_offset to avoid touching the pageframes
2954 * of this compound page.
2959 spin_unlock(&mm->page_table_lock);
2960 *length = remainder;
2963 return i ? i : -EFAULT;
2966 void hugetlb_change_protection(struct vm_area_struct *vma,
2967 unsigned long address, unsigned long end, pgprot_t newprot)
2969 struct mm_struct *mm = vma->vm_mm;
2970 unsigned long start = address;
2973 struct hstate *h = hstate_vma(vma);
2975 BUG_ON(address >= end);
2976 flush_cache_range(vma, address, end);
2978 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2979 spin_lock(&mm->page_table_lock);
2980 for (; address < end; address += huge_page_size(h)) {
2981 ptep = huge_pte_offset(mm, address);
2984 if (huge_pmd_unshare(mm, &address, ptep))
2986 if (!huge_pte_none(huge_ptep_get(ptep))) {
2987 pte = huge_ptep_get_and_clear(mm, address, ptep);
2988 pte = pte_mkhuge(pte_modify(pte, newprot));
2989 set_huge_pte_at(mm, address, ptep, pte);
2992 spin_unlock(&mm->page_table_lock);
2994 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
2995 * may have cleared our pud entry and done put_page on the page table:
2996 * once we release i_mmap_mutex, another task can do the final put_page
2997 * and that page table be reused and filled with junk.
2999 flush_tlb_range(vma, start, end);
3000 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3003 int hugetlb_reserve_pages(struct inode *inode,
3005 struct vm_area_struct *vma,
3006 vm_flags_t vm_flags)
3009 struct hstate *h = hstate_inode(inode);
3010 struct hugepage_subpool *spool = subpool_inode(inode);
3013 * Only apply hugepage reservation if asked. At fault time, an
3014 * attempt will be made for VM_NORESERVE to allocate a page
3015 * without using reserves
3017 if (vm_flags & VM_NORESERVE)
3021 * Shared mappings base their reservation on the number of pages that
3022 * are already allocated on behalf of the file. Private mappings need
3023 * to reserve the full area even if read-only as mprotect() may be
3024 * called to make the mapping read-write. Assume !vma is a shm mapping
3026 if (!vma || vma->vm_flags & VM_MAYSHARE)
3027 chg = region_chg(&inode->i_mapping->private_list, from, to);
3029 struct resv_map *resv_map = resv_map_alloc();
3035 set_vma_resv_map(vma, resv_map);
3036 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3044 /* There must be enough pages in the subpool for the mapping */
3045 if (hugepage_subpool_get_pages(spool, chg)) {
3051 * Check enough hugepages are available for the reservation.
3052 * Hand the pages back to the subpool if there are not
3054 ret = hugetlb_acct_memory(h, chg);
3056 hugepage_subpool_put_pages(spool, chg);
3061 * Account for the reservations made. Shared mappings record regions
3062 * that have reservations as they are shared by multiple VMAs.
3063 * When the last VMA disappears, the region map says how much
3064 * the reservation was and the page cache tells how much of
3065 * the reservation was consumed. Private mappings are per-VMA and
3066 * only the consumed reservations are tracked. When the VMA
3067 * disappears, the original reservation is the VMA size and the
3068 * consumed reservations are stored in the map. Hence, nothing
3069 * else has to be done for private mappings here
3071 if (!vma || vma->vm_flags & VM_MAYSHARE)
3072 region_add(&inode->i_mapping->private_list, from, to);
3080 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3082 struct hstate *h = hstate_inode(inode);
3083 long chg = region_truncate(&inode->i_mapping->private_list, offset);
3084 struct hugepage_subpool *spool = subpool_inode(inode);
3086 spin_lock(&inode->i_lock);
3087 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3088 spin_unlock(&inode->i_lock);
3090 hugepage_subpool_put_pages(spool, (chg - freed));
3091 hugetlb_acct_memory(h, -(chg - freed));
3094 #ifdef CONFIG_MEMORY_FAILURE
3096 /* Should be called in hugetlb_lock */
3097 static int is_hugepage_on_freelist(struct page *hpage)
3101 struct hstate *h = page_hstate(hpage);
3102 int nid = page_to_nid(hpage);
3104 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3111 * This function is called from memory failure code.
3112 * Assume the caller holds page lock of the head page.
3114 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3116 struct hstate *h = page_hstate(hpage);
3117 int nid = page_to_nid(hpage);
3120 spin_lock(&hugetlb_lock);
3121 if (is_hugepage_on_freelist(hpage)) {
3122 list_del(&hpage->lru);
3123 set_page_refcounted(hpage);
3124 h->free_huge_pages--;
3125 h->free_huge_pages_node[nid]--;
3128 spin_unlock(&hugetlb_lock);