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/seq_file.h>
11 #include <linux/sysctl.h>
12 #include <linux/highmem.h>
13 #include <linux/mmu_notifier.h>
14 #include <linux/nodemask.h>
15 #include <linux/pagemap.h>
16 #include <linux/mempolicy.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/bootmem.h>
20 #include <linux/sysfs.h>
23 #include <asm/pgtable.h>
26 #include <linux/hugetlb.h>
29 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
30 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
31 unsigned long hugepages_treat_as_movable;
33 static int max_hstate;
34 unsigned int default_hstate_idx;
35 struct hstate hstates[HUGE_MAX_HSTATE];
37 __initdata LIST_HEAD(huge_boot_pages);
39 /* for command line parsing */
40 static struct hstate * __initdata parsed_hstate;
41 static unsigned long __initdata default_hstate_max_huge_pages;
42 static unsigned long __initdata default_hstate_size;
44 #define for_each_hstate(h) \
45 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
48 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
50 static DEFINE_SPINLOCK(hugetlb_lock);
53 * Region tracking -- allows tracking of reservations and instantiated pages
54 * across the pages in a mapping.
56 * The region data structures are protected by a combination of the mmap_sem
57 * and the hugetlb_instantion_mutex. To access or modify a region the caller
58 * must either hold the mmap_sem for write, or the mmap_sem for read and
59 * the hugetlb_instantiation mutex:
61 * down_write(&mm->mmap_sem);
63 * down_read(&mm->mmap_sem);
64 * mutex_lock(&hugetlb_instantiation_mutex);
67 struct list_head link;
72 static long region_add(struct list_head *head, long f, long t)
74 struct file_region *rg, *nrg, *trg;
76 /* Locate the region we are either in or before. */
77 list_for_each_entry(rg, head, link)
81 /* Round our left edge to the current segment if it encloses us. */
85 /* Check for and consume any regions we now overlap with. */
87 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
88 if (&rg->link == head)
93 /* If this area reaches higher then extend our area to
94 * include it completely. If this is not the first area
95 * which we intend to reuse, free it. */
108 static long region_chg(struct list_head *head, long f, long t)
110 struct file_region *rg, *nrg;
113 /* Locate the region we are before or in. */
114 list_for_each_entry(rg, head, link)
118 /* If we are below the current region then a new region is required.
119 * Subtle, allocate a new region at the position but make it zero
120 * size such that we can guarantee to record the reservation. */
121 if (&rg->link == head || t < rg->from) {
122 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
127 INIT_LIST_HEAD(&nrg->link);
128 list_add(&nrg->link, rg->link.prev);
133 /* Round our left edge to the current segment if it encloses us. */
138 /* Check for and consume any regions we now overlap with. */
139 list_for_each_entry(rg, rg->link.prev, link) {
140 if (&rg->link == head)
145 /* We overlap with this area, if it extends futher than
146 * us then we must extend ourselves. Account for its
147 * existing reservation. */
152 chg -= rg->to - rg->from;
157 static long region_truncate(struct list_head *head, long end)
159 struct file_region *rg, *trg;
162 /* Locate the region we are either in or before. */
163 list_for_each_entry(rg, head, link)
166 if (&rg->link == head)
169 /* If we are in the middle of a region then adjust it. */
170 if (end > rg->from) {
173 rg = list_entry(rg->link.next, typeof(*rg), link);
176 /* Drop any remaining regions. */
177 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
178 if (&rg->link == head)
180 chg += rg->to - rg->from;
187 static long region_count(struct list_head *head, long f, long t)
189 struct file_region *rg;
192 /* Locate each segment we overlap with, and count that overlap. */
193 list_for_each_entry(rg, head, link) {
202 seg_from = max(rg->from, f);
203 seg_to = min(rg->to, t);
205 chg += seg_to - seg_from;
212 * Convert the address within this vma to the page offset within
213 * the mapping, in pagecache page units; huge pages here.
215 static pgoff_t vma_hugecache_offset(struct hstate *h,
216 struct vm_area_struct *vma, unsigned long address)
218 return ((address - vma->vm_start) >> huge_page_shift(h)) +
219 (vma->vm_pgoff >> huge_page_order(h));
223 * Return the size of the pages allocated when backing a VMA. In the majority
224 * cases this will be same size as used by the page table entries.
226 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
228 struct hstate *hstate;
230 if (!is_vm_hugetlb_page(vma))
233 hstate = hstate_vma(vma);
235 return 1UL << (hstate->order + PAGE_SHIFT);
237 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
240 * Return the page size being used by the MMU to back a VMA. In the majority
241 * of cases, the page size used by the kernel matches the MMU size. On
242 * architectures where it differs, an architecture-specific version of this
243 * function is required.
245 #ifndef vma_mmu_pagesize
246 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
248 return vma_kernel_pagesize(vma);
253 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
254 * bits of the reservation map pointer, which are always clear due to
257 #define HPAGE_RESV_OWNER (1UL << 0)
258 #define HPAGE_RESV_UNMAPPED (1UL << 1)
259 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
262 * These helpers are used to track how many pages are reserved for
263 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
264 * is guaranteed to have their future faults succeed.
266 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
267 * the reserve counters are updated with the hugetlb_lock held. It is safe
268 * to reset the VMA at fork() time as it is not in use yet and there is no
269 * chance of the global counters getting corrupted as a result of the values.
271 * The private mapping reservation is represented in a subtly different
272 * manner to a shared mapping. A shared mapping has a region map associated
273 * with the underlying file, this region map represents the backing file
274 * pages which have ever had a reservation assigned which this persists even
275 * after the page is instantiated. A private mapping has a region map
276 * associated with the original mmap which is attached to all VMAs which
277 * reference it, this region map represents those offsets which have consumed
278 * reservation ie. where pages have been instantiated.
280 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
282 return (unsigned long)vma->vm_private_data;
285 static void set_vma_private_data(struct vm_area_struct *vma,
288 vma->vm_private_data = (void *)value;
293 struct list_head regions;
296 static struct resv_map *resv_map_alloc(void)
298 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
302 kref_init(&resv_map->refs);
303 INIT_LIST_HEAD(&resv_map->regions);
308 static void resv_map_release(struct kref *ref)
310 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
312 /* Clear out any active regions before we release the map. */
313 region_truncate(&resv_map->regions, 0);
317 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
319 VM_BUG_ON(!is_vm_hugetlb_page(vma));
320 if (!(vma->vm_flags & VM_MAYSHARE))
321 return (struct resv_map *)(get_vma_private_data(vma) &
326 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
328 VM_BUG_ON(!is_vm_hugetlb_page(vma));
329 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
331 set_vma_private_data(vma, (get_vma_private_data(vma) &
332 HPAGE_RESV_MASK) | (unsigned long)map);
335 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
337 VM_BUG_ON(!is_vm_hugetlb_page(vma));
338 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
340 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
343 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
345 VM_BUG_ON(!is_vm_hugetlb_page(vma));
347 return (get_vma_private_data(vma) & flag) != 0;
350 /* Decrement the reserved pages in the hugepage pool by one */
351 static void decrement_hugepage_resv_vma(struct hstate *h,
352 struct vm_area_struct *vma)
354 if (vma->vm_flags & VM_NORESERVE)
357 if (vma->vm_flags & VM_MAYSHARE) {
358 /* Shared mappings always use reserves */
359 h->resv_huge_pages--;
360 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
362 * Only the process that called mmap() has reserves for
365 h->resv_huge_pages--;
369 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
370 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
372 VM_BUG_ON(!is_vm_hugetlb_page(vma));
373 if (!(vma->vm_flags & VM_MAYSHARE))
374 vma->vm_private_data = (void *)0;
377 /* Returns true if the VMA has associated reserve pages */
378 static int vma_has_reserves(struct vm_area_struct *vma)
380 if (vma->vm_flags & VM_MAYSHARE)
382 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
387 static void clear_gigantic_page(struct page *page,
388 unsigned long addr, unsigned long sz)
391 struct page *p = page;
394 for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) {
396 clear_user_highpage(p, addr + i * PAGE_SIZE);
399 static void clear_huge_page(struct page *page,
400 unsigned long addr, unsigned long sz)
404 if (unlikely(sz > MAX_ORDER_NR_PAGES)) {
405 clear_gigantic_page(page, addr, sz);
410 for (i = 0; i < sz/PAGE_SIZE; i++) {
412 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
416 static void copy_gigantic_page(struct page *dst, struct page *src,
417 unsigned long addr, struct vm_area_struct *vma)
420 struct hstate *h = hstate_vma(vma);
421 struct page *dst_base = dst;
422 struct page *src_base = src;
424 for (i = 0; i < pages_per_huge_page(h); ) {
426 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
429 dst = mem_map_next(dst, dst_base, i);
430 src = mem_map_next(src, src_base, i);
433 static void copy_huge_page(struct page *dst, struct page *src,
434 unsigned long addr, struct vm_area_struct *vma)
437 struct hstate *h = hstate_vma(vma);
439 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
440 copy_gigantic_page(dst, src, addr, vma);
445 for (i = 0; i < pages_per_huge_page(h); i++) {
447 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
451 static void enqueue_huge_page(struct hstate *h, struct page *page)
453 int nid = page_to_nid(page);
454 list_add(&page->lru, &h->hugepage_freelists[nid]);
455 h->free_huge_pages++;
456 h->free_huge_pages_node[nid]++;
459 static struct page *dequeue_huge_page_vma(struct hstate *h,
460 struct vm_area_struct *vma,
461 unsigned long address, int avoid_reserve)
464 struct page *page = NULL;
465 struct mempolicy *mpol;
466 nodemask_t *nodemask;
467 struct zonelist *zonelist = huge_zonelist(vma, address,
468 htlb_alloc_mask, &mpol, &nodemask);
473 * A child process with MAP_PRIVATE mappings created by their parent
474 * have no page reserves. This check ensures that reservations are
475 * not "stolen". The child may still get SIGKILLed
477 if (!vma_has_reserves(vma) &&
478 h->free_huge_pages - h->resv_huge_pages == 0)
481 /* If reserves cannot be used, ensure enough pages are in the pool */
482 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
485 for_each_zone_zonelist_nodemask(zone, z, zonelist,
486 MAX_NR_ZONES - 1, nodemask) {
487 nid = zone_to_nid(zone);
488 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
489 !list_empty(&h->hugepage_freelists[nid])) {
490 page = list_entry(h->hugepage_freelists[nid].next,
492 list_del(&page->lru);
493 h->free_huge_pages--;
494 h->free_huge_pages_node[nid]--;
497 decrement_hugepage_resv_vma(h, vma);
506 static void update_and_free_page(struct hstate *h, struct page *page)
510 VM_BUG_ON(h->order >= MAX_ORDER);
513 h->nr_huge_pages_node[page_to_nid(page)]--;
514 for (i = 0; i < pages_per_huge_page(h); i++) {
515 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
516 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
517 1 << PG_private | 1<< PG_writeback);
519 set_compound_page_dtor(page, NULL);
520 set_page_refcounted(page);
521 arch_release_hugepage(page);
522 __free_pages(page, huge_page_order(h));
525 struct hstate *size_to_hstate(unsigned long size)
530 if (huge_page_size(h) == size)
536 static void free_huge_page(struct page *page)
539 * Can't pass hstate in here because it is called from the
540 * compound page destructor.
542 struct hstate *h = page_hstate(page);
543 int nid = page_to_nid(page);
544 struct address_space *mapping;
546 mapping = (struct address_space *) page_private(page);
547 set_page_private(page, 0);
548 BUG_ON(page_count(page));
549 INIT_LIST_HEAD(&page->lru);
551 spin_lock(&hugetlb_lock);
552 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
553 update_and_free_page(h, page);
554 h->surplus_huge_pages--;
555 h->surplus_huge_pages_node[nid]--;
557 enqueue_huge_page(h, page);
559 spin_unlock(&hugetlb_lock);
561 hugetlb_put_quota(mapping, 1);
564 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
566 set_compound_page_dtor(page, free_huge_page);
567 spin_lock(&hugetlb_lock);
569 h->nr_huge_pages_node[nid]++;
570 spin_unlock(&hugetlb_lock);
571 put_page(page); /* free it into the hugepage allocator */
574 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
577 int nr_pages = 1 << order;
578 struct page *p = page + 1;
580 /* we rely on prep_new_huge_page to set the destructor */
581 set_compound_order(page, order);
583 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
585 p->first_page = page;
589 int PageHuge(struct page *page)
591 compound_page_dtor *dtor;
593 if (!PageCompound(page))
596 page = compound_head(page);
597 dtor = get_compound_page_dtor(page);
599 return dtor == free_huge_page;
602 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
606 if (h->order >= MAX_ORDER)
609 page = alloc_pages_exact_node(nid,
610 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
611 __GFP_REPEAT|__GFP_NOWARN,
614 if (arch_prepare_hugepage(page)) {
615 __free_pages(page, huge_page_order(h));
618 prep_new_huge_page(h, page, nid);
625 * common helper functions for hstate_next_node_to_{alloc|free}.
626 * We may have allocated or freed a huge page based on a different
627 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
628 * be outside of *nodes_allowed. Ensure that we use an allowed
629 * node for alloc or free.
631 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
633 nid = next_node(nid, *nodes_allowed);
634 if (nid == MAX_NUMNODES)
635 nid = first_node(*nodes_allowed);
636 VM_BUG_ON(nid >= MAX_NUMNODES);
641 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
643 if (!node_isset(nid, *nodes_allowed))
644 nid = next_node_allowed(nid, nodes_allowed);
649 * returns the previously saved node ["this node"] from which to
650 * allocate a persistent huge page for the pool and advance the
651 * next node from which to allocate, handling wrap at end of node
654 static int hstate_next_node_to_alloc(struct hstate *h,
655 nodemask_t *nodes_allowed)
659 VM_BUG_ON(!nodes_allowed);
661 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
662 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
667 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
674 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
675 next_nid = start_nid;
678 page = alloc_fresh_huge_page_node(h, next_nid);
683 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
684 } while (next_nid != start_nid);
687 count_vm_event(HTLB_BUDDY_PGALLOC);
689 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
695 * helper for free_pool_huge_page() - return the previously saved
696 * node ["this node"] from which to free a huge page. Advance the
697 * next node id whether or not we find a free huge page to free so
698 * that the next attempt to free addresses the next node.
700 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
704 VM_BUG_ON(!nodes_allowed);
706 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
707 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
713 * Free huge page from pool from next node to free.
714 * Attempt to keep persistent huge pages more or less
715 * balanced over allowed nodes.
716 * Called with hugetlb_lock locked.
718 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
725 start_nid = hstate_next_node_to_free(h, nodes_allowed);
726 next_nid = start_nid;
730 * If we're returning unused surplus pages, only examine
731 * nodes with surplus pages.
733 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
734 !list_empty(&h->hugepage_freelists[next_nid])) {
736 list_entry(h->hugepage_freelists[next_nid].next,
738 list_del(&page->lru);
739 h->free_huge_pages--;
740 h->free_huge_pages_node[next_nid]--;
742 h->surplus_huge_pages--;
743 h->surplus_huge_pages_node[next_nid]--;
745 update_and_free_page(h, page);
749 next_nid = hstate_next_node_to_free(h, nodes_allowed);
750 } while (next_nid != start_nid);
755 static struct page *alloc_buddy_huge_page(struct hstate *h,
756 struct vm_area_struct *vma, unsigned long address)
761 if (h->order >= MAX_ORDER)
765 * Assume we will successfully allocate the surplus page to
766 * prevent racing processes from causing the surplus to exceed
769 * This however introduces a different race, where a process B
770 * tries to grow the static hugepage pool while alloc_pages() is
771 * called by process A. B will only examine the per-node
772 * counters in determining if surplus huge pages can be
773 * converted to normal huge pages in adjust_pool_surplus(). A
774 * won't be able to increment the per-node counter, until the
775 * lock is dropped by B, but B doesn't drop hugetlb_lock until
776 * no more huge pages can be converted from surplus to normal
777 * state (and doesn't try to convert again). Thus, we have a
778 * case where a surplus huge page exists, the pool is grown, and
779 * the surplus huge page still exists after, even though it
780 * should just have been converted to a normal huge page. This
781 * does not leak memory, though, as the hugepage will be freed
782 * once it is out of use. It also does not allow the counters to
783 * go out of whack in adjust_pool_surplus() as we don't modify
784 * the node values until we've gotten the hugepage and only the
785 * per-node value is checked there.
787 spin_lock(&hugetlb_lock);
788 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
789 spin_unlock(&hugetlb_lock);
793 h->surplus_huge_pages++;
795 spin_unlock(&hugetlb_lock);
797 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
798 __GFP_REPEAT|__GFP_NOWARN,
801 if (page && arch_prepare_hugepage(page)) {
802 __free_pages(page, huge_page_order(h));
806 spin_lock(&hugetlb_lock);
809 * This page is now managed by the hugetlb allocator and has
810 * no users -- drop the buddy allocator's reference.
812 put_page_testzero(page);
813 VM_BUG_ON(page_count(page));
814 nid = page_to_nid(page);
815 set_compound_page_dtor(page, free_huge_page);
817 * We incremented the global counters already
819 h->nr_huge_pages_node[nid]++;
820 h->surplus_huge_pages_node[nid]++;
821 __count_vm_event(HTLB_BUDDY_PGALLOC);
824 h->surplus_huge_pages--;
825 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
827 spin_unlock(&hugetlb_lock);
833 * Increase the hugetlb pool such that it can accomodate a reservation
836 static int gather_surplus_pages(struct hstate *h, int delta)
838 struct list_head surplus_list;
839 struct page *page, *tmp;
841 int needed, allocated;
843 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
845 h->resv_huge_pages += delta;
850 INIT_LIST_HEAD(&surplus_list);
854 spin_unlock(&hugetlb_lock);
855 for (i = 0; i < needed; i++) {
856 page = alloc_buddy_huge_page(h, NULL, 0);
859 * We were not able to allocate enough pages to
860 * satisfy the entire reservation so we free what
861 * we've allocated so far.
863 spin_lock(&hugetlb_lock);
868 list_add(&page->lru, &surplus_list);
873 * After retaking hugetlb_lock, we need to recalculate 'needed'
874 * because either resv_huge_pages or free_huge_pages may have changed.
876 spin_lock(&hugetlb_lock);
877 needed = (h->resv_huge_pages + delta) -
878 (h->free_huge_pages + allocated);
883 * The surplus_list now contains _at_least_ the number of extra pages
884 * needed to accomodate the reservation. Add the appropriate number
885 * of pages to the hugetlb pool and free the extras back to the buddy
886 * allocator. Commit the entire reservation here to prevent another
887 * process from stealing the pages as they are added to the pool but
888 * before they are reserved.
891 h->resv_huge_pages += delta;
894 /* Free the needed pages to the hugetlb pool */
895 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
898 list_del(&page->lru);
899 enqueue_huge_page(h, page);
902 /* Free unnecessary surplus pages to the buddy allocator */
903 if (!list_empty(&surplus_list)) {
904 spin_unlock(&hugetlb_lock);
905 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
906 list_del(&page->lru);
908 * The page has a reference count of zero already, so
909 * call free_huge_page directly instead of using
910 * put_page. This must be done with hugetlb_lock
911 * unlocked which is safe because free_huge_page takes
912 * hugetlb_lock before deciding how to free the page.
914 free_huge_page(page);
916 spin_lock(&hugetlb_lock);
923 * When releasing a hugetlb pool reservation, any surplus pages that were
924 * allocated to satisfy the reservation must be explicitly freed if they were
926 * Called with hugetlb_lock held.
928 static void return_unused_surplus_pages(struct hstate *h,
929 unsigned long unused_resv_pages)
931 unsigned long nr_pages;
933 /* Uncommit the reservation */
934 h->resv_huge_pages -= unused_resv_pages;
936 /* Cannot return gigantic pages currently */
937 if (h->order >= MAX_ORDER)
940 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
943 * We want to release as many surplus pages as possible, spread
944 * evenly across all nodes. Iterate across all nodes until we
945 * can no longer free unreserved surplus pages. This occurs when
946 * the nodes with surplus pages have no free pages.
947 * free_pool_huge_page() will balance the the frees across the
948 * on-line nodes for us and will handle the hstate accounting.
951 if (!free_pool_huge_page(h, &node_online_map, 1))
957 * Determine if the huge page at addr within the vma has an associated
958 * reservation. Where it does not we will need to logically increase
959 * reservation and actually increase quota before an allocation can occur.
960 * Where any new reservation would be required the reservation change is
961 * prepared, but not committed. Once the page has been quota'd allocated
962 * an instantiated the change should be committed via vma_commit_reservation.
963 * No action is required on failure.
965 static long vma_needs_reservation(struct hstate *h,
966 struct vm_area_struct *vma, unsigned long addr)
968 struct address_space *mapping = vma->vm_file->f_mapping;
969 struct inode *inode = mapping->host;
971 if (vma->vm_flags & VM_MAYSHARE) {
972 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
973 return region_chg(&inode->i_mapping->private_list,
976 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
981 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
982 struct resv_map *reservations = vma_resv_map(vma);
984 err = region_chg(&reservations->regions, idx, idx + 1);
990 static void vma_commit_reservation(struct hstate *h,
991 struct vm_area_struct *vma, unsigned long addr)
993 struct address_space *mapping = vma->vm_file->f_mapping;
994 struct inode *inode = mapping->host;
996 if (vma->vm_flags & VM_MAYSHARE) {
997 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
998 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1000 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1001 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1002 struct resv_map *reservations = vma_resv_map(vma);
1004 /* Mark this page used in the map. */
1005 region_add(&reservations->regions, idx, idx + 1);
1009 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1010 unsigned long addr, int avoid_reserve)
1012 struct hstate *h = hstate_vma(vma);
1014 struct address_space *mapping = vma->vm_file->f_mapping;
1015 struct inode *inode = mapping->host;
1019 * Processes that did not create the mapping will have no reserves and
1020 * will not have accounted against quota. Check that the quota can be
1021 * made before satisfying the allocation
1022 * MAP_NORESERVE mappings may also need pages and quota allocated
1023 * if no reserve mapping overlaps.
1025 chg = vma_needs_reservation(h, vma, addr);
1027 return ERR_PTR(chg);
1029 if (hugetlb_get_quota(inode->i_mapping, chg))
1030 return ERR_PTR(-ENOSPC);
1032 spin_lock(&hugetlb_lock);
1033 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1034 spin_unlock(&hugetlb_lock);
1037 page = alloc_buddy_huge_page(h, vma, addr);
1039 hugetlb_put_quota(inode->i_mapping, chg);
1040 return ERR_PTR(-VM_FAULT_OOM);
1044 set_page_refcounted(page);
1045 set_page_private(page, (unsigned long) mapping);
1047 vma_commit_reservation(h, vma, addr);
1052 int __weak alloc_bootmem_huge_page(struct hstate *h)
1054 struct huge_bootmem_page *m;
1055 int nr_nodes = nodes_weight(node_online_map);
1060 addr = __alloc_bootmem_node_nopanic(
1061 NODE_DATA(hstate_next_node_to_alloc(h,
1063 huge_page_size(h), huge_page_size(h), 0);
1067 * Use the beginning of the huge page to store the
1068 * huge_bootmem_page struct (until gather_bootmem
1069 * puts them into the mem_map).
1079 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1080 /* Put them into a private list first because mem_map is not up yet */
1081 list_add(&m->list, &huge_boot_pages);
1086 static void prep_compound_huge_page(struct page *page, int order)
1088 if (unlikely(order > (MAX_ORDER - 1)))
1089 prep_compound_gigantic_page(page, order);
1091 prep_compound_page(page, order);
1094 /* Put bootmem huge pages into the standard lists after mem_map is up */
1095 static void __init gather_bootmem_prealloc(void)
1097 struct huge_bootmem_page *m;
1099 list_for_each_entry(m, &huge_boot_pages, list) {
1100 struct page *page = virt_to_page(m);
1101 struct hstate *h = m->hstate;
1102 __ClearPageReserved(page);
1103 WARN_ON(page_count(page) != 1);
1104 prep_compound_huge_page(page, h->order);
1105 prep_new_huge_page(h, page, page_to_nid(page));
1109 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1113 for (i = 0; i < h->max_huge_pages; ++i) {
1114 if (h->order >= MAX_ORDER) {
1115 if (!alloc_bootmem_huge_page(h))
1117 } else if (!alloc_fresh_huge_page(h, &node_online_map))
1120 h->max_huge_pages = i;
1123 static void __init hugetlb_init_hstates(void)
1127 for_each_hstate(h) {
1128 /* oversize hugepages were init'ed in early boot */
1129 if (h->order < MAX_ORDER)
1130 hugetlb_hstate_alloc_pages(h);
1134 static char * __init memfmt(char *buf, unsigned long n)
1136 if (n >= (1UL << 30))
1137 sprintf(buf, "%lu GB", n >> 30);
1138 else if (n >= (1UL << 20))
1139 sprintf(buf, "%lu MB", n >> 20);
1141 sprintf(buf, "%lu KB", n >> 10);
1145 static void __init report_hugepages(void)
1149 for_each_hstate(h) {
1151 printk(KERN_INFO "HugeTLB registered %s page size, "
1152 "pre-allocated %ld pages\n",
1153 memfmt(buf, huge_page_size(h)),
1154 h->free_huge_pages);
1158 #ifdef CONFIG_HIGHMEM
1159 static void try_to_free_low(struct hstate *h, unsigned long count,
1160 nodemask_t *nodes_allowed)
1164 if (h->order >= MAX_ORDER)
1167 for_each_node_mask(i, *nodes_allowed) {
1168 struct page *page, *next;
1169 struct list_head *freel = &h->hugepage_freelists[i];
1170 list_for_each_entry_safe(page, next, freel, lru) {
1171 if (count >= h->nr_huge_pages)
1173 if (PageHighMem(page))
1175 list_del(&page->lru);
1176 update_and_free_page(h, page);
1177 h->free_huge_pages--;
1178 h->free_huge_pages_node[page_to_nid(page)]--;
1183 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1184 nodemask_t *nodes_allowed)
1190 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1191 * balanced by operating on them in a round-robin fashion.
1192 * Returns 1 if an adjustment was made.
1194 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1197 int start_nid, next_nid;
1200 VM_BUG_ON(delta != -1 && delta != 1);
1203 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1205 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1206 next_nid = start_nid;
1212 * To shrink on this node, there must be a surplus page
1214 if (!h->surplus_huge_pages_node[nid]) {
1215 next_nid = hstate_next_node_to_alloc(h,
1222 * Surplus cannot exceed the total number of pages
1224 if (h->surplus_huge_pages_node[nid] >=
1225 h->nr_huge_pages_node[nid]) {
1226 next_nid = hstate_next_node_to_free(h,
1232 h->surplus_huge_pages += delta;
1233 h->surplus_huge_pages_node[nid] += delta;
1236 } while (next_nid != start_nid);
1241 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1242 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1243 nodemask_t *nodes_allowed)
1245 unsigned long min_count, ret;
1247 if (h->order >= MAX_ORDER)
1248 return h->max_huge_pages;
1251 * Increase the pool size
1252 * First take pages out of surplus state. Then make up the
1253 * remaining difference by allocating fresh huge pages.
1255 * We might race with alloc_buddy_huge_page() here and be unable
1256 * to convert a surplus huge page to a normal huge page. That is
1257 * not critical, though, it just means the overall size of the
1258 * pool might be one hugepage larger than it needs to be, but
1259 * within all the constraints specified by the sysctls.
1261 spin_lock(&hugetlb_lock);
1262 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1263 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1267 while (count > persistent_huge_pages(h)) {
1269 * If this allocation races such that we no longer need the
1270 * page, free_huge_page will handle it by freeing the page
1271 * and reducing the surplus.
1273 spin_unlock(&hugetlb_lock);
1274 ret = alloc_fresh_huge_page(h, nodes_allowed);
1275 spin_lock(&hugetlb_lock);
1282 * Decrease the pool size
1283 * First return free pages to the buddy allocator (being careful
1284 * to keep enough around to satisfy reservations). Then place
1285 * pages into surplus state as needed so the pool will shrink
1286 * to the desired size as pages become free.
1288 * By placing pages into the surplus state independent of the
1289 * overcommit value, we are allowing the surplus pool size to
1290 * exceed overcommit. There are few sane options here. Since
1291 * alloc_buddy_huge_page() is checking the global counter,
1292 * though, we'll note that we're not allowed to exceed surplus
1293 * and won't grow the pool anywhere else. Not until one of the
1294 * sysctls are changed, or the surplus pages go out of use.
1296 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1297 min_count = max(count, min_count);
1298 try_to_free_low(h, min_count, nodes_allowed);
1299 while (min_count < persistent_huge_pages(h)) {
1300 if (!free_pool_huge_page(h, nodes_allowed, 0))
1303 while (count < persistent_huge_pages(h)) {
1304 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1308 ret = persistent_huge_pages(h);
1309 spin_unlock(&hugetlb_lock);
1313 #define HSTATE_ATTR_RO(_name) \
1314 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1316 #define HSTATE_ATTR(_name) \
1317 static struct kobj_attribute _name##_attr = \
1318 __ATTR(_name, 0644, _name##_show, _name##_store)
1320 static struct kobject *hugepages_kobj;
1321 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1323 static struct hstate *kobj_to_hstate(struct kobject *kobj)
1326 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1327 if (hstate_kobjs[i] == kobj)
1333 static ssize_t nr_hugepages_show(struct kobject *kobj,
1334 struct kobj_attribute *attr, char *buf)
1336 struct hstate *h = kobj_to_hstate(kobj);
1337 return sprintf(buf, "%lu\n", h->nr_huge_pages);
1339 static ssize_t nr_hugepages_store(struct kobject *kobj,
1340 struct kobj_attribute *attr, const char *buf, size_t count)
1343 unsigned long input;
1344 struct hstate *h = kobj_to_hstate(kobj);
1346 err = strict_strtoul(buf, 10, &input);
1350 h->max_huge_pages = set_max_huge_pages(h, input, &node_online_map);
1354 HSTATE_ATTR(nr_hugepages);
1356 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1357 struct kobj_attribute *attr, char *buf)
1359 struct hstate *h = kobj_to_hstate(kobj);
1360 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1362 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1363 struct kobj_attribute *attr, const char *buf, size_t count)
1366 unsigned long input;
1367 struct hstate *h = kobj_to_hstate(kobj);
1369 err = strict_strtoul(buf, 10, &input);
1373 spin_lock(&hugetlb_lock);
1374 h->nr_overcommit_huge_pages = input;
1375 spin_unlock(&hugetlb_lock);
1379 HSTATE_ATTR(nr_overcommit_hugepages);
1381 static ssize_t free_hugepages_show(struct kobject *kobj,
1382 struct kobj_attribute *attr, char *buf)
1384 struct hstate *h = kobj_to_hstate(kobj);
1385 return sprintf(buf, "%lu\n", h->free_huge_pages);
1387 HSTATE_ATTR_RO(free_hugepages);
1389 static ssize_t resv_hugepages_show(struct kobject *kobj,
1390 struct kobj_attribute *attr, char *buf)
1392 struct hstate *h = kobj_to_hstate(kobj);
1393 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1395 HSTATE_ATTR_RO(resv_hugepages);
1397 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1398 struct kobj_attribute *attr, char *buf)
1400 struct hstate *h = kobj_to_hstate(kobj);
1401 return sprintf(buf, "%lu\n", h->surplus_huge_pages);
1403 HSTATE_ATTR_RO(surplus_hugepages);
1405 static struct attribute *hstate_attrs[] = {
1406 &nr_hugepages_attr.attr,
1407 &nr_overcommit_hugepages_attr.attr,
1408 &free_hugepages_attr.attr,
1409 &resv_hugepages_attr.attr,
1410 &surplus_hugepages_attr.attr,
1414 static struct attribute_group hstate_attr_group = {
1415 .attrs = hstate_attrs,
1418 static int __init hugetlb_sysfs_add_hstate(struct hstate *h)
1422 hstate_kobjs[h - hstates] = kobject_create_and_add(h->name,
1424 if (!hstate_kobjs[h - hstates])
1427 retval = sysfs_create_group(hstate_kobjs[h - hstates],
1428 &hstate_attr_group);
1430 kobject_put(hstate_kobjs[h - hstates]);
1435 static void __init hugetlb_sysfs_init(void)
1440 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1441 if (!hugepages_kobj)
1444 for_each_hstate(h) {
1445 err = hugetlb_sysfs_add_hstate(h);
1447 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1452 static void __exit hugetlb_exit(void)
1456 for_each_hstate(h) {
1457 kobject_put(hstate_kobjs[h - hstates]);
1460 kobject_put(hugepages_kobj);
1462 module_exit(hugetlb_exit);
1464 static int __init hugetlb_init(void)
1466 /* Some platform decide whether they support huge pages at boot
1467 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1468 * there is no such support
1470 if (HPAGE_SHIFT == 0)
1473 if (!size_to_hstate(default_hstate_size)) {
1474 default_hstate_size = HPAGE_SIZE;
1475 if (!size_to_hstate(default_hstate_size))
1476 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1478 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1479 if (default_hstate_max_huge_pages)
1480 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1482 hugetlb_init_hstates();
1484 gather_bootmem_prealloc();
1488 hugetlb_sysfs_init();
1492 module_init(hugetlb_init);
1494 /* Should be called on processing a hugepagesz=... option */
1495 void __init hugetlb_add_hstate(unsigned order)
1500 if (size_to_hstate(PAGE_SIZE << order)) {
1501 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1504 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1506 h = &hstates[max_hstate++];
1508 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1509 h->nr_huge_pages = 0;
1510 h->free_huge_pages = 0;
1511 for (i = 0; i < MAX_NUMNODES; ++i)
1512 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1513 h->next_nid_to_alloc = first_node(node_online_map);
1514 h->next_nid_to_free = first_node(node_online_map);
1515 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1516 huge_page_size(h)/1024);
1521 static int __init hugetlb_nrpages_setup(char *s)
1524 static unsigned long *last_mhp;
1527 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1528 * so this hugepages= parameter goes to the "default hstate".
1531 mhp = &default_hstate_max_huge_pages;
1533 mhp = &parsed_hstate->max_huge_pages;
1535 if (mhp == last_mhp) {
1536 printk(KERN_WARNING "hugepages= specified twice without "
1537 "interleaving hugepagesz=, ignoring\n");
1541 if (sscanf(s, "%lu", mhp) <= 0)
1545 * Global state is always initialized later in hugetlb_init.
1546 * But we need to allocate >= MAX_ORDER hstates here early to still
1547 * use the bootmem allocator.
1549 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1550 hugetlb_hstate_alloc_pages(parsed_hstate);
1556 __setup("hugepages=", hugetlb_nrpages_setup);
1558 static int __init hugetlb_default_setup(char *s)
1560 default_hstate_size = memparse(s, &s);
1563 __setup("default_hugepagesz=", hugetlb_default_setup);
1565 static unsigned int cpuset_mems_nr(unsigned int *array)
1568 unsigned int nr = 0;
1570 for_each_node_mask(node, cpuset_current_mems_allowed)
1576 #ifdef CONFIG_SYSCTL
1577 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1578 void __user *buffer,
1579 size_t *length, loff_t *ppos)
1581 struct hstate *h = &default_hstate;
1585 tmp = h->max_huge_pages;
1588 table->maxlen = sizeof(unsigned long);
1589 proc_doulongvec_minmax(table, write, buffer, length, ppos);
1592 h->max_huge_pages = set_max_huge_pages(h, tmp,
1598 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1599 void __user *buffer,
1600 size_t *length, loff_t *ppos)
1602 proc_dointvec(table, write, buffer, length, ppos);
1603 if (hugepages_treat_as_movable)
1604 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1606 htlb_alloc_mask = GFP_HIGHUSER;
1610 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1611 void __user *buffer,
1612 size_t *length, loff_t *ppos)
1614 struct hstate *h = &default_hstate;
1618 tmp = h->nr_overcommit_huge_pages;
1621 table->maxlen = sizeof(unsigned long);
1622 proc_doulongvec_minmax(table, write, buffer, length, ppos);
1625 spin_lock(&hugetlb_lock);
1626 h->nr_overcommit_huge_pages = tmp;
1627 spin_unlock(&hugetlb_lock);
1633 #endif /* CONFIG_SYSCTL */
1635 void hugetlb_report_meminfo(struct seq_file *m)
1637 struct hstate *h = &default_hstate;
1639 "HugePages_Total: %5lu\n"
1640 "HugePages_Free: %5lu\n"
1641 "HugePages_Rsvd: %5lu\n"
1642 "HugePages_Surp: %5lu\n"
1643 "Hugepagesize: %8lu kB\n",
1647 h->surplus_huge_pages,
1648 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1651 int hugetlb_report_node_meminfo(int nid, char *buf)
1653 struct hstate *h = &default_hstate;
1655 "Node %d HugePages_Total: %5u\n"
1656 "Node %d HugePages_Free: %5u\n"
1657 "Node %d HugePages_Surp: %5u\n",
1658 nid, h->nr_huge_pages_node[nid],
1659 nid, h->free_huge_pages_node[nid],
1660 nid, h->surplus_huge_pages_node[nid]);
1663 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1664 unsigned long hugetlb_total_pages(void)
1666 struct hstate *h = &default_hstate;
1667 return h->nr_huge_pages * pages_per_huge_page(h);
1670 static int hugetlb_acct_memory(struct hstate *h, long delta)
1674 spin_lock(&hugetlb_lock);
1676 * When cpuset is configured, it breaks the strict hugetlb page
1677 * reservation as the accounting is done on a global variable. Such
1678 * reservation is completely rubbish in the presence of cpuset because
1679 * the reservation is not checked against page availability for the
1680 * current cpuset. Application can still potentially OOM'ed by kernel
1681 * with lack of free htlb page in cpuset that the task is in.
1682 * Attempt to enforce strict accounting with cpuset is almost
1683 * impossible (or too ugly) because cpuset is too fluid that
1684 * task or memory node can be dynamically moved between cpusets.
1686 * The change of semantics for shared hugetlb mapping with cpuset is
1687 * undesirable. However, in order to preserve some of the semantics,
1688 * we fall back to check against current free page availability as
1689 * a best attempt and hopefully to minimize the impact of changing
1690 * semantics that cpuset has.
1693 if (gather_surplus_pages(h, delta) < 0)
1696 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1697 return_unused_surplus_pages(h, delta);
1704 return_unused_surplus_pages(h, (unsigned long) -delta);
1707 spin_unlock(&hugetlb_lock);
1711 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
1713 struct resv_map *reservations = vma_resv_map(vma);
1716 * This new VMA should share its siblings reservation map if present.
1717 * The VMA will only ever have a valid reservation map pointer where
1718 * it is being copied for another still existing VMA. As that VMA
1719 * has a reference to the reservation map it cannot dissappear until
1720 * after this open call completes. It is therefore safe to take a
1721 * new reference here without additional locking.
1724 kref_get(&reservations->refs);
1727 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
1729 struct hstate *h = hstate_vma(vma);
1730 struct resv_map *reservations = vma_resv_map(vma);
1731 unsigned long reserve;
1732 unsigned long start;
1736 start = vma_hugecache_offset(h, vma, vma->vm_start);
1737 end = vma_hugecache_offset(h, vma, vma->vm_end);
1739 reserve = (end - start) -
1740 region_count(&reservations->regions, start, end);
1742 kref_put(&reservations->refs, resv_map_release);
1745 hugetlb_acct_memory(h, -reserve);
1746 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
1752 * We cannot handle pagefaults against hugetlb pages at all. They cause
1753 * handle_mm_fault() to try to instantiate regular-sized pages in the
1754 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
1757 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1763 const struct vm_operations_struct hugetlb_vm_ops = {
1764 .fault = hugetlb_vm_op_fault,
1765 .open = hugetlb_vm_op_open,
1766 .close = hugetlb_vm_op_close,
1769 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
1776 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1778 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
1780 entry = pte_mkyoung(entry);
1781 entry = pte_mkhuge(entry);
1786 static void set_huge_ptep_writable(struct vm_area_struct *vma,
1787 unsigned long address, pte_t *ptep)
1791 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
1792 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
1793 update_mmu_cache(vma, address, entry);
1798 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
1799 struct vm_area_struct *vma)
1801 pte_t *src_pte, *dst_pte, entry;
1802 struct page *ptepage;
1805 struct hstate *h = hstate_vma(vma);
1806 unsigned long sz = huge_page_size(h);
1808 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
1810 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
1811 src_pte = huge_pte_offset(src, addr);
1814 dst_pte = huge_pte_alloc(dst, addr, sz);
1818 /* If the pagetables are shared don't copy or take references */
1819 if (dst_pte == src_pte)
1822 spin_lock(&dst->page_table_lock);
1823 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
1824 if (!huge_pte_none(huge_ptep_get(src_pte))) {
1826 huge_ptep_set_wrprotect(src, addr, src_pte);
1827 entry = huge_ptep_get(src_pte);
1828 ptepage = pte_page(entry);
1830 set_huge_pte_at(dst, addr, dst_pte, entry);
1832 spin_unlock(&src->page_table_lock);
1833 spin_unlock(&dst->page_table_lock);
1841 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1842 unsigned long end, struct page *ref_page)
1844 struct mm_struct *mm = vma->vm_mm;
1845 unsigned long address;
1850 struct hstate *h = hstate_vma(vma);
1851 unsigned long sz = huge_page_size(h);
1854 * A page gathering list, protected by per file i_mmap_lock. The
1855 * lock is used to avoid list corruption from multiple unmapping
1856 * of the same page since we are using page->lru.
1858 LIST_HEAD(page_list);
1860 WARN_ON(!is_vm_hugetlb_page(vma));
1861 BUG_ON(start & ~huge_page_mask(h));
1862 BUG_ON(end & ~huge_page_mask(h));
1864 mmu_notifier_invalidate_range_start(mm, start, end);
1865 spin_lock(&mm->page_table_lock);
1866 for (address = start; address < end; address += sz) {
1867 ptep = huge_pte_offset(mm, address);
1871 if (huge_pmd_unshare(mm, &address, ptep))
1875 * If a reference page is supplied, it is because a specific
1876 * page is being unmapped, not a range. Ensure the page we
1877 * are about to unmap is the actual page of interest.
1880 pte = huge_ptep_get(ptep);
1881 if (huge_pte_none(pte))
1883 page = pte_page(pte);
1884 if (page != ref_page)
1888 * Mark the VMA as having unmapped its page so that
1889 * future faults in this VMA will fail rather than
1890 * looking like data was lost
1892 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
1895 pte = huge_ptep_get_and_clear(mm, address, ptep);
1896 if (huge_pte_none(pte))
1899 page = pte_page(pte);
1901 set_page_dirty(page);
1902 list_add(&page->lru, &page_list);
1904 spin_unlock(&mm->page_table_lock);
1905 flush_tlb_range(vma, start, end);
1906 mmu_notifier_invalidate_range_end(mm, start, end);
1907 list_for_each_entry_safe(page, tmp, &page_list, lru) {
1908 list_del(&page->lru);
1913 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1914 unsigned long end, struct page *ref_page)
1916 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1917 __unmap_hugepage_range(vma, start, end, ref_page);
1918 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1922 * This is called when the original mapper is failing to COW a MAP_PRIVATE
1923 * mappping it owns the reserve page for. The intention is to unmap the page
1924 * from other VMAs and let the children be SIGKILLed if they are faulting the
1927 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
1928 struct page *page, unsigned long address)
1930 struct hstate *h = hstate_vma(vma);
1931 struct vm_area_struct *iter_vma;
1932 struct address_space *mapping;
1933 struct prio_tree_iter iter;
1937 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1938 * from page cache lookup which is in HPAGE_SIZE units.
1940 address = address & huge_page_mask(h);
1941 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
1942 + (vma->vm_pgoff >> PAGE_SHIFT);
1943 mapping = (struct address_space *)page_private(page);
1945 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1946 /* Do not unmap the current VMA */
1947 if (iter_vma == vma)
1951 * Unmap the page from other VMAs without their own reserves.
1952 * They get marked to be SIGKILLed if they fault in these
1953 * areas. This is because a future no-page fault on this VMA
1954 * could insert a zeroed page instead of the data existing
1955 * from the time of fork. This would look like data corruption
1957 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
1958 unmap_hugepage_range(iter_vma,
1959 address, address + huge_page_size(h),
1966 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
1967 unsigned long address, pte_t *ptep, pte_t pte,
1968 struct page *pagecache_page)
1970 struct hstate *h = hstate_vma(vma);
1971 struct page *old_page, *new_page;
1973 int outside_reserve = 0;
1975 old_page = pte_page(pte);
1978 /* If no-one else is actually using this page, avoid the copy
1979 * and just make the page writable */
1980 avoidcopy = (page_count(old_page) == 1);
1982 set_huge_ptep_writable(vma, address, ptep);
1987 * If the process that created a MAP_PRIVATE mapping is about to
1988 * perform a COW due to a shared page count, attempt to satisfy
1989 * the allocation without using the existing reserves. The pagecache
1990 * page is used to determine if the reserve at this address was
1991 * consumed or not. If reserves were used, a partial faulted mapping
1992 * at the time of fork() could consume its reserves on COW instead
1993 * of the full address range.
1995 if (!(vma->vm_flags & VM_MAYSHARE) &&
1996 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
1997 old_page != pagecache_page)
1998 outside_reserve = 1;
2000 page_cache_get(old_page);
2001 new_page = alloc_huge_page(vma, address, outside_reserve);
2003 if (IS_ERR(new_page)) {
2004 page_cache_release(old_page);
2007 * If a process owning a MAP_PRIVATE mapping fails to COW,
2008 * it is due to references held by a child and an insufficient
2009 * huge page pool. To guarantee the original mappers
2010 * reliability, unmap the page from child processes. The child
2011 * may get SIGKILLed if it later faults.
2013 if (outside_reserve) {
2014 BUG_ON(huge_pte_none(pte));
2015 if (unmap_ref_private(mm, vma, old_page, address)) {
2016 BUG_ON(page_count(old_page) != 1);
2017 BUG_ON(huge_pte_none(pte));
2018 goto retry_avoidcopy;
2023 return -PTR_ERR(new_page);
2026 spin_unlock(&mm->page_table_lock);
2027 copy_huge_page(new_page, old_page, address, vma);
2028 __SetPageUptodate(new_page);
2029 spin_lock(&mm->page_table_lock);
2031 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2032 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2034 huge_ptep_clear_flush(vma, address, ptep);
2035 set_huge_pte_at(mm, address, ptep,
2036 make_huge_pte(vma, new_page, 1));
2037 /* Make the old page be freed below */
2038 new_page = old_page;
2040 page_cache_release(new_page);
2041 page_cache_release(old_page);
2045 /* Return the pagecache page at a given address within a VMA */
2046 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2047 struct vm_area_struct *vma, unsigned long address)
2049 struct address_space *mapping;
2052 mapping = vma->vm_file->f_mapping;
2053 idx = vma_hugecache_offset(h, vma, address);
2055 return find_lock_page(mapping, idx);
2059 * Return whether there is a pagecache page to back given address within VMA.
2060 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2062 static bool hugetlbfs_pagecache_present(struct hstate *h,
2063 struct vm_area_struct *vma, unsigned long address)
2065 struct address_space *mapping;
2069 mapping = vma->vm_file->f_mapping;
2070 idx = vma_hugecache_offset(h, vma, address);
2072 page = find_get_page(mapping, idx);
2075 return page != NULL;
2078 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2079 unsigned long address, pte_t *ptep, unsigned int flags)
2081 struct hstate *h = hstate_vma(vma);
2082 int ret = VM_FAULT_SIGBUS;
2086 struct address_space *mapping;
2090 * Currently, we are forced to kill the process in the event the
2091 * original mapper has unmapped pages from the child due to a failed
2092 * COW. Warn that such a situation has occured as it may not be obvious
2094 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2096 "PID %d killed due to inadequate hugepage pool\n",
2101 mapping = vma->vm_file->f_mapping;
2102 idx = vma_hugecache_offset(h, vma, address);
2105 * Use page lock to guard against racing truncation
2106 * before we get page_table_lock.
2109 page = find_lock_page(mapping, idx);
2111 size = i_size_read(mapping->host) >> huge_page_shift(h);
2114 page = alloc_huge_page(vma, address, 0);
2116 ret = -PTR_ERR(page);
2119 clear_huge_page(page, address, huge_page_size(h));
2120 __SetPageUptodate(page);
2122 if (vma->vm_flags & VM_MAYSHARE) {
2124 struct inode *inode = mapping->host;
2126 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2134 spin_lock(&inode->i_lock);
2135 inode->i_blocks += blocks_per_huge_page(h);
2136 spin_unlock(&inode->i_lock);
2142 * If we are going to COW a private mapping later, we examine the
2143 * pending reservations for this page now. This will ensure that
2144 * any allocations necessary to record that reservation occur outside
2147 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2148 if (vma_needs_reservation(h, vma, address) < 0) {
2150 goto backout_unlocked;
2153 spin_lock(&mm->page_table_lock);
2154 size = i_size_read(mapping->host) >> huge_page_shift(h);
2159 if (!huge_pte_none(huge_ptep_get(ptep)))
2162 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2163 && (vma->vm_flags & VM_SHARED)));
2164 set_huge_pte_at(mm, address, ptep, new_pte);
2166 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2167 /* Optimization, do the COW without a second fault */
2168 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2171 spin_unlock(&mm->page_table_lock);
2177 spin_unlock(&mm->page_table_lock);
2184 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2185 unsigned long address, unsigned int flags)
2190 struct page *pagecache_page = NULL;
2191 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2192 struct hstate *h = hstate_vma(vma);
2194 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2196 return VM_FAULT_OOM;
2199 * Serialize hugepage allocation and instantiation, so that we don't
2200 * get spurious allocation failures if two CPUs race to instantiate
2201 * the same page in the page cache.
2203 mutex_lock(&hugetlb_instantiation_mutex);
2204 entry = huge_ptep_get(ptep);
2205 if (huge_pte_none(entry)) {
2206 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2213 * If we are going to COW the mapping later, we examine the pending
2214 * reservations for this page now. This will ensure that any
2215 * allocations necessary to record that reservation occur outside the
2216 * spinlock. For private mappings, we also lookup the pagecache
2217 * page now as it is used to determine if a reservation has been
2220 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2221 if (vma_needs_reservation(h, vma, address) < 0) {
2226 if (!(vma->vm_flags & VM_MAYSHARE))
2227 pagecache_page = hugetlbfs_pagecache_page(h,
2231 spin_lock(&mm->page_table_lock);
2232 /* Check for a racing update before calling hugetlb_cow */
2233 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2234 goto out_page_table_lock;
2237 if (flags & FAULT_FLAG_WRITE) {
2238 if (!pte_write(entry)) {
2239 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2241 goto out_page_table_lock;
2243 entry = pte_mkdirty(entry);
2245 entry = pte_mkyoung(entry);
2246 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2247 flags & FAULT_FLAG_WRITE))
2248 update_mmu_cache(vma, address, entry);
2250 out_page_table_lock:
2251 spin_unlock(&mm->page_table_lock);
2253 if (pagecache_page) {
2254 unlock_page(pagecache_page);
2255 put_page(pagecache_page);
2259 mutex_unlock(&hugetlb_instantiation_mutex);
2264 /* Can be overriden by architectures */
2265 __attribute__((weak)) struct page *
2266 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2267 pud_t *pud, int write)
2273 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2274 struct page **pages, struct vm_area_struct **vmas,
2275 unsigned long *position, int *length, int i,
2278 unsigned long pfn_offset;
2279 unsigned long vaddr = *position;
2280 int remainder = *length;
2281 struct hstate *h = hstate_vma(vma);
2283 spin_lock(&mm->page_table_lock);
2284 while (vaddr < vma->vm_end && remainder) {
2290 * Some archs (sparc64, sh*) have multiple pte_ts to
2291 * each hugepage. We have to make sure we get the
2292 * first, for the page indexing below to work.
2294 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2295 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2298 * When coredumping, it suits get_dump_page if we just return
2299 * an error where there's an empty slot with no huge pagecache
2300 * to back it. This way, we avoid allocating a hugepage, and
2301 * the sparse dumpfile avoids allocating disk blocks, but its
2302 * huge holes still show up with zeroes where they need to be.
2304 if (absent && (flags & FOLL_DUMP) &&
2305 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2311 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2314 spin_unlock(&mm->page_table_lock);
2315 ret = hugetlb_fault(mm, vma, vaddr,
2316 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2317 spin_lock(&mm->page_table_lock);
2318 if (!(ret & VM_FAULT_ERROR))
2325 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2326 page = pte_page(huge_ptep_get(pte));
2329 pages[i] = mem_map_offset(page, pfn_offset);
2340 if (vaddr < vma->vm_end && remainder &&
2341 pfn_offset < pages_per_huge_page(h)) {
2343 * We use pfn_offset to avoid touching the pageframes
2344 * of this compound page.
2349 spin_unlock(&mm->page_table_lock);
2350 *length = remainder;
2353 return i ? i : -EFAULT;
2356 void hugetlb_change_protection(struct vm_area_struct *vma,
2357 unsigned long address, unsigned long end, pgprot_t newprot)
2359 struct mm_struct *mm = vma->vm_mm;
2360 unsigned long start = address;
2363 struct hstate *h = hstate_vma(vma);
2365 BUG_ON(address >= end);
2366 flush_cache_range(vma, address, end);
2368 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2369 spin_lock(&mm->page_table_lock);
2370 for (; address < end; address += huge_page_size(h)) {
2371 ptep = huge_pte_offset(mm, address);
2374 if (huge_pmd_unshare(mm, &address, ptep))
2376 if (!huge_pte_none(huge_ptep_get(ptep))) {
2377 pte = huge_ptep_get_and_clear(mm, address, ptep);
2378 pte = pte_mkhuge(pte_modify(pte, newprot));
2379 set_huge_pte_at(mm, address, ptep, pte);
2382 spin_unlock(&mm->page_table_lock);
2383 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2385 flush_tlb_range(vma, start, end);
2388 int hugetlb_reserve_pages(struct inode *inode,
2390 struct vm_area_struct *vma,
2394 struct hstate *h = hstate_inode(inode);
2397 * Only apply hugepage reservation if asked. At fault time, an
2398 * attempt will be made for VM_NORESERVE to allocate a page
2399 * and filesystem quota without using reserves
2401 if (acctflag & VM_NORESERVE)
2405 * Shared mappings base their reservation on the number of pages that
2406 * are already allocated on behalf of the file. Private mappings need
2407 * to reserve the full area even if read-only as mprotect() may be
2408 * called to make the mapping read-write. Assume !vma is a shm mapping
2410 if (!vma || vma->vm_flags & VM_MAYSHARE)
2411 chg = region_chg(&inode->i_mapping->private_list, from, to);
2413 struct resv_map *resv_map = resv_map_alloc();
2419 set_vma_resv_map(vma, resv_map);
2420 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2426 /* There must be enough filesystem quota for the mapping */
2427 if (hugetlb_get_quota(inode->i_mapping, chg))
2431 * Check enough hugepages are available for the reservation.
2432 * Hand back the quota if there are not
2434 ret = hugetlb_acct_memory(h, chg);
2436 hugetlb_put_quota(inode->i_mapping, chg);
2441 * Account for the reservations made. Shared mappings record regions
2442 * that have reservations as they are shared by multiple VMAs.
2443 * When the last VMA disappears, the region map says how much
2444 * the reservation was and the page cache tells how much of
2445 * the reservation was consumed. Private mappings are per-VMA and
2446 * only the consumed reservations are tracked. When the VMA
2447 * disappears, the original reservation is the VMA size and the
2448 * consumed reservations are stored in the map. Hence, nothing
2449 * else has to be done for private mappings here
2451 if (!vma || vma->vm_flags & VM_MAYSHARE)
2452 region_add(&inode->i_mapping->private_list, from, to);
2456 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2458 struct hstate *h = hstate_inode(inode);
2459 long chg = region_truncate(&inode->i_mapping->private_list, offset);
2461 spin_lock(&inode->i_lock);
2462 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
2463 spin_unlock(&inode->i_lock);
2465 hugetlb_put_quota(inode->i_mapping, (chg - freed));
2466 hugetlb_acct_memory(h, -(chg - freed));