2 * Generic hugetlb support.
3 * (C) William Irwin, April 2004
6 #include <linux/list.h>
7 #include <linux/init.h>
8 #include <linux/module.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>
22 #include <asm/pgtable.h>
25 #include <linux/hugetlb.h>
28 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
29 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
30 unsigned long hugepages_treat_as_movable;
32 static int max_hstate;
33 unsigned int default_hstate_idx;
34 struct hstate hstates[HUGE_MAX_HSTATE];
36 __initdata LIST_HEAD(huge_boot_pages);
38 /* for command line parsing */
39 static struct hstate * __initdata parsed_hstate;
40 static unsigned long __initdata default_hstate_max_huge_pages;
41 static unsigned long __initdata default_hstate_size;
43 #define for_each_hstate(h) \
44 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
47 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
49 static DEFINE_SPINLOCK(hugetlb_lock);
52 * Region tracking -- allows tracking of reservations and instantiated pages
53 * across the pages in a mapping.
55 * The region data structures are protected by a combination of the mmap_sem
56 * and the hugetlb_instantion_mutex. To access or modify a region the caller
57 * must either hold the mmap_sem for write, or the mmap_sem for read and
58 * the hugetlb_instantiation mutex:
60 * down_write(&mm->mmap_sem);
62 * down_read(&mm->mmap_sem);
63 * mutex_lock(&hugetlb_instantiation_mutex);
66 struct list_head link;
71 static long region_add(struct list_head *head, long f, long t)
73 struct file_region *rg, *nrg, *trg;
75 /* Locate the region we are either in or before. */
76 list_for_each_entry(rg, head, link)
80 /* Round our left edge to the current segment if it encloses us. */
84 /* Check for and consume any regions we now overlap with. */
86 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
87 if (&rg->link == head)
92 /* If this area reaches higher then extend our area to
93 * include it completely. If this is not the first area
94 * which we intend to reuse, free it. */
107 static long region_chg(struct list_head *head, long f, long t)
109 struct file_region *rg, *nrg;
112 /* Locate the region we are before or in. */
113 list_for_each_entry(rg, head, link)
117 /* If we are below the current region then a new region is required.
118 * Subtle, allocate a new region at the position but make it zero
119 * size such that we can guarantee to record the reservation. */
120 if (&rg->link == head || t < rg->from) {
121 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
126 INIT_LIST_HEAD(&nrg->link);
127 list_add(&nrg->link, rg->link.prev);
132 /* Round our left edge to the current segment if it encloses us. */
137 /* Check for and consume any regions we now overlap with. */
138 list_for_each_entry(rg, rg->link.prev, link) {
139 if (&rg->link == head)
144 /* We overlap with this area, if it extends futher than
145 * us then we must extend ourselves. Account for its
146 * existing reservation. */
151 chg -= rg->to - rg->from;
156 static long region_truncate(struct list_head *head, long end)
158 struct file_region *rg, *trg;
161 /* Locate the region we are either in or before. */
162 list_for_each_entry(rg, head, link)
165 if (&rg->link == head)
168 /* If we are in the middle of a region then adjust it. */
169 if (end > rg->from) {
172 rg = list_entry(rg->link.next, typeof(*rg), link);
175 /* Drop any remaining regions. */
176 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
177 if (&rg->link == head)
179 chg += rg->to - rg->from;
186 static long region_count(struct list_head *head, long f, long t)
188 struct file_region *rg;
191 /* Locate each segment we overlap with, and count that overlap. */
192 list_for_each_entry(rg, head, link) {
201 seg_from = max(rg->from, f);
202 seg_to = min(rg->to, t);
204 chg += seg_to - seg_from;
211 * Convert the address within this vma to the page offset within
212 * the mapping, in pagecache page units; huge pages here.
214 static pgoff_t vma_hugecache_offset(struct hstate *h,
215 struct vm_area_struct *vma, unsigned long address)
217 return ((address - vma->vm_start) >> huge_page_shift(h)) +
218 (vma->vm_pgoff >> huge_page_order(h));
222 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
223 * bits of the reservation map pointer, which are always clear due to
226 #define HPAGE_RESV_OWNER (1UL << 0)
227 #define HPAGE_RESV_UNMAPPED (1UL << 1)
228 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
231 * These helpers are used to track how many pages are reserved for
232 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
233 * is guaranteed to have their future faults succeed.
235 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
236 * the reserve counters are updated with the hugetlb_lock held. It is safe
237 * to reset the VMA at fork() time as it is not in use yet and there is no
238 * chance of the global counters getting corrupted as a result of the values.
240 * The private mapping reservation is represented in a subtly different
241 * manner to a shared mapping. A shared mapping has a region map associated
242 * with the underlying file, this region map represents the backing file
243 * pages which have ever had a reservation assigned which this persists even
244 * after the page is instantiated. A private mapping has a region map
245 * associated with the original mmap which is attached to all VMAs which
246 * reference it, this region map represents those offsets which have consumed
247 * reservation ie. where pages have been instantiated.
249 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
251 return (unsigned long)vma->vm_private_data;
254 static void set_vma_private_data(struct vm_area_struct *vma,
257 vma->vm_private_data = (void *)value;
262 struct list_head regions;
265 struct resv_map *resv_map_alloc(void)
267 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
271 kref_init(&resv_map->refs);
272 INIT_LIST_HEAD(&resv_map->regions);
277 void resv_map_release(struct kref *ref)
279 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
281 /* Clear out any active regions before we release the map. */
282 region_truncate(&resv_map->regions, 0);
286 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
288 VM_BUG_ON(!is_vm_hugetlb_page(vma));
289 if (!(vma->vm_flags & VM_SHARED))
290 return (struct resv_map *)(get_vma_private_data(vma) &
295 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
297 VM_BUG_ON(!is_vm_hugetlb_page(vma));
298 VM_BUG_ON(vma->vm_flags & VM_SHARED);
300 set_vma_private_data(vma, (get_vma_private_data(vma) &
301 HPAGE_RESV_MASK) | (unsigned long)map);
304 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
306 VM_BUG_ON(!is_vm_hugetlb_page(vma));
307 VM_BUG_ON(vma->vm_flags & VM_SHARED);
309 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
312 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
314 VM_BUG_ON(!is_vm_hugetlb_page(vma));
316 return (get_vma_private_data(vma) & flag) != 0;
319 /* Decrement the reserved pages in the hugepage pool by one */
320 static void decrement_hugepage_resv_vma(struct hstate *h,
321 struct vm_area_struct *vma)
323 if (vma->vm_flags & VM_NORESERVE)
326 if (vma->vm_flags & VM_SHARED) {
327 /* Shared mappings always use reserves */
328 h->resv_huge_pages--;
329 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
331 * Only the process that called mmap() has reserves for
334 h->resv_huge_pages--;
338 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
339 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
341 VM_BUG_ON(!is_vm_hugetlb_page(vma));
342 if (!(vma->vm_flags & VM_SHARED))
343 vma->vm_private_data = (void *)0;
346 /* Returns true if the VMA has associated reserve pages */
347 static int vma_has_reserves(struct vm_area_struct *vma)
349 if (vma->vm_flags & VM_SHARED)
351 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
356 static void clear_gigantic_page(struct page *page,
357 unsigned long addr, unsigned long sz)
360 struct page *p = page;
363 for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) {
365 clear_user_highpage(p, addr + i * PAGE_SIZE);
368 static void clear_huge_page(struct page *page,
369 unsigned long addr, unsigned long sz)
373 if (unlikely(sz > MAX_ORDER_NR_PAGES))
374 return clear_gigantic_page(page, addr, sz);
377 for (i = 0; i < sz/PAGE_SIZE; i++) {
379 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
383 static void copy_gigantic_page(struct page *dst, struct page *src,
384 unsigned long addr, struct vm_area_struct *vma)
387 struct hstate *h = hstate_vma(vma);
388 struct page *dst_base = dst;
389 struct page *src_base = src;
391 for (i = 0; i < pages_per_huge_page(h); ) {
393 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
396 dst = mem_map_next(dst, dst_base, i);
397 src = mem_map_next(src, src_base, i);
400 static void copy_huge_page(struct page *dst, struct page *src,
401 unsigned long addr, struct vm_area_struct *vma)
404 struct hstate *h = hstate_vma(vma);
406 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES))
407 return copy_gigantic_page(dst, src, addr, vma);
410 for (i = 0; i < pages_per_huge_page(h); i++) {
412 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
416 static void enqueue_huge_page(struct hstate *h, struct page *page)
418 int nid = page_to_nid(page);
419 list_add(&page->lru, &h->hugepage_freelists[nid]);
420 h->free_huge_pages++;
421 h->free_huge_pages_node[nid]++;
424 static struct page *dequeue_huge_page(struct hstate *h)
427 struct page *page = NULL;
429 for (nid = 0; nid < MAX_NUMNODES; ++nid) {
430 if (!list_empty(&h->hugepage_freelists[nid])) {
431 page = list_entry(h->hugepage_freelists[nid].next,
433 list_del(&page->lru);
434 h->free_huge_pages--;
435 h->free_huge_pages_node[nid]--;
442 static struct page *dequeue_huge_page_vma(struct hstate *h,
443 struct vm_area_struct *vma,
444 unsigned long address, int avoid_reserve)
447 struct page *page = NULL;
448 struct mempolicy *mpol;
449 nodemask_t *nodemask;
450 struct zonelist *zonelist = huge_zonelist(vma, address,
451 htlb_alloc_mask, &mpol, &nodemask);
456 * A child process with MAP_PRIVATE mappings created by their parent
457 * have no page reserves. This check ensures that reservations are
458 * not "stolen". The child may still get SIGKILLed
460 if (!vma_has_reserves(vma) &&
461 h->free_huge_pages - h->resv_huge_pages == 0)
464 /* If reserves cannot be used, ensure enough pages are in the pool */
465 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
468 for_each_zone_zonelist_nodemask(zone, z, zonelist,
469 MAX_NR_ZONES - 1, nodemask) {
470 nid = zone_to_nid(zone);
471 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
472 !list_empty(&h->hugepage_freelists[nid])) {
473 page = list_entry(h->hugepage_freelists[nid].next,
475 list_del(&page->lru);
476 h->free_huge_pages--;
477 h->free_huge_pages_node[nid]--;
480 decrement_hugepage_resv_vma(h, vma);
489 static void update_and_free_page(struct hstate *h, struct page *page)
493 VM_BUG_ON(h->order >= MAX_ORDER);
496 h->nr_huge_pages_node[page_to_nid(page)]--;
497 for (i = 0; i < pages_per_huge_page(h); i++) {
498 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
499 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
500 1 << PG_private | 1<< PG_writeback);
502 set_compound_page_dtor(page, NULL);
503 set_page_refcounted(page);
504 arch_release_hugepage(page);
505 __free_pages(page, huge_page_order(h));
508 struct hstate *size_to_hstate(unsigned long size)
513 if (huge_page_size(h) == size)
519 static void free_huge_page(struct page *page)
522 * Can't pass hstate in here because it is called from the
523 * compound page destructor.
525 struct hstate *h = page_hstate(page);
526 int nid = page_to_nid(page);
527 struct address_space *mapping;
529 mapping = (struct address_space *) page_private(page);
530 set_page_private(page, 0);
531 BUG_ON(page_count(page));
532 INIT_LIST_HEAD(&page->lru);
534 spin_lock(&hugetlb_lock);
535 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
536 update_and_free_page(h, page);
537 h->surplus_huge_pages--;
538 h->surplus_huge_pages_node[nid]--;
540 enqueue_huge_page(h, page);
542 spin_unlock(&hugetlb_lock);
544 hugetlb_put_quota(mapping, 1);
548 * Increment or decrement surplus_huge_pages. Keep node-specific counters
549 * balanced by operating on them in a round-robin fashion.
550 * Returns 1 if an adjustment was made.
552 static int adjust_pool_surplus(struct hstate *h, int delta)
558 VM_BUG_ON(delta != -1 && delta != 1);
560 nid = next_node(nid, node_online_map);
561 if (nid == MAX_NUMNODES)
562 nid = first_node(node_online_map);
564 /* To shrink on this node, there must be a surplus page */
565 if (delta < 0 && !h->surplus_huge_pages_node[nid])
567 /* Surplus cannot exceed the total number of pages */
568 if (delta > 0 && h->surplus_huge_pages_node[nid] >=
569 h->nr_huge_pages_node[nid])
572 h->surplus_huge_pages += delta;
573 h->surplus_huge_pages_node[nid] += delta;
576 } while (nid != prev_nid);
582 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
584 set_compound_page_dtor(page, free_huge_page);
585 spin_lock(&hugetlb_lock);
587 h->nr_huge_pages_node[nid]++;
588 spin_unlock(&hugetlb_lock);
589 put_page(page); /* free it into the hugepage allocator */
592 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
596 if (h->order >= MAX_ORDER)
599 page = alloc_pages_node(nid,
600 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
601 __GFP_REPEAT|__GFP_NOWARN,
604 if (arch_prepare_hugepage(page)) {
605 __free_pages(page, huge_page_order(h));
608 prep_new_huge_page(h, page, nid);
615 * Use a helper variable to find the next node and then
616 * copy it back to hugetlb_next_nid afterwards:
617 * otherwise there's a window in which a racer might
618 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
619 * But we don't need to use a spin_lock here: it really
620 * doesn't matter if occasionally a racer chooses the
621 * same nid as we do. Move nid forward in the mask even
622 * if we just successfully allocated a hugepage so that
623 * the next caller gets hugepages on the next node.
625 static int hstate_next_node(struct hstate *h)
628 next_nid = next_node(h->hugetlb_next_nid, node_online_map);
629 if (next_nid == MAX_NUMNODES)
630 next_nid = first_node(node_online_map);
631 h->hugetlb_next_nid = next_nid;
635 static int alloc_fresh_huge_page(struct hstate *h)
642 start_nid = h->hugetlb_next_nid;
645 page = alloc_fresh_huge_page_node(h, h->hugetlb_next_nid);
648 next_nid = hstate_next_node(h);
649 } while (!page && h->hugetlb_next_nid != start_nid);
652 count_vm_event(HTLB_BUDDY_PGALLOC);
654 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
659 static struct page *alloc_buddy_huge_page(struct hstate *h,
660 struct vm_area_struct *vma, unsigned long address)
665 if (h->order >= MAX_ORDER)
669 * Assume we will successfully allocate the surplus page to
670 * prevent racing processes from causing the surplus to exceed
673 * This however introduces a different race, where a process B
674 * tries to grow the static hugepage pool while alloc_pages() is
675 * called by process A. B will only examine the per-node
676 * counters in determining if surplus huge pages can be
677 * converted to normal huge pages in adjust_pool_surplus(). A
678 * won't be able to increment the per-node counter, until the
679 * lock is dropped by B, but B doesn't drop hugetlb_lock until
680 * no more huge pages can be converted from surplus to normal
681 * state (and doesn't try to convert again). Thus, we have a
682 * case where a surplus huge page exists, the pool is grown, and
683 * the surplus huge page still exists after, even though it
684 * should just have been converted to a normal huge page. This
685 * does not leak memory, though, as the hugepage will be freed
686 * once it is out of use. It also does not allow the counters to
687 * go out of whack in adjust_pool_surplus() as we don't modify
688 * the node values until we've gotten the hugepage and only the
689 * per-node value is checked there.
691 spin_lock(&hugetlb_lock);
692 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
693 spin_unlock(&hugetlb_lock);
697 h->surplus_huge_pages++;
699 spin_unlock(&hugetlb_lock);
701 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
702 __GFP_REPEAT|__GFP_NOWARN,
705 if (page && arch_prepare_hugepage(page)) {
706 __free_pages(page, huge_page_order(h));
710 spin_lock(&hugetlb_lock);
713 * This page is now managed by the hugetlb allocator and has
714 * no users -- drop the buddy allocator's reference.
716 put_page_testzero(page);
717 VM_BUG_ON(page_count(page));
718 nid = page_to_nid(page);
719 set_compound_page_dtor(page, free_huge_page);
721 * We incremented the global counters already
723 h->nr_huge_pages_node[nid]++;
724 h->surplus_huge_pages_node[nid]++;
725 __count_vm_event(HTLB_BUDDY_PGALLOC);
728 h->surplus_huge_pages--;
729 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
731 spin_unlock(&hugetlb_lock);
737 * Increase the hugetlb pool such that it can accomodate a reservation
740 static int gather_surplus_pages(struct hstate *h, int delta)
742 struct list_head surplus_list;
743 struct page *page, *tmp;
745 int needed, allocated;
747 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
749 h->resv_huge_pages += delta;
754 INIT_LIST_HEAD(&surplus_list);
758 spin_unlock(&hugetlb_lock);
759 for (i = 0; i < needed; i++) {
760 page = alloc_buddy_huge_page(h, NULL, 0);
763 * We were not able to allocate enough pages to
764 * satisfy the entire reservation so we free what
765 * we've allocated so far.
767 spin_lock(&hugetlb_lock);
772 list_add(&page->lru, &surplus_list);
777 * After retaking hugetlb_lock, we need to recalculate 'needed'
778 * because either resv_huge_pages or free_huge_pages may have changed.
780 spin_lock(&hugetlb_lock);
781 needed = (h->resv_huge_pages + delta) -
782 (h->free_huge_pages + allocated);
787 * The surplus_list now contains _at_least_ the number of extra pages
788 * needed to accomodate the reservation. Add the appropriate number
789 * of pages to the hugetlb pool and free the extras back to the buddy
790 * allocator. Commit the entire reservation here to prevent another
791 * process from stealing the pages as they are added to the pool but
792 * before they are reserved.
795 h->resv_huge_pages += delta;
798 /* Free the needed pages to the hugetlb pool */
799 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
802 list_del(&page->lru);
803 enqueue_huge_page(h, page);
806 /* Free unnecessary surplus pages to the buddy allocator */
807 if (!list_empty(&surplus_list)) {
808 spin_unlock(&hugetlb_lock);
809 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
810 list_del(&page->lru);
812 * The page has a reference count of zero already, so
813 * call free_huge_page directly instead of using
814 * put_page. This must be done with hugetlb_lock
815 * unlocked which is safe because free_huge_page takes
816 * hugetlb_lock before deciding how to free the page.
818 free_huge_page(page);
820 spin_lock(&hugetlb_lock);
827 * When releasing a hugetlb pool reservation, any surplus pages that were
828 * allocated to satisfy the reservation must be explicitly freed if they were
831 static void return_unused_surplus_pages(struct hstate *h,
832 unsigned long unused_resv_pages)
836 unsigned long nr_pages;
839 * We want to release as many surplus pages as possible, spread
840 * evenly across all nodes. Iterate across all nodes until we
841 * can no longer free unreserved surplus pages. This occurs when
842 * the nodes with surplus pages have no free pages.
844 unsigned long remaining_iterations = num_online_nodes();
846 /* Uncommit the reservation */
847 h->resv_huge_pages -= unused_resv_pages;
849 /* Cannot return gigantic pages currently */
850 if (h->order >= MAX_ORDER)
853 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
855 while (remaining_iterations-- && nr_pages) {
856 nid = next_node(nid, node_online_map);
857 if (nid == MAX_NUMNODES)
858 nid = first_node(node_online_map);
860 if (!h->surplus_huge_pages_node[nid])
863 if (!list_empty(&h->hugepage_freelists[nid])) {
864 page = list_entry(h->hugepage_freelists[nid].next,
866 list_del(&page->lru);
867 update_and_free_page(h, page);
868 h->free_huge_pages--;
869 h->free_huge_pages_node[nid]--;
870 h->surplus_huge_pages--;
871 h->surplus_huge_pages_node[nid]--;
873 remaining_iterations = num_online_nodes();
879 * Determine if the huge page at addr within the vma has an associated
880 * reservation. Where it does not we will need to logically increase
881 * reservation and actually increase quota before an allocation can occur.
882 * Where any new reservation would be required the reservation change is
883 * prepared, but not committed. Once the page has been quota'd allocated
884 * an instantiated the change should be committed via vma_commit_reservation.
885 * No action is required on failure.
887 static int vma_needs_reservation(struct hstate *h,
888 struct vm_area_struct *vma, unsigned long addr)
890 struct address_space *mapping = vma->vm_file->f_mapping;
891 struct inode *inode = mapping->host;
893 if (vma->vm_flags & VM_SHARED) {
894 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
895 return region_chg(&inode->i_mapping->private_list,
898 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
903 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
904 struct resv_map *reservations = vma_resv_map(vma);
906 err = region_chg(&reservations->regions, idx, idx + 1);
912 static void vma_commit_reservation(struct hstate *h,
913 struct vm_area_struct *vma, unsigned long addr)
915 struct address_space *mapping = vma->vm_file->f_mapping;
916 struct inode *inode = mapping->host;
918 if (vma->vm_flags & VM_SHARED) {
919 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
920 region_add(&inode->i_mapping->private_list, idx, idx + 1);
922 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
923 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
924 struct resv_map *reservations = vma_resv_map(vma);
926 /* Mark this page used in the map. */
927 region_add(&reservations->regions, idx, idx + 1);
931 static struct page *alloc_huge_page(struct vm_area_struct *vma,
932 unsigned long addr, int avoid_reserve)
934 struct hstate *h = hstate_vma(vma);
936 struct address_space *mapping = vma->vm_file->f_mapping;
937 struct inode *inode = mapping->host;
941 * Processes that did not create the mapping will have no reserves and
942 * will not have accounted against quota. Check that the quota can be
943 * made before satisfying the allocation
944 * MAP_NORESERVE mappings may also need pages and quota allocated
945 * if no reserve mapping overlaps.
947 chg = vma_needs_reservation(h, vma, addr);
951 if (hugetlb_get_quota(inode->i_mapping, chg))
952 return ERR_PTR(-ENOSPC);
954 spin_lock(&hugetlb_lock);
955 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
956 spin_unlock(&hugetlb_lock);
959 page = alloc_buddy_huge_page(h, vma, addr);
961 hugetlb_put_quota(inode->i_mapping, chg);
962 return ERR_PTR(-VM_FAULT_OOM);
966 set_page_refcounted(page);
967 set_page_private(page, (unsigned long) mapping);
969 vma_commit_reservation(h, vma, addr);
974 __attribute__((weak)) int alloc_bootmem_huge_page(struct hstate *h)
976 struct huge_bootmem_page *m;
977 int nr_nodes = nodes_weight(node_online_map);
982 addr = __alloc_bootmem_node_nopanic(
983 NODE_DATA(h->hugetlb_next_nid),
984 huge_page_size(h), huge_page_size(h), 0);
988 * Use the beginning of the huge page to store the
989 * huge_bootmem_page struct (until gather_bootmem
990 * puts them into the mem_map).
1002 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1003 /* Put them into a private list first because mem_map is not up yet */
1004 list_add(&m->list, &huge_boot_pages);
1009 static void prep_compound_huge_page(struct page *page, int order)
1011 if (unlikely(order > (MAX_ORDER - 1)))
1012 prep_compound_gigantic_page(page, order);
1014 prep_compound_page(page, order);
1017 /* Put bootmem huge pages into the standard lists after mem_map is up */
1018 static void __init gather_bootmem_prealloc(void)
1020 struct huge_bootmem_page *m;
1022 list_for_each_entry(m, &huge_boot_pages, list) {
1023 struct page *page = virt_to_page(m);
1024 struct hstate *h = m->hstate;
1025 __ClearPageReserved(page);
1026 WARN_ON(page_count(page) != 1);
1027 prep_compound_huge_page(page, h->order);
1028 prep_new_huge_page(h, page, page_to_nid(page));
1032 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1036 for (i = 0; i < h->max_huge_pages; ++i) {
1037 if (h->order >= MAX_ORDER) {
1038 if (!alloc_bootmem_huge_page(h))
1040 } else if (!alloc_fresh_huge_page(h))
1043 h->max_huge_pages = i;
1046 static void __init hugetlb_init_hstates(void)
1050 for_each_hstate(h) {
1051 /* oversize hugepages were init'ed in early boot */
1052 if (h->order < MAX_ORDER)
1053 hugetlb_hstate_alloc_pages(h);
1057 static char * __init memfmt(char *buf, unsigned long n)
1059 if (n >= (1UL << 30))
1060 sprintf(buf, "%lu GB", n >> 30);
1061 else if (n >= (1UL << 20))
1062 sprintf(buf, "%lu MB", n >> 20);
1064 sprintf(buf, "%lu KB", n >> 10);
1068 static void __init report_hugepages(void)
1072 for_each_hstate(h) {
1074 printk(KERN_INFO "HugeTLB registered %s page size, "
1075 "pre-allocated %ld pages\n",
1076 memfmt(buf, huge_page_size(h)),
1077 h->free_huge_pages);
1081 #ifdef CONFIG_HIGHMEM
1082 static void try_to_free_low(struct hstate *h, unsigned long count)
1086 if (h->order >= MAX_ORDER)
1089 for (i = 0; i < MAX_NUMNODES; ++i) {
1090 struct page *page, *next;
1091 struct list_head *freel = &h->hugepage_freelists[i];
1092 list_for_each_entry_safe(page, next, freel, lru) {
1093 if (count >= h->nr_huge_pages)
1095 if (PageHighMem(page))
1097 list_del(&page->lru);
1098 update_and_free_page(h, page);
1099 h->free_huge_pages--;
1100 h->free_huge_pages_node[page_to_nid(page)]--;
1105 static inline void try_to_free_low(struct hstate *h, unsigned long count)
1110 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1111 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count)
1113 unsigned long min_count, ret;
1115 if (h->order >= MAX_ORDER)
1116 return h->max_huge_pages;
1119 * Increase the pool size
1120 * First take pages out of surplus state. Then make up the
1121 * remaining difference by allocating fresh huge pages.
1123 * We might race with alloc_buddy_huge_page() here and be unable
1124 * to convert a surplus huge page to a normal huge page. That is
1125 * not critical, though, it just means the overall size of the
1126 * pool might be one hugepage larger than it needs to be, but
1127 * within all the constraints specified by the sysctls.
1129 spin_lock(&hugetlb_lock);
1130 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1131 if (!adjust_pool_surplus(h, -1))
1135 while (count > persistent_huge_pages(h)) {
1137 * If this allocation races such that we no longer need the
1138 * page, free_huge_page will handle it by freeing the page
1139 * and reducing the surplus.
1141 spin_unlock(&hugetlb_lock);
1142 ret = alloc_fresh_huge_page(h);
1143 spin_lock(&hugetlb_lock);
1150 * Decrease the pool size
1151 * First return free pages to the buddy allocator (being careful
1152 * to keep enough around to satisfy reservations). Then place
1153 * pages into surplus state as needed so the pool will shrink
1154 * to the desired size as pages become free.
1156 * By placing pages into the surplus state independent of the
1157 * overcommit value, we are allowing the surplus pool size to
1158 * exceed overcommit. There are few sane options here. Since
1159 * alloc_buddy_huge_page() is checking the global counter,
1160 * though, we'll note that we're not allowed to exceed surplus
1161 * and won't grow the pool anywhere else. Not until one of the
1162 * sysctls are changed, or the surplus pages go out of use.
1164 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1165 min_count = max(count, min_count);
1166 try_to_free_low(h, min_count);
1167 while (min_count < persistent_huge_pages(h)) {
1168 struct page *page = dequeue_huge_page(h);
1171 update_and_free_page(h, page);
1173 while (count < persistent_huge_pages(h)) {
1174 if (!adjust_pool_surplus(h, 1))
1178 ret = persistent_huge_pages(h);
1179 spin_unlock(&hugetlb_lock);
1183 #define HSTATE_ATTR_RO(_name) \
1184 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1186 #define HSTATE_ATTR(_name) \
1187 static struct kobj_attribute _name##_attr = \
1188 __ATTR(_name, 0644, _name##_show, _name##_store)
1190 static struct kobject *hugepages_kobj;
1191 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1193 static struct hstate *kobj_to_hstate(struct kobject *kobj)
1196 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1197 if (hstate_kobjs[i] == kobj)
1203 static ssize_t nr_hugepages_show(struct kobject *kobj,
1204 struct kobj_attribute *attr, char *buf)
1206 struct hstate *h = kobj_to_hstate(kobj);
1207 return sprintf(buf, "%lu\n", h->nr_huge_pages);
1209 static ssize_t nr_hugepages_store(struct kobject *kobj,
1210 struct kobj_attribute *attr, const char *buf, size_t count)
1213 unsigned long input;
1214 struct hstate *h = kobj_to_hstate(kobj);
1216 err = strict_strtoul(buf, 10, &input);
1220 h->max_huge_pages = set_max_huge_pages(h, input);
1224 HSTATE_ATTR(nr_hugepages);
1226 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1227 struct kobj_attribute *attr, char *buf)
1229 struct hstate *h = kobj_to_hstate(kobj);
1230 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1232 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1233 struct kobj_attribute *attr, const char *buf, size_t count)
1236 unsigned long input;
1237 struct hstate *h = kobj_to_hstate(kobj);
1239 err = strict_strtoul(buf, 10, &input);
1243 spin_lock(&hugetlb_lock);
1244 h->nr_overcommit_huge_pages = input;
1245 spin_unlock(&hugetlb_lock);
1249 HSTATE_ATTR(nr_overcommit_hugepages);
1251 static ssize_t free_hugepages_show(struct kobject *kobj,
1252 struct kobj_attribute *attr, char *buf)
1254 struct hstate *h = kobj_to_hstate(kobj);
1255 return sprintf(buf, "%lu\n", h->free_huge_pages);
1257 HSTATE_ATTR_RO(free_hugepages);
1259 static ssize_t resv_hugepages_show(struct kobject *kobj,
1260 struct kobj_attribute *attr, char *buf)
1262 struct hstate *h = kobj_to_hstate(kobj);
1263 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1265 HSTATE_ATTR_RO(resv_hugepages);
1267 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1268 struct kobj_attribute *attr, char *buf)
1270 struct hstate *h = kobj_to_hstate(kobj);
1271 return sprintf(buf, "%lu\n", h->surplus_huge_pages);
1273 HSTATE_ATTR_RO(surplus_hugepages);
1275 static struct attribute *hstate_attrs[] = {
1276 &nr_hugepages_attr.attr,
1277 &nr_overcommit_hugepages_attr.attr,
1278 &free_hugepages_attr.attr,
1279 &resv_hugepages_attr.attr,
1280 &surplus_hugepages_attr.attr,
1284 static struct attribute_group hstate_attr_group = {
1285 .attrs = hstate_attrs,
1288 static int __init hugetlb_sysfs_add_hstate(struct hstate *h)
1292 hstate_kobjs[h - hstates] = kobject_create_and_add(h->name,
1294 if (!hstate_kobjs[h - hstates])
1297 retval = sysfs_create_group(hstate_kobjs[h - hstates],
1298 &hstate_attr_group);
1300 kobject_put(hstate_kobjs[h - hstates]);
1305 static void __init hugetlb_sysfs_init(void)
1310 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1311 if (!hugepages_kobj)
1314 for_each_hstate(h) {
1315 err = hugetlb_sysfs_add_hstate(h);
1317 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1322 static void __exit hugetlb_exit(void)
1326 for_each_hstate(h) {
1327 kobject_put(hstate_kobjs[h - hstates]);
1330 kobject_put(hugepages_kobj);
1332 module_exit(hugetlb_exit);
1334 static int __init hugetlb_init(void)
1336 /* Some platform decide whether they support huge pages at boot
1337 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1338 * there is no such support
1340 if (HPAGE_SHIFT == 0)
1343 if (!size_to_hstate(default_hstate_size)) {
1344 default_hstate_size = HPAGE_SIZE;
1345 if (!size_to_hstate(default_hstate_size))
1346 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1348 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1349 if (default_hstate_max_huge_pages)
1350 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1352 hugetlb_init_hstates();
1354 gather_bootmem_prealloc();
1358 hugetlb_sysfs_init();
1362 module_init(hugetlb_init);
1364 /* Should be called on processing a hugepagesz=... option */
1365 void __init hugetlb_add_hstate(unsigned order)
1370 if (size_to_hstate(PAGE_SIZE << order)) {
1371 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1374 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1376 h = &hstates[max_hstate++];
1378 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1379 h->nr_huge_pages = 0;
1380 h->free_huge_pages = 0;
1381 for (i = 0; i < MAX_NUMNODES; ++i)
1382 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1383 h->hugetlb_next_nid = first_node(node_online_map);
1384 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1385 huge_page_size(h)/1024);
1390 static int __init hugetlb_nrpages_setup(char *s)
1393 static unsigned long *last_mhp;
1396 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1397 * so this hugepages= parameter goes to the "default hstate".
1400 mhp = &default_hstate_max_huge_pages;
1402 mhp = &parsed_hstate->max_huge_pages;
1404 if (mhp == last_mhp) {
1405 printk(KERN_WARNING "hugepages= specified twice without "
1406 "interleaving hugepagesz=, ignoring\n");
1410 if (sscanf(s, "%lu", mhp) <= 0)
1414 * Global state is always initialized later in hugetlb_init.
1415 * But we need to allocate >= MAX_ORDER hstates here early to still
1416 * use the bootmem allocator.
1418 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1419 hugetlb_hstate_alloc_pages(parsed_hstate);
1425 __setup("hugepages=", hugetlb_nrpages_setup);
1427 static int __init hugetlb_default_setup(char *s)
1429 default_hstate_size = memparse(s, &s);
1432 __setup("default_hugepagesz=", hugetlb_default_setup);
1434 static unsigned int cpuset_mems_nr(unsigned int *array)
1437 unsigned int nr = 0;
1439 for_each_node_mask(node, cpuset_current_mems_allowed)
1445 #ifdef CONFIG_SYSCTL
1446 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1447 struct file *file, void __user *buffer,
1448 size_t *length, loff_t *ppos)
1450 struct hstate *h = &default_hstate;
1454 tmp = h->max_huge_pages;
1457 table->maxlen = sizeof(unsigned long);
1458 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1461 h->max_huge_pages = set_max_huge_pages(h, tmp);
1466 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1467 struct file *file, void __user *buffer,
1468 size_t *length, loff_t *ppos)
1470 proc_dointvec(table, write, file, buffer, length, ppos);
1471 if (hugepages_treat_as_movable)
1472 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1474 htlb_alloc_mask = GFP_HIGHUSER;
1478 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1479 struct file *file, void __user *buffer,
1480 size_t *length, loff_t *ppos)
1482 struct hstate *h = &default_hstate;
1486 tmp = h->nr_overcommit_huge_pages;
1489 table->maxlen = sizeof(unsigned long);
1490 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1493 spin_lock(&hugetlb_lock);
1494 h->nr_overcommit_huge_pages = tmp;
1495 spin_unlock(&hugetlb_lock);
1501 #endif /* CONFIG_SYSCTL */
1503 int hugetlb_report_meminfo(char *buf)
1505 struct hstate *h = &default_hstate;
1507 "HugePages_Total: %5lu\n"
1508 "HugePages_Free: %5lu\n"
1509 "HugePages_Rsvd: %5lu\n"
1510 "HugePages_Surp: %5lu\n"
1511 "Hugepagesize: %5lu kB\n",
1515 h->surplus_huge_pages,
1516 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1519 int hugetlb_report_node_meminfo(int nid, char *buf)
1521 struct hstate *h = &default_hstate;
1523 "Node %d HugePages_Total: %5u\n"
1524 "Node %d HugePages_Free: %5u\n"
1525 "Node %d HugePages_Surp: %5u\n",
1526 nid, h->nr_huge_pages_node[nid],
1527 nid, h->free_huge_pages_node[nid],
1528 nid, h->surplus_huge_pages_node[nid]);
1531 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1532 unsigned long hugetlb_total_pages(void)
1534 struct hstate *h = &default_hstate;
1535 return h->nr_huge_pages * pages_per_huge_page(h);
1538 static int hugetlb_acct_memory(struct hstate *h, long delta)
1542 spin_lock(&hugetlb_lock);
1544 * When cpuset is configured, it breaks the strict hugetlb page
1545 * reservation as the accounting is done on a global variable. Such
1546 * reservation is completely rubbish in the presence of cpuset because
1547 * the reservation is not checked against page availability for the
1548 * current cpuset. Application can still potentially OOM'ed by kernel
1549 * with lack of free htlb page in cpuset that the task is in.
1550 * Attempt to enforce strict accounting with cpuset is almost
1551 * impossible (or too ugly) because cpuset is too fluid that
1552 * task or memory node can be dynamically moved between cpusets.
1554 * The change of semantics for shared hugetlb mapping with cpuset is
1555 * undesirable. However, in order to preserve some of the semantics,
1556 * we fall back to check against current free page availability as
1557 * a best attempt and hopefully to minimize the impact of changing
1558 * semantics that cpuset has.
1561 if (gather_surplus_pages(h, delta) < 0)
1564 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1565 return_unused_surplus_pages(h, delta);
1572 return_unused_surplus_pages(h, (unsigned long) -delta);
1575 spin_unlock(&hugetlb_lock);
1579 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
1581 struct resv_map *reservations = vma_resv_map(vma);
1584 * This new VMA should share its siblings reservation map if present.
1585 * The VMA will only ever have a valid reservation map pointer where
1586 * it is being copied for another still existing VMA. As that VMA
1587 * has a reference to the reservation map it cannot dissappear until
1588 * after this open call completes. It is therefore safe to take a
1589 * new reference here without additional locking.
1592 kref_get(&reservations->refs);
1595 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
1597 struct hstate *h = hstate_vma(vma);
1598 struct resv_map *reservations = vma_resv_map(vma);
1599 unsigned long reserve;
1600 unsigned long start;
1604 start = vma_hugecache_offset(h, vma, vma->vm_start);
1605 end = vma_hugecache_offset(h, vma, vma->vm_end);
1607 reserve = (end - start) -
1608 region_count(&reservations->regions, start, end);
1610 kref_put(&reservations->refs, resv_map_release);
1613 hugetlb_acct_memory(h, -reserve);
1614 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
1620 * We cannot handle pagefaults against hugetlb pages at all. They cause
1621 * handle_mm_fault() to try to instantiate regular-sized pages in the
1622 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
1625 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1631 struct vm_operations_struct hugetlb_vm_ops = {
1632 .fault = hugetlb_vm_op_fault,
1633 .open = hugetlb_vm_op_open,
1634 .close = hugetlb_vm_op_close,
1637 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
1644 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1646 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
1648 entry = pte_mkyoung(entry);
1649 entry = pte_mkhuge(entry);
1654 static void set_huge_ptep_writable(struct vm_area_struct *vma,
1655 unsigned long address, pte_t *ptep)
1659 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
1660 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
1661 update_mmu_cache(vma, address, entry);
1666 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
1667 struct vm_area_struct *vma)
1669 pte_t *src_pte, *dst_pte, entry;
1670 struct page *ptepage;
1673 struct hstate *h = hstate_vma(vma);
1674 unsigned long sz = huge_page_size(h);
1676 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
1678 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
1679 src_pte = huge_pte_offset(src, addr);
1682 dst_pte = huge_pte_alloc(dst, addr, sz);
1686 /* If the pagetables are shared don't copy or take references */
1687 if (dst_pte == src_pte)
1690 spin_lock(&dst->page_table_lock);
1691 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
1692 if (!huge_pte_none(huge_ptep_get(src_pte))) {
1694 huge_ptep_set_wrprotect(src, addr, src_pte);
1695 entry = huge_ptep_get(src_pte);
1696 ptepage = pte_page(entry);
1698 set_huge_pte_at(dst, addr, dst_pte, entry);
1700 spin_unlock(&src->page_table_lock);
1701 spin_unlock(&dst->page_table_lock);
1709 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1710 unsigned long end, struct page *ref_page)
1712 struct mm_struct *mm = vma->vm_mm;
1713 unsigned long address;
1718 struct hstate *h = hstate_vma(vma);
1719 unsigned long sz = huge_page_size(h);
1722 * A page gathering list, protected by per file i_mmap_lock. The
1723 * lock is used to avoid list corruption from multiple unmapping
1724 * of the same page since we are using page->lru.
1726 LIST_HEAD(page_list);
1728 WARN_ON(!is_vm_hugetlb_page(vma));
1729 BUG_ON(start & ~huge_page_mask(h));
1730 BUG_ON(end & ~huge_page_mask(h));
1732 mmu_notifier_invalidate_range_start(mm, start, end);
1733 spin_lock(&mm->page_table_lock);
1734 for (address = start; address < end; address += sz) {
1735 ptep = huge_pte_offset(mm, address);
1739 if (huge_pmd_unshare(mm, &address, ptep))
1743 * If a reference page is supplied, it is because a specific
1744 * page is being unmapped, not a range. Ensure the page we
1745 * are about to unmap is the actual page of interest.
1748 pte = huge_ptep_get(ptep);
1749 if (huge_pte_none(pte))
1751 page = pte_page(pte);
1752 if (page != ref_page)
1756 * Mark the VMA as having unmapped its page so that
1757 * future faults in this VMA will fail rather than
1758 * looking like data was lost
1760 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
1763 pte = huge_ptep_get_and_clear(mm, address, ptep);
1764 if (huge_pte_none(pte))
1767 page = pte_page(pte);
1769 set_page_dirty(page);
1770 list_add(&page->lru, &page_list);
1772 spin_unlock(&mm->page_table_lock);
1773 flush_tlb_range(vma, start, end);
1774 mmu_notifier_invalidate_range_end(mm, start, end);
1775 list_for_each_entry_safe(page, tmp, &page_list, lru) {
1776 list_del(&page->lru);
1781 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1782 unsigned long end, struct page *ref_page)
1784 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1785 __unmap_hugepage_range(vma, start, end, ref_page);
1786 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1790 * This is called when the original mapper is failing to COW a MAP_PRIVATE
1791 * mappping it owns the reserve page for. The intention is to unmap the page
1792 * from other VMAs and let the children be SIGKILLed if they are faulting the
1795 int unmap_ref_private(struct mm_struct *mm,
1796 struct vm_area_struct *vma,
1798 unsigned long address)
1800 struct vm_area_struct *iter_vma;
1801 struct address_space *mapping;
1802 struct prio_tree_iter iter;
1806 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1807 * from page cache lookup which is in HPAGE_SIZE units.
1809 address = address & huge_page_mask(hstate_vma(vma));
1810 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
1811 + (vma->vm_pgoff >> PAGE_SHIFT);
1812 mapping = (struct address_space *)page_private(page);
1814 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1815 /* Do not unmap the current VMA */
1816 if (iter_vma == vma)
1820 * Unmap the page from other VMAs without their own reserves.
1821 * They get marked to be SIGKILLed if they fault in these
1822 * areas. This is because a future no-page fault on this VMA
1823 * could insert a zeroed page instead of the data existing
1824 * from the time of fork. This would look like data corruption
1826 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
1827 unmap_hugepage_range(iter_vma,
1828 address, address + HPAGE_SIZE,
1835 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
1836 unsigned long address, pte_t *ptep, pte_t pte,
1837 struct page *pagecache_page)
1839 struct hstate *h = hstate_vma(vma);
1840 struct page *old_page, *new_page;
1842 int outside_reserve = 0;
1844 old_page = pte_page(pte);
1847 /* If no-one else is actually using this page, avoid the copy
1848 * and just make the page writable */
1849 avoidcopy = (page_count(old_page) == 1);
1851 set_huge_ptep_writable(vma, address, ptep);
1856 * If the process that created a MAP_PRIVATE mapping is about to
1857 * perform a COW due to a shared page count, attempt to satisfy
1858 * the allocation without using the existing reserves. The pagecache
1859 * page is used to determine if the reserve at this address was
1860 * consumed or not. If reserves were used, a partial faulted mapping
1861 * at the time of fork() could consume its reserves on COW instead
1862 * of the full address range.
1864 if (!(vma->vm_flags & VM_SHARED) &&
1865 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
1866 old_page != pagecache_page)
1867 outside_reserve = 1;
1869 page_cache_get(old_page);
1870 new_page = alloc_huge_page(vma, address, outside_reserve);
1872 if (IS_ERR(new_page)) {
1873 page_cache_release(old_page);
1876 * If a process owning a MAP_PRIVATE mapping fails to COW,
1877 * it is due to references held by a child and an insufficient
1878 * huge page pool. To guarantee the original mappers
1879 * reliability, unmap the page from child processes. The child
1880 * may get SIGKILLed if it later faults.
1882 if (outside_reserve) {
1883 BUG_ON(huge_pte_none(pte));
1884 if (unmap_ref_private(mm, vma, old_page, address)) {
1885 BUG_ON(page_count(old_page) != 1);
1886 BUG_ON(huge_pte_none(pte));
1887 goto retry_avoidcopy;
1892 return -PTR_ERR(new_page);
1895 spin_unlock(&mm->page_table_lock);
1896 copy_huge_page(new_page, old_page, address, vma);
1897 __SetPageUptodate(new_page);
1898 spin_lock(&mm->page_table_lock);
1900 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
1901 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
1903 huge_ptep_clear_flush(vma, address, ptep);
1904 set_huge_pte_at(mm, address, ptep,
1905 make_huge_pte(vma, new_page, 1));
1906 /* Make the old page be freed below */
1907 new_page = old_page;
1909 page_cache_release(new_page);
1910 page_cache_release(old_page);
1914 /* Return the pagecache page at a given address within a VMA */
1915 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
1916 struct vm_area_struct *vma, unsigned long address)
1918 struct address_space *mapping;
1921 mapping = vma->vm_file->f_mapping;
1922 idx = vma_hugecache_offset(h, vma, address);
1924 return find_lock_page(mapping, idx);
1927 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1928 unsigned long address, pte_t *ptep, int write_access)
1930 struct hstate *h = hstate_vma(vma);
1931 int ret = VM_FAULT_SIGBUS;
1935 struct address_space *mapping;
1939 * Currently, we are forced to kill the process in the event the
1940 * original mapper has unmapped pages from the child due to a failed
1941 * COW. Warn that such a situation has occured as it may not be obvious
1943 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
1945 "PID %d killed due to inadequate hugepage pool\n",
1950 mapping = vma->vm_file->f_mapping;
1951 idx = vma_hugecache_offset(h, vma, address);
1954 * Use page lock to guard against racing truncation
1955 * before we get page_table_lock.
1958 page = find_lock_page(mapping, idx);
1960 size = i_size_read(mapping->host) >> huge_page_shift(h);
1963 page = alloc_huge_page(vma, address, 0);
1965 ret = -PTR_ERR(page);
1968 clear_huge_page(page, address, huge_page_size(h));
1969 __SetPageUptodate(page);
1971 if (vma->vm_flags & VM_SHARED) {
1973 struct inode *inode = mapping->host;
1975 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
1983 spin_lock(&inode->i_lock);
1984 inode->i_blocks += blocks_per_huge_page(h);
1985 spin_unlock(&inode->i_lock);
1991 * If we are going to COW a private mapping later, we examine the
1992 * pending reservations for this page now. This will ensure that
1993 * any allocations necessary to record that reservation occur outside
1996 if (write_access && !(vma->vm_flags & VM_SHARED))
1997 if (vma_needs_reservation(h, vma, address) < 0) {
1999 goto backout_unlocked;
2002 spin_lock(&mm->page_table_lock);
2003 size = i_size_read(mapping->host) >> huge_page_shift(h);
2008 if (!huge_pte_none(huge_ptep_get(ptep)))
2011 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2012 && (vma->vm_flags & VM_SHARED)));
2013 set_huge_pte_at(mm, address, ptep, new_pte);
2015 if (write_access && !(vma->vm_flags & VM_SHARED)) {
2016 /* Optimization, do the COW without a second fault */
2017 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2020 spin_unlock(&mm->page_table_lock);
2026 spin_unlock(&mm->page_table_lock);
2033 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2034 unsigned long address, int write_access)
2039 struct page *pagecache_page = NULL;
2040 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2041 struct hstate *h = hstate_vma(vma);
2043 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2045 return VM_FAULT_OOM;
2048 * Serialize hugepage allocation and instantiation, so that we don't
2049 * get spurious allocation failures if two CPUs race to instantiate
2050 * the same page in the page cache.
2052 mutex_lock(&hugetlb_instantiation_mutex);
2053 entry = huge_ptep_get(ptep);
2054 if (huge_pte_none(entry)) {
2055 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
2062 * If we are going to COW the mapping later, we examine the pending
2063 * reservations for this page now. This will ensure that any
2064 * allocations necessary to record that reservation occur outside the
2065 * spinlock. For private mappings, we also lookup the pagecache
2066 * page now as it is used to determine if a reservation has been
2069 if (write_access && !pte_write(entry)) {
2070 if (vma_needs_reservation(h, vma, address) < 0) {
2075 if (!(vma->vm_flags & VM_SHARED))
2076 pagecache_page = hugetlbfs_pagecache_page(h,
2080 spin_lock(&mm->page_table_lock);
2081 /* Check for a racing update before calling hugetlb_cow */
2082 if (likely(pte_same(entry, huge_ptep_get(ptep))))
2083 if (write_access && !pte_write(entry))
2084 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2086 spin_unlock(&mm->page_table_lock);
2088 if (pagecache_page) {
2089 unlock_page(pagecache_page);
2090 put_page(pagecache_page);
2094 mutex_unlock(&hugetlb_instantiation_mutex);
2099 /* Can be overriden by architectures */
2100 __attribute__((weak)) struct page *
2101 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2102 pud_t *pud, int write)
2108 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2109 struct page **pages, struct vm_area_struct **vmas,
2110 unsigned long *position, int *length, int i,
2113 unsigned long pfn_offset;
2114 unsigned long vaddr = *position;
2115 int remainder = *length;
2116 struct hstate *h = hstate_vma(vma);
2118 spin_lock(&mm->page_table_lock);
2119 while (vaddr < vma->vm_end && remainder) {
2124 * Some archs (sparc64, sh*) have multiple pte_ts to
2125 * each hugepage. We have to make * sure we get the
2126 * first, for the page indexing below to work.
2128 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2130 if (!pte || huge_pte_none(huge_ptep_get(pte)) ||
2131 (write && !pte_write(huge_ptep_get(pte)))) {
2134 spin_unlock(&mm->page_table_lock);
2135 ret = hugetlb_fault(mm, vma, vaddr, write);
2136 spin_lock(&mm->page_table_lock);
2137 if (!(ret & VM_FAULT_ERROR))
2146 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2147 page = pte_page(huge_ptep_get(pte));
2151 pages[i] = mem_map_offset(page, pfn_offset);
2161 if (vaddr < vma->vm_end && remainder &&
2162 pfn_offset < pages_per_huge_page(h)) {
2164 * We use pfn_offset to avoid touching the pageframes
2165 * of this compound page.
2170 spin_unlock(&mm->page_table_lock);
2171 *length = remainder;
2177 void hugetlb_change_protection(struct vm_area_struct *vma,
2178 unsigned long address, unsigned long end, pgprot_t newprot)
2180 struct mm_struct *mm = vma->vm_mm;
2181 unsigned long start = address;
2184 struct hstate *h = hstate_vma(vma);
2186 BUG_ON(address >= end);
2187 flush_cache_range(vma, address, end);
2189 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2190 spin_lock(&mm->page_table_lock);
2191 for (; address < end; address += huge_page_size(h)) {
2192 ptep = huge_pte_offset(mm, address);
2195 if (huge_pmd_unshare(mm, &address, ptep))
2197 if (!huge_pte_none(huge_ptep_get(ptep))) {
2198 pte = huge_ptep_get_and_clear(mm, address, ptep);
2199 pte = pte_mkhuge(pte_modify(pte, newprot));
2200 set_huge_pte_at(mm, address, ptep, pte);
2203 spin_unlock(&mm->page_table_lock);
2204 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2206 flush_tlb_range(vma, start, end);
2209 int hugetlb_reserve_pages(struct inode *inode,
2211 struct vm_area_struct *vma)
2214 struct hstate *h = hstate_inode(inode);
2216 if (vma && vma->vm_flags & VM_NORESERVE)
2220 * Shared mappings base their reservation on the number of pages that
2221 * are already allocated on behalf of the file. Private mappings need
2222 * to reserve the full area even if read-only as mprotect() may be
2223 * called to make the mapping read-write. Assume !vma is a shm mapping
2225 if (!vma || vma->vm_flags & VM_SHARED)
2226 chg = region_chg(&inode->i_mapping->private_list, from, to);
2228 struct resv_map *resv_map = resv_map_alloc();
2234 set_vma_resv_map(vma, resv_map);
2235 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2241 if (hugetlb_get_quota(inode->i_mapping, chg))
2243 ret = hugetlb_acct_memory(h, chg);
2245 hugetlb_put_quota(inode->i_mapping, chg);
2248 if (!vma || vma->vm_flags & VM_SHARED)
2249 region_add(&inode->i_mapping->private_list, from, to);
2253 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2255 struct hstate *h = hstate_inode(inode);
2256 long chg = region_truncate(&inode->i_mapping->private_list, offset);
2258 spin_lock(&inode->i_lock);
2259 inode->i_blocks -= blocks_per_huge_page(h);
2260 spin_unlock(&inode->i_lock);
2262 hugetlb_put_quota(inode->i_mapping, (chg - freed));
2263 hugetlb_acct_memory(h, -(chg - freed));