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 base page units.
206 static pgoff_t vma_page_offset(struct vm_area_struct *vma,
207 unsigned long address)
209 return ((address - vma->vm_start) >> PAGE_SHIFT) +
210 (vma->vm_pgoff >> PAGE_SHIFT);
214 * Convert the address within this vma to the page offset within
215 * the mapping, in pagecache page units; huge pages here.
217 static pgoff_t vma_pagecache_offset(struct vm_area_struct *vma,
218 unsigned long address)
220 return ((address - vma->vm_start) >> HPAGE_SHIFT) +
221 (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
225 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
226 * bits of the reservation map pointer, which are always clear due to
229 #define HPAGE_RESV_OWNER (1UL << 0)
230 #define HPAGE_RESV_UNMAPPED (1UL << 1)
231 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
234 * These helpers are used to track how many pages are reserved for
235 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
236 * is guaranteed to have their future faults succeed.
238 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
239 * the reserve counters are updated with the hugetlb_lock held. It is safe
240 * to reset the VMA at fork() time as it is not in use yet and there is no
241 * chance of the global counters getting corrupted as a result of the values.
243 * The private mapping reservation is represented in a subtly different
244 * manner to a shared mapping. A shared mapping has a region map associated
245 * with the underlying file, this region map represents the backing file
246 * pages which have ever had a reservation assigned which this persists even
247 * after the page is instantiated. A private mapping has a region map
248 * associated with the original mmap which is attached to all VMAs which
249 * reference it, this region map represents those offsets which have consumed
250 * reservation ie. where pages have been instantiated.
252 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
254 return (unsigned long)vma->vm_private_data;
257 static void set_vma_private_data(struct vm_area_struct *vma,
260 vma->vm_private_data = (void *)value;
265 struct list_head regions;
268 struct resv_map *resv_map_alloc(void)
270 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
274 kref_init(&resv_map->refs);
275 INIT_LIST_HEAD(&resv_map->regions);
280 void resv_map_release(struct kref *ref)
282 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
284 /* Clear out any active regions before we release the map. */
285 region_truncate(&resv_map->regions, 0);
289 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
291 VM_BUG_ON(!is_vm_hugetlb_page(vma));
292 if (!(vma->vm_flags & VM_SHARED))
293 return (struct resv_map *)(get_vma_private_data(vma) &
298 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
300 VM_BUG_ON(!is_vm_hugetlb_page(vma));
301 VM_BUG_ON(vma->vm_flags & VM_SHARED);
303 set_vma_private_data(vma, (get_vma_private_data(vma) &
304 HPAGE_RESV_MASK) | (unsigned long)map);
307 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
309 VM_BUG_ON(!is_vm_hugetlb_page(vma));
310 VM_BUG_ON(vma->vm_flags & VM_SHARED);
312 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
315 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
317 VM_BUG_ON(!is_vm_hugetlb_page(vma));
319 return (get_vma_private_data(vma) & flag) != 0;
322 /* Decrement the reserved pages in the hugepage pool by one */
323 static void decrement_hugepage_resv_vma(struct vm_area_struct *vma)
325 if (vma->vm_flags & VM_NORESERVE)
328 if (vma->vm_flags & VM_SHARED) {
329 /* Shared mappings always use reserves */
331 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
333 * Only the process that called mmap() has reserves for
340 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
341 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
343 VM_BUG_ON(!is_vm_hugetlb_page(vma));
344 if (!(vma->vm_flags & VM_SHARED))
345 vma->vm_private_data = (void *)0;
348 /* Returns true if the VMA has associated reserve pages */
349 static int vma_has_private_reserves(struct vm_area_struct *vma)
351 if (vma->vm_flags & VM_SHARED)
353 if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER))
358 static void clear_huge_page(struct page *page, unsigned long addr)
363 for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
365 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
369 static void copy_huge_page(struct page *dst, struct page *src,
370 unsigned long addr, struct vm_area_struct *vma)
375 for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
377 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
381 static void enqueue_huge_page(struct page *page)
383 int nid = page_to_nid(page);
384 list_add(&page->lru, &hugepage_freelists[nid]);
386 free_huge_pages_node[nid]++;
389 static struct page *dequeue_huge_page(void)
392 struct page *page = NULL;
394 for (nid = 0; nid < MAX_NUMNODES; ++nid) {
395 if (!list_empty(&hugepage_freelists[nid])) {
396 page = list_entry(hugepage_freelists[nid].next,
398 list_del(&page->lru);
400 free_huge_pages_node[nid]--;
407 static struct page *dequeue_huge_page_vma(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_private_reserves(vma) &&
425 free_huge_pages - resv_huge_pages == 0)
428 /* If reserves cannot be used, ensure enough pages are in the pool */
429 if (avoid_reserve && free_huge_pages - 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(&hugepage_freelists[nid])) {
437 page = list_entry(hugepage_freelists[nid].next,
439 list_del(&page->lru);
441 free_huge_pages_node[nid]--;
444 decrement_hugepage_resv_vma(vma);
453 static void update_and_free_page(struct page *page)
457 nr_huge_pages_node[page_to_nid(page)]--;
458 for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
459 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
460 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
461 1 << PG_private | 1<< PG_writeback);
463 set_compound_page_dtor(page, NULL);
464 set_page_refcounted(page);
465 arch_release_hugepage(page);
466 __free_pages(page, HUGETLB_PAGE_ORDER);
469 static void free_huge_page(struct page *page)
471 int nid = page_to_nid(page);
472 struct address_space *mapping;
474 mapping = (struct address_space *) page_private(page);
475 set_page_private(page, 0);
476 BUG_ON(page_count(page));
477 INIT_LIST_HEAD(&page->lru);
479 spin_lock(&hugetlb_lock);
480 if (surplus_huge_pages_node[nid]) {
481 update_and_free_page(page);
482 surplus_huge_pages--;
483 surplus_huge_pages_node[nid]--;
485 enqueue_huge_page(page);
487 spin_unlock(&hugetlb_lock);
489 hugetlb_put_quota(mapping, 1);
493 * Increment or decrement surplus_huge_pages. Keep node-specific counters
494 * balanced by operating on them in a round-robin fashion.
495 * Returns 1 if an adjustment was made.
497 static int adjust_pool_surplus(int delta)
503 VM_BUG_ON(delta != -1 && delta != 1);
505 nid = next_node(nid, node_online_map);
506 if (nid == MAX_NUMNODES)
507 nid = first_node(node_online_map);
509 /* To shrink on this node, there must be a surplus page */
510 if (delta < 0 && !surplus_huge_pages_node[nid])
512 /* Surplus cannot exceed the total number of pages */
513 if (delta > 0 && surplus_huge_pages_node[nid] >=
514 nr_huge_pages_node[nid])
517 surplus_huge_pages += delta;
518 surplus_huge_pages_node[nid] += delta;
521 } while (nid != prev_nid);
527 static struct page *alloc_fresh_huge_page_node(int nid)
531 page = alloc_pages_node(nid,
532 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
533 __GFP_REPEAT|__GFP_NOWARN,
536 if (arch_prepare_hugepage(page)) {
537 __free_pages(page, HUGETLB_PAGE_ORDER);
540 set_compound_page_dtor(page, free_huge_page);
541 spin_lock(&hugetlb_lock);
543 nr_huge_pages_node[nid]++;
544 spin_unlock(&hugetlb_lock);
545 put_page(page); /* free it into the hugepage allocator */
551 static int alloc_fresh_huge_page(void)
558 start_nid = hugetlb_next_nid;
561 page = alloc_fresh_huge_page_node(hugetlb_next_nid);
565 * Use a helper variable to find the next node and then
566 * copy it back to hugetlb_next_nid afterwards:
567 * otherwise there's a window in which a racer might
568 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
569 * But we don't need to use a spin_lock here: it really
570 * doesn't matter if occasionally a racer chooses the
571 * same nid as we do. Move nid forward in the mask even
572 * if we just successfully allocated a hugepage so that
573 * the next caller gets hugepages on the next node.
575 next_nid = next_node(hugetlb_next_nid, node_online_map);
576 if (next_nid == MAX_NUMNODES)
577 next_nid = first_node(node_online_map);
578 hugetlb_next_nid = next_nid;
579 } while (!page && hugetlb_next_nid != start_nid);
582 count_vm_event(HTLB_BUDDY_PGALLOC);
584 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
589 static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
590 unsigned long address)
596 * Assume we will successfully allocate the surplus page to
597 * prevent racing processes from causing the surplus to exceed
600 * This however introduces a different race, where a process B
601 * tries to grow the static hugepage pool while alloc_pages() is
602 * called by process A. B will only examine the per-node
603 * counters in determining if surplus huge pages can be
604 * converted to normal huge pages in adjust_pool_surplus(). A
605 * won't be able to increment the per-node counter, until the
606 * lock is dropped by B, but B doesn't drop hugetlb_lock until
607 * no more huge pages can be converted from surplus to normal
608 * state (and doesn't try to convert again). Thus, we have a
609 * case where a surplus huge page exists, the pool is grown, and
610 * the surplus huge page still exists after, even though it
611 * should just have been converted to a normal huge page. This
612 * does not leak memory, though, as the hugepage will be freed
613 * once it is out of use. It also does not allow the counters to
614 * go out of whack in adjust_pool_surplus() as we don't modify
615 * the node values until we've gotten the hugepage and only the
616 * per-node value is checked there.
618 spin_lock(&hugetlb_lock);
619 if (surplus_huge_pages >= nr_overcommit_huge_pages) {
620 spin_unlock(&hugetlb_lock);
624 surplus_huge_pages++;
626 spin_unlock(&hugetlb_lock);
628 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
629 __GFP_REPEAT|__GFP_NOWARN,
632 spin_lock(&hugetlb_lock);
635 * This page is now managed by the hugetlb allocator and has
636 * no users -- drop the buddy allocator's reference.
638 put_page_testzero(page);
639 VM_BUG_ON(page_count(page));
640 nid = page_to_nid(page);
641 set_compound_page_dtor(page, free_huge_page);
643 * We incremented the global counters already
645 nr_huge_pages_node[nid]++;
646 surplus_huge_pages_node[nid]++;
647 __count_vm_event(HTLB_BUDDY_PGALLOC);
650 surplus_huge_pages--;
651 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
653 spin_unlock(&hugetlb_lock);
659 * Increase the hugetlb pool such that it can accomodate a reservation
662 static int gather_surplus_pages(int delta)
664 struct list_head surplus_list;
665 struct page *page, *tmp;
667 int needed, allocated;
669 needed = (resv_huge_pages + delta) - free_huge_pages;
671 resv_huge_pages += delta;
676 INIT_LIST_HEAD(&surplus_list);
680 spin_unlock(&hugetlb_lock);
681 for (i = 0; i < needed; i++) {
682 page = alloc_buddy_huge_page(NULL, 0);
685 * We were not able to allocate enough pages to
686 * satisfy the entire reservation so we free what
687 * we've allocated so far.
689 spin_lock(&hugetlb_lock);
694 list_add(&page->lru, &surplus_list);
699 * After retaking hugetlb_lock, we need to recalculate 'needed'
700 * because either resv_huge_pages or free_huge_pages may have changed.
702 spin_lock(&hugetlb_lock);
703 needed = (resv_huge_pages + delta) - (free_huge_pages + allocated);
708 * The surplus_list now contains _at_least_ the number of extra pages
709 * needed to accomodate the reservation. Add the appropriate number
710 * of pages to the hugetlb pool and free the extras back to the buddy
711 * allocator. Commit the entire reservation here to prevent another
712 * process from stealing the pages as they are added to the pool but
713 * before they are reserved.
716 resv_huge_pages += delta;
719 /* Free the needed pages to the hugetlb pool */
720 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
723 list_del(&page->lru);
724 enqueue_huge_page(page);
727 /* Free unnecessary surplus pages to the buddy allocator */
728 if (!list_empty(&surplus_list)) {
729 spin_unlock(&hugetlb_lock);
730 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
731 list_del(&page->lru);
733 * The page has a reference count of zero already, so
734 * call free_huge_page directly instead of using
735 * put_page. This must be done with hugetlb_lock
736 * unlocked which is safe because free_huge_page takes
737 * hugetlb_lock before deciding how to free the page.
739 free_huge_page(page);
741 spin_lock(&hugetlb_lock);
748 * When releasing a hugetlb pool reservation, any surplus pages that were
749 * allocated to satisfy the reservation must be explicitly freed if they were
752 static void return_unused_surplus_pages(unsigned long unused_resv_pages)
756 unsigned long nr_pages;
759 * We want to release as many surplus pages as possible, spread
760 * evenly across all nodes. Iterate across all nodes until we
761 * can no longer free unreserved surplus pages. This occurs when
762 * the nodes with surplus pages have no free pages.
764 unsigned long remaining_iterations = num_online_nodes();
766 /* Uncommit the reservation */
767 resv_huge_pages -= unused_resv_pages;
769 nr_pages = min(unused_resv_pages, surplus_huge_pages);
771 while (remaining_iterations-- && nr_pages) {
772 nid = next_node(nid, node_online_map);
773 if (nid == MAX_NUMNODES)
774 nid = first_node(node_online_map);
776 if (!surplus_huge_pages_node[nid])
779 if (!list_empty(&hugepage_freelists[nid])) {
780 page = list_entry(hugepage_freelists[nid].next,
782 list_del(&page->lru);
783 update_and_free_page(page);
785 free_huge_pages_node[nid]--;
786 surplus_huge_pages--;
787 surplus_huge_pages_node[nid]--;
789 remaining_iterations = num_online_nodes();
795 * Determine if the huge page at addr within the vma has an associated
796 * reservation. Where it does not we will need to logically increase
797 * reservation and actually increase quota before an allocation can occur.
798 * Where any new reservation would be required the reservation change is
799 * prepared, but not committed. Once the page has been quota'd allocated
800 * an instantiated the change should be committed via vma_commit_reservation.
801 * No action is required on failure.
803 static int vma_needs_reservation(struct vm_area_struct *vma, unsigned long addr)
805 struct address_space *mapping = vma->vm_file->f_mapping;
806 struct inode *inode = mapping->host;
808 if (vma->vm_flags & VM_SHARED) {
809 pgoff_t idx = vma_pagecache_offset(vma, addr);
810 return region_chg(&inode->i_mapping->private_list,
813 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
818 pgoff_t idx = vma_pagecache_offset(vma, addr);
819 struct resv_map *reservations = vma_resv_map(vma);
821 err = region_chg(&reservations->regions, idx, idx + 1);
827 static void vma_commit_reservation(struct vm_area_struct *vma,
830 struct address_space *mapping = vma->vm_file->f_mapping;
831 struct inode *inode = mapping->host;
833 if (vma->vm_flags & VM_SHARED) {
834 pgoff_t idx = vma_pagecache_offset(vma, addr);
835 region_add(&inode->i_mapping->private_list, idx, idx + 1);
837 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
838 pgoff_t idx = vma_pagecache_offset(vma, addr);
839 struct resv_map *reservations = vma_resv_map(vma);
841 /* Mark this page used in the map. */
842 region_add(&reservations->regions, idx, idx + 1);
846 static struct page *alloc_huge_page(struct vm_area_struct *vma,
847 unsigned long addr, int avoid_reserve)
850 struct address_space *mapping = vma->vm_file->f_mapping;
851 struct inode *inode = mapping->host;
855 * Processes that did not create the mapping will have no reserves and
856 * will not have accounted against quota. Check that the quota can be
857 * made before satisfying the allocation
858 * MAP_NORESERVE mappings may also need pages and quota allocated
859 * if no reserve mapping overlaps.
861 chg = vma_needs_reservation(vma, addr);
865 if (hugetlb_get_quota(inode->i_mapping, chg))
866 return ERR_PTR(-ENOSPC);
868 spin_lock(&hugetlb_lock);
869 page = dequeue_huge_page_vma(vma, addr, avoid_reserve);
870 spin_unlock(&hugetlb_lock);
873 page = alloc_buddy_huge_page(vma, addr);
875 hugetlb_put_quota(inode->i_mapping, chg);
876 return ERR_PTR(-VM_FAULT_OOM);
880 set_page_refcounted(page);
881 set_page_private(page, (unsigned long) mapping);
883 vma_commit_reservation(vma, addr);
888 static int __init hugetlb_init(void)
892 if (HPAGE_SHIFT == 0)
895 for (i = 0; i < MAX_NUMNODES; ++i)
896 INIT_LIST_HEAD(&hugepage_freelists[i]);
898 hugetlb_next_nid = first_node(node_online_map);
900 for (i = 0; i < max_huge_pages; ++i) {
901 if (!alloc_fresh_huge_page())
904 max_huge_pages = free_huge_pages = nr_huge_pages = i;
905 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
908 module_init(hugetlb_init);
910 static int __init hugetlb_setup(char *s)
912 if (sscanf(s, "%lu", &max_huge_pages) <= 0)
916 __setup("hugepages=", hugetlb_setup);
918 static unsigned int cpuset_mems_nr(unsigned int *array)
923 for_each_node_mask(node, cpuset_current_mems_allowed)
930 #ifdef CONFIG_HIGHMEM
931 static void try_to_free_low(unsigned long count)
935 for (i = 0; i < MAX_NUMNODES; ++i) {
936 struct page *page, *next;
937 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
938 if (count >= nr_huge_pages)
940 if (PageHighMem(page))
942 list_del(&page->lru);
943 update_and_free_page(page);
945 free_huge_pages_node[page_to_nid(page)]--;
950 static inline void try_to_free_low(unsigned long count)
955 #define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
956 static unsigned long set_max_huge_pages(unsigned long count)
958 unsigned long min_count, ret;
961 * Increase the pool size
962 * First take pages out of surplus state. Then make up the
963 * remaining difference by allocating fresh huge pages.
965 * We might race with alloc_buddy_huge_page() here and be unable
966 * to convert a surplus huge page to a normal huge page. That is
967 * not critical, though, it just means the overall size of the
968 * pool might be one hugepage larger than it needs to be, but
969 * within all the constraints specified by the sysctls.
971 spin_lock(&hugetlb_lock);
972 while (surplus_huge_pages && count > persistent_huge_pages) {
973 if (!adjust_pool_surplus(-1))
977 while (count > persistent_huge_pages) {
979 * If this allocation races such that we no longer need the
980 * page, free_huge_page will handle it by freeing the page
981 * and reducing the surplus.
983 spin_unlock(&hugetlb_lock);
984 ret = alloc_fresh_huge_page();
985 spin_lock(&hugetlb_lock);
992 * Decrease the pool size
993 * First return free pages to the buddy allocator (being careful
994 * to keep enough around to satisfy reservations). Then place
995 * pages into surplus state as needed so the pool will shrink
996 * to the desired size as pages become free.
998 * By placing pages into the surplus state independent of the
999 * overcommit value, we are allowing the surplus pool size to
1000 * exceed overcommit. There are few sane options here. Since
1001 * alloc_buddy_huge_page() is checking the global counter,
1002 * though, we'll note that we're not allowed to exceed surplus
1003 * and won't grow the pool anywhere else. Not until one of the
1004 * sysctls are changed, or the surplus pages go out of use.
1006 min_count = resv_huge_pages + nr_huge_pages - free_huge_pages;
1007 min_count = max(count, min_count);
1008 try_to_free_low(min_count);
1009 while (min_count < persistent_huge_pages) {
1010 struct page *page = dequeue_huge_page();
1013 update_and_free_page(page);
1015 while (count < persistent_huge_pages) {
1016 if (!adjust_pool_surplus(1))
1020 ret = persistent_huge_pages;
1021 spin_unlock(&hugetlb_lock);
1025 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1026 struct file *file, void __user *buffer,
1027 size_t *length, loff_t *ppos)
1029 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1030 max_huge_pages = set_max_huge_pages(max_huge_pages);
1034 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1035 struct file *file, void __user *buffer,
1036 size_t *length, loff_t *ppos)
1038 proc_dointvec(table, write, file, buffer, length, ppos);
1039 if (hugepages_treat_as_movable)
1040 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1042 htlb_alloc_mask = GFP_HIGHUSER;
1046 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1047 struct file *file, void __user *buffer,
1048 size_t *length, loff_t *ppos)
1050 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1051 spin_lock(&hugetlb_lock);
1052 nr_overcommit_huge_pages = sysctl_overcommit_huge_pages;
1053 spin_unlock(&hugetlb_lock);
1057 #endif /* CONFIG_SYSCTL */
1059 int hugetlb_report_meminfo(char *buf)
1062 "HugePages_Total: %5lu\n"
1063 "HugePages_Free: %5lu\n"
1064 "HugePages_Rsvd: %5lu\n"
1065 "HugePages_Surp: %5lu\n"
1066 "Hugepagesize: %5lu kB\n",
1074 int hugetlb_report_node_meminfo(int nid, char *buf)
1077 "Node %d HugePages_Total: %5u\n"
1078 "Node %d HugePages_Free: %5u\n"
1079 "Node %d HugePages_Surp: %5u\n",
1080 nid, nr_huge_pages_node[nid],
1081 nid, free_huge_pages_node[nid],
1082 nid, surplus_huge_pages_node[nid]);
1085 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1086 unsigned long hugetlb_total_pages(void)
1088 return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
1091 static int hugetlb_acct_memory(long delta)
1095 spin_lock(&hugetlb_lock);
1097 * When cpuset is configured, it breaks the strict hugetlb page
1098 * reservation as the accounting is done on a global variable. Such
1099 * reservation is completely rubbish in the presence of cpuset because
1100 * the reservation is not checked against page availability for the
1101 * current cpuset. Application can still potentially OOM'ed by kernel
1102 * with lack of free htlb page in cpuset that the task is in.
1103 * Attempt to enforce strict accounting with cpuset is almost
1104 * impossible (or too ugly) because cpuset is too fluid that
1105 * task or memory node can be dynamically moved between cpusets.
1107 * The change of semantics for shared hugetlb mapping with cpuset is
1108 * undesirable. However, in order to preserve some of the semantics,
1109 * we fall back to check against current free page availability as
1110 * a best attempt and hopefully to minimize the impact of changing
1111 * semantics that cpuset has.
1114 if (gather_surplus_pages(delta) < 0)
1117 if (delta > cpuset_mems_nr(free_huge_pages_node)) {
1118 return_unused_surplus_pages(delta);
1125 return_unused_surplus_pages((unsigned long) -delta);
1128 spin_unlock(&hugetlb_lock);
1132 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
1134 struct resv_map *reservations = vma_resv_map(vma);
1137 * This new VMA should share its siblings reservation map if present.
1138 * The VMA will only ever have a valid reservation map pointer where
1139 * it is being copied for another still existing VMA. As that VMA
1140 * has a reference to the reservation map it cannot dissappear until
1141 * after this open call completes. It is therefore safe to take a
1142 * new reference here without additional locking.
1145 kref_get(&reservations->refs);
1148 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
1150 struct resv_map *reservations = vma_resv_map(vma);
1151 unsigned long reserve;
1152 unsigned long start;
1156 start = vma_pagecache_offset(vma, vma->vm_start);
1157 end = vma_pagecache_offset(vma, vma->vm_end);
1159 reserve = (end - start) -
1160 region_count(&reservations->regions, start, end);
1162 kref_put(&reservations->refs, resv_map_release);
1165 hugetlb_acct_memory(-reserve);
1170 * We cannot handle pagefaults against hugetlb pages at all. They cause
1171 * handle_mm_fault() to try to instantiate regular-sized pages in the
1172 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
1175 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1181 struct vm_operations_struct hugetlb_vm_ops = {
1182 .fault = hugetlb_vm_op_fault,
1183 .open = hugetlb_vm_op_open,
1184 .close = hugetlb_vm_op_close,
1187 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
1194 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1196 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
1198 entry = pte_mkyoung(entry);
1199 entry = pte_mkhuge(entry);
1204 static void set_huge_ptep_writable(struct vm_area_struct *vma,
1205 unsigned long address, pte_t *ptep)
1209 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
1210 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
1211 update_mmu_cache(vma, address, entry);
1216 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
1217 struct vm_area_struct *vma)
1219 pte_t *src_pte, *dst_pte, entry;
1220 struct page *ptepage;
1224 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
1226 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
1227 src_pte = huge_pte_offset(src, addr);
1230 dst_pte = huge_pte_alloc(dst, addr);
1234 /* If the pagetables are shared don't copy or take references */
1235 if (dst_pte == src_pte)
1238 spin_lock(&dst->page_table_lock);
1239 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
1240 if (!huge_pte_none(huge_ptep_get(src_pte))) {
1242 huge_ptep_set_wrprotect(src, addr, src_pte);
1243 entry = huge_ptep_get(src_pte);
1244 ptepage = pte_page(entry);
1246 set_huge_pte_at(dst, addr, dst_pte, entry);
1248 spin_unlock(&src->page_table_lock);
1249 spin_unlock(&dst->page_table_lock);
1257 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1258 unsigned long end, struct page *ref_page)
1260 struct mm_struct *mm = vma->vm_mm;
1261 unsigned long address;
1267 * A page gathering list, protected by per file i_mmap_lock. The
1268 * lock is used to avoid list corruption from multiple unmapping
1269 * of the same page since we are using page->lru.
1271 LIST_HEAD(page_list);
1273 WARN_ON(!is_vm_hugetlb_page(vma));
1274 BUG_ON(start & ~HPAGE_MASK);
1275 BUG_ON(end & ~HPAGE_MASK);
1277 spin_lock(&mm->page_table_lock);
1278 for (address = start; address < end; address += HPAGE_SIZE) {
1279 ptep = huge_pte_offset(mm, address);
1283 if (huge_pmd_unshare(mm, &address, ptep))
1287 * If a reference page is supplied, it is because a specific
1288 * page is being unmapped, not a range. Ensure the page we
1289 * are about to unmap is the actual page of interest.
1292 pte = huge_ptep_get(ptep);
1293 if (huge_pte_none(pte))
1295 page = pte_page(pte);
1296 if (page != ref_page)
1300 * Mark the VMA as having unmapped its page so that
1301 * future faults in this VMA will fail rather than
1302 * looking like data was lost
1304 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
1307 pte = huge_ptep_get_and_clear(mm, address, ptep);
1308 if (huge_pte_none(pte))
1311 page = pte_page(pte);
1313 set_page_dirty(page);
1314 list_add(&page->lru, &page_list);
1316 spin_unlock(&mm->page_table_lock);
1317 flush_tlb_range(vma, start, end);
1318 list_for_each_entry_safe(page, tmp, &page_list, lru) {
1319 list_del(&page->lru);
1324 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1325 unsigned long end, struct page *ref_page)
1328 * It is undesirable to test vma->vm_file as it should be non-null
1329 * for valid hugetlb area. However, vm_file will be NULL in the error
1330 * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
1331 * do_mmap_pgoff() nullifies vma->vm_file before calling this function
1332 * to clean up. Since no pte has actually been setup, it is safe to
1333 * do nothing in this case.
1336 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1337 __unmap_hugepage_range(vma, start, end, ref_page);
1338 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1343 * This is called when the original mapper is failing to COW a MAP_PRIVATE
1344 * mappping it owns the reserve page for. The intention is to unmap the page
1345 * from other VMAs and let the children be SIGKILLed if they are faulting the
1348 int unmap_ref_private(struct mm_struct *mm,
1349 struct vm_area_struct *vma,
1351 unsigned long address)
1353 struct vm_area_struct *iter_vma;
1354 struct address_space *mapping;
1355 struct prio_tree_iter iter;
1359 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1360 * from page cache lookup which is in HPAGE_SIZE units.
1362 address = address & huge_page_mask(hstate_vma(vma));
1363 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
1364 + (vma->vm_pgoff >> PAGE_SHIFT);
1365 mapping = (struct address_space *)page_private(page);
1367 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1368 /* Do not unmap the current VMA */
1369 if (iter_vma == vma)
1373 * Unmap the page from other VMAs without their own reserves.
1374 * They get marked to be SIGKILLed if they fault in these
1375 * areas. This is because a future no-page fault on this VMA
1376 * could insert a zeroed page instead of the data existing
1377 * from the time of fork. This would look like data corruption
1379 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
1380 unmap_hugepage_range(iter_vma,
1381 address, address + HPAGE_SIZE,
1388 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
1389 unsigned long address, pte_t *ptep, pte_t pte,
1390 struct page *pagecache_page)
1392 struct page *old_page, *new_page;
1394 int outside_reserve = 0;
1396 old_page = pte_page(pte);
1399 /* If no-one else is actually using this page, avoid the copy
1400 * and just make the page writable */
1401 avoidcopy = (page_count(old_page) == 1);
1403 set_huge_ptep_writable(vma, address, ptep);
1408 * If the process that created a MAP_PRIVATE mapping is about to
1409 * perform a COW due to a shared page count, attempt to satisfy
1410 * the allocation without using the existing reserves. The pagecache
1411 * page is used to determine if the reserve at this address was
1412 * consumed or not. If reserves were used, a partial faulted mapping
1413 * at the time of fork() could consume its reserves on COW instead
1414 * of the full address range.
1416 if (!(vma->vm_flags & VM_SHARED) &&
1417 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
1418 old_page != pagecache_page)
1419 outside_reserve = 1;
1421 page_cache_get(old_page);
1422 new_page = alloc_huge_page(vma, address, outside_reserve);
1424 if (IS_ERR(new_page)) {
1425 page_cache_release(old_page);
1428 * If a process owning a MAP_PRIVATE mapping fails to COW,
1429 * it is due to references held by a child and an insufficient
1430 * huge page pool. To guarantee the original mappers
1431 * reliability, unmap the page from child processes. The child
1432 * may get SIGKILLed if it later faults.
1434 if (outside_reserve) {
1435 BUG_ON(huge_pte_none(pte));
1436 if (unmap_ref_private(mm, vma, old_page, address)) {
1437 BUG_ON(page_count(old_page) != 1);
1438 BUG_ON(huge_pte_none(pte));
1439 goto retry_avoidcopy;
1444 return -PTR_ERR(new_page);
1447 spin_unlock(&mm->page_table_lock);
1448 copy_huge_page(new_page, old_page, address, vma);
1449 __SetPageUptodate(new_page);
1450 spin_lock(&mm->page_table_lock);
1452 ptep = huge_pte_offset(mm, address & HPAGE_MASK);
1453 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
1455 huge_ptep_clear_flush(vma, address, ptep);
1456 set_huge_pte_at(mm, address, ptep,
1457 make_huge_pte(vma, new_page, 1));
1458 /* Make the old page be freed below */
1459 new_page = old_page;
1461 page_cache_release(new_page);
1462 page_cache_release(old_page);
1466 /* Return the pagecache page at a given address within a VMA */
1467 static struct page *hugetlbfs_pagecache_page(struct vm_area_struct *vma,
1468 unsigned long address)
1470 struct address_space *mapping;
1473 mapping = vma->vm_file->f_mapping;
1474 idx = vma_pagecache_offset(vma, address);
1476 return find_lock_page(mapping, idx);
1479 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1480 unsigned long address, pte_t *ptep, int write_access)
1482 int ret = VM_FAULT_SIGBUS;
1486 struct address_space *mapping;
1490 * Currently, we are forced to kill the process in the event the
1491 * original mapper has unmapped pages from the child due to a failed
1492 * COW. Warn that such a situation has occured as it may not be obvious
1494 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
1496 "PID %d killed due to inadequate hugepage pool\n",
1501 mapping = vma->vm_file->f_mapping;
1502 idx = vma_pagecache_offset(vma, address);
1505 * Use page lock to guard against racing truncation
1506 * before we get page_table_lock.
1509 page = find_lock_page(mapping, idx);
1511 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
1514 page = alloc_huge_page(vma, address, 0);
1516 ret = -PTR_ERR(page);
1519 clear_huge_page(page, address);
1520 __SetPageUptodate(page);
1522 if (vma->vm_flags & VM_SHARED) {
1524 struct inode *inode = mapping->host;
1526 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
1534 spin_lock(&inode->i_lock);
1535 inode->i_blocks += BLOCKS_PER_HUGEPAGE;
1536 spin_unlock(&inode->i_lock);
1541 spin_lock(&mm->page_table_lock);
1542 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
1547 if (!huge_pte_none(huge_ptep_get(ptep)))
1550 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
1551 && (vma->vm_flags & VM_SHARED)));
1552 set_huge_pte_at(mm, address, ptep, new_pte);
1554 if (write_access && !(vma->vm_flags & VM_SHARED)) {
1555 /* Optimization, do the COW without a second fault */
1556 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
1559 spin_unlock(&mm->page_table_lock);
1565 spin_unlock(&mm->page_table_lock);
1571 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
1572 unsigned long address, int write_access)
1577 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
1579 ptep = huge_pte_alloc(mm, address);
1581 return VM_FAULT_OOM;
1584 * Serialize hugepage allocation and instantiation, so that we don't
1585 * get spurious allocation failures if two CPUs race to instantiate
1586 * the same page in the page cache.
1588 mutex_lock(&hugetlb_instantiation_mutex);
1589 entry = huge_ptep_get(ptep);
1590 if (huge_pte_none(entry)) {
1591 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
1592 mutex_unlock(&hugetlb_instantiation_mutex);
1598 spin_lock(&mm->page_table_lock);
1599 /* Check for a racing update before calling hugetlb_cow */
1600 if (likely(pte_same(entry, huge_ptep_get(ptep))))
1601 if (write_access && !pte_write(entry)) {
1603 page = hugetlbfs_pagecache_page(vma, address);
1604 ret = hugetlb_cow(mm, vma, address, ptep, entry, page);
1610 spin_unlock(&mm->page_table_lock);
1611 mutex_unlock(&hugetlb_instantiation_mutex);
1616 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
1617 struct page **pages, struct vm_area_struct **vmas,
1618 unsigned long *position, int *length, int i,
1621 unsigned long pfn_offset;
1622 unsigned long vaddr = *position;
1623 int remainder = *length;
1625 spin_lock(&mm->page_table_lock);
1626 while (vaddr < vma->vm_end && remainder) {
1631 * Some archs (sparc64, sh*) have multiple pte_ts to
1632 * each hugepage. We have to make * sure we get the
1633 * first, for the page indexing below to work.
1635 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
1637 if (!pte || huge_pte_none(huge_ptep_get(pte)) ||
1638 (write && !pte_write(huge_ptep_get(pte)))) {
1641 spin_unlock(&mm->page_table_lock);
1642 ret = hugetlb_fault(mm, vma, vaddr, write);
1643 spin_lock(&mm->page_table_lock);
1644 if (!(ret & VM_FAULT_ERROR))
1653 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
1654 page = pte_page(huge_ptep_get(pte));
1658 pages[i] = page + pfn_offset;
1668 if (vaddr < vma->vm_end && remainder &&
1669 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
1671 * We use pfn_offset to avoid touching the pageframes
1672 * of this compound page.
1677 spin_unlock(&mm->page_table_lock);
1678 *length = remainder;
1684 void hugetlb_change_protection(struct vm_area_struct *vma,
1685 unsigned long address, unsigned long end, pgprot_t newprot)
1687 struct mm_struct *mm = vma->vm_mm;
1688 unsigned long start = address;
1692 BUG_ON(address >= end);
1693 flush_cache_range(vma, address, end);
1695 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1696 spin_lock(&mm->page_table_lock);
1697 for (; address < end; address += HPAGE_SIZE) {
1698 ptep = huge_pte_offset(mm, address);
1701 if (huge_pmd_unshare(mm, &address, ptep))
1703 if (!huge_pte_none(huge_ptep_get(ptep))) {
1704 pte = huge_ptep_get_and_clear(mm, address, ptep);
1705 pte = pte_mkhuge(pte_modify(pte, newprot));
1706 set_huge_pte_at(mm, address, ptep, pte);
1709 spin_unlock(&mm->page_table_lock);
1710 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1712 flush_tlb_range(vma, start, end);
1715 int hugetlb_reserve_pages(struct inode *inode,
1717 struct vm_area_struct *vma)
1721 if (vma && vma->vm_flags & VM_NORESERVE)
1725 * Shared mappings base their reservation on the number of pages that
1726 * are already allocated on behalf of the file. Private mappings need
1727 * to reserve the full area even if read-only as mprotect() may be
1728 * called to make the mapping read-write. Assume !vma is a shm mapping
1730 if (!vma || vma->vm_flags & VM_SHARED)
1731 chg = region_chg(&inode->i_mapping->private_list, from, to);
1733 struct resv_map *resv_map = resv_map_alloc();
1739 set_vma_resv_map(vma, resv_map);
1740 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
1746 if (hugetlb_get_quota(inode->i_mapping, chg))
1748 ret = hugetlb_acct_memory(chg);
1750 hugetlb_put_quota(inode->i_mapping, chg);
1753 if (!vma || vma->vm_flags & VM_SHARED)
1754 region_add(&inode->i_mapping->private_list, from, to);
1758 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
1760 long chg = region_truncate(&inode->i_mapping->private_list, offset);
1762 spin_lock(&inode->i_lock);
1763 inode->i_blocks -= BLOCKS_PER_HUGEPAGE * freed;
1764 spin_unlock(&inode->i_lock);
1766 hugetlb_put_quota(inode->i_mapping, (chg - freed));
1767 hugetlb_acct_memory(-(chg - freed));