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>
24 #include <linux/hugetlb.h>
27 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
28 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
29 unsigned long hugepages_treat_as_movable;
31 static int max_hstate;
32 unsigned int default_hstate_idx;
33 struct hstate hstates[HUGE_MAX_HSTATE];
35 __initdata LIST_HEAD(huge_boot_pages);
37 /* for command line parsing */
38 static struct hstate * __initdata parsed_hstate;
39 static unsigned long __initdata default_hstate_max_huge_pages;
40 static unsigned long __initdata default_hstate_size;
42 #define for_each_hstate(h) \
43 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
46 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
48 static DEFINE_SPINLOCK(hugetlb_lock);
51 * Region tracking -- allows tracking of reservations and instantiated pages
52 * across the pages in a mapping.
54 * The region data structures are protected by a combination of the mmap_sem
55 * and the hugetlb_instantion_mutex. To access or modify a region the caller
56 * must either hold the mmap_sem for write, or the mmap_sem for read and
57 * the hugetlb_instantiation mutex:
59 * down_write(&mm->mmap_sem);
61 * down_read(&mm->mmap_sem);
62 * mutex_lock(&hugetlb_instantiation_mutex);
65 struct list_head link;
70 static long region_add(struct list_head *head, long f, long t)
72 struct file_region *rg, *nrg, *trg;
74 /* Locate the region we are either in or before. */
75 list_for_each_entry(rg, head, link)
79 /* Round our left edge to the current segment if it encloses us. */
83 /* Check for and consume any regions we now overlap with. */
85 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
86 if (&rg->link == head)
91 /* If this area reaches higher then extend our area to
92 * include it completely. If this is not the first area
93 * which we intend to reuse, free it. */
106 static long region_chg(struct list_head *head, long f, long t)
108 struct file_region *rg, *nrg;
111 /* Locate the region we are before or in. */
112 list_for_each_entry(rg, head, link)
116 /* If we are below the current region then a new region is required.
117 * Subtle, allocate a new region at the position but make it zero
118 * size such that we can guarantee to record the reservation. */
119 if (&rg->link == head || t < rg->from) {
120 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
125 INIT_LIST_HEAD(&nrg->link);
126 list_add(&nrg->link, rg->link.prev);
131 /* Round our left edge to the current segment if it encloses us. */
136 /* Check for and consume any regions we now overlap with. */
137 list_for_each_entry(rg, rg->link.prev, link) {
138 if (&rg->link == head)
143 /* We overlap with this area, if it extends futher than
144 * us then we must extend ourselves. Account for its
145 * existing reservation. */
150 chg -= rg->to - rg->from;
155 static long region_truncate(struct list_head *head, long end)
157 struct file_region *rg, *trg;
160 /* Locate the region we are either in or before. */
161 list_for_each_entry(rg, head, link)
164 if (&rg->link == head)
167 /* If we are in the middle of a region then adjust it. */
168 if (end > rg->from) {
171 rg = list_entry(rg->link.next, typeof(*rg), link);
174 /* Drop any remaining regions. */
175 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
176 if (&rg->link == head)
178 chg += rg->to - rg->from;
185 static long region_count(struct list_head *head, long f, long t)
187 struct file_region *rg;
190 /* Locate each segment we overlap with, and count that overlap. */
191 list_for_each_entry(rg, head, link) {
200 seg_from = max(rg->from, f);
201 seg_to = min(rg->to, t);
203 chg += seg_to - seg_from;
210 * Convert the address within this vma to the page offset within
211 * the mapping, in pagecache page units; huge pages here.
213 static pgoff_t vma_hugecache_offset(struct hstate *h,
214 struct vm_area_struct *vma, unsigned long address)
216 return ((address - vma->vm_start) >> huge_page_shift(h)) +
217 (vma->vm_pgoff >> huge_page_order(h));
221 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
222 * bits of the reservation map pointer, which are always clear due to
225 #define HPAGE_RESV_OWNER (1UL << 0)
226 #define HPAGE_RESV_UNMAPPED (1UL << 1)
227 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
230 * These helpers are used to track how many pages are reserved for
231 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
232 * is guaranteed to have their future faults succeed.
234 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
235 * the reserve counters are updated with the hugetlb_lock held. It is safe
236 * to reset the VMA at fork() time as it is not in use yet and there is no
237 * chance of the global counters getting corrupted as a result of the values.
239 * The private mapping reservation is represented in a subtly different
240 * manner to a shared mapping. A shared mapping has a region map associated
241 * with the underlying file, this region map represents the backing file
242 * pages which have ever had a reservation assigned which this persists even
243 * after the page is instantiated. A private mapping has a region map
244 * associated with the original mmap which is attached to all VMAs which
245 * reference it, this region map represents those offsets which have consumed
246 * reservation ie. where pages have been instantiated.
248 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
250 return (unsigned long)vma->vm_private_data;
253 static void set_vma_private_data(struct vm_area_struct *vma,
256 vma->vm_private_data = (void *)value;
261 struct list_head regions;
264 struct resv_map *resv_map_alloc(void)
266 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
270 kref_init(&resv_map->refs);
271 INIT_LIST_HEAD(&resv_map->regions);
276 void resv_map_release(struct kref *ref)
278 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
280 /* Clear out any active regions before we release the map. */
281 region_truncate(&resv_map->regions, 0);
285 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
287 VM_BUG_ON(!is_vm_hugetlb_page(vma));
288 if (!(vma->vm_flags & VM_SHARED))
289 return (struct resv_map *)(get_vma_private_data(vma) &
294 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
296 VM_BUG_ON(!is_vm_hugetlb_page(vma));
297 VM_BUG_ON(vma->vm_flags & VM_SHARED);
299 set_vma_private_data(vma, (get_vma_private_data(vma) &
300 HPAGE_RESV_MASK) | (unsigned long)map);
303 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
305 VM_BUG_ON(!is_vm_hugetlb_page(vma));
306 VM_BUG_ON(vma->vm_flags & VM_SHARED);
308 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
311 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
313 VM_BUG_ON(!is_vm_hugetlb_page(vma));
315 return (get_vma_private_data(vma) & flag) != 0;
318 /* Decrement the reserved pages in the hugepage pool by one */
319 static void decrement_hugepage_resv_vma(struct hstate *h,
320 struct vm_area_struct *vma)
322 if (vma->vm_flags & VM_NORESERVE)
325 if (vma->vm_flags & VM_SHARED) {
326 /* Shared mappings always use reserves */
327 h->resv_huge_pages--;
328 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
330 * Only the process that called mmap() has reserves for
333 h->resv_huge_pages--;
337 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
338 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
340 VM_BUG_ON(!is_vm_hugetlb_page(vma));
341 if (!(vma->vm_flags & VM_SHARED))
342 vma->vm_private_data = (void *)0;
345 /* Returns true if the VMA has associated reserve pages */
346 static int vma_has_reserves(struct vm_area_struct *vma)
348 if (vma->vm_flags & VM_SHARED)
350 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
355 static void clear_huge_page(struct page *page,
356 unsigned long addr, unsigned long sz)
361 for (i = 0; i < sz/PAGE_SIZE; i++) {
363 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
367 static void copy_huge_page(struct page *dst, struct page *src,
368 unsigned long addr, struct vm_area_struct *vma)
371 struct hstate *h = hstate_vma(vma);
374 for (i = 0; i < pages_per_huge_page(h); i++) {
376 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
380 static void enqueue_huge_page(struct hstate *h, struct page *page)
382 int nid = page_to_nid(page);
383 list_add(&page->lru, &h->hugepage_freelists[nid]);
384 h->free_huge_pages++;
385 h->free_huge_pages_node[nid]++;
388 static struct page *dequeue_huge_page(struct hstate *h)
391 struct page *page = NULL;
393 for (nid = 0; nid < MAX_NUMNODES; ++nid) {
394 if (!list_empty(&h->hugepage_freelists[nid])) {
395 page = list_entry(h->hugepage_freelists[nid].next,
397 list_del(&page->lru);
398 h->free_huge_pages--;
399 h->free_huge_pages_node[nid]--;
406 static struct page *dequeue_huge_page_vma(struct hstate *h,
407 struct vm_area_struct *vma,
408 unsigned long address, int avoid_reserve)
411 struct page *page = NULL;
412 struct mempolicy *mpol;
413 nodemask_t *nodemask;
414 struct zonelist *zonelist = huge_zonelist(vma, address,
415 htlb_alloc_mask, &mpol, &nodemask);
420 * A child process with MAP_PRIVATE mappings created by their parent
421 * have no page reserves. This check ensures that reservations are
422 * not "stolen". The child may still get SIGKILLed
424 if (!vma_has_reserves(vma) &&
425 h->free_huge_pages - h->resv_huge_pages == 0)
428 /* If reserves cannot be used, ensure enough pages are in the pool */
429 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
432 for_each_zone_zonelist_nodemask(zone, z, zonelist,
433 MAX_NR_ZONES - 1, nodemask) {
434 nid = zone_to_nid(zone);
435 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
436 !list_empty(&h->hugepage_freelists[nid])) {
437 page = list_entry(h->hugepage_freelists[nid].next,
439 list_del(&page->lru);
440 h->free_huge_pages--;
441 h->free_huge_pages_node[nid]--;
444 decrement_hugepage_resv_vma(h, vma);
453 static void update_and_free_page(struct hstate *h, struct page *page)
458 h->nr_huge_pages_node[page_to_nid(page)]--;
459 for (i = 0; i < pages_per_huge_page(h); i++) {
460 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
461 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
462 1 << PG_private | 1<< PG_writeback);
464 set_compound_page_dtor(page, NULL);
465 set_page_refcounted(page);
466 arch_release_hugepage(page);
467 __free_pages(page, huge_page_order(h));
470 struct hstate *size_to_hstate(unsigned long size)
475 if (huge_page_size(h) == size)
481 static void free_huge_page(struct page *page)
484 * Can't pass hstate in here because it is called from the
485 * compound page destructor.
487 struct hstate *h = page_hstate(page);
488 int nid = page_to_nid(page);
489 struct address_space *mapping;
491 mapping = (struct address_space *) page_private(page);
492 set_page_private(page, 0);
493 BUG_ON(page_count(page));
494 INIT_LIST_HEAD(&page->lru);
496 spin_lock(&hugetlb_lock);
497 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
498 update_and_free_page(h, page);
499 h->surplus_huge_pages--;
500 h->surplus_huge_pages_node[nid]--;
502 enqueue_huge_page(h, page);
504 spin_unlock(&hugetlb_lock);
506 hugetlb_put_quota(mapping, 1);
510 * Increment or decrement surplus_huge_pages. Keep node-specific counters
511 * balanced by operating on them in a round-robin fashion.
512 * Returns 1 if an adjustment was made.
514 static int adjust_pool_surplus(struct hstate *h, int delta)
520 VM_BUG_ON(delta != -1 && delta != 1);
522 nid = next_node(nid, node_online_map);
523 if (nid == MAX_NUMNODES)
524 nid = first_node(node_online_map);
526 /* To shrink on this node, there must be a surplus page */
527 if (delta < 0 && !h->surplus_huge_pages_node[nid])
529 /* Surplus cannot exceed the total number of pages */
530 if (delta > 0 && h->surplus_huge_pages_node[nid] >=
531 h->nr_huge_pages_node[nid])
534 h->surplus_huge_pages += delta;
535 h->surplus_huge_pages_node[nid] += delta;
538 } while (nid != prev_nid);
544 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
546 set_compound_page_dtor(page, free_huge_page);
547 spin_lock(&hugetlb_lock);
549 h->nr_huge_pages_node[nid]++;
550 spin_unlock(&hugetlb_lock);
551 put_page(page); /* free it into the hugepage allocator */
554 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
558 if (h->order >= MAX_ORDER)
561 page = alloc_pages_node(nid,
562 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
563 __GFP_REPEAT|__GFP_NOWARN,
566 if (arch_prepare_hugepage(page)) {
567 __free_pages(page, HUGETLB_PAGE_ORDER);
570 prep_new_huge_page(h, page, nid);
577 * Use a helper variable to find the next node and then
578 * copy it back to hugetlb_next_nid afterwards:
579 * otherwise there's a window in which a racer might
580 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
581 * But we don't need to use a spin_lock here: it really
582 * doesn't matter if occasionally a racer chooses the
583 * same nid as we do. Move nid forward in the mask even
584 * if we just successfully allocated a hugepage so that
585 * the next caller gets hugepages on the next node.
587 static int hstate_next_node(struct hstate *h)
590 next_nid = next_node(h->hugetlb_next_nid, node_online_map);
591 if (next_nid == MAX_NUMNODES)
592 next_nid = first_node(node_online_map);
593 h->hugetlb_next_nid = next_nid;
597 static int alloc_fresh_huge_page(struct hstate *h)
604 start_nid = h->hugetlb_next_nid;
607 page = alloc_fresh_huge_page_node(h, h->hugetlb_next_nid);
610 next_nid = hstate_next_node(h);
611 } while (!page && h->hugetlb_next_nid != start_nid);
614 count_vm_event(HTLB_BUDDY_PGALLOC);
616 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
621 static struct page *alloc_buddy_huge_page(struct hstate *h,
622 struct vm_area_struct *vma, unsigned long address)
627 if (h->order >= MAX_ORDER)
631 * Assume we will successfully allocate the surplus page to
632 * prevent racing processes from causing the surplus to exceed
635 * This however introduces a different race, where a process B
636 * tries to grow the static hugepage pool while alloc_pages() is
637 * called by process A. B will only examine the per-node
638 * counters in determining if surplus huge pages can be
639 * converted to normal huge pages in adjust_pool_surplus(). A
640 * won't be able to increment the per-node counter, until the
641 * lock is dropped by B, but B doesn't drop hugetlb_lock until
642 * no more huge pages can be converted from surplus to normal
643 * state (and doesn't try to convert again). Thus, we have a
644 * case where a surplus huge page exists, the pool is grown, and
645 * the surplus huge page still exists after, even though it
646 * should just have been converted to a normal huge page. This
647 * does not leak memory, though, as the hugepage will be freed
648 * once it is out of use. It also does not allow the counters to
649 * go out of whack in adjust_pool_surplus() as we don't modify
650 * the node values until we've gotten the hugepage and only the
651 * per-node value is checked there.
653 spin_lock(&hugetlb_lock);
654 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
655 spin_unlock(&hugetlb_lock);
659 h->surplus_huge_pages++;
661 spin_unlock(&hugetlb_lock);
663 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
664 __GFP_REPEAT|__GFP_NOWARN,
667 spin_lock(&hugetlb_lock);
670 * This page is now managed by the hugetlb allocator and has
671 * no users -- drop the buddy allocator's reference.
673 put_page_testzero(page);
674 VM_BUG_ON(page_count(page));
675 nid = page_to_nid(page);
676 set_compound_page_dtor(page, free_huge_page);
678 * We incremented the global counters already
680 h->nr_huge_pages_node[nid]++;
681 h->surplus_huge_pages_node[nid]++;
682 __count_vm_event(HTLB_BUDDY_PGALLOC);
685 h->surplus_huge_pages--;
686 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
688 spin_unlock(&hugetlb_lock);
694 * Increase the hugetlb pool such that it can accomodate a reservation
697 static int gather_surplus_pages(struct hstate *h, int delta)
699 struct list_head surplus_list;
700 struct page *page, *tmp;
702 int needed, allocated;
704 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
706 h->resv_huge_pages += delta;
711 INIT_LIST_HEAD(&surplus_list);
715 spin_unlock(&hugetlb_lock);
716 for (i = 0; i < needed; i++) {
717 page = alloc_buddy_huge_page(h, NULL, 0);
720 * We were not able to allocate enough pages to
721 * satisfy the entire reservation so we free what
722 * we've allocated so far.
724 spin_lock(&hugetlb_lock);
729 list_add(&page->lru, &surplus_list);
734 * After retaking hugetlb_lock, we need to recalculate 'needed'
735 * because either resv_huge_pages or free_huge_pages may have changed.
737 spin_lock(&hugetlb_lock);
738 needed = (h->resv_huge_pages + delta) -
739 (h->free_huge_pages + allocated);
744 * The surplus_list now contains _at_least_ the number of extra pages
745 * needed to accomodate the reservation. Add the appropriate number
746 * of pages to the hugetlb pool and free the extras back to the buddy
747 * allocator. Commit the entire reservation here to prevent another
748 * process from stealing the pages as they are added to the pool but
749 * before they are reserved.
752 h->resv_huge_pages += delta;
755 /* Free the needed pages to the hugetlb pool */
756 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
759 list_del(&page->lru);
760 enqueue_huge_page(h, page);
763 /* Free unnecessary surplus pages to the buddy allocator */
764 if (!list_empty(&surplus_list)) {
765 spin_unlock(&hugetlb_lock);
766 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
767 list_del(&page->lru);
769 * The page has a reference count of zero already, so
770 * call free_huge_page directly instead of using
771 * put_page. This must be done with hugetlb_lock
772 * unlocked which is safe because free_huge_page takes
773 * hugetlb_lock before deciding how to free the page.
775 free_huge_page(page);
777 spin_lock(&hugetlb_lock);
784 * When releasing a hugetlb pool reservation, any surplus pages that were
785 * allocated to satisfy the reservation must be explicitly freed if they were
788 static void return_unused_surplus_pages(struct hstate *h,
789 unsigned long unused_resv_pages)
793 unsigned long nr_pages;
796 * We want to release as many surplus pages as possible, spread
797 * evenly across all nodes. Iterate across all nodes until we
798 * can no longer free unreserved surplus pages. This occurs when
799 * the nodes with surplus pages have no free pages.
801 unsigned long remaining_iterations = num_online_nodes();
803 /* Uncommit the reservation */
804 h->resv_huge_pages -= unused_resv_pages;
806 /* Cannot return gigantic pages currently */
807 if (h->order >= MAX_ORDER)
810 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
812 while (remaining_iterations-- && nr_pages) {
813 nid = next_node(nid, node_online_map);
814 if (nid == MAX_NUMNODES)
815 nid = first_node(node_online_map);
817 if (!h->surplus_huge_pages_node[nid])
820 if (!list_empty(&h->hugepage_freelists[nid])) {
821 page = list_entry(h->hugepage_freelists[nid].next,
823 list_del(&page->lru);
824 update_and_free_page(h, page);
825 h->free_huge_pages--;
826 h->free_huge_pages_node[nid]--;
827 h->surplus_huge_pages--;
828 h->surplus_huge_pages_node[nid]--;
830 remaining_iterations = num_online_nodes();
836 * Determine if the huge page at addr within the vma has an associated
837 * reservation. Where it does not we will need to logically increase
838 * reservation and actually increase quota before an allocation can occur.
839 * Where any new reservation would be required the reservation change is
840 * prepared, but not committed. Once the page has been quota'd allocated
841 * an instantiated the change should be committed via vma_commit_reservation.
842 * No action is required on failure.
844 static int vma_needs_reservation(struct hstate *h,
845 struct vm_area_struct *vma, unsigned long addr)
847 struct address_space *mapping = vma->vm_file->f_mapping;
848 struct inode *inode = mapping->host;
850 if (vma->vm_flags & VM_SHARED) {
851 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
852 return region_chg(&inode->i_mapping->private_list,
855 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
860 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
861 struct resv_map *reservations = vma_resv_map(vma);
863 err = region_chg(&reservations->regions, idx, idx + 1);
869 static void vma_commit_reservation(struct hstate *h,
870 struct vm_area_struct *vma, unsigned long addr)
872 struct address_space *mapping = vma->vm_file->f_mapping;
873 struct inode *inode = mapping->host;
875 if (vma->vm_flags & VM_SHARED) {
876 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
877 region_add(&inode->i_mapping->private_list, idx, idx + 1);
879 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
880 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
881 struct resv_map *reservations = vma_resv_map(vma);
883 /* Mark this page used in the map. */
884 region_add(&reservations->regions, idx, idx + 1);
888 static struct page *alloc_huge_page(struct vm_area_struct *vma,
889 unsigned long addr, int avoid_reserve)
891 struct hstate *h = hstate_vma(vma);
893 struct address_space *mapping = vma->vm_file->f_mapping;
894 struct inode *inode = mapping->host;
898 * Processes that did not create the mapping will have no reserves and
899 * will not have accounted against quota. Check that the quota can be
900 * made before satisfying the allocation
901 * MAP_NORESERVE mappings may also need pages and quota allocated
902 * if no reserve mapping overlaps.
904 chg = vma_needs_reservation(h, vma, addr);
908 if (hugetlb_get_quota(inode->i_mapping, chg))
909 return ERR_PTR(-ENOSPC);
911 spin_lock(&hugetlb_lock);
912 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
913 spin_unlock(&hugetlb_lock);
916 page = alloc_buddy_huge_page(h, vma, addr);
918 hugetlb_put_quota(inode->i_mapping, chg);
919 return ERR_PTR(-VM_FAULT_OOM);
923 set_page_refcounted(page);
924 set_page_private(page, (unsigned long) mapping);
926 vma_commit_reservation(h, vma, addr);
931 __attribute__((weak)) int alloc_bootmem_huge_page(struct hstate *h)
933 struct huge_bootmem_page *m;
934 int nr_nodes = nodes_weight(node_online_map);
939 addr = __alloc_bootmem_node_nopanic(
940 NODE_DATA(h->hugetlb_next_nid),
941 huge_page_size(h), huge_page_size(h), 0);
945 * Use the beginning of the huge page to store the
946 * huge_bootmem_page struct (until gather_bootmem
947 * puts them into the mem_map).
959 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
960 /* Put them into a private list first because mem_map is not up yet */
961 list_add(&m->list, &huge_boot_pages);
966 /* Put bootmem huge pages into the standard lists after mem_map is up */
967 static void __init gather_bootmem_prealloc(void)
969 struct huge_bootmem_page *m;
971 list_for_each_entry(m, &huge_boot_pages, list) {
972 struct page *page = virt_to_page(m);
973 struct hstate *h = m->hstate;
974 __ClearPageReserved(page);
975 WARN_ON(page_count(page) != 1);
976 prep_compound_page(page, h->order);
977 prep_new_huge_page(h, page, page_to_nid(page));
981 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
985 for (i = 0; i < h->max_huge_pages; ++i) {
986 if (h->order >= MAX_ORDER) {
987 if (!alloc_bootmem_huge_page(h))
989 } else if (!alloc_fresh_huge_page(h))
992 h->max_huge_pages = i;
995 static void __init hugetlb_init_hstates(void)
1000 /* oversize hugepages were init'ed in early boot */
1001 if (h->order < MAX_ORDER)
1002 hugetlb_hstate_alloc_pages(h);
1006 static char * __init memfmt(char *buf, unsigned long n)
1008 if (n >= (1UL << 30))
1009 sprintf(buf, "%lu GB", n >> 30);
1010 else if (n >= (1UL << 20))
1011 sprintf(buf, "%lu MB", n >> 20);
1013 sprintf(buf, "%lu KB", n >> 10);
1017 static void __init report_hugepages(void)
1021 for_each_hstate(h) {
1023 printk(KERN_INFO "HugeTLB registered %s page size, "
1024 "pre-allocated %ld pages\n",
1025 memfmt(buf, huge_page_size(h)),
1026 h->free_huge_pages);
1030 #ifdef CONFIG_HIGHMEM
1031 static void try_to_free_low(struct hstate *h, unsigned long count)
1035 if (h->order >= MAX_ORDER)
1038 for (i = 0; i < MAX_NUMNODES; ++i) {
1039 struct page *page, *next;
1040 struct list_head *freel = &h->hugepage_freelists[i];
1041 list_for_each_entry_safe(page, next, freel, lru) {
1042 if (count >= h->nr_huge_pages)
1044 if (PageHighMem(page))
1046 list_del(&page->lru);
1047 update_and_free_page(h, page);
1048 h->free_huge_pages--;
1049 h->free_huge_pages_node[page_to_nid(page)]--;
1054 static inline void try_to_free_low(struct hstate *h, unsigned long count)
1059 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1060 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count)
1062 unsigned long min_count, ret;
1064 if (h->order >= MAX_ORDER)
1065 return h->max_huge_pages;
1068 * Increase the pool size
1069 * First take pages out of surplus state. Then make up the
1070 * remaining difference by allocating fresh huge pages.
1072 * We might race with alloc_buddy_huge_page() here and be unable
1073 * to convert a surplus huge page to a normal huge page. That is
1074 * not critical, though, it just means the overall size of the
1075 * pool might be one hugepage larger than it needs to be, but
1076 * within all the constraints specified by the sysctls.
1078 spin_lock(&hugetlb_lock);
1079 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1080 if (!adjust_pool_surplus(h, -1))
1084 while (count > persistent_huge_pages(h)) {
1086 * If this allocation races such that we no longer need the
1087 * page, free_huge_page will handle it by freeing the page
1088 * and reducing the surplus.
1090 spin_unlock(&hugetlb_lock);
1091 ret = alloc_fresh_huge_page(h);
1092 spin_lock(&hugetlb_lock);
1099 * Decrease the pool size
1100 * First return free pages to the buddy allocator (being careful
1101 * to keep enough around to satisfy reservations). Then place
1102 * pages into surplus state as needed so the pool will shrink
1103 * to the desired size as pages become free.
1105 * By placing pages into the surplus state independent of the
1106 * overcommit value, we are allowing the surplus pool size to
1107 * exceed overcommit. There are few sane options here. Since
1108 * alloc_buddy_huge_page() is checking the global counter,
1109 * though, we'll note that we're not allowed to exceed surplus
1110 * and won't grow the pool anywhere else. Not until one of the
1111 * sysctls are changed, or the surplus pages go out of use.
1113 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1114 min_count = max(count, min_count);
1115 try_to_free_low(h, min_count);
1116 while (min_count < persistent_huge_pages(h)) {
1117 struct page *page = dequeue_huge_page(h);
1120 update_and_free_page(h, page);
1122 while (count < persistent_huge_pages(h)) {
1123 if (!adjust_pool_surplus(h, 1))
1127 ret = persistent_huge_pages(h);
1128 spin_unlock(&hugetlb_lock);
1132 #define HSTATE_ATTR_RO(_name) \
1133 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1135 #define HSTATE_ATTR(_name) \
1136 static struct kobj_attribute _name##_attr = \
1137 __ATTR(_name, 0644, _name##_show, _name##_store)
1139 static struct kobject *hugepages_kobj;
1140 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1142 static struct hstate *kobj_to_hstate(struct kobject *kobj)
1145 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1146 if (hstate_kobjs[i] == kobj)
1152 static ssize_t nr_hugepages_show(struct kobject *kobj,
1153 struct kobj_attribute *attr, char *buf)
1155 struct hstate *h = kobj_to_hstate(kobj);
1156 return sprintf(buf, "%lu\n", h->nr_huge_pages);
1158 static ssize_t nr_hugepages_store(struct kobject *kobj,
1159 struct kobj_attribute *attr, const char *buf, size_t count)
1162 unsigned long input;
1163 struct hstate *h = kobj_to_hstate(kobj);
1165 err = strict_strtoul(buf, 10, &input);
1169 h->max_huge_pages = set_max_huge_pages(h, input);
1173 HSTATE_ATTR(nr_hugepages);
1175 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1176 struct kobj_attribute *attr, char *buf)
1178 struct hstate *h = kobj_to_hstate(kobj);
1179 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1181 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1182 struct kobj_attribute *attr, const char *buf, size_t count)
1185 unsigned long input;
1186 struct hstate *h = kobj_to_hstate(kobj);
1188 err = strict_strtoul(buf, 10, &input);
1192 spin_lock(&hugetlb_lock);
1193 h->nr_overcommit_huge_pages = input;
1194 spin_unlock(&hugetlb_lock);
1198 HSTATE_ATTR(nr_overcommit_hugepages);
1200 static ssize_t free_hugepages_show(struct kobject *kobj,
1201 struct kobj_attribute *attr, char *buf)
1203 struct hstate *h = kobj_to_hstate(kobj);
1204 return sprintf(buf, "%lu\n", h->free_huge_pages);
1206 HSTATE_ATTR_RO(free_hugepages);
1208 static ssize_t resv_hugepages_show(struct kobject *kobj,
1209 struct kobj_attribute *attr, char *buf)
1211 struct hstate *h = kobj_to_hstate(kobj);
1212 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1214 HSTATE_ATTR_RO(resv_hugepages);
1216 static ssize_t surplus_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->surplus_huge_pages);
1222 HSTATE_ATTR_RO(surplus_hugepages);
1224 static struct attribute *hstate_attrs[] = {
1225 &nr_hugepages_attr.attr,
1226 &nr_overcommit_hugepages_attr.attr,
1227 &free_hugepages_attr.attr,
1228 &resv_hugepages_attr.attr,
1229 &surplus_hugepages_attr.attr,
1233 static struct attribute_group hstate_attr_group = {
1234 .attrs = hstate_attrs,
1237 static int __init hugetlb_sysfs_add_hstate(struct hstate *h)
1241 hstate_kobjs[h - hstates] = kobject_create_and_add(h->name,
1243 if (!hstate_kobjs[h - hstates])
1246 retval = sysfs_create_group(hstate_kobjs[h - hstates],
1247 &hstate_attr_group);
1249 kobject_put(hstate_kobjs[h - hstates]);
1254 static void __init hugetlb_sysfs_init(void)
1259 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1260 if (!hugepages_kobj)
1263 for_each_hstate(h) {
1264 err = hugetlb_sysfs_add_hstate(h);
1266 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1271 static void __exit hugetlb_exit(void)
1275 for_each_hstate(h) {
1276 kobject_put(hstate_kobjs[h - hstates]);
1279 kobject_put(hugepages_kobj);
1281 module_exit(hugetlb_exit);
1283 static int __init hugetlb_init(void)
1285 BUILD_BUG_ON(HPAGE_SHIFT == 0);
1287 if (!size_to_hstate(default_hstate_size)) {
1288 default_hstate_size = HPAGE_SIZE;
1289 if (!size_to_hstate(default_hstate_size))
1290 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1292 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1293 if (default_hstate_max_huge_pages)
1294 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1296 hugetlb_init_hstates();
1298 gather_bootmem_prealloc();
1302 hugetlb_sysfs_init();
1306 module_init(hugetlb_init);
1308 /* Should be called on processing a hugepagesz=... option */
1309 void __init hugetlb_add_hstate(unsigned order)
1314 if (size_to_hstate(PAGE_SIZE << order)) {
1315 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1318 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1320 h = &hstates[max_hstate++];
1322 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1323 h->nr_huge_pages = 0;
1324 h->free_huge_pages = 0;
1325 for (i = 0; i < MAX_NUMNODES; ++i)
1326 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1327 h->hugetlb_next_nid = first_node(node_online_map);
1328 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1329 huge_page_size(h)/1024);
1334 static int __init hugetlb_nrpages_setup(char *s)
1337 static unsigned long *last_mhp;
1340 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1341 * so this hugepages= parameter goes to the "default hstate".
1344 mhp = &default_hstate_max_huge_pages;
1346 mhp = &parsed_hstate->max_huge_pages;
1348 if (mhp == last_mhp) {
1349 printk(KERN_WARNING "hugepages= specified twice without "
1350 "interleaving hugepagesz=, ignoring\n");
1354 if (sscanf(s, "%lu", mhp) <= 0)
1358 * Global state is always initialized later in hugetlb_init.
1359 * But we need to allocate >= MAX_ORDER hstates here early to still
1360 * use the bootmem allocator.
1362 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1363 hugetlb_hstate_alloc_pages(parsed_hstate);
1369 __setup("hugepages=", hugetlb_nrpages_setup);
1371 static int __init hugetlb_default_setup(char *s)
1373 default_hstate_size = memparse(s, &s);
1376 __setup("default_hugepagesz=", hugetlb_default_setup);
1378 static unsigned int cpuset_mems_nr(unsigned int *array)
1381 unsigned int nr = 0;
1383 for_each_node_mask(node, cpuset_current_mems_allowed)
1389 #ifdef CONFIG_SYSCTL
1390 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1391 struct file *file, void __user *buffer,
1392 size_t *length, loff_t *ppos)
1394 struct hstate *h = &default_hstate;
1398 tmp = h->max_huge_pages;
1401 table->maxlen = sizeof(unsigned long);
1402 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1405 h->max_huge_pages = set_max_huge_pages(h, tmp);
1410 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1411 struct file *file, void __user *buffer,
1412 size_t *length, loff_t *ppos)
1414 proc_dointvec(table, write, file, buffer, length, ppos);
1415 if (hugepages_treat_as_movable)
1416 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1418 htlb_alloc_mask = GFP_HIGHUSER;
1422 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1423 struct file *file, void __user *buffer,
1424 size_t *length, loff_t *ppos)
1426 struct hstate *h = &default_hstate;
1430 tmp = h->nr_overcommit_huge_pages;
1433 table->maxlen = sizeof(unsigned long);
1434 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1437 spin_lock(&hugetlb_lock);
1438 h->nr_overcommit_huge_pages = tmp;
1439 spin_unlock(&hugetlb_lock);
1445 #endif /* CONFIG_SYSCTL */
1447 int hugetlb_report_meminfo(char *buf)
1449 struct hstate *h = &default_hstate;
1451 "HugePages_Total: %5lu\n"
1452 "HugePages_Free: %5lu\n"
1453 "HugePages_Rsvd: %5lu\n"
1454 "HugePages_Surp: %5lu\n"
1455 "Hugepagesize: %5lu kB\n",
1459 h->surplus_huge_pages,
1460 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1463 int hugetlb_report_node_meminfo(int nid, char *buf)
1465 struct hstate *h = &default_hstate;
1467 "Node %d HugePages_Total: %5u\n"
1468 "Node %d HugePages_Free: %5u\n"
1469 "Node %d HugePages_Surp: %5u\n",
1470 nid, h->nr_huge_pages_node[nid],
1471 nid, h->free_huge_pages_node[nid],
1472 nid, h->surplus_huge_pages_node[nid]);
1475 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1476 unsigned long hugetlb_total_pages(void)
1478 struct hstate *h = &default_hstate;
1479 return h->nr_huge_pages * pages_per_huge_page(h);
1482 static int hugetlb_acct_memory(struct hstate *h, long delta)
1486 spin_lock(&hugetlb_lock);
1488 * When cpuset is configured, it breaks the strict hugetlb page
1489 * reservation as the accounting is done on a global variable. Such
1490 * reservation is completely rubbish in the presence of cpuset because
1491 * the reservation is not checked against page availability for the
1492 * current cpuset. Application can still potentially OOM'ed by kernel
1493 * with lack of free htlb page in cpuset that the task is in.
1494 * Attempt to enforce strict accounting with cpuset is almost
1495 * impossible (or too ugly) because cpuset is too fluid that
1496 * task or memory node can be dynamically moved between cpusets.
1498 * The change of semantics for shared hugetlb mapping with cpuset is
1499 * undesirable. However, in order to preserve some of the semantics,
1500 * we fall back to check against current free page availability as
1501 * a best attempt and hopefully to minimize the impact of changing
1502 * semantics that cpuset has.
1505 if (gather_surplus_pages(h, delta) < 0)
1508 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1509 return_unused_surplus_pages(h, delta);
1516 return_unused_surplus_pages(h, (unsigned long) -delta);
1519 spin_unlock(&hugetlb_lock);
1523 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
1525 struct resv_map *reservations = vma_resv_map(vma);
1528 * This new VMA should share its siblings reservation map if present.
1529 * The VMA will only ever have a valid reservation map pointer where
1530 * it is being copied for another still existing VMA. As that VMA
1531 * has a reference to the reservation map it cannot dissappear until
1532 * after this open call completes. It is therefore safe to take a
1533 * new reference here without additional locking.
1536 kref_get(&reservations->refs);
1539 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
1541 struct hstate *h = hstate_vma(vma);
1542 struct resv_map *reservations = vma_resv_map(vma);
1543 unsigned long reserve;
1544 unsigned long start;
1548 start = vma_hugecache_offset(h, vma, vma->vm_start);
1549 end = vma_hugecache_offset(h, vma, vma->vm_end);
1551 reserve = (end - start) -
1552 region_count(&reservations->regions, start, end);
1554 kref_put(&reservations->refs, resv_map_release);
1557 hugetlb_acct_memory(h, -reserve);
1558 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
1564 * We cannot handle pagefaults against hugetlb pages at all. They cause
1565 * handle_mm_fault() to try to instantiate regular-sized pages in the
1566 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
1569 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1575 struct vm_operations_struct hugetlb_vm_ops = {
1576 .fault = hugetlb_vm_op_fault,
1577 .open = hugetlb_vm_op_open,
1578 .close = hugetlb_vm_op_close,
1581 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
1588 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1590 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
1592 entry = pte_mkyoung(entry);
1593 entry = pte_mkhuge(entry);
1598 static void set_huge_ptep_writable(struct vm_area_struct *vma,
1599 unsigned long address, pte_t *ptep)
1603 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
1604 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
1605 update_mmu_cache(vma, address, entry);
1610 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
1611 struct vm_area_struct *vma)
1613 pte_t *src_pte, *dst_pte, entry;
1614 struct page *ptepage;
1617 struct hstate *h = hstate_vma(vma);
1618 unsigned long sz = huge_page_size(h);
1620 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
1622 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
1623 src_pte = huge_pte_offset(src, addr);
1626 dst_pte = huge_pte_alloc(dst, addr, sz);
1630 /* If the pagetables are shared don't copy or take references */
1631 if (dst_pte == src_pte)
1634 spin_lock(&dst->page_table_lock);
1635 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
1636 if (!huge_pte_none(huge_ptep_get(src_pte))) {
1638 huge_ptep_set_wrprotect(src, addr, src_pte);
1639 entry = huge_ptep_get(src_pte);
1640 ptepage = pte_page(entry);
1642 set_huge_pte_at(dst, addr, dst_pte, entry);
1644 spin_unlock(&src->page_table_lock);
1645 spin_unlock(&dst->page_table_lock);
1653 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1654 unsigned long end, struct page *ref_page)
1656 struct mm_struct *mm = vma->vm_mm;
1657 unsigned long address;
1662 struct hstate *h = hstate_vma(vma);
1663 unsigned long sz = huge_page_size(h);
1666 * A page gathering list, protected by per file i_mmap_lock. The
1667 * lock is used to avoid list corruption from multiple unmapping
1668 * of the same page since we are using page->lru.
1670 LIST_HEAD(page_list);
1672 WARN_ON(!is_vm_hugetlb_page(vma));
1673 BUG_ON(start & ~huge_page_mask(h));
1674 BUG_ON(end & ~huge_page_mask(h));
1676 mmu_notifier_invalidate_range_start(mm, start, end);
1677 spin_lock(&mm->page_table_lock);
1678 for (address = start; address < end; address += sz) {
1679 ptep = huge_pte_offset(mm, address);
1683 if (huge_pmd_unshare(mm, &address, ptep))
1687 * If a reference page is supplied, it is because a specific
1688 * page is being unmapped, not a range. Ensure the page we
1689 * are about to unmap is the actual page of interest.
1692 pte = huge_ptep_get(ptep);
1693 if (huge_pte_none(pte))
1695 page = pte_page(pte);
1696 if (page != ref_page)
1700 * Mark the VMA as having unmapped its page so that
1701 * future faults in this VMA will fail rather than
1702 * looking like data was lost
1704 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
1707 pte = huge_ptep_get_and_clear(mm, address, ptep);
1708 if (huge_pte_none(pte))
1711 page = pte_page(pte);
1713 set_page_dirty(page);
1714 list_add(&page->lru, &page_list);
1716 spin_unlock(&mm->page_table_lock);
1717 flush_tlb_range(vma, start, end);
1718 mmu_notifier_invalidate_range_end(mm, start, end);
1719 list_for_each_entry_safe(page, tmp, &page_list, lru) {
1720 list_del(&page->lru);
1725 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1726 unsigned long end, struct page *ref_page)
1728 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1729 __unmap_hugepage_range(vma, start, end, ref_page);
1730 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1734 * This is called when the original mapper is failing to COW a MAP_PRIVATE
1735 * mappping it owns the reserve page for. The intention is to unmap the page
1736 * from other VMAs and let the children be SIGKILLed if they are faulting the
1739 int unmap_ref_private(struct mm_struct *mm,
1740 struct vm_area_struct *vma,
1742 unsigned long address)
1744 struct vm_area_struct *iter_vma;
1745 struct address_space *mapping;
1746 struct prio_tree_iter iter;
1750 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1751 * from page cache lookup which is in HPAGE_SIZE units.
1753 address = address & huge_page_mask(hstate_vma(vma));
1754 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
1755 + (vma->vm_pgoff >> PAGE_SHIFT);
1756 mapping = (struct address_space *)page_private(page);
1758 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1759 /* Do not unmap the current VMA */
1760 if (iter_vma == vma)
1764 * Unmap the page from other VMAs without their own reserves.
1765 * They get marked to be SIGKILLed if they fault in these
1766 * areas. This is because a future no-page fault on this VMA
1767 * could insert a zeroed page instead of the data existing
1768 * from the time of fork. This would look like data corruption
1770 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
1771 unmap_hugepage_range(iter_vma,
1772 address, address + HPAGE_SIZE,
1779 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
1780 unsigned long address, pte_t *ptep, pte_t pte,
1781 struct page *pagecache_page)
1783 struct hstate *h = hstate_vma(vma);
1784 struct page *old_page, *new_page;
1786 int outside_reserve = 0;
1788 old_page = pte_page(pte);
1791 /* If no-one else is actually using this page, avoid the copy
1792 * and just make the page writable */
1793 avoidcopy = (page_count(old_page) == 1);
1795 set_huge_ptep_writable(vma, address, ptep);
1800 * If the process that created a MAP_PRIVATE mapping is about to
1801 * perform a COW due to a shared page count, attempt to satisfy
1802 * the allocation without using the existing reserves. The pagecache
1803 * page is used to determine if the reserve at this address was
1804 * consumed or not. If reserves were used, a partial faulted mapping
1805 * at the time of fork() could consume its reserves on COW instead
1806 * of the full address range.
1808 if (!(vma->vm_flags & VM_SHARED) &&
1809 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
1810 old_page != pagecache_page)
1811 outside_reserve = 1;
1813 page_cache_get(old_page);
1814 new_page = alloc_huge_page(vma, address, outside_reserve);
1816 if (IS_ERR(new_page)) {
1817 page_cache_release(old_page);
1820 * If a process owning a MAP_PRIVATE mapping fails to COW,
1821 * it is due to references held by a child and an insufficient
1822 * huge page pool. To guarantee the original mappers
1823 * reliability, unmap the page from child processes. The child
1824 * may get SIGKILLed if it later faults.
1826 if (outside_reserve) {
1827 BUG_ON(huge_pte_none(pte));
1828 if (unmap_ref_private(mm, vma, old_page, address)) {
1829 BUG_ON(page_count(old_page) != 1);
1830 BUG_ON(huge_pte_none(pte));
1831 goto retry_avoidcopy;
1836 return -PTR_ERR(new_page);
1839 spin_unlock(&mm->page_table_lock);
1840 copy_huge_page(new_page, old_page, address, vma);
1841 __SetPageUptodate(new_page);
1842 spin_lock(&mm->page_table_lock);
1844 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
1845 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
1847 huge_ptep_clear_flush(vma, address, ptep);
1848 set_huge_pte_at(mm, address, ptep,
1849 make_huge_pte(vma, new_page, 1));
1850 /* Make the old page be freed below */
1851 new_page = old_page;
1853 page_cache_release(new_page);
1854 page_cache_release(old_page);
1858 /* Return the pagecache page at a given address within a VMA */
1859 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
1860 struct vm_area_struct *vma, unsigned long address)
1862 struct address_space *mapping;
1865 mapping = vma->vm_file->f_mapping;
1866 idx = vma_hugecache_offset(h, vma, address);
1868 return find_lock_page(mapping, idx);
1871 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1872 unsigned long address, pte_t *ptep, int write_access)
1874 struct hstate *h = hstate_vma(vma);
1875 int ret = VM_FAULT_SIGBUS;
1879 struct address_space *mapping;
1883 * Currently, we are forced to kill the process in the event the
1884 * original mapper has unmapped pages from the child due to a failed
1885 * COW. Warn that such a situation has occured as it may not be obvious
1887 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
1889 "PID %d killed due to inadequate hugepage pool\n",
1894 mapping = vma->vm_file->f_mapping;
1895 idx = vma_hugecache_offset(h, vma, address);
1898 * Use page lock to guard against racing truncation
1899 * before we get page_table_lock.
1902 page = find_lock_page(mapping, idx);
1904 size = i_size_read(mapping->host) >> huge_page_shift(h);
1907 page = alloc_huge_page(vma, address, 0);
1909 ret = -PTR_ERR(page);
1912 clear_huge_page(page, address, huge_page_size(h));
1913 __SetPageUptodate(page);
1915 if (vma->vm_flags & VM_SHARED) {
1917 struct inode *inode = mapping->host;
1919 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
1927 spin_lock(&inode->i_lock);
1928 inode->i_blocks += blocks_per_huge_page(h);
1929 spin_unlock(&inode->i_lock);
1934 spin_lock(&mm->page_table_lock);
1935 size = i_size_read(mapping->host) >> huge_page_shift(h);
1940 if (!huge_pte_none(huge_ptep_get(ptep)))
1943 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
1944 && (vma->vm_flags & VM_SHARED)));
1945 set_huge_pte_at(mm, address, ptep, new_pte);
1947 if (write_access && !(vma->vm_flags & VM_SHARED)) {
1948 /* Optimization, do the COW without a second fault */
1949 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
1952 spin_unlock(&mm->page_table_lock);
1958 spin_unlock(&mm->page_table_lock);
1964 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
1965 unsigned long address, int write_access)
1970 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
1971 struct hstate *h = hstate_vma(vma);
1973 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
1975 return VM_FAULT_OOM;
1978 * Serialize hugepage allocation and instantiation, so that we don't
1979 * get spurious allocation failures if two CPUs race to instantiate
1980 * the same page in the page cache.
1982 mutex_lock(&hugetlb_instantiation_mutex);
1983 entry = huge_ptep_get(ptep);
1984 if (huge_pte_none(entry)) {
1985 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
1986 mutex_unlock(&hugetlb_instantiation_mutex);
1992 spin_lock(&mm->page_table_lock);
1993 /* Check for a racing update before calling hugetlb_cow */
1994 if (likely(pte_same(entry, huge_ptep_get(ptep))))
1995 if (write_access && !pte_write(entry)) {
1997 page = hugetlbfs_pagecache_page(h, vma, address);
1998 ret = hugetlb_cow(mm, vma, address, ptep, entry, page);
2004 spin_unlock(&mm->page_table_lock);
2005 mutex_unlock(&hugetlb_instantiation_mutex);
2010 /* Can be overriden by architectures */
2011 __attribute__((weak)) struct page *
2012 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2013 pud_t *pud, int write)
2019 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2020 struct page **pages, struct vm_area_struct **vmas,
2021 unsigned long *position, int *length, int i,
2024 unsigned long pfn_offset;
2025 unsigned long vaddr = *position;
2026 int remainder = *length;
2027 struct hstate *h = hstate_vma(vma);
2029 spin_lock(&mm->page_table_lock);
2030 while (vaddr < vma->vm_end && remainder) {
2035 * Some archs (sparc64, sh*) have multiple pte_ts to
2036 * each hugepage. We have to make * sure we get the
2037 * first, for the page indexing below to work.
2039 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2041 if (!pte || huge_pte_none(huge_ptep_get(pte)) ||
2042 (write && !pte_write(huge_ptep_get(pte)))) {
2045 spin_unlock(&mm->page_table_lock);
2046 ret = hugetlb_fault(mm, vma, vaddr, write);
2047 spin_lock(&mm->page_table_lock);
2048 if (!(ret & VM_FAULT_ERROR))
2057 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2058 page = pte_page(huge_ptep_get(pte));
2062 pages[i] = page + pfn_offset;
2072 if (vaddr < vma->vm_end && remainder &&
2073 pfn_offset < pages_per_huge_page(h)) {
2075 * We use pfn_offset to avoid touching the pageframes
2076 * of this compound page.
2081 spin_unlock(&mm->page_table_lock);
2082 *length = remainder;
2088 void hugetlb_change_protection(struct vm_area_struct *vma,
2089 unsigned long address, unsigned long end, pgprot_t newprot)
2091 struct mm_struct *mm = vma->vm_mm;
2092 unsigned long start = address;
2095 struct hstate *h = hstate_vma(vma);
2097 BUG_ON(address >= end);
2098 flush_cache_range(vma, address, end);
2100 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2101 spin_lock(&mm->page_table_lock);
2102 for (; address < end; address += huge_page_size(h)) {
2103 ptep = huge_pte_offset(mm, address);
2106 if (huge_pmd_unshare(mm, &address, ptep))
2108 if (!huge_pte_none(huge_ptep_get(ptep))) {
2109 pte = huge_ptep_get_and_clear(mm, address, ptep);
2110 pte = pte_mkhuge(pte_modify(pte, newprot));
2111 set_huge_pte_at(mm, address, ptep, pte);
2114 spin_unlock(&mm->page_table_lock);
2115 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2117 flush_tlb_range(vma, start, end);
2120 int hugetlb_reserve_pages(struct inode *inode,
2122 struct vm_area_struct *vma)
2125 struct hstate *h = hstate_inode(inode);
2127 if (vma && vma->vm_flags & VM_NORESERVE)
2131 * Shared mappings base their reservation on the number of pages that
2132 * are already allocated on behalf of the file. Private mappings need
2133 * to reserve the full area even if read-only as mprotect() may be
2134 * called to make the mapping read-write. Assume !vma is a shm mapping
2136 if (!vma || vma->vm_flags & VM_SHARED)
2137 chg = region_chg(&inode->i_mapping->private_list, from, to);
2139 struct resv_map *resv_map = resv_map_alloc();
2145 set_vma_resv_map(vma, resv_map);
2146 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2152 if (hugetlb_get_quota(inode->i_mapping, chg))
2154 ret = hugetlb_acct_memory(h, chg);
2156 hugetlb_put_quota(inode->i_mapping, chg);
2159 if (!vma || vma->vm_flags & VM_SHARED)
2160 region_add(&inode->i_mapping->private_list, from, to);
2164 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2166 struct hstate *h = hstate_inode(inode);
2167 long chg = region_truncate(&inode->i_mapping->private_list, offset);
2169 spin_lock(&inode->i_lock);
2170 inode->i_blocks -= blocks_per_huge_page(h);
2171 spin_unlock(&inode->i_lock);
2173 hugetlb_put_quota(inode->i_mapping, (chg - freed));
2174 hugetlb_acct_memory(h, -(chg - freed));