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_huge_page(struct page *page,
357 unsigned long addr, unsigned long sz)
362 for (i = 0; i < sz/PAGE_SIZE; i++) {
364 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
368 static void copy_huge_page(struct page *dst, struct page *src,
369 unsigned long addr, struct vm_area_struct *vma)
372 struct hstate *h = hstate_vma(vma);
375 for (i = 0; i < pages_per_huge_page(h); i++) {
377 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
381 static void enqueue_huge_page(struct hstate *h, struct page *page)
383 int nid = page_to_nid(page);
384 list_add(&page->lru, &h->hugepage_freelists[nid]);
385 h->free_huge_pages++;
386 h->free_huge_pages_node[nid]++;
389 static struct page *dequeue_huge_page(struct hstate *h)
392 struct page *page = NULL;
394 for (nid = 0; nid < MAX_NUMNODES; ++nid) {
395 if (!list_empty(&h->hugepage_freelists[nid])) {
396 page = list_entry(h->hugepage_freelists[nid].next,
398 list_del(&page->lru);
399 h->free_huge_pages--;
400 h->free_huge_pages_node[nid]--;
407 static struct page *dequeue_huge_page_vma(struct hstate *h,
408 struct vm_area_struct *vma,
409 unsigned long address, int avoid_reserve)
412 struct page *page = NULL;
413 struct mempolicy *mpol;
414 nodemask_t *nodemask;
415 struct zonelist *zonelist = huge_zonelist(vma, address,
416 htlb_alloc_mask, &mpol, &nodemask);
421 * A child process with MAP_PRIVATE mappings created by their parent
422 * have no page reserves. This check ensures that reservations are
423 * not "stolen". The child may still get SIGKILLed
425 if (!vma_has_reserves(vma) &&
426 h->free_huge_pages - h->resv_huge_pages == 0)
429 /* If reserves cannot be used, ensure enough pages are in the pool */
430 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
433 for_each_zone_zonelist_nodemask(zone, z, zonelist,
434 MAX_NR_ZONES - 1, nodemask) {
435 nid = zone_to_nid(zone);
436 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
437 !list_empty(&h->hugepage_freelists[nid])) {
438 page = list_entry(h->hugepage_freelists[nid].next,
440 list_del(&page->lru);
441 h->free_huge_pages--;
442 h->free_huge_pages_node[nid]--;
445 decrement_hugepage_resv_vma(h, vma);
454 static void update_and_free_page(struct hstate *h, struct page *page)
458 VM_BUG_ON(h->order >= MAX_ORDER);
461 h->nr_huge_pages_node[page_to_nid(page)]--;
462 for (i = 0; i < pages_per_huge_page(h); i++) {
463 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
464 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
465 1 << PG_private | 1<< PG_writeback);
467 set_compound_page_dtor(page, NULL);
468 set_page_refcounted(page);
469 arch_release_hugepage(page);
470 __free_pages(page, huge_page_order(h));
473 struct hstate *size_to_hstate(unsigned long size)
478 if (huge_page_size(h) == size)
484 static void free_huge_page(struct page *page)
487 * Can't pass hstate in here because it is called from the
488 * compound page destructor.
490 struct hstate *h = page_hstate(page);
491 int nid = page_to_nid(page);
492 struct address_space *mapping;
494 mapping = (struct address_space *) page_private(page);
495 set_page_private(page, 0);
496 BUG_ON(page_count(page));
497 INIT_LIST_HEAD(&page->lru);
499 spin_lock(&hugetlb_lock);
500 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
501 update_and_free_page(h, page);
502 h->surplus_huge_pages--;
503 h->surplus_huge_pages_node[nid]--;
505 enqueue_huge_page(h, page);
507 spin_unlock(&hugetlb_lock);
509 hugetlb_put_quota(mapping, 1);
513 * Increment or decrement surplus_huge_pages. Keep node-specific counters
514 * balanced by operating on them in a round-robin fashion.
515 * Returns 1 if an adjustment was made.
517 static int adjust_pool_surplus(struct hstate *h, int delta)
523 VM_BUG_ON(delta != -1 && delta != 1);
525 nid = next_node(nid, node_online_map);
526 if (nid == MAX_NUMNODES)
527 nid = first_node(node_online_map);
529 /* To shrink on this node, there must be a surplus page */
530 if (delta < 0 && !h->surplus_huge_pages_node[nid])
532 /* Surplus cannot exceed the total number of pages */
533 if (delta > 0 && h->surplus_huge_pages_node[nid] >=
534 h->nr_huge_pages_node[nid])
537 h->surplus_huge_pages += delta;
538 h->surplus_huge_pages_node[nid] += delta;
541 } while (nid != prev_nid);
547 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
549 set_compound_page_dtor(page, free_huge_page);
550 spin_lock(&hugetlb_lock);
552 h->nr_huge_pages_node[nid]++;
553 spin_unlock(&hugetlb_lock);
554 put_page(page); /* free it into the hugepage allocator */
557 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
561 if (h->order >= MAX_ORDER)
564 page = alloc_pages_node(nid,
565 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
566 __GFP_REPEAT|__GFP_NOWARN,
569 if (arch_prepare_hugepage(page)) {
570 __free_pages(page, huge_page_order(h));
573 prep_new_huge_page(h, page, nid);
580 * Use a helper variable to find the next node and then
581 * copy it back to hugetlb_next_nid afterwards:
582 * otherwise there's a window in which a racer might
583 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
584 * But we don't need to use a spin_lock here: it really
585 * doesn't matter if occasionally a racer chooses the
586 * same nid as we do. Move nid forward in the mask even
587 * if we just successfully allocated a hugepage so that
588 * the next caller gets hugepages on the next node.
590 static int hstate_next_node(struct hstate *h)
593 next_nid = next_node(h->hugetlb_next_nid, node_online_map);
594 if (next_nid == MAX_NUMNODES)
595 next_nid = first_node(node_online_map);
596 h->hugetlb_next_nid = next_nid;
600 static int alloc_fresh_huge_page(struct hstate *h)
607 start_nid = h->hugetlb_next_nid;
610 page = alloc_fresh_huge_page_node(h, h->hugetlb_next_nid);
613 next_nid = hstate_next_node(h);
614 } while (!page && h->hugetlb_next_nid != start_nid);
617 count_vm_event(HTLB_BUDDY_PGALLOC);
619 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
624 static struct page *alloc_buddy_huge_page(struct hstate *h,
625 struct vm_area_struct *vma, unsigned long address)
630 if (h->order >= MAX_ORDER)
634 * Assume we will successfully allocate the surplus page to
635 * prevent racing processes from causing the surplus to exceed
638 * This however introduces a different race, where a process B
639 * tries to grow the static hugepage pool while alloc_pages() is
640 * called by process A. B will only examine the per-node
641 * counters in determining if surplus huge pages can be
642 * converted to normal huge pages in adjust_pool_surplus(). A
643 * won't be able to increment the per-node counter, until the
644 * lock is dropped by B, but B doesn't drop hugetlb_lock until
645 * no more huge pages can be converted from surplus to normal
646 * state (and doesn't try to convert again). Thus, we have a
647 * case where a surplus huge page exists, the pool is grown, and
648 * the surplus huge page still exists after, even though it
649 * should just have been converted to a normal huge page. This
650 * does not leak memory, though, as the hugepage will be freed
651 * once it is out of use. It also does not allow the counters to
652 * go out of whack in adjust_pool_surplus() as we don't modify
653 * the node values until we've gotten the hugepage and only the
654 * per-node value is checked there.
656 spin_lock(&hugetlb_lock);
657 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
658 spin_unlock(&hugetlb_lock);
662 h->surplus_huge_pages++;
664 spin_unlock(&hugetlb_lock);
666 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
667 __GFP_REPEAT|__GFP_NOWARN,
670 if (page && arch_prepare_hugepage(page)) {
671 __free_pages(page, huge_page_order(h));
675 spin_lock(&hugetlb_lock);
678 * This page is now managed by the hugetlb allocator and has
679 * no users -- drop the buddy allocator's reference.
681 put_page_testzero(page);
682 VM_BUG_ON(page_count(page));
683 nid = page_to_nid(page);
684 set_compound_page_dtor(page, free_huge_page);
686 * We incremented the global counters already
688 h->nr_huge_pages_node[nid]++;
689 h->surplus_huge_pages_node[nid]++;
690 __count_vm_event(HTLB_BUDDY_PGALLOC);
693 h->surplus_huge_pages--;
694 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
696 spin_unlock(&hugetlb_lock);
702 * Increase the hugetlb pool such that it can accomodate a reservation
705 static int gather_surplus_pages(struct hstate *h, int delta)
707 struct list_head surplus_list;
708 struct page *page, *tmp;
710 int needed, allocated;
712 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
714 h->resv_huge_pages += delta;
719 INIT_LIST_HEAD(&surplus_list);
723 spin_unlock(&hugetlb_lock);
724 for (i = 0; i < needed; i++) {
725 page = alloc_buddy_huge_page(h, NULL, 0);
728 * We were not able to allocate enough pages to
729 * satisfy the entire reservation so we free what
730 * we've allocated so far.
732 spin_lock(&hugetlb_lock);
737 list_add(&page->lru, &surplus_list);
742 * After retaking hugetlb_lock, we need to recalculate 'needed'
743 * because either resv_huge_pages or free_huge_pages may have changed.
745 spin_lock(&hugetlb_lock);
746 needed = (h->resv_huge_pages + delta) -
747 (h->free_huge_pages + allocated);
752 * The surplus_list now contains _at_least_ the number of extra pages
753 * needed to accomodate the reservation. Add the appropriate number
754 * of pages to the hugetlb pool and free the extras back to the buddy
755 * allocator. Commit the entire reservation here to prevent another
756 * process from stealing the pages as they are added to the pool but
757 * before they are reserved.
760 h->resv_huge_pages += delta;
763 /* Free the needed pages to the hugetlb pool */
764 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
767 list_del(&page->lru);
768 enqueue_huge_page(h, page);
771 /* Free unnecessary surplus pages to the buddy allocator */
772 if (!list_empty(&surplus_list)) {
773 spin_unlock(&hugetlb_lock);
774 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
775 list_del(&page->lru);
777 * The page has a reference count of zero already, so
778 * call free_huge_page directly instead of using
779 * put_page. This must be done with hugetlb_lock
780 * unlocked which is safe because free_huge_page takes
781 * hugetlb_lock before deciding how to free the page.
783 free_huge_page(page);
785 spin_lock(&hugetlb_lock);
792 * When releasing a hugetlb pool reservation, any surplus pages that were
793 * allocated to satisfy the reservation must be explicitly freed if they were
796 static void return_unused_surplus_pages(struct hstate *h,
797 unsigned long unused_resv_pages)
801 unsigned long nr_pages;
804 * We want to release as many surplus pages as possible, spread
805 * evenly across all nodes. Iterate across all nodes until we
806 * can no longer free unreserved surplus pages. This occurs when
807 * the nodes with surplus pages have no free pages.
809 unsigned long remaining_iterations = num_online_nodes();
811 /* Uncommit the reservation */
812 h->resv_huge_pages -= unused_resv_pages;
814 /* Cannot return gigantic pages currently */
815 if (h->order >= MAX_ORDER)
818 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
820 while (remaining_iterations-- && nr_pages) {
821 nid = next_node(nid, node_online_map);
822 if (nid == MAX_NUMNODES)
823 nid = first_node(node_online_map);
825 if (!h->surplus_huge_pages_node[nid])
828 if (!list_empty(&h->hugepage_freelists[nid])) {
829 page = list_entry(h->hugepage_freelists[nid].next,
831 list_del(&page->lru);
832 update_and_free_page(h, page);
833 h->free_huge_pages--;
834 h->free_huge_pages_node[nid]--;
835 h->surplus_huge_pages--;
836 h->surplus_huge_pages_node[nid]--;
838 remaining_iterations = num_online_nodes();
844 * Determine if the huge page at addr within the vma has an associated
845 * reservation. Where it does not we will need to logically increase
846 * reservation and actually increase quota before an allocation can occur.
847 * Where any new reservation would be required the reservation change is
848 * prepared, but not committed. Once the page has been quota'd allocated
849 * an instantiated the change should be committed via vma_commit_reservation.
850 * No action is required on failure.
852 static int vma_needs_reservation(struct hstate *h,
853 struct vm_area_struct *vma, unsigned long addr)
855 struct address_space *mapping = vma->vm_file->f_mapping;
856 struct inode *inode = mapping->host;
858 if (vma->vm_flags & VM_SHARED) {
859 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
860 return region_chg(&inode->i_mapping->private_list,
863 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
868 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
869 struct resv_map *reservations = vma_resv_map(vma);
871 err = region_chg(&reservations->regions, idx, idx + 1);
877 static void vma_commit_reservation(struct hstate *h,
878 struct vm_area_struct *vma, unsigned long addr)
880 struct address_space *mapping = vma->vm_file->f_mapping;
881 struct inode *inode = mapping->host;
883 if (vma->vm_flags & VM_SHARED) {
884 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
885 region_add(&inode->i_mapping->private_list, idx, idx + 1);
887 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
888 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
889 struct resv_map *reservations = vma_resv_map(vma);
891 /* Mark this page used in the map. */
892 region_add(&reservations->regions, idx, idx + 1);
896 static struct page *alloc_huge_page(struct vm_area_struct *vma,
897 unsigned long addr, int avoid_reserve)
899 struct hstate *h = hstate_vma(vma);
901 struct address_space *mapping = vma->vm_file->f_mapping;
902 struct inode *inode = mapping->host;
906 * Processes that did not create the mapping will have no reserves and
907 * will not have accounted against quota. Check that the quota can be
908 * made before satisfying the allocation
909 * MAP_NORESERVE mappings may also need pages and quota allocated
910 * if no reserve mapping overlaps.
912 chg = vma_needs_reservation(h, vma, addr);
916 if (hugetlb_get_quota(inode->i_mapping, chg))
917 return ERR_PTR(-ENOSPC);
919 spin_lock(&hugetlb_lock);
920 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
921 spin_unlock(&hugetlb_lock);
924 page = alloc_buddy_huge_page(h, vma, addr);
926 hugetlb_put_quota(inode->i_mapping, chg);
927 return ERR_PTR(-VM_FAULT_OOM);
931 set_page_refcounted(page);
932 set_page_private(page, (unsigned long) mapping);
934 vma_commit_reservation(h, vma, addr);
939 __attribute__((weak)) int alloc_bootmem_huge_page(struct hstate *h)
941 struct huge_bootmem_page *m;
942 int nr_nodes = nodes_weight(node_online_map);
947 addr = __alloc_bootmem_node_nopanic(
948 NODE_DATA(h->hugetlb_next_nid),
949 huge_page_size(h), huge_page_size(h), 0);
953 * Use the beginning of the huge page to store the
954 * huge_bootmem_page struct (until gather_bootmem
955 * puts them into the mem_map).
967 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
968 /* Put them into a private list first because mem_map is not up yet */
969 list_add(&m->list, &huge_boot_pages);
974 static void prep_compound_huge_page(struct page *page, int order)
976 if (unlikely(order > (MAX_ORDER - 1)))
977 prep_compound_gigantic_page(page, order);
979 prep_compound_page(page, order);
982 /* Put bootmem huge pages into the standard lists after mem_map is up */
983 static void __init gather_bootmem_prealloc(void)
985 struct huge_bootmem_page *m;
987 list_for_each_entry(m, &huge_boot_pages, list) {
988 struct page *page = virt_to_page(m);
989 struct hstate *h = m->hstate;
990 __ClearPageReserved(page);
991 WARN_ON(page_count(page) != 1);
992 prep_compound_huge_page(page, h->order);
993 prep_new_huge_page(h, page, page_to_nid(page));
997 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1001 for (i = 0; i < h->max_huge_pages; ++i) {
1002 if (h->order >= MAX_ORDER) {
1003 if (!alloc_bootmem_huge_page(h))
1005 } else if (!alloc_fresh_huge_page(h))
1008 h->max_huge_pages = i;
1011 static void __init hugetlb_init_hstates(void)
1015 for_each_hstate(h) {
1016 /* oversize hugepages were init'ed in early boot */
1017 if (h->order < MAX_ORDER)
1018 hugetlb_hstate_alloc_pages(h);
1022 static char * __init memfmt(char *buf, unsigned long n)
1024 if (n >= (1UL << 30))
1025 sprintf(buf, "%lu GB", n >> 30);
1026 else if (n >= (1UL << 20))
1027 sprintf(buf, "%lu MB", n >> 20);
1029 sprintf(buf, "%lu KB", n >> 10);
1033 static void __init report_hugepages(void)
1037 for_each_hstate(h) {
1039 printk(KERN_INFO "HugeTLB registered %s page size, "
1040 "pre-allocated %ld pages\n",
1041 memfmt(buf, huge_page_size(h)),
1042 h->free_huge_pages);
1046 #ifdef CONFIG_HIGHMEM
1047 static void try_to_free_low(struct hstate *h, unsigned long count)
1051 if (h->order >= MAX_ORDER)
1054 for (i = 0; i < MAX_NUMNODES; ++i) {
1055 struct page *page, *next;
1056 struct list_head *freel = &h->hugepage_freelists[i];
1057 list_for_each_entry_safe(page, next, freel, lru) {
1058 if (count >= h->nr_huge_pages)
1060 if (PageHighMem(page))
1062 list_del(&page->lru);
1063 update_and_free_page(h, page);
1064 h->free_huge_pages--;
1065 h->free_huge_pages_node[page_to_nid(page)]--;
1070 static inline void try_to_free_low(struct hstate *h, unsigned long count)
1075 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1076 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count)
1078 unsigned long min_count, ret;
1080 if (h->order >= MAX_ORDER)
1081 return h->max_huge_pages;
1084 * Increase the pool size
1085 * First take pages out of surplus state. Then make up the
1086 * remaining difference by allocating fresh huge pages.
1088 * We might race with alloc_buddy_huge_page() here and be unable
1089 * to convert a surplus huge page to a normal huge page. That is
1090 * not critical, though, it just means the overall size of the
1091 * pool might be one hugepage larger than it needs to be, but
1092 * within all the constraints specified by the sysctls.
1094 spin_lock(&hugetlb_lock);
1095 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1096 if (!adjust_pool_surplus(h, -1))
1100 while (count > persistent_huge_pages(h)) {
1102 * If this allocation races such that we no longer need the
1103 * page, free_huge_page will handle it by freeing the page
1104 * and reducing the surplus.
1106 spin_unlock(&hugetlb_lock);
1107 ret = alloc_fresh_huge_page(h);
1108 spin_lock(&hugetlb_lock);
1115 * Decrease the pool size
1116 * First return free pages to the buddy allocator (being careful
1117 * to keep enough around to satisfy reservations). Then place
1118 * pages into surplus state as needed so the pool will shrink
1119 * to the desired size as pages become free.
1121 * By placing pages into the surplus state independent of the
1122 * overcommit value, we are allowing the surplus pool size to
1123 * exceed overcommit. There are few sane options here. Since
1124 * alloc_buddy_huge_page() is checking the global counter,
1125 * though, we'll note that we're not allowed to exceed surplus
1126 * and won't grow the pool anywhere else. Not until one of the
1127 * sysctls are changed, or the surplus pages go out of use.
1129 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1130 min_count = max(count, min_count);
1131 try_to_free_low(h, min_count);
1132 while (min_count < persistent_huge_pages(h)) {
1133 struct page *page = dequeue_huge_page(h);
1136 update_and_free_page(h, page);
1138 while (count < persistent_huge_pages(h)) {
1139 if (!adjust_pool_surplus(h, 1))
1143 ret = persistent_huge_pages(h);
1144 spin_unlock(&hugetlb_lock);
1148 #define HSTATE_ATTR_RO(_name) \
1149 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1151 #define HSTATE_ATTR(_name) \
1152 static struct kobj_attribute _name##_attr = \
1153 __ATTR(_name, 0644, _name##_show, _name##_store)
1155 static struct kobject *hugepages_kobj;
1156 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1158 static struct hstate *kobj_to_hstate(struct kobject *kobj)
1161 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1162 if (hstate_kobjs[i] == kobj)
1168 static ssize_t nr_hugepages_show(struct kobject *kobj,
1169 struct kobj_attribute *attr, char *buf)
1171 struct hstate *h = kobj_to_hstate(kobj);
1172 return sprintf(buf, "%lu\n", h->nr_huge_pages);
1174 static ssize_t nr_hugepages_store(struct kobject *kobj,
1175 struct kobj_attribute *attr, const char *buf, size_t count)
1178 unsigned long input;
1179 struct hstate *h = kobj_to_hstate(kobj);
1181 err = strict_strtoul(buf, 10, &input);
1185 h->max_huge_pages = set_max_huge_pages(h, input);
1189 HSTATE_ATTR(nr_hugepages);
1191 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1192 struct kobj_attribute *attr, char *buf)
1194 struct hstate *h = kobj_to_hstate(kobj);
1195 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1197 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1198 struct kobj_attribute *attr, const char *buf, size_t count)
1201 unsigned long input;
1202 struct hstate *h = kobj_to_hstate(kobj);
1204 err = strict_strtoul(buf, 10, &input);
1208 spin_lock(&hugetlb_lock);
1209 h->nr_overcommit_huge_pages = input;
1210 spin_unlock(&hugetlb_lock);
1214 HSTATE_ATTR(nr_overcommit_hugepages);
1216 static ssize_t free_hugepages_show(struct kobject *kobj,
1217 struct kobj_attribute *attr, char *buf)
1219 struct hstate *h = kobj_to_hstate(kobj);
1220 return sprintf(buf, "%lu\n", h->free_huge_pages);
1222 HSTATE_ATTR_RO(free_hugepages);
1224 static ssize_t resv_hugepages_show(struct kobject *kobj,
1225 struct kobj_attribute *attr, char *buf)
1227 struct hstate *h = kobj_to_hstate(kobj);
1228 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1230 HSTATE_ATTR_RO(resv_hugepages);
1232 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1233 struct kobj_attribute *attr, char *buf)
1235 struct hstate *h = kobj_to_hstate(kobj);
1236 return sprintf(buf, "%lu\n", h->surplus_huge_pages);
1238 HSTATE_ATTR_RO(surplus_hugepages);
1240 static struct attribute *hstate_attrs[] = {
1241 &nr_hugepages_attr.attr,
1242 &nr_overcommit_hugepages_attr.attr,
1243 &free_hugepages_attr.attr,
1244 &resv_hugepages_attr.attr,
1245 &surplus_hugepages_attr.attr,
1249 static struct attribute_group hstate_attr_group = {
1250 .attrs = hstate_attrs,
1253 static int __init hugetlb_sysfs_add_hstate(struct hstate *h)
1257 hstate_kobjs[h - hstates] = kobject_create_and_add(h->name,
1259 if (!hstate_kobjs[h - hstates])
1262 retval = sysfs_create_group(hstate_kobjs[h - hstates],
1263 &hstate_attr_group);
1265 kobject_put(hstate_kobjs[h - hstates]);
1270 static void __init hugetlb_sysfs_init(void)
1275 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1276 if (!hugepages_kobj)
1279 for_each_hstate(h) {
1280 err = hugetlb_sysfs_add_hstate(h);
1282 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1287 static void __exit hugetlb_exit(void)
1291 for_each_hstate(h) {
1292 kobject_put(hstate_kobjs[h - hstates]);
1295 kobject_put(hugepages_kobj);
1297 module_exit(hugetlb_exit);
1299 static int __init hugetlb_init(void)
1301 /* Some platform decide whether they support huge pages at boot
1302 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1303 * there is no such support
1305 if (HPAGE_SHIFT == 0)
1308 if (!size_to_hstate(default_hstate_size)) {
1309 default_hstate_size = HPAGE_SIZE;
1310 if (!size_to_hstate(default_hstate_size))
1311 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1313 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1314 if (default_hstate_max_huge_pages)
1315 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1317 hugetlb_init_hstates();
1319 gather_bootmem_prealloc();
1323 hugetlb_sysfs_init();
1327 module_init(hugetlb_init);
1329 /* Should be called on processing a hugepagesz=... option */
1330 void __init hugetlb_add_hstate(unsigned order)
1335 if (size_to_hstate(PAGE_SIZE << order)) {
1336 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1339 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1341 h = &hstates[max_hstate++];
1343 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1344 h->nr_huge_pages = 0;
1345 h->free_huge_pages = 0;
1346 for (i = 0; i < MAX_NUMNODES; ++i)
1347 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1348 h->hugetlb_next_nid = first_node(node_online_map);
1349 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1350 huge_page_size(h)/1024);
1355 static int __init hugetlb_nrpages_setup(char *s)
1358 static unsigned long *last_mhp;
1361 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1362 * so this hugepages= parameter goes to the "default hstate".
1365 mhp = &default_hstate_max_huge_pages;
1367 mhp = &parsed_hstate->max_huge_pages;
1369 if (mhp == last_mhp) {
1370 printk(KERN_WARNING "hugepages= specified twice without "
1371 "interleaving hugepagesz=, ignoring\n");
1375 if (sscanf(s, "%lu", mhp) <= 0)
1379 * Global state is always initialized later in hugetlb_init.
1380 * But we need to allocate >= MAX_ORDER hstates here early to still
1381 * use the bootmem allocator.
1383 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1384 hugetlb_hstate_alloc_pages(parsed_hstate);
1390 __setup("hugepages=", hugetlb_nrpages_setup);
1392 static int __init hugetlb_default_setup(char *s)
1394 default_hstate_size = memparse(s, &s);
1397 __setup("default_hugepagesz=", hugetlb_default_setup);
1399 static unsigned int cpuset_mems_nr(unsigned int *array)
1402 unsigned int nr = 0;
1404 for_each_node_mask(node, cpuset_current_mems_allowed)
1410 #ifdef CONFIG_SYSCTL
1411 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1412 struct file *file, void __user *buffer,
1413 size_t *length, loff_t *ppos)
1415 struct hstate *h = &default_hstate;
1419 tmp = h->max_huge_pages;
1422 table->maxlen = sizeof(unsigned long);
1423 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1426 h->max_huge_pages = set_max_huge_pages(h, tmp);
1431 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1432 struct file *file, void __user *buffer,
1433 size_t *length, loff_t *ppos)
1435 proc_dointvec(table, write, file, buffer, length, ppos);
1436 if (hugepages_treat_as_movable)
1437 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1439 htlb_alloc_mask = GFP_HIGHUSER;
1443 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1444 struct file *file, void __user *buffer,
1445 size_t *length, loff_t *ppos)
1447 struct hstate *h = &default_hstate;
1451 tmp = h->nr_overcommit_huge_pages;
1454 table->maxlen = sizeof(unsigned long);
1455 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1458 spin_lock(&hugetlb_lock);
1459 h->nr_overcommit_huge_pages = tmp;
1460 spin_unlock(&hugetlb_lock);
1466 #endif /* CONFIG_SYSCTL */
1468 int hugetlb_report_meminfo(char *buf)
1470 struct hstate *h = &default_hstate;
1472 "HugePages_Total: %5lu\n"
1473 "HugePages_Free: %5lu\n"
1474 "HugePages_Rsvd: %5lu\n"
1475 "HugePages_Surp: %5lu\n"
1476 "Hugepagesize: %5lu kB\n",
1480 h->surplus_huge_pages,
1481 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1484 int hugetlb_report_node_meminfo(int nid, char *buf)
1486 struct hstate *h = &default_hstate;
1488 "Node %d HugePages_Total: %5u\n"
1489 "Node %d HugePages_Free: %5u\n"
1490 "Node %d HugePages_Surp: %5u\n",
1491 nid, h->nr_huge_pages_node[nid],
1492 nid, h->free_huge_pages_node[nid],
1493 nid, h->surplus_huge_pages_node[nid]);
1496 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1497 unsigned long hugetlb_total_pages(void)
1499 struct hstate *h = &default_hstate;
1500 return h->nr_huge_pages * pages_per_huge_page(h);
1503 static int hugetlb_acct_memory(struct hstate *h, long delta)
1507 spin_lock(&hugetlb_lock);
1509 * When cpuset is configured, it breaks the strict hugetlb page
1510 * reservation as the accounting is done on a global variable. Such
1511 * reservation is completely rubbish in the presence of cpuset because
1512 * the reservation is not checked against page availability for the
1513 * current cpuset. Application can still potentially OOM'ed by kernel
1514 * with lack of free htlb page in cpuset that the task is in.
1515 * Attempt to enforce strict accounting with cpuset is almost
1516 * impossible (or too ugly) because cpuset is too fluid that
1517 * task or memory node can be dynamically moved between cpusets.
1519 * The change of semantics for shared hugetlb mapping with cpuset is
1520 * undesirable. However, in order to preserve some of the semantics,
1521 * we fall back to check against current free page availability as
1522 * a best attempt and hopefully to minimize the impact of changing
1523 * semantics that cpuset has.
1526 if (gather_surplus_pages(h, delta) < 0)
1529 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1530 return_unused_surplus_pages(h, delta);
1537 return_unused_surplus_pages(h, (unsigned long) -delta);
1540 spin_unlock(&hugetlb_lock);
1544 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
1546 struct resv_map *reservations = vma_resv_map(vma);
1549 * This new VMA should share its siblings reservation map if present.
1550 * The VMA will only ever have a valid reservation map pointer where
1551 * it is being copied for another still existing VMA. As that VMA
1552 * has a reference to the reservation map it cannot dissappear until
1553 * after this open call completes. It is therefore safe to take a
1554 * new reference here without additional locking.
1557 kref_get(&reservations->refs);
1560 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
1562 struct hstate *h = hstate_vma(vma);
1563 struct resv_map *reservations = vma_resv_map(vma);
1564 unsigned long reserve;
1565 unsigned long start;
1569 start = vma_hugecache_offset(h, vma, vma->vm_start);
1570 end = vma_hugecache_offset(h, vma, vma->vm_end);
1572 reserve = (end - start) -
1573 region_count(&reservations->regions, start, end);
1575 kref_put(&reservations->refs, resv_map_release);
1578 hugetlb_acct_memory(h, -reserve);
1579 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
1585 * We cannot handle pagefaults against hugetlb pages at all. They cause
1586 * handle_mm_fault() to try to instantiate regular-sized pages in the
1587 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
1590 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1596 struct vm_operations_struct hugetlb_vm_ops = {
1597 .fault = hugetlb_vm_op_fault,
1598 .open = hugetlb_vm_op_open,
1599 .close = hugetlb_vm_op_close,
1602 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
1609 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1611 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
1613 entry = pte_mkyoung(entry);
1614 entry = pte_mkhuge(entry);
1619 static void set_huge_ptep_writable(struct vm_area_struct *vma,
1620 unsigned long address, pte_t *ptep)
1624 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
1625 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
1626 update_mmu_cache(vma, address, entry);
1631 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
1632 struct vm_area_struct *vma)
1634 pte_t *src_pte, *dst_pte, entry;
1635 struct page *ptepage;
1638 struct hstate *h = hstate_vma(vma);
1639 unsigned long sz = huge_page_size(h);
1641 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
1643 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
1644 src_pte = huge_pte_offset(src, addr);
1647 dst_pte = huge_pte_alloc(dst, addr, sz);
1651 /* If the pagetables are shared don't copy or take references */
1652 if (dst_pte == src_pte)
1655 spin_lock(&dst->page_table_lock);
1656 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
1657 if (!huge_pte_none(huge_ptep_get(src_pte))) {
1659 huge_ptep_set_wrprotect(src, addr, src_pte);
1660 entry = huge_ptep_get(src_pte);
1661 ptepage = pte_page(entry);
1663 set_huge_pte_at(dst, addr, dst_pte, entry);
1665 spin_unlock(&src->page_table_lock);
1666 spin_unlock(&dst->page_table_lock);
1674 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1675 unsigned long end, struct page *ref_page)
1677 struct mm_struct *mm = vma->vm_mm;
1678 unsigned long address;
1683 struct hstate *h = hstate_vma(vma);
1684 unsigned long sz = huge_page_size(h);
1687 * A page gathering list, protected by per file i_mmap_lock. The
1688 * lock is used to avoid list corruption from multiple unmapping
1689 * of the same page since we are using page->lru.
1691 LIST_HEAD(page_list);
1693 WARN_ON(!is_vm_hugetlb_page(vma));
1694 BUG_ON(start & ~huge_page_mask(h));
1695 BUG_ON(end & ~huge_page_mask(h));
1697 mmu_notifier_invalidate_range_start(mm, start, end);
1698 spin_lock(&mm->page_table_lock);
1699 for (address = start; address < end; address += sz) {
1700 ptep = huge_pte_offset(mm, address);
1704 if (huge_pmd_unshare(mm, &address, ptep))
1708 * If a reference page is supplied, it is because a specific
1709 * page is being unmapped, not a range. Ensure the page we
1710 * are about to unmap is the actual page of interest.
1713 pte = huge_ptep_get(ptep);
1714 if (huge_pte_none(pte))
1716 page = pte_page(pte);
1717 if (page != ref_page)
1721 * Mark the VMA as having unmapped its page so that
1722 * future faults in this VMA will fail rather than
1723 * looking like data was lost
1725 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
1728 pte = huge_ptep_get_and_clear(mm, address, ptep);
1729 if (huge_pte_none(pte))
1732 page = pte_page(pte);
1734 set_page_dirty(page);
1735 list_add(&page->lru, &page_list);
1737 spin_unlock(&mm->page_table_lock);
1738 flush_tlb_range(vma, start, end);
1739 mmu_notifier_invalidate_range_end(mm, start, end);
1740 list_for_each_entry_safe(page, tmp, &page_list, lru) {
1741 list_del(&page->lru);
1746 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1747 unsigned long end, struct page *ref_page)
1749 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1750 __unmap_hugepage_range(vma, start, end, ref_page);
1751 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1755 * This is called when the original mapper is failing to COW a MAP_PRIVATE
1756 * mappping it owns the reserve page for. The intention is to unmap the page
1757 * from other VMAs and let the children be SIGKILLed if they are faulting the
1760 int unmap_ref_private(struct mm_struct *mm,
1761 struct vm_area_struct *vma,
1763 unsigned long address)
1765 struct vm_area_struct *iter_vma;
1766 struct address_space *mapping;
1767 struct prio_tree_iter iter;
1771 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1772 * from page cache lookup which is in HPAGE_SIZE units.
1774 address = address & huge_page_mask(hstate_vma(vma));
1775 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
1776 + (vma->vm_pgoff >> PAGE_SHIFT);
1777 mapping = (struct address_space *)page_private(page);
1779 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1780 /* Do not unmap the current VMA */
1781 if (iter_vma == vma)
1785 * Unmap the page from other VMAs without their own reserves.
1786 * They get marked to be SIGKILLed if they fault in these
1787 * areas. This is because a future no-page fault on this VMA
1788 * could insert a zeroed page instead of the data existing
1789 * from the time of fork. This would look like data corruption
1791 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
1792 unmap_hugepage_range(iter_vma,
1793 address, address + HPAGE_SIZE,
1800 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
1801 unsigned long address, pte_t *ptep, pte_t pte,
1802 struct page *pagecache_page)
1804 struct hstate *h = hstate_vma(vma);
1805 struct page *old_page, *new_page;
1807 int outside_reserve = 0;
1809 old_page = pte_page(pte);
1812 /* If no-one else is actually using this page, avoid the copy
1813 * and just make the page writable */
1814 avoidcopy = (page_count(old_page) == 1);
1816 set_huge_ptep_writable(vma, address, ptep);
1821 * If the process that created a MAP_PRIVATE mapping is about to
1822 * perform a COW due to a shared page count, attempt to satisfy
1823 * the allocation without using the existing reserves. The pagecache
1824 * page is used to determine if the reserve at this address was
1825 * consumed or not. If reserves were used, a partial faulted mapping
1826 * at the time of fork() could consume its reserves on COW instead
1827 * of the full address range.
1829 if (!(vma->vm_flags & VM_SHARED) &&
1830 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
1831 old_page != pagecache_page)
1832 outside_reserve = 1;
1834 page_cache_get(old_page);
1835 new_page = alloc_huge_page(vma, address, outside_reserve);
1837 if (IS_ERR(new_page)) {
1838 page_cache_release(old_page);
1841 * If a process owning a MAP_PRIVATE mapping fails to COW,
1842 * it is due to references held by a child and an insufficient
1843 * huge page pool. To guarantee the original mappers
1844 * reliability, unmap the page from child processes. The child
1845 * may get SIGKILLed if it later faults.
1847 if (outside_reserve) {
1848 BUG_ON(huge_pte_none(pte));
1849 if (unmap_ref_private(mm, vma, old_page, address)) {
1850 BUG_ON(page_count(old_page) != 1);
1851 BUG_ON(huge_pte_none(pte));
1852 goto retry_avoidcopy;
1857 return -PTR_ERR(new_page);
1860 spin_unlock(&mm->page_table_lock);
1861 copy_huge_page(new_page, old_page, address, vma);
1862 __SetPageUptodate(new_page);
1863 spin_lock(&mm->page_table_lock);
1865 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
1866 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
1868 huge_ptep_clear_flush(vma, address, ptep);
1869 set_huge_pte_at(mm, address, ptep,
1870 make_huge_pte(vma, new_page, 1));
1871 /* Make the old page be freed below */
1872 new_page = old_page;
1874 page_cache_release(new_page);
1875 page_cache_release(old_page);
1879 /* Return the pagecache page at a given address within a VMA */
1880 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
1881 struct vm_area_struct *vma, unsigned long address)
1883 struct address_space *mapping;
1886 mapping = vma->vm_file->f_mapping;
1887 idx = vma_hugecache_offset(h, vma, address);
1889 return find_lock_page(mapping, idx);
1892 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1893 unsigned long address, pte_t *ptep, int write_access)
1895 struct hstate *h = hstate_vma(vma);
1896 int ret = VM_FAULT_SIGBUS;
1900 struct address_space *mapping;
1904 * Currently, we are forced to kill the process in the event the
1905 * original mapper has unmapped pages from the child due to a failed
1906 * COW. Warn that such a situation has occured as it may not be obvious
1908 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
1910 "PID %d killed due to inadequate hugepage pool\n",
1915 mapping = vma->vm_file->f_mapping;
1916 idx = vma_hugecache_offset(h, vma, address);
1919 * Use page lock to guard against racing truncation
1920 * before we get page_table_lock.
1923 page = find_lock_page(mapping, idx);
1925 size = i_size_read(mapping->host) >> huge_page_shift(h);
1928 page = alloc_huge_page(vma, address, 0);
1930 ret = -PTR_ERR(page);
1933 clear_huge_page(page, address, huge_page_size(h));
1934 __SetPageUptodate(page);
1936 if (vma->vm_flags & VM_SHARED) {
1938 struct inode *inode = mapping->host;
1940 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
1948 spin_lock(&inode->i_lock);
1949 inode->i_blocks += blocks_per_huge_page(h);
1950 spin_unlock(&inode->i_lock);
1956 * If we are going to COW a private mapping later, we examine the
1957 * pending reservations for this page now. This will ensure that
1958 * any allocations necessary to record that reservation occur outside
1961 if (write_access && !(vma->vm_flags & VM_SHARED))
1962 if (vma_needs_reservation(h, vma, address) < 0) {
1964 goto backout_unlocked;
1967 spin_lock(&mm->page_table_lock);
1968 size = i_size_read(mapping->host) >> huge_page_shift(h);
1973 if (!huge_pte_none(huge_ptep_get(ptep)))
1976 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
1977 && (vma->vm_flags & VM_SHARED)));
1978 set_huge_pte_at(mm, address, ptep, new_pte);
1980 if (write_access && !(vma->vm_flags & VM_SHARED)) {
1981 /* Optimization, do the COW without a second fault */
1982 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
1985 spin_unlock(&mm->page_table_lock);
1991 spin_unlock(&mm->page_table_lock);
1998 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
1999 unsigned long address, int write_access)
2004 struct page *pagecache_page = NULL;
2005 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2006 struct hstate *h = hstate_vma(vma);
2008 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2010 return VM_FAULT_OOM;
2013 * Serialize hugepage allocation and instantiation, so that we don't
2014 * get spurious allocation failures if two CPUs race to instantiate
2015 * the same page in the page cache.
2017 mutex_lock(&hugetlb_instantiation_mutex);
2018 entry = huge_ptep_get(ptep);
2019 if (huge_pte_none(entry)) {
2020 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
2027 * If we are going to COW the mapping later, we examine the pending
2028 * reservations for this page now. This will ensure that any
2029 * allocations necessary to record that reservation occur outside the
2030 * spinlock. For private mappings, we also lookup the pagecache
2031 * page now as it is used to determine if a reservation has been
2034 if (write_access && !pte_write(entry)) {
2035 if (vma_needs_reservation(h, vma, address) < 0) {
2040 if (!(vma->vm_flags & VM_SHARED))
2041 pagecache_page = hugetlbfs_pagecache_page(h,
2045 spin_lock(&mm->page_table_lock);
2046 /* Check for a racing update before calling hugetlb_cow */
2047 if (likely(pte_same(entry, huge_ptep_get(ptep))))
2048 if (write_access && !pte_write(entry))
2049 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2051 spin_unlock(&mm->page_table_lock);
2053 if (pagecache_page) {
2054 unlock_page(pagecache_page);
2055 put_page(pagecache_page);
2059 mutex_unlock(&hugetlb_instantiation_mutex);
2064 /* Can be overriden by architectures */
2065 __attribute__((weak)) struct page *
2066 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2067 pud_t *pud, int write)
2073 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2074 struct page **pages, struct vm_area_struct **vmas,
2075 unsigned long *position, int *length, int i,
2078 unsigned long pfn_offset;
2079 unsigned long vaddr = *position;
2080 int remainder = *length;
2081 struct hstate *h = hstate_vma(vma);
2083 spin_lock(&mm->page_table_lock);
2084 while (vaddr < vma->vm_end && remainder) {
2089 * Some archs (sparc64, sh*) have multiple pte_ts to
2090 * each hugepage. We have to make * sure we get the
2091 * first, for the page indexing below to work.
2093 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2095 if (!pte || huge_pte_none(huge_ptep_get(pte)) ||
2096 (write && !pte_write(huge_ptep_get(pte)))) {
2099 spin_unlock(&mm->page_table_lock);
2100 ret = hugetlb_fault(mm, vma, vaddr, write);
2101 spin_lock(&mm->page_table_lock);
2102 if (!(ret & VM_FAULT_ERROR))
2111 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2112 page = pte_page(huge_ptep_get(pte));
2116 pages[i] = page + pfn_offset;
2126 if (vaddr < vma->vm_end && remainder &&
2127 pfn_offset < pages_per_huge_page(h)) {
2129 * We use pfn_offset to avoid touching the pageframes
2130 * of this compound page.
2135 spin_unlock(&mm->page_table_lock);
2136 *length = remainder;
2142 void hugetlb_change_protection(struct vm_area_struct *vma,
2143 unsigned long address, unsigned long end, pgprot_t newprot)
2145 struct mm_struct *mm = vma->vm_mm;
2146 unsigned long start = address;
2149 struct hstate *h = hstate_vma(vma);
2151 BUG_ON(address >= end);
2152 flush_cache_range(vma, address, end);
2154 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2155 spin_lock(&mm->page_table_lock);
2156 for (; address < end; address += huge_page_size(h)) {
2157 ptep = huge_pte_offset(mm, address);
2160 if (huge_pmd_unshare(mm, &address, ptep))
2162 if (!huge_pte_none(huge_ptep_get(ptep))) {
2163 pte = huge_ptep_get_and_clear(mm, address, ptep);
2164 pte = pte_mkhuge(pte_modify(pte, newprot));
2165 set_huge_pte_at(mm, address, ptep, pte);
2168 spin_unlock(&mm->page_table_lock);
2169 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2171 flush_tlb_range(vma, start, end);
2174 int hugetlb_reserve_pages(struct inode *inode,
2176 struct vm_area_struct *vma)
2179 struct hstate *h = hstate_inode(inode);
2181 if (vma && vma->vm_flags & VM_NORESERVE)
2185 * Shared mappings base their reservation on the number of pages that
2186 * are already allocated on behalf of the file. Private mappings need
2187 * to reserve the full area even if read-only as mprotect() may be
2188 * called to make the mapping read-write. Assume !vma is a shm mapping
2190 if (!vma || vma->vm_flags & VM_SHARED)
2191 chg = region_chg(&inode->i_mapping->private_list, from, to);
2193 struct resv_map *resv_map = resv_map_alloc();
2199 set_vma_resv_map(vma, resv_map);
2200 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2206 if (hugetlb_get_quota(inode->i_mapping, chg))
2208 ret = hugetlb_acct_memory(h, chg);
2210 hugetlb_put_quota(inode->i_mapping, chg);
2213 if (!vma || vma->vm_flags & VM_SHARED)
2214 region_add(&inode->i_mapping->private_list, from, to);
2218 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2220 struct hstate *h = hstate_inode(inode);
2221 long chg = region_truncate(&inode->i_mapping->private_list, offset);
2223 spin_lock(&inode->i_lock);
2224 inode->i_blocks -= blocks_per_huge_page(h);
2225 spin_unlock(&inode->i_lock);
2227 hugetlb_put_quota(inode->i_mapping, (chg - freed));
2228 hugetlb_acct_memory(h, -(chg - freed));