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/nodemask.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/cpuset.h>
16 #include <linux/mutex.h>
19 #include <asm/pgtable.h>
21 #include <linux/hugetlb.h>
24 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
25 static unsigned long nr_huge_pages, free_huge_pages, resv_huge_pages;
26 static unsigned long surplus_huge_pages;
27 static unsigned long nr_overcommit_huge_pages;
28 unsigned long max_huge_pages;
29 unsigned long sysctl_overcommit_huge_pages;
30 static struct list_head hugepage_freelists[MAX_NUMNODES];
31 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
32 static unsigned int free_huge_pages_node[MAX_NUMNODES];
33 static unsigned int surplus_huge_pages_node[MAX_NUMNODES];
34 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
35 unsigned long hugepages_treat_as_movable;
36 static int hugetlb_next_nid;
39 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
41 static DEFINE_SPINLOCK(hugetlb_lock);
44 * Region tracking -- allows tracking of reservations and instantiated pages
45 * across the pages in a mapping.
47 * The region data structures are protected by a combination of the mmap_sem
48 * and the hugetlb_instantion_mutex. To access or modify a region the caller
49 * must either hold the mmap_sem for write, or the mmap_sem for read and
50 * the hugetlb_instantiation mutex:
52 * down_write(&mm->mmap_sem);
54 * down_read(&mm->mmap_sem);
55 * mutex_lock(&hugetlb_instantiation_mutex);
58 struct list_head link;
63 static long region_add(struct list_head *head, long f, long t)
65 struct file_region *rg, *nrg, *trg;
67 /* Locate the region we are either in or before. */
68 list_for_each_entry(rg, head, link)
72 /* Round our left edge to the current segment if it encloses us. */
76 /* Check for and consume any regions we now overlap with. */
78 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
79 if (&rg->link == head)
84 /* If this area reaches higher then extend our area to
85 * include it completely. If this is not the first area
86 * which we intend to reuse, free it. */
99 static long region_chg(struct list_head *head, long f, long t)
101 struct file_region *rg, *nrg;
104 /* Locate the region we are before or in. */
105 list_for_each_entry(rg, head, link)
109 /* If we are below the current region then a new region is required.
110 * Subtle, allocate a new region at the position but make it zero
111 * size such that we can guarantee to record the reservation. */
112 if (&rg->link == head || t < rg->from) {
113 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
118 INIT_LIST_HEAD(&nrg->link);
119 list_add(&nrg->link, rg->link.prev);
124 /* Round our left edge to the current segment if it encloses us. */
129 /* Check for and consume any regions we now overlap with. */
130 list_for_each_entry(rg, rg->link.prev, link) {
131 if (&rg->link == head)
136 /* We overlap with this area, if it extends futher than
137 * us then we must extend ourselves. Account for its
138 * existing reservation. */
143 chg -= rg->to - rg->from;
148 static long region_truncate(struct list_head *head, long end)
150 struct file_region *rg, *trg;
153 /* Locate the region we are either in or before. */
154 list_for_each_entry(rg, head, link)
157 if (&rg->link == head)
160 /* If we are in the middle of a region then adjust it. */
161 if (end > rg->from) {
164 rg = list_entry(rg->link.next, typeof(*rg), link);
167 /* Drop any remaining regions. */
168 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
169 if (&rg->link == head)
171 chg += rg->to - rg->from;
178 static long region_count(struct list_head *head, long f, long t)
180 struct file_region *rg;
183 /* Locate each segment we overlap with, and count that overlap. */
184 list_for_each_entry(rg, head, link) {
193 seg_from = max(rg->from, f);
194 seg_to = min(rg->to, t);
196 chg += seg_to - seg_from;
203 * Convert the address within this vma to the page offset within
204 * the mapping, in pagecache page units; huge pages here.
206 static pgoff_t vma_hugecache_offset(struct vm_area_struct *vma,
207 unsigned long address)
209 return ((address - vma->vm_start) >> HPAGE_SHIFT) +
210 (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
214 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
215 * bits of the reservation map pointer, which are always clear due to
218 #define HPAGE_RESV_OWNER (1UL << 0)
219 #define HPAGE_RESV_UNMAPPED (1UL << 1)
220 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
223 * These helpers are used to track how many pages are reserved for
224 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
225 * is guaranteed to have their future faults succeed.
227 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
228 * the reserve counters are updated with the hugetlb_lock held. It is safe
229 * to reset the VMA at fork() time as it is not in use yet and there is no
230 * chance of the global counters getting corrupted as a result of the values.
232 * The private mapping reservation is represented in a subtly different
233 * manner to a shared mapping. A shared mapping has a region map associated
234 * with the underlying file, this region map represents the backing file
235 * pages which have ever had a reservation assigned which this persists even
236 * after the page is instantiated. A private mapping has a region map
237 * associated with the original mmap which is attached to all VMAs which
238 * reference it, this region map represents those offsets which have consumed
239 * reservation ie. where pages have been instantiated.
241 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
243 return (unsigned long)vma->vm_private_data;
246 static void set_vma_private_data(struct vm_area_struct *vma,
249 vma->vm_private_data = (void *)value;
254 struct list_head regions;
257 struct resv_map *resv_map_alloc(void)
259 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
263 kref_init(&resv_map->refs);
264 INIT_LIST_HEAD(&resv_map->regions);
269 void resv_map_release(struct kref *ref)
271 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
273 /* Clear out any active regions before we release the map. */
274 region_truncate(&resv_map->regions, 0);
278 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
280 VM_BUG_ON(!is_vm_hugetlb_page(vma));
281 if (!(vma->vm_flags & VM_SHARED))
282 return (struct resv_map *)(get_vma_private_data(vma) &
287 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
289 VM_BUG_ON(!is_vm_hugetlb_page(vma));
290 VM_BUG_ON(vma->vm_flags & VM_SHARED);
292 set_vma_private_data(vma, (get_vma_private_data(vma) &
293 HPAGE_RESV_MASK) | (unsigned long)map);
296 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
298 VM_BUG_ON(!is_vm_hugetlb_page(vma));
299 VM_BUG_ON(vma->vm_flags & VM_SHARED);
301 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
304 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
306 VM_BUG_ON(!is_vm_hugetlb_page(vma));
308 return (get_vma_private_data(vma) & flag) != 0;
311 /* Decrement the reserved pages in the hugepage pool by one */
312 static void decrement_hugepage_resv_vma(struct vm_area_struct *vma)
314 if (vma->vm_flags & VM_NORESERVE)
317 if (vma->vm_flags & VM_SHARED) {
318 /* Shared mappings always use reserves */
320 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
322 * Only the process that called mmap() has reserves for
329 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
330 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
332 VM_BUG_ON(!is_vm_hugetlb_page(vma));
333 if (!(vma->vm_flags & VM_SHARED))
334 vma->vm_private_data = (void *)0;
337 /* Returns true if the VMA has associated reserve pages */
338 static int vma_has_private_reserves(struct vm_area_struct *vma)
340 if (vma->vm_flags & VM_SHARED)
342 if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER))
347 static void clear_huge_page(struct page *page, unsigned long addr)
352 for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
354 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
358 static void copy_huge_page(struct page *dst, struct page *src,
359 unsigned long addr, struct vm_area_struct *vma)
364 for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
366 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
370 static void enqueue_huge_page(struct page *page)
372 int nid = page_to_nid(page);
373 list_add(&page->lru, &hugepage_freelists[nid]);
375 free_huge_pages_node[nid]++;
378 static struct page *dequeue_huge_page(void)
381 struct page *page = NULL;
383 for (nid = 0; nid < MAX_NUMNODES; ++nid) {
384 if (!list_empty(&hugepage_freelists[nid])) {
385 page = list_entry(hugepage_freelists[nid].next,
387 list_del(&page->lru);
389 free_huge_pages_node[nid]--;
396 static struct page *dequeue_huge_page_vma(struct vm_area_struct *vma,
397 unsigned long address, int avoid_reserve)
400 struct page *page = NULL;
401 struct mempolicy *mpol;
402 nodemask_t *nodemask;
403 struct zonelist *zonelist = huge_zonelist(vma, address,
404 htlb_alloc_mask, &mpol, &nodemask);
409 * A child process with MAP_PRIVATE mappings created by their parent
410 * have no page reserves. This check ensures that reservations are
411 * not "stolen". The child may still get SIGKILLed
413 if (!vma_has_private_reserves(vma) &&
414 free_huge_pages - resv_huge_pages == 0)
417 /* If reserves cannot be used, ensure enough pages are in the pool */
418 if (avoid_reserve && free_huge_pages - resv_huge_pages == 0)
421 for_each_zone_zonelist_nodemask(zone, z, zonelist,
422 MAX_NR_ZONES - 1, nodemask) {
423 nid = zone_to_nid(zone);
424 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
425 !list_empty(&hugepage_freelists[nid])) {
426 page = list_entry(hugepage_freelists[nid].next,
428 list_del(&page->lru);
430 free_huge_pages_node[nid]--;
433 decrement_hugepage_resv_vma(vma);
442 static void update_and_free_page(struct page *page)
446 nr_huge_pages_node[page_to_nid(page)]--;
447 for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
448 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
449 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
450 1 << PG_private | 1<< PG_writeback);
452 set_compound_page_dtor(page, NULL);
453 set_page_refcounted(page);
454 arch_release_hugepage(page);
455 __free_pages(page, HUGETLB_PAGE_ORDER);
458 static void free_huge_page(struct page *page)
460 int nid = page_to_nid(page);
461 struct address_space *mapping;
463 mapping = (struct address_space *) page_private(page);
464 set_page_private(page, 0);
465 BUG_ON(page_count(page));
466 INIT_LIST_HEAD(&page->lru);
468 spin_lock(&hugetlb_lock);
469 if (surplus_huge_pages_node[nid]) {
470 update_and_free_page(page);
471 surplus_huge_pages--;
472 surplus_huge_pages_node[nid]--;
474 enqueue_huge_page(page);
476 spin_unlock(&hugetlb_lock);
478 hugetlb_put_quota(mapping, 1);
482 * Increment or decrement surplus_huge_pages. Keep node-specific counters
483 * balanced by operating on them in a round-robin fashion.
484 * Returns 1 if an adjustment was made.
486 static int adjust_pool_surplus(int delta)
492 VM_BUG_ON(delta != -1 && delta != 1);
494 nid = next_node(nid, node_online_map);
495 if (nid == MAX_NUMNODES)
496 nid = first_node(node_online_map);
498 /* To shrink on this node, there must be a surplus page */
499 if (delta < 0 && !surplus_huge_pages_node[nid])
501 /* Surplus cannot exceed the total number of pages */
502 if (delta > 0 && surplus_huge_pages_node[nid] >=
503 nr_huge_pages_node[nid])
506 surplus_huge_pages += delta;
507 surplus_huge_pages_node[nid] += delta;
510 } while (nid != prev_nid);
516 static struct page *alloc_fresh_huge_page_node(int nid)
520 page = alloc_pages_node(nid,
521 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
522 __GFP_REPEAT|__GFP_NOWARN,
525 if (arch_prepare_hugepage(page)) {
526 __free_pages(page, HUGETLB_PAGE_ORDER);
529 set_compound_page_dtor(page, free_huge_page);
530 spin_lock(&hugetlb_lock);
532 nr_huge_pages_node[nid]++;
533 spin_unlock(&hugetlb_lock);
534 put_page(page); /* free it into the hugepage allocator */
540 static int alloc_fresh_huge_page(void)
547 start_nid = hugetlb_next_nid;
550 page = alloc_fresh_huge_page_node(hugetlb_next_nid);
554 * Use a helper variable to find the next node and then
555 * copy it back to hugetlb_next_nid afterwards:
556 * otherwise there's a window in which a racer might
557 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
558 * But we don't need to use a spin_lock here: it really
559 * doesn't matter if occasionally a racer chooses the
560 * same nid as we do. Move nid forward in the mask even
561 * if we just successfully allocated a hugepage so that
562 * the next caller gets hugepages on the next node.
564 next_nid = next_node(hugetlb_next_nid, node_online_map);
565 if (next_nid == MAX_NUMNODES)
566 next_nid = first_node(node_online_map);
567 hugetlb_next_nid = next_nid;
568 } while (!page && hugetlb_next_nid != start_nid);
571 count_vm_event(HTLB_BUDDY_PGALLOC);
573 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
578 static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
579 unsigned long address)
585 * Assume we will successfully allocate the surplus page to
586 * prevent racing processes from causing the surplus to exceed
589 * This however introduces a different race, where a process B
590 * tries to grow the static hugepage pool while alloc_pages() is
591 * called by process A. B will only examine the per-node
592 * counters in determining if surplus huge pages can be
593 * converted to normal huge pages in adjust_pool_surplus(). A
594 * won't be able to increment the per-node counter, until the
595 * lock is dropped by B, but B doesn't drop hugetlb_lock until
596 * no more huge pages can be converted from surplus to normal
597 * state (and doesn't try to convert again). Thus, we have a
598 * case where a surplus huge page exists, the pool is grown, and
599 * the surplus huge page still exists after, even though it
600 * should just have been converted to a normal huge page. This
601 * does not leak memory, though, as the hugepage will be freed
602 * once it is out of use. It also does not allow the counters to
603 * go out of whack in adjust_pool_surplus() as we don't modify
604 * the node values until we've gotten the hugepage and only the
605 * per-node value is checked there.
607 spin_lock(&hugetlb_lock);
608 if (surplus_huge_pages >= nr_overcommit_huge_pages) {
609 spin_unlock(&hugetlb_lock);
613 surplus_huge_pages++;
615 spin_unlock(&hugetlb_lock);
617 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
618 __GFP_REPEAT|__GFP_NOWARN,
621 spin_lock(&hugetlb_lock);
624 * This page is now managed by the hugetlb allocator and has
625 * no users -- drop the buddy allocator's reference.
627 put_page_testzero(page);
628 VM_BUG_ON(page_count(page));
629 nid = page_to_nid(page);
630 set_compound_page_dtor(page, free_huge_page);
632 * We incremented the global counters already
634 nr_huge_pages_node[nid]++;
635 surplus_huge_pages_node[nid]++;
636 __count_vm_event(HTLB_BUDDY_PGALLOC);
639 surplus_huge_pages--;
640 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
642 spin_unlock(&hugetlb_lock);
648 * Increase the hugetlb pool such that it can accomodate a reservation
651 static int gather_surplus_pages(int delta)
653 struct list_head surplus_list;
654 struct page *page, *tmp;
656 int needed, allocated;
658 needed = (resv_huge_pages + delta) - free_huge_pages;
660 resv_huge_pages += delta;
665 INIT_LIST_HEAD(&surplus_list);
669 spin_unlock(&hugetlb_lock);
670 for (i = 0; i < needed; i++) {
671 page = alloc_buddy_huge_page(NULL, 0);
674 * We were not able to allocate enough pages to
675 * satisfy the entire reservation so we free what
676 * we've allocated so far.
678 spin_lock(&hugetlb_lock);
683 list_add(&page->lru, &surplus_list);
688 * After retaking hugetlb_lock, we need to recalculate 'needed'
689 * because either resv_huge_pages or free_huge_pages may have changed.
691 spin_lock(&hugetlb_lock);
692 needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
697 * The surplus_list now contains _at_least_ the number of extra pages
698 * needed to accomodate the reservation. Add the appropriate number
699 * of pages to the hugetlb pool and free the extras back to the buddy
700 * allocator. Commit the entire reservation here to prevent another
701 * process from stealing the pages as they are added to the pool but
702 * before they are reserved.
705 resv_huge_pages += delta;
708 /* Free the needed pages to the hugetlb pool */
709 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
712 list_del(&page->lru);
713 enqueue_huge_page(page);
716 /* Free unnecessary surplus pages to the buddy allocator */
717 if (!list_empty(&surplus_list)) {
718 spin_unlock(&hugetlb_lock);
719 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
720 list_del(&page->lru);
722 * The page has a reference count of zero already, so
723 * call free_huge_page directly instead of using
724 * put_page. This must be done with hugetlb_lock
725 * unlocked which is safe because free_huge_page takes
726 * hugetlb_lock before deciding how to free the page.
728 free_huge_page(page);
730 spin_lock(&hugetlb_lock);
737 * When releasing a hugetlb pool reservation, any surplus pages that were
738 * allocated to satisfy the reservation must be explicitly freed if they were
741 static void return_unused_surplus_pages(unsigned long unused_resv_pages)
745 unsigned long nr_pages;
748 * We want to release as many surplus pages as possible, spread
749 * evenly across all nodes. Iterate across all nodes until we
750 * can no longer free unreserved surplus pages. This occurs when
751 * the nodes with surplus pages have no free pages.
753 unsigned long remaining_iterations = num_online_nodes();
755 /* Uncommit the reservation */
756 resv_huge_pages -= unused_resv_pages;
758 nr_pages = min(unused_resv_pages, surplus_huge_pages);
760 while (remaining_iterations-- && nr_pages) {
761 nid = next_node(nid, node_online_map);
762 if (nid == MAX_NUMNODES)
763 nid = first_node(node_online_map);
765 if (!surplus_huge_pages_node[nid])
768 if (!list_empty(&hugepage_freelists[nid])) {
769 page = list_entry(hugepage_freelists[nid].next,
771 list_del(&page->lru);
772 update_and_free_page(page);
774 free_huge_pages_node[nid]--;
775 surplus_huge_pages--;
776 surplus_huge_pages_node[nid]--;
778 remaining_iterations = num_online_nodes();
784 * Determine if the huge page at addr within the vma has an associated
785 * reservation. Where it does not we will need to logically increase
786 * reservation and actually increase quota before an allocation can occur.
787 * Where any new reservation would be required the reservation change is
788 * prepared, but not committed. Once the page has been quota'd allocated
789 * an instantiated the change should be committed via vma_commit_reservation.
790 * No action is required on failure.
792 static int vma_needs_reservation(struct vm_area_struct *vma, unsigned long addr)
794 struct address_space *mapping = vma->vm_file->f_mapping;
795 struct inode *inode = mapping->host;
797 if (vma->vm_flags & VM_SHARED) {
798 pgoff_t idx = vma_hugecache_offset(vma, addr);
799 return region_chg(&inode->i_mapping->private_list,
802 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
807 pgoff_t idx = vma_hugecache_offset(vma, addr);
808 struct resv_map *reservations = vma_resv_map(vma);
810 err = region_chg(&reservations->regions, idx, idx + 1);
816 static void vma_commit_reservation(struct vm_area_struct *vma,
819 struct address_space *mapping = vma->vm_file->f_mapping;
820 struct inode *inode = mapping->host;
822 if (vma->vm_flags & VM_SHARED) {
823 pgoff_t idx = vma_hugecache_offset(vma, addr);
824 region_add(&inode->i_mapping->private_list, idx, idx + 1);
826 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
827 pgoff_t idx = vma_hugecache_offset(vma, addr);
828 struct resv_map *reservations = vma_resv_map(vma);
830 /* Mark this page used in the map. */
831 region_add(&reservations->regions, idx, idx + 1);
835 static struct page *alloc_huge_page(struct vm_area_struct *vma,
836 unsigned long addr, int avoid_reserve)
839 struct address_space *mapping = vma->vm_file->f_mapping;
840 struct inode *inode = mapping->host;
844 * Processes that did not create the mapping will have no reserves and
845 * will not have accounted against quota. Check that the quota can be
846 * made before satisfying the allocation
847 * MAP_NORESERVE mappings may also need pages and quota allocated
848 * if no reserve mapping overlaps.
850 chg = vma_needs_reservation(vma, addr);
854 if (hugetlb_get_quota(inode->i_mapping, chg))
855 return ERR_PTR(-ENOSPC);
857 spin_lock(&hugetlb_lock);
858 page = dequeue_huge_page_vma(vma, addr, avoid_reserve);
859 spin_unlock(&hugetlb_lock);
862 page = alloc_buddy_huge_page(vma, addr);
864 hugetlb_put_quota(inode->i_mapping, chg);
865 return ERR_PTR(-VM_FAULT_OOM);
869 set_page_refcounted(page);
870 set_page_private(page, (unsigned long) mapping);
872 vma_commit_reservation(vma, addr);
877 static int __init hugetlb_init(void)
881 if (HPAGE_SHIFT == 0)
884 for (i = 0; i < MAX_NUMNODES; ++i)
885 INIT_LIST_HEAD(&hugepage_freelists[i]);
887 hugetlb_next_nid = first_node(node_online_map);
889 for (i = 0; i < max_huge_pages; ++i) {
890 if (!alloc_fresh_huge_page())
893 max_huge_pages = free_huge_pages = nr_huge_pages = i;
894 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
897 module_init(hugetlb_init);
899 static int __init hugetlb_setup(char *s)
901 if (sscanf(s, "%lu", &max_huge_pages) <= 0)
905 __setup("hugepages=", hugetlb_setup);
907 static unsigned int cpuset_mems_nr(unsigned int *array)
912 for_each_node_mask(node, cpuset_current_mems_allowed)
919 #ifdef CONFIG_HIGHMEM
920 static void try_to_free_low(unsigned long count)
924 for (i = 0; i < MAX_NUMNODES; ++i) {
925 struct page *page, *next;
926 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
927 if (count >= nr_huge_pages)
929 if (PageHighMem(page))
931 list_del(&page->lru);
932 update_and_free_page(page);
934 free_huge_pages_node[page_to_nid(page)]--;
939 static inline void try_to_free_low(unsigned long count)
944 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
945 static unsigned long set_max_huge_pages(unsigned long count)
947 unsigned long min_count, ret;
950 * Increase the pool size
951 * First take pages out of surplus state. Then make up the
952 * remaining difference by allocating fresh huge pages.
954 * We might race with alloc_buddy_huge_page() here and be unable
955 * to convert a surplus huge page to a normal huge page. That is
956 * not critical, though, it just means the overall size of the
957 * pool might be one hugepage larger than it needs to be, but
958 * within all the constraints specified by the sysctls.
960 spin_lock(&hugetlb_lock);
961 while (surplus_huge_pages && count > persistent_huge_pages) {
962 if (!adjust_pool_surplus(-1))
966 while (count > persistent_huge_pages) {
968 * If this allocation races such that we no longer need the
969 * page, free_huge_page will handle it by freeing the page
970 * and reducing the surplus.
972 spin_unlock(&hugetlb_lock);
973 ret = alloc_fresh_huge_page();
974 spin_lock(&hugetlb_lock);
981 * Decrease the pool size
982 * First return free pages to the buddy allocator (being careful
983 * to keep enough around to satisfy reservations). Then place
984 * pages into surplus state as needed so the pool will shrink
985 * to the desired size as pages become free.
987 * By placing pages into the surplus state independent of the
988 * overcommit value, we are allowing the surplus pool size to
989 * exceed overcommit. There are few sane options here. Since
990 * alloc_buddy_huge_page() is checking the global counter,
991 * though, we'll note that we're not allowed to exceed surplus
992 * and won't grow the pool anywhere else. Not until one of the
993 * sysctls are changed, or the surplus pages go out of use.
995 min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
996 min_count = max(count, min_count);
997 try_to_free_low(min_count);
998 while (min_count < persistent_huge_pages) {
999 struct page *page = dequeue_huge_page();
1002 update_and_free_page(page);
1004 while (count < persistent_huge_pages) {
1005 if (!adjust_pool_surplus(1))
1009 ret = persistent_huge_pages;
1010 spin_unlock(&hugetlb_lock);
1014 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1015 struct file *file, void __user *buffer,
1016 size_t *length, loff_t *ppos)
1018 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1019 max_huge_pages = set_max_huge_pages(max_huge_pages);
1023 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1024 struct file *file, void __user *buffer,
1025 size_t *length, loff_t *ppos)
1027 proc_dointvec(table, write, file, buffer, length, ppos);
1028 if (hugepages_treat_as_movable)
1029 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1031 htlb_alloc_mask = GFP_HIGHUSER;
1035 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1036 struct file *file, void __user *buffer,
1037 size_t *length, loff_t *ppos)
1039 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1040 spin_lock(&hugetlb_lock);
1041 nr_overcommit_huge_pages = sysctl_overcommit_huge_pages;
1042 spin_unlock(&hugetlb_lock);
1046 #endif /* CONFIG_SYSCTL */
1048 int hugetlb_report_meminfo(char *buf)
1051 "HugePages_Total: %5lu\n"
1052 "HugePages_Free: %5lu\n"
1053 "HugePages_Rsvd: %5lu\n"
1054 "HugePages_Surp: %5lu\n"
1055 "Hugepagesize: %5lu kB\n",
1063 int hugetlb_report_node_meminfo(int nid, char *buf)
1066 "Node %d HugePages_Total: %5u\n"
1067 "Node %d HugePages_Free: %5u\n"
1068 "Node %d HugePages_Surp: %5u\n",
1069 nid, nr_huge_pages_node[nid],
1070 nid, free_huge_pages_node[nid],
1071 nid, surplus_huge_pages_node[nid]);
1074 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1075 unsigned long hugetlb_total_pages(void)
1077 return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
1080 static int hugetlb_acct_memory(long delta)
1084 spin_lock(&hugetlb_lock);
1086 * When cpuset is configured, it breaks the strict hugetlb page
1087 * reservation as the accounting is done on a global variable. Such
1088 * reservation is completely rubbish in the presence of cpuset because
1089 * the reservation is not checked against page availability for the
1090 * current cpuset. Application can still potentially OOM'ed by kernel
1091 * with lack of free htlb page in cpuset that the task is in.
1092 * Attempt to enforce strict accounting with cpuset is almost
1093 * impossible (or too ugly) because cpuset is too fluid that
1094 * task or memory node can be dynamically moved between cpusets.
1096 * The change of semantics for shared hugetlb mapping with cpuset is
1097 * undesirable. However, in order to preserve some of the semantics,
1098 * we fall back to check against current free page availability as
1099 * a best attempt and hopefully to minimize the impact of changing
1100 * semantics that cpuset has.
1103 if (gather_surplus_pages(delta) < 0)
1106 if (delta > cpuset_mems_nr(free_huge_pages_node)) {
1107 return_unused_surplus_pages(delta);
1114 return_unused_surplus_pages((unsigned long) -delta);
1117 spin_unlock(&hugetlb_lock);
1121 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
1123 struct resv_map *reservations = vma_resv_map(vma);
1126 * This new VMA should share its siblings reservation map if present.
1127 * The VMA will only ever have a valid reservation map pointer where
1128 * it is being copied for another still existing VMA. As that VMA
1129 * has a reference to the reservation map it cannot dissappear until
1130 * after this open call completes. It is therefore safe to take a
1131 * new reference here without additional locking.
1134 kref_get(&reservations->refs);
1137 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
1139 struct resv_map *reservations = vma_resv_map(vma);
1140 unsigned long reserve;
1141 unsigned long start;
1145 start = vma_hugecache_offset(vma, vma->vm_start);
1146 end = vma_hugecache_offset(vma, vma->vm_end);
1148 reserve = (end - start) -
1149 region_count(&reservations->regions, start, end);
1151 kref_put(&reservations->refs, resv_map_release);
1154 hugetlb_acct_memory(-reserve);
1159 * We cannot handle pagefaults against hugetlb pages at all. They cause
1160 * handle_mm_fault() to try to instantiate regular-sized pages in the
1161 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
1164 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1170 struct vm_operations_struct hugetlb_vm_ops = {
1171 .fault = hugetlb_vm_op_fault,
1172 .open = hugetlb_vm_op_open,
1173 .close = hugetlb_vm_op_close,
1176 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
1183 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1185 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
1187 entry = pte_mkyoung(entry);
1188 entry = pte_mkhuge(entry);
1193 static void set_huge_ptep_writable(struct vm_area_struct *vma,
1194 unsigned long address, pte_t *ptep)
1198 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
1199 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
1200 update_mmu_cache(vma, address, entry);
1205 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
1206 struct vm_area_struct *vma)
1208 pte_t *src_pte, *dst_pte, entry;
1209 struct page *ptepage;
1213 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
1215 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
1216 src_pte = huge_pte_offset(src, addr);
1219 dst_pte = huge_pte_alloc(dst, addr);
1223 /* If the pagetables are shared don't copy or take references */
1224 if (dst_pte == src_pte)
1227 spin_lock(&dst->page_table_lock);
1228 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
1229 if (!huge_pte_none(huge_ptep_get(src_pte))) {
1231 huge_ptep_set_wrprotect(src, addr, src_pte);
1232 entry = huge_ptep_get(src_pte);
1233 ptepage = pte_page(entry);
1235 set_huge_pte_at(dst, addr, dst_pte, entry);
1237 spin_unlock(&src->page_table_lock);
1238 spin_unlock(&dst->page_table_lock);
1246 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1247 unsigned long end, struct page *ref_page)
1249 struct mm_struct *mm = vma->vm_mm;
1250 unsigned long address;
1256 * A page gathering list, protected by per file i_mmap_lock. The
1257 * lock is used to avoid list corruption from multiple unmapping
1258 * of the same page since we are using page->lru.
1260 LIST_HEAD(page_list);
1262 WARN_ON(!is_vm_hugetlb_page(vma));
1263 BUG_ON(start & ~HPAGE_MASK);
1264 BUG_ON(end & ~HPAGE_MASK);
1266 spin_lock(&mm->page_table_lock);
1267 for (address = start; address < end; address += HPAGE_SIZE) {
1268 ptep = huge_pte_offset(mm, address);
1272 if (huge_pmd_unshare(mm, &address, ptep))
1276 * If a reference page is supplied, it is because a specific
1277 * page is being unmapped, not a range. Ensure the page we
1278 * are about to unmap is the actual page of interest.
1281 pte = huge_ptep_get(ptep);
1282 if (huge_pte_none(pte))
1284 page = pte_page(pte);
1285 if (page != ref_page)
1289 * Mark the VMA as having unmapped its page so that
1290 * future faults in this VMA will fail rather than
1291 * looking like data was lost
1293 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
1296 pte = huge_ptep_get_and_clear(mm, address, ptep);
1297 if (huge_pte_none(pte))
1300 page = pte_page(pte);
1302 set_page_dirty(page);
1303 list_add(&page->lru, &page_list);
1305 spin_unlock(&mm->page_table_lock);
1306 flush_tlb_range(vma, start, end);
1307 list_for_each_entry_safe(page, tmp, &page_list, lru) {
1308 list_del(&page->lru);
1313 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1314 unsigned long end, struct page *ref_page)
1317 * It is undesirable to test vma->vm_file as it should be non-null
1318 * for valid hugetlb area. However, vm_file will be NULL in the error
1319 * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
1320 * do_mmap_pgoff() nullifies vma->vm_file before calling this function
1321 * to clean up. Since no pte has actually been setup, it is safe to
1322 * do nothing in this case.
1325 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1326 __unmap_hugepage_range(vma, start, end, ref_page);
1327 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1332 * This is called when the original mapper is failing to COW a MAP_PRIVATE
1333 * mappping it owns the reserve page for. The intention is to unmap the page
1334 * from other VMAs and let the children be SIGKILLed if they are faulting the
1337 int unmap_ref_private(struct mm_struct *mm,
1338 struct vm_area_struct *vma,
1340 unsigned long address)
1342 struct vm_area_struct *iter_vma;
1343 struct address_space *mapping;
1344 struct prio_tree_iter iter;
1348 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1349 * from page cache lookup which is in HPAGE_SIZE units.
1351 address = address & huge_page_mask(hstate_vma(vma));
1352 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
1353 + (vma->vm_pgoff >> PAGE_SHIFT);
1354 mapping = (struct address_space *)page_private(page);
1356 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1357 /* Do not unmap the current VMA */
1358 if (iter_vma == vma)
1362 * Unmap the page from other VMAs without their own reserves.
1363 * They get marked to be SIGKILLed if they fault in these
1364 * areas. This is because a future no-page fault on this VMA
1365 * could insert a zeroed page instead of the data existing
1366 * from the time of fork. This would look like data corruption
1368 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
1369 unmap_hugepage_range(iter_vma,
1370 address, address + HPAGE_SIZE,
1377 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
1378 unsigned long address, pte_t *ptep, pte_t pte,
1379 struct page *pagecache_page)
1381 struct page *old_page, *new_page;
1383 int outside_reserve = 0;
1385 old_page = pte_page(pte);
1388 /* If no-one else is actually using this page, avoid the copy
1389 * and just make the page writable */
1390 avoidcopy = (page_count(old_page) == 1);
1392 set_huge_ptep_writable(vma, address, ptep);
1397 * If the process that created a MAP_PRIVATE mapping is about to
1398 * perform a COW due to a shared page count, attempt to satisfy
1399 * the allocation without using the existing reserves. The pagecache
1400 * page is used to determine if the reserve at this address was
1401 * consumed or not. If reserves were used, a partial faulted mapping
1402 * at the time of fork() could consume its reserves on COW instead
1403 * of the full address range.
1405 if (!(vma->vm_flags & VM_SHARED) &&
1406 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
1407 old_page != pagecache_page)
1408 outside_reserve = 1;
1410 page_cache_get(old_page);
1411 new_page = alloc_huge_page(vma, address, outside_reserve);
1413 if (IS_ERR(new_page)) {
1414 page_cache_release(old_page);
1417 * If a process owning a MAP_PRIVATE mapping fails to COW,
1418 * it is due to references held by a child and an insufficient
1419 * huge page pool. To guarantee the original mappers
1420 * reliability, unmap the page from child processes. The child
1421 * may get SIGKILLed if it later faults.
1423 if (outside_reserve) {
1424 BUG_ON(huge_pte_none(pte));
1425 if (unmap_ref_private(mm, vma, old_page, address)) {
1426 BUG_ON(page_count(old_page) != 1);
1427 BUG_ON(huge_pte_none(pte));
1428 goto retry_avoidcopy;
1433 return -PTR_ERR(new_page);
1436 spin_unlock(&mm->page_table_lock);
1437 copy_huge_page(new_page, old_page, address, vma);
1438 __SetPageUptodate(new_page);
1439 spin_lock(&mm->page_table_lock);
1441 ptep = huge_pte_offset(mm, address & HPAGE_MASK);
1442 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
1444 huge_ptep_clear_flush(vma, address, ptep);
1445 set_huge_pte_at(mm, address, ptep,
1446 make_huge_pte(vma, new_page, 1));
1447 /* Make the old page be freed below */
1448 new_page = old_page;
1450 page_cache_release(new_page);
1451 page_cache_release(old_page);
1455 /* Return the pagecache page at a given address within a VMA */
1456 static struct page *hugetlbfs_pagecache_page(struct vm_area_struct *vma,
1457 unsigned long address)
1459 struct address_space *mapping;
1462 mapping = vma->vm_file->f_mapping;
1463 idx = vma_hugecache_offset(vma, address);
1465 return find_lock_page(mapping, idx);
1468 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1469 unsigned long address, pte_t *ptep, int write_access)
1471 int ret = VM_FAULT_SIGBUS;
1475 struct address_space *mapping;
1479 * Currently, we are forced to kill the process in the event the
1480 * original mapper has unmapped pages from the child due to a failed
1481 * COW. Warn that such a situation has occured as it may not be obvious
1483 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
1485 "PID %d killed due to inadequate hugepage pool\n",
1490 mapping = vma->vm_file->f_mapping;
1491 idx = vma_hugecache_offset(vma, address);
1494 * Use page lock to guard against racing truncation
1495 * before we get page_table_lock.
1498 page = find_lock_page(mapping, idx);
1500 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
1503 page = alloc_huge_page(vma, address, 0);
1505 ret = -PTR_ERR(page);
1508 clear_huge_page(page, address);
1509 __SetPageUptodate(page);
1511 if (vma->vm_flags & VM_SHARED) {
1513 struct inode *inode = mapping->host;
1515 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
1523 spin_lock(&inode->i_lock);
1524 inode->i_blocks += BLOCKS_PER_HUGEPAGE;
1525 spin_unlock(&inode->i_lock);
1530 spin_lock(&mm->page_table_lock);
1531 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
1536 if (!huge_pte_none(huge_ptep_get(ptep)))
1539 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
1540 && (vma->vm_flags & VM_SHARED)));
1541 set_huge_pte_at(mm, address, ptep, new_pte);
1543 if (write_access && !(vma->vm_flags & VM_SHARED)) {
1544 /* Optimization, do the COW without a second fault */
1545 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
1548 spin_unlock(&mm->page_table_lock);
1554 spin_unlock(&mm->page_table_lock);
1560 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
1561 unsigned long address, int write_access)
1566 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
1568 ptep = huge_pte_alloc(mm, address);
1570 return VM_FAULT_OOM;
1573 * Serialize hugepage allocation and instantiation, so that we don't
1574 * get spurious allocation failures if two CPUs race to instantiate
1575 * the same page in the page cache.
1577 mutex_lock(&hugetlb_instantiation_mutex);
1578 entry = huge_ptep_get(ptep);
1579 if (huge_pte_none(entry)) {
1580 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
1581 mutex_unlock(&hugetlb_instantiation_mutex);
1587 spin_lock(&mm->page_table_lock);
1588 /* Check for a racing update before calling hugetlb_cow */
1589 if (likely(pte_same(entry, huge_ptep_get(ptep))))
1590 if (write_access && !pte_write(entry)) {
1592 page = hugetlbfs_pagecache_page(vma, address);
1593 ret = hugetlb_cow(mm, vma, address, ptep, entry, page);
1599 spin_unlock(&mm->page_table_lock);
1600 mutex_unlock(&hugetlb_instantiation_mutex);
1605 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
1606 struct page **pages, struct vm_area_struct **vmas,
1607 unsigned long *position, int *length, int i,
1610 unsigned long pfn_offset;
1611 unsigned long vaddr = *position;
1612 int remainder = *length;
1614 spin_lock(&mm->page_table_lock);
1615 while (vaddr < vma->vm_end && remainder) {
1620 * Some archs (sparc64, sh*) have multiple pte_ts to
1621 * each hugepage. We have to make * sure we get the
1622 * first, for the page indexing below to work.
1624 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
1626 if (!pte || huge_pte_none(huge_ptep_get(pte)) ||
1627 (write && !pte_write(huge_ptep_get(pte)))) {
1630 spin_unlock(&mm->page_table_lock);
1631 ret = hugetlb_fault(mm, vma, vaddr, write);
1632 spin_lock(&mm->page_table_lock);
1633 if (!(ret & VM_FAULT_ERROR))
1642 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
1643 page = pte_page(huge_ptep_get(pte));
1647 pages[i] = page + pfn_offset;
1657 if (vaddr < vma->vm_end && remainder &&
1658 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
1660 * We use pfn_offset to avoid touching the pageframes
1661 * of this compound page.
1666 spin_unlock(&mm->page_table_lock);
1667 *length = remainder;
1673 void hugetlb_change_protection(struct vm_area_struct *vma,
1674 unsigned long address, unsigned long end, pgprot_t newprot)
1676 struct mm_struct *mm = vma->vm_mm;
1677 unsigned long start = address;
1681 BUG_ON(address >= end);
1682 flush_cache_range(vma, address, end);
1684 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1685 spin_lock(&mm->page_table_lock);
1686 for (; address < end; address += HPAGE_SIZE) {
1687 ptep = huge_pte_offset(mm, address);
1690 if (huge_pmd_unshare(mm, &address, ptep))
1692 if (!huge_pte_none(huge_ptep_get(ptep))) {
1693 pte = huge_ptep_get_and_clear(mm, address, ptep);
1694 pte = pte_mkhuge(pte_modify(pte, newprot));
1695 set_huge_pte_at(mm, address, ptep, pte);
1698 spin_unlock(&mm->page_table_lock);
1699 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1701 flush_tlb_range(vma, start, end);
1704 int hugetlb_reserve_pages(struct inode *inode,
1706 struct vm_area_struct *vma)
1710 if (vma && vma->vm_flags & VM_NORESERVE)
1714 * Shared mappings base their reservation on the number of pages that
1715 * are already allocated on behalf of the file. Private mappings need
1716 * to reserve the full area even if read-only as mprotect() may be
1717 * called to make the mapping read-write. Assume !vma is a shm mapping
1719 if (!vma || vma->vm_flags & VM_SHARED)
1720 chg = region_chg(&inode->i_mapping->private_list, from, to);
1722 struct resv_map *resv_map = resv_map_alloc();
1728 set_vma_resv_map(vma, resv_map);
1729 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
1735 if (hugetlb_get_quota(inode->i_mapping, chg))
1737 ret = hugetlb_acct_memory(chg);
1739 hugetlb_put_quota(inode->i_mapping, chg);
1742 if (!vma || vma->vm_flags & VM_SHARED)
1743 region_add(&inode->i_mapping->private_list, from, to);
1747 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1749 long chg = region_truncate(&inode->i_mapping->private_list, offset);
1751 spin_lock(&inode->i_lock);
1752 inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed;
1753 spin_unlock(&inode->i_lock);
1755 hugetlb_put_quota(inode->i_mapping, (chg - freed));
1756 hugetlb_acct_memory(-(chg - freed));