2 * Generic hugetlb support.
3 * (C) William Irwin, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/module.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/rmap.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
26 #include <asm/pgtable.h>
29 #include <linux/hugetlb.h>
30 #include <linux/node.h>
33 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
34 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
35 unsigned long hugepages_treat_as_movable;
37 static int max_hstate;
38 unsigned int default_hstate_idx;
39 struct hstate hstates[HUGE_MAX_HSTATE];
41 __initdata LIST_HEAD(huge_boot_pages);
43 /* for command line parsing */
44 static struct hstate * __initdata parsed_hstate;
45 static unsigned long __initdata default_hstate_max_huge_pages;
46 static unsigned long __initdata default_hstate_size;
48 #define for_each_hstate(h) \
49 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
52 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
54 static DEFINE_SPINLOCK(hugetlb_lock);
57 * Region tracking -- allows tracking of reservations and instantiated pages
58 * across the pages in a mapping.
60 * The region data structures are protected by a combination of the mmap_sem
61 * and the hugetlb_instantion_mutex. To access or modify a region the caller
62 * must either hold the mmap_sem for write, or the mmap_sem for read and
63 * the hugetlb_instantiation mutex:
65 * down_write(&mm->mmap_sem);
67 * down_read(&mm->mmap_sem);
68 * mutex_lock(&hugetlb_instantiation_mutex);
71 struct list_head link;
76 static long region_add(struct list_head *head, long f, long t)
78 struct file_region *rg, *nrg, *trg;
80 /* Locate the region we are either in or before. */
81 list_for_each_entry(rg, head, link)
85 /* Round our left edge to the current segment if it encloses us. */
89 /* Check for and consume any regions we now overlap with. */
91 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
92 if (&rg->link == head)
97 /* If this area reaches higher then extend our area to
98 * include it completely. If this is not the first area
99 * which we intend to reuse, free it. */
112 static long region_chg(struct list_head *head, long f, long t)
114 struct file_region *rg, *nrg;
117 /* Locate the region we are before or in. */
118 list_for_each_entry(rg, head, link)
122 /* If we are below the current region then a new region is required.
123 * Subtle, allocate a new region at the position but make it zero
124 * size such that we can guarantee to record the reservation. */
125 if (&rg->link == head || t < rg->from) {
126 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
131 INIT_LIST_HEAD(&nrg->link);
132 list_add(&nrg->link, rg->link.prev);
137 /* Round our left edge to the current segment if it encloses us. */
142 /* Check for and consume any regions we now overlap with. */
143 list_for_each_entry(rg, rg->link.prev, link) {
144 if (&rg->link == head)
149 /* We overlap with this area, if it extends futher than
150 * us then we must extend ourselves. Account for its
151 * existing reservation. */
156 chg -= rg->to - rg->from;
161 static long region_truncate(struct list_head *head, long end)
163 struct file_region *rg, *trg;
166 /* Locate the region we are either in or before. */
167 list_for_each_entry(rg, head, link)
170 if (&rg->link == head)
173 /* If we are in the middle of a region then adjust it. */
174 if (end > rg->from) {
177 rg = list_entry(rg->link.next, typeof(*rg), link);
180 /* Drop any remaining regions. */
181 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
182 if (&rg->link == head)
184 chg += rg->to - rg->from;
191 static long region_count(struct list_head *head, long f, long t)
193 struct file_region *rg;
196 /* Locate each segment we overlap with, and count that overlap. */
197 list_for_each_entry(rg, head, link) {
206 seg_from = max(rg->from, f);
207 seg_to = min(rg->to, t);
209 chg += seg_to - seg_from;
216 * Convert the address within this vma to the page offset within
217 * the mapping, in pagecache page units; huge pages here.
219 static pgoff_t vma_hugecache_offset(struct hstate *h,
220 struct vm_area_struct *vma, unsigned long address)
222 return ((address - vma->vm_start) >> huge_page_shift(h)) +
223 (vma->vm_pgoff >> huge_page_order(h));
226 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
227 unsigned long address)
229 return vma_hugecache_offset(hstate_vma(vma), vma, address);
233 * Return the size of the pages allocated when backing a VMA. In the majority
234 * cases this will be same size as used by the page table entries.
236 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
238 struct hstate *hstate;
240 if (!is_vm_hugetlb_page(vma))
243 hstate = hstate_vma(vma);
245 return 1UL << (hstate->order + PAGE_SHIFT);
247 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
250 * Return the page size being used by the MMU to back a VMA. In the majority
251 * of cases, the page size used by the kernel matches the MMU size. On
252 * architectures where it differs, an architecture-specific version of this
253 * function is required.
255 #ifndef vma_mmu_pagesize
256 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
258 return vma_kernel_pagesize(vma);
263 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
264 * bits of the reservation map pointer, which are always clear due to
267 #define HPAGE_RESV_OWNER (1UL << 0)
268 #define HPAGE_RESV_UNMAPPED (1UL << 1)
269 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
272 * These helpers are used to track how many pages are reserved for
273 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
274 * is guaranteed to have their future faults succeed.
276 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
277 * the reserve counters are updated with the hugetlb_lock held. It is safe
278 * to reset the VMA at fork() time as it is not in use yet and there is no
279 * chance of the global counters getting corrupted as a result of the values.
281 * The private mapping reservation is represented in a subtly different
282 * manner to a shared mapping. A shared mapping has a region map associated
283 * with the underlying file, this region map represents the backing file
284 * pages which have ever had a reservation assigned which this persists even
285 * after the page is instantiated. A private mapping has a region map
286 * associated with the original mmap which is attached to all VMAs which
287 * reference it, this region map represents those offsets which have consumed
288 * reservation ie. where pages have been instantiated.
290 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
292 return (unsigned long)vma->vm_private_data;
295 static void set_vma_private_data(struct vm_area_struct *vma,
298 vma->vm_private_data = (void *)value;
303 struct list_head regions;
306 static struct resv_map *resv_map_alloc(void)
308 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
312 kref_init(&resv_map->refs);
313 INIT_LIST_HEAD(&resv_map->regions);
318 static void resv_map_release(struct kref *ref)
320 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
322 /* Clear out any active regions before we release the map. */
323 region_truncate(&resv_map->regions, 0);
327 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
329 VM_BUG_ON(!is_vm_hugetlb_page(vma));
330 if (!(vma->vm_flags & VM_MAYSHARE))
331 return (struct resv_map *)(get_vma_private_data(vma) &
336 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
338 VM_BUG_ON(!is_vm_hugetlb_page(vma));
339 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
341 set_vma_private_data(vma, (get_vma_private_data(vma) &
342 HPAGE_RESV_MASK) | (unsigned long)map);
345 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
347 VM_BUG_ON(!is_vm_hugetlb_page(vma));
348 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
350 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
353 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
355 VM_BUG_ON(!is_vm_hugetlb_page(vma));
357 return (get_vma_private_data(vma) & flag) != 0;
360 /* Decrement the reserved pages in the hugepage pool by one */
361 static void decrement_hugepage_resv_vma(struct hstate *h,
362 struct vm_area_struct *vma)
364 if (vma->vm_flags & VM_NORESERVE)
367 if (vma->vm_flags & VM_MAYSHARE) {
368 /* Shared mappings always use reserves */
369 h->resv_huge_pages--;
370 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
372 * Only the process that called mmap() has reserves for
375 h->resv_huge_pages--;
379 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
380 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
382 VM_BUG_ON(!is_vm_hugetlb_page(vma));
383 if (!(vma->vm_flags & VM_MAYSHARE))
384 vma->vm_private_data = (void *)0;
387 /* Returns true if the VMA has associated reserve pages */
388 static int vma_has_reserves(struct vm_area_struct *vma)
390 if (vma->vm_flags & VM_MAYSHARE)
392 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
397 static void clear_gigantic_page(struct page *page,
398 unsigned long addr, unsigned long sz)
401 struct page *p = page;
404 for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) {
406 clear_user_highpage(p, addr + i * PAGE_SIZE);
409 static void clear_huge_page(struct page *page,
410 unsigned long addr, unsigned long sz)
414 if (unlikely(sz/PAGE_SIZE > MAX_ORDER_NR_PAGES)) {
415 clear_gigantic_page(page, addr, sz);
420 for (i = 0; i < sz/PAGE_SIZE; i++) {
422 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
426 static void copy_gigantic_page(struct page *dst, struct page *src,
427 unsigned long addr, struct vm_area_struct *vma)
430 struct hstate *h = hstate_vma(vma);
431 struct page *dst_base = dst;
432 struct page *src_base = src;
434 for (i = 0; i < pages_per_huge_page(h); ) {
436 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
439 dst = mem_map_next(dst, dst_base, i);
440 src = mem_map_next(src, src_base, i);
443 static void copy_huge_page(struct page *dst, struct page *src,
444 unsigned long addr, struct vm_area_struct *vma)
447 struct hstate *h = hstate_vma(vma);
449 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
450 copy_gigantic_page(dst, src, addr, vma);
455 for (i = 0; i < pages_per_huge_page(h); i++) {
457 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
461 static void enqueue_huge_page(struct hstate *h, struct page *page)
463 int nid = page_to_nid(page);
464 list_add(&page->lru, &h->hugepage_freelists[nid]);
465 h->free_huge_pages++;
466 h->free_huge_pages_node[nid]++;
469 static struct page *dequeue_huge_page_vma(struct hstate *h,
470 struct vm_area_struct *vma,
471 unsigned long address, int avoid_reserve)
474 struct page *page = NULL;
475 struct mempolicy *mpol;
476 nodemask_t *nodemask;
477 struct zonelist *zonelist;
482 zonelist = huge_zonelist(vma, address,
483 htlb_alloc_mask, &mpol, &nodemask);
485 * A child process with MAP_PRIVATE mappings created by their parent
486 * have no page reserves. This check ensures that reservations are
487 * not "stolen". The child may still get SIGKILLed
489 if (!vma_has_reserves(vma) &&
490 h->free_huge_pages - h->resv_huge_pages == 0)
493 /* If reserves cannot be used, ensure enough pages are in the pool */
494 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
497 for_each_zone_zonelist_nodemask(zone, z, zonelist,
498 MAX_NR_ZONES - 1, nodemask) {
499 nid = zone_to_nid(zone);
500 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
501 !list_empty(&h->hugepage_freelists[nid])) {
502 page = list_entry(h->hugepage_freelists[nid].next,
504 list_del(&page->lru);
505 h->free_huge_pages--;
506 h->free_huge_pages_node[nid]--;
509 decrement_hugepage_resv_vma(h, vma);
520 static void update_and_free_page(struct hstate *h, struct page *page)
524 VM_BUG_ON(h->order >= MAX_ORDER);
527 h->nr_huge_pages_node[page_to_nid(page)]--;
528 for (i = 0; i < pages_per_huge_page(h); i++) {
529 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
530 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
531 1 << PG_private | 1<< PG_writeback);
533 set_compound_page_dtor(page, NULL);
534 set_page_refcounted(page);
535 arch_release_hugepage(page);
536 __free_pages(page, huge_page_order(h));
539 struct hstate *size_to_hstate(unsigned long size)
544 if (huge_page_size(h) == size)
550 static void free_huge_page(struct page *page)
553 * Can't pass hstate in here because it is called from the
554 * compound page destructor.
556 struct hstate *h = page_hstate(page);
557 int nid = page_to_nid(page);
558 struct address_space *mapping;
560 mapping = (struct address_space *) page_private(page);
561 set_page_private(page, 0);
562 page->mapping = NULL;
563 BUG_ON(page_count(page));
564 BUG_ON(page_mapcount(page));
565 INIT_LIST_HEAD(&page->lru);
567 spin_lock(&hugetlb_lock);
568 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
569 update_and_free_page(h, page);
570 h->surplus_huge_pages--;
571 h->surplus_huge_pages_node[nid]--;
573 enqueue_huge_page(h, page);
575 spin_unlock(&hugetlb_lock);
577 hugetlb_put_quota(mapping, 1);
580 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
582 set_compound_page_dtor(page, free_huge_page);
583 spin_lock(&hugetlb_lock);
585 h->nr_huge_pages_node[nid]++;
586 spin_unlock(&hugetlb_lock);
587 put_page(page); /* free it into the hugepage allocator */
590 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
593 int nr_pages = 1 << order;
594 struct page *p = page + 1;
596 /* we rely on prep_new_huge_page to set the destructor */
597 set_compound_order(page, order);
599 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
601 p->first_page = page;
605 int PageHuge(struct page *page)
607 compound_page_dtor *dtor;
609 if (!PageCompound(page))
612 page = compound_head(page);
613 dtor = get_compound_page_dtor(page);
615 return dtor == free_huge_page;
618 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
622 if (h->order >= MAX_ORDER)
625 page = alloc_pages_exact_node(nid,
626 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
627 __GFP_REPEAT|__GFP_NOWARN,
630 if (arch_prepare_hugepage(page)) {
631 __free_pages(page, huge_page_order(h));
634 prep_new_huge_page(h, page, nid);
641 * common helper functions for hstate_next_node_to_{alloc|free}.
642 * We may have allocated or freed a huge page based on a different
643 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
644 * be outside of *nodes_allowed. Ensure that we use an allowed
645 * node for alloc or free.
647 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
649 nid = next_node(nid, *nodes_allowed);
650 if (nid == MAX_NUMNODES)
651 nid = first_node(*nodes_allowed);
652 VM_BUG_ON(nid >= MAX_NUMNODES);
657 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
659 if (!node_isset(nid, *nodes_allowed))
660 nid = next_node_allowed(nid, nodes_allowed);
665 * returns the previously saved node ["this node"] from which to
666 * allocate a persistent huge page for the pool and advance the
667 * next node from which to allocate, handling wrap at end of node
670 static int hstate_next_node_to_alloc(struct hstate *h,
671 nodemask_t *nodes_allowed)
675 VM_BUG_ON(!nodes_allowed);
677 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
678 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
683 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
690 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
691 next_nid = start_nid;
694 page = alloc_fresh_huge_page_node(h, next_nid);
699 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
700 } while (next_nid != start_nid);
703 count_vm_event(HTLB_BUDDY_PGALLOC);
705 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
711 * helper for free_pool_huge_page() - return the previously saved
712 * node ["this node"] from which to free a huge page. Advance the
713 * next node id whether or not we find a free huge page to free so
714 * that the next attempt to free addresses the next node.
716 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
720 VM_BUG_ON(!nodes_allowed);
722 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
723 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
729 * Free huge page from pool from next node to free.
730 * Attempt to keep persistent huge pages more or less
731 * balanced over allowed nodes.
732 * Called with hugetlb_lock locked.
734 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
741 start_nid = hstate_next_node_to_free(h, nodes_allowed);
742 next_nid = start_nid;
746 * If we're returning unused surplus pages, only examine
747 * nodes with surplus pages.
749 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
750 !list_empty(&h->hugepage_freelists[next_nid])) {
752 list_entry(h->hugepage_freelists[next_nid].next,
754 list_del(&page->lru);
755 h->free_huge_pages--;
756 h->free_huge_pages_node[next_nid]--;
758 h->surplus_huge_pages--;
759 h->surplus_huge_pages_node[next_nid]--;
761 update_and_free_page(h, page);
765 next_nid = hstate_next_node_to_free(h, nodes_allowed);
766 } while (next_nid != start_nid);
771 static struct page *alloc_buddy_huge_page(struct hstate *h,
772 struct vm_area_struct *vma, unsigned long address)
777 if (h->order >= MAX_ORDER)
781 * Assume we will successfully allocate the surplus page to
782 * prevent racing processes from causing the surplus to exceed
785 * This however introduces a different race, where a process B
786 * tries to grow the static hugepage pool while alloc_pages() is
787 * called by process A. B will only examine the per-node
788 * counters in determining if surplus huge pages can be
789 * converted to normal huge pages in adjust_pool_surplus(). A
790 * won't be able to increment the per-node counter, until the
791 * lock is dropped by B, but B doesn't drop hugetlb_lock until
792 * no more huge pages can be converted from surplus to normal
793 * state (and doesn't try to convert again). Thus, we have a
794 * case where a surplus huge page exists, the pool is grown, and
795 * the surplus huge page still exists after, even though it
796 * should just have been converted to a normal huge page. This
797 * does not leak memory, though, as the hugepage will be freed
798 * once it is out of use. It also does not allow the counters to
799 * go out of whack in adjust_pool_surplus() as we don't modify
800 * the node values until we've gotten the hugepage and only the
801 * per-node value is checked there.
803 spin_lock(&hugetlb_lock);
804 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
805 spin_unlock(&hugetlb_lock);
809 h->surplus_huge_pages++;
811 spin_unlock(&hugetlb_lock);
813 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
814 __GFP_REPEAT|__GFP_NOWARN,
817 if (page && arch_prepare_hugepage(page)) {
818 __free_pages(page, huge_page_order(h));
822 spin_lock(&hugetlb_lock);
825 * This page is now managed by the hugetlb allocator and has
826 * no users -- drop the buddy allocator's reference.
828 put_page_testzero(page);
829 VM_BUG_ON(page_count(page));
830 nid = page_to_nid(page);
831 set_compound_page_dtor(page, free_huge_page);
833 * We incremented the global counters already
835 h->nr_huge_pages_node[nid]++;
836 h->surplus_huge_pages_node[nid]++;
837 __count_vm_event(HTLB_BUDDY_PGALLOC);
840 h->surplus_huge_pages--;
841 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
843 spin_unlock(&hugetlb_lock);
849 * Increase the hugetlb pool such that it can accomodate a reservation
852 static int gather_surplus_pages(struct hstate *h, int delta)
854 struct list_head surplus_list;
855 struct page *page, *tmp;
857 int needed, allocated;
859 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
861 h->resv_huge_pages += delta;
866 INIT_LIST_HEAD(&surplus_list);
870 spin_unlock(&hugetlb_lock);
871 for (i = 0; i < needed; i++) {
872 page = alloc_buddy_huge_page(h, NULL, 0);
875 * We were not able to allocate enough pages to
876 * satisfy the entire reservation so we free what
877 * we've allocated so far.
879 spin_lock(&hugetlb_lock);
884 list_add(&page->lru, &surplus_list);
889 * After retaking hugetlb_lock, we need to recalculate 'needed'
890 * because either resv_huge_pages or free_huge_pages may have changed.
892 spin_lock(&hugetlb_lock);
893 needed = (h->resv_huge_pages + delta) -
894 (h->free_huge_pages + allocated);
899 * The surplus_list now contains _at_least_ the number of extra pages
900 * needed to accomodate the reservation. Add the appropriate number
901 * of pages to the hugetlb pool and free the extras back to the buddy
902 * allocator. Commit the entire reservation here to prevent another
903 * process from stealing the pages as they are added to the pool but
904 * before they are reserved.
907 h->resv_huge_pages += delta;
910 /* Free the needed pages to the hugetlb pool */
911 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
914 list_del(&page->lru);
915 enqueue_huge_page(h, page);
918 /* Free unnecessary surplus pages to the buddy allocator */
919 if (!list_empty(&surplus_list)) {
920 spin_unlock(&hugetlb_lock);
921 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
922 list_del(&page->lru);
924 * The page has a reference count of zero already, so
925 * call free_huge_page directly instead of using
926 * put_page. This must be done with hugetlb_lock
927 * unlocked which is safe because free_huge_page takes
928 * hugetlb_lock before deciding how to free the page.
930 free_huge_page(page);
932 spin_lock(&hugetlb_lock);
939 * When releasing a hugetlb pool reservation, any surplus pages that were
940 * allocated to satisfy the reservation must be explicitly freed if they were
942 * Called with hugetlb_lock held.
944 static void return_unused_surplus_pages(struct hstate *h,
945 unsigned long unused_resv_pages)
947 unsigned long nr_pages;
949 /* Uncommit the reservation */
950 h->resv_huge_pages -= unused_resv_pages;
952 /* Cannot return gigantic pages currently */
953 if (h->order >= MAX_ORDER)
956 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
959 * We want to release as many surplus pages as possible, spread
960 * evenly across all nodes with memory. Iterate across these nodes
961 * until we can no longer free unreserved surplus pages. This occurs
962 * when the nodes with surplus pages have no free pages.
963 * free_pool_huge_page() will balance the the freed pages across the
964 * on-line nodes with memory and will handle the hstate accounting.
967 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
973 * Determine if the huge page at addr within the vma has an associated
974 * reservation. Where it does not we will need to logically increase
975 * reservation and actually increase quota before an allocation can occur.
976 * Where any new reservation would be required the reservation change is
977 * prepared, but not committed. Once the page has been quota'd allocated
978 * an instantiated the change should be committed via vma_commit_reservation.
979 * No action is required on failure.
981 static long vma_needs_reservation(struct hstate *h,
982 struct vm_area_struct *vma, unsigned long addr)
984 struct address_space *mapping = vma->vm_file->f_mapping;
985 struct inode *inode = mapping->host;
987 if (vma->vm_flags & VM_MAYSHARE) {
988 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
989 return region_chg(&inode->i_mapping->private_list,
992 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
997 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
998 struct resv_map *reservations = vma_resv_map(vma);
1000 err = region_chg(&reservations->regions, idx, idx + 1);
1006 static void vma_commit_reservation(struct hstate *h,
1007 struct vm_area_struct *vma, unsigned long addr)
1009 struct address_space *mapping = vma->vm_file->f_mapping;
1010 struct inode *inode = mapping->host;
1012 if (vma->vm_flags & VM_MAYSHARE) {
1013 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1014 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1016 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1017 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1018 struct resv_map *reservations = vma_resv_map(vma);
1020 /* Mark this page used in the map. */
1021 region_add(&reservations->regions, idx, idx + 1);
1025 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1026 unsigned long addr, int avoid_reserve)
1028 struct hstate *h = hstate_vma(vma);
1030 struct address_space *mapping = vma->vm_file->f_mapping;
1031 struct inode *inode = mapping->host;
1035 * Processes that did not create the mapping will have no reserves and
1036 * will not have accounted against quota. Check that the quota can be
1037 * made before satisfying the allocation
1038 * MAP_NORESERVE mappings may also need pages and quota allocated
1039 * if no reserve mapping overlaps.
1041 chg = vma_needs_reservation(h, vma, addr);
1043 return ERR_PTR(chg);
1045 if (hugetlb_get_quota(inode->i_mapping, chg))
1046 return ERR_PTR(-ENOSPC);
1048 spin_lock(&hugetlb_lock);
1049 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1050 spin_unlock(&hugetlb_lock);
1053 page = alloc_buddy_huge_page(h, vma, addr);
1055 hugetlb_put_quota(inode->i_mapping, chg);
1056 return ERR_PTR(-VM_FAULT_SIGBUS);
1060 set_page_refcounted(page);
1061 set_page_private(page, (unsigned long) mapping);
1063 vma_commit_reservation(h, vma, addr);
1068 int __weak alloc_bootmem_huge_page(struct hstate *h)
1070 struct huge_bootmem_page *m;
1071 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1076 addr = __alloc_bootmem_node_nopanic(
1077 NODE_DATA(hstate_next_node_to_alloc(h,
1078 &node_states[N_HIGH_MEMORY])),
1079 huge_page_size(h), huge_page_size(h), 0);
1083 * Use the beginning of the huge page to store the
1084 * huge_bootmem_page struct (until gather_bootmem
1085 * puts them into the mem_map).
1095 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1096 /* Put them into a private list first because mem_map is not up yet */
1097 list_add(&m->list, &huge_boot_pages);
1102 static void prep_compound_huge_page(struct page *page, int order)
1104 if (unlikely(order > (MAX_ORDER - 1)))
1105 prep_compound_gigantic_page(page, order);
1107 prep_compound_page(page, order);
1110 /* Put bootmem huge pages into the standard lists after mem_map is up */
1111 static void __init gather_bootmem_prealloc(void)
1113 struct huge_bootmem_page *m;
1115 list_for_each_entry(m, &huge_boot_pages, list) {
1116 struct page *page = virt_to_page(m);
1117 struct hstate *h = m->hstate;
1118 __ClearPageReserved(page);
1119 WARN_ON(page_count(page) != 1);
1120 prep_compound_huge_page(page, h->order);
1121 prep_new_huge_page(h, page, page_to_nid(page));
1125 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1129 for (i = 0; i < h->max_huge_pages; ++i) {
1130 if (h->order >= MAX_ORDER) {
1131 if (!alloc_bootmem_huge_page(h))
1133 } else if (!alloc_fresh_huge_page(h,
1134 &node_states[N_HIGH_MEMORY]))
1137 h->max_huge_pages = i;
1140 static void __init hugetlb_init_hstates(void)
1144 for_each_hstate(h) {
1145 /* oversize hugepages were init'ed in early boot */
1146 if (h->order < MAX_ORDER)
1147 hugetlb_hstate_alloc_pages(h);
1151 static char * __init memfmt(char *buf, unsigned long n)
1153 if (n >= (1UL << 30))
1154 sprintf(buf, "%lu GB", n >> 30);
1155 else if (n >= (1UL << 20))
1156 sprintf(buf, "%lu MB", n >> 20);
1158 sprintf(buf, "%lu KB", n >> 10);
1162 static void __init report_hugepages(void)
1166 for_each_hstate(h) {
1168 printk(KERN_INFO "HugeTLB registered %s page size, "
1169 "pre-allocated %ld pages\n",
1170 memfmt(buf, huge_page_size(h)),
1171 h->free_huge_pages);
1175 #ifdef CONFIG_HIGHMEM
1176 static void try_to_free_low(struct hstate *h, unsigned long count,
1177 nodemask_t *nodes_allowed)
1181 if (h->order >= MAX_ORDER)
1184 for_each_node_mask(i, *nodes_allowed) {
1185 struct page *page, *next;
1186 struct list_head *freel = &h->hugepage_freelists[i];
1187 list_for_each_entry_safe(page, next, freel, lru) {
1188 if (count >= h->nr_huge_pages)
1190 if (PageHighMem(page))
1192 list_del(&page->lru);
1193 update_and_free_page(h, page);
1194 h->free_huge_pages--;
1195 h->free_huge_pages_node[page_to_nid(page)]--;
1200 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1201 nodemask_t *nodes_allowed)
1207 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1208 * balanced by operating on them in a round-robin fashion.
1209 * Returns 1 if an adjustment was made.
1211 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1214 int start_nid, next_nid;
1217 VM_BUG_ON(delta != -1 && delta != 1);
1220 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1222 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1223 next_nid = start_nid;
1229 * To shrink on this node, there must be a surplus page
1231 if (!h->surplus_huge_pages_node[nid]) {
1232 next_nid = hstate_next_node_to_alloc(h,
1239 * Surplus cannot exceed the total number of pages
1241 if (h->surplus_huge_pages_node[nid] >=
1242 h->nr_huge_pages_node[nid]) {
1243 next_nid = hstate_next_node_to_free(h,
1249 h->surplus_huge_pages += delta;
1250 h->surplus_huge_pages_node[nid] += delta;
1253 } while (next_nid != start_nid);
1258 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1259 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1260 nodemask_t *nodes_allowed)
1262 unsigned long min_count, ret;
1264 if (h->order >= MAX_ORDER)
1265 return h->max_huge_pages;
1268 * Increase the pool size
1269 * First take pages out of surplus state. Then make up the
1270 * remaining difference by allocating fresh huge pages.
1272 * We might race with alloc_buddy_huge_page() here and be unable
1273 * to convert a surplus huge page to a normal huge page. That is
1274 * not critical, though, it just means the overall size of the
1275 * pool might be one hugepage larger than it needs to be, but
1276 * within all the constraints specified by the sysctls.
1278 spin_lock(&hugetlb_lock);
1279 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1280 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1284 while (count > persistent_huge_pages(h)) {
1286 * If this allocation races such that we no longer need the
1287 * page, free_huge_page will handle it by freeing the page
1288 * and reducing the surplus.
1290 spin_unlock(&hugetlb_lock);
1291 ret = alloc_fresh_huge_page(h, nodes_allowed);
1292 spin_lock(&hugetlb_lock);
1296 /* Bail for signals. Probably ctrl-c from user */
1297 if (signal_pending(current))
1302 * Decrease the pool size
1303 * First return free pages to the buddy allocator (being careful
1304 * to keep enough around to satisfy reservations). Then place
1305 * pages into surplus state as needed so the pool will shrink
1306 * to the desired size as pages become free.
1308 * By placing pages into the surplus state independent of the
1309 * overcommit value, we are allowing the surplus pool size to
1310 * exceed overcommit. There are few sane options here. Since
1311 * alloc_buddy_huge_page() is checking the global counter,
1312 * though, we'll note that we're not allowed to exceed surplus
1313 * and won't grow the pool anywhere else. Not until one of the
1314 * sysctls are changed, or the surplus pages go out of use.
1316 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1317 min_count = max(count, min_count);
1318 try_to_free_low(h, min_count, nodes_allowed);
1319 while (min_count < persistent_huge_pages(h)) {
1320 if (!free_pool_huge_page(h, nodes_allowed, 0))
1323 while (count < persistent_huge_pages(h)) {
1324 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1328 ret = persistent_huge_pages(h);
1329 spin_unlock(&hugetlb_lock);
1333 #define HSTATE_ATTR_RO(_name) \
1334 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1336 #define HSTATE_ATTR(_name) \
1337 static struct kobj_attribute _name##_attr = \
1338 __ATTR(_name, 0644, _name##_show, _name##_store)
1340 static struct kobject *hugepages_kobj;
1341 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1343 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1345 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1349 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1350 if (hstate_kobjs[i] == kobj) {
1352 *nidp = NUMA_NO_NODE;
1356 return kobj_to_node_hstate(kobj, nidp);
1359 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1360 struct kobj_attribute *attr, char *buf)
1363 unsigned long nr_huge_pages;
1366 h = kobj_to_hstate(kobj, &nid);
1367 if (nid == NUMA_NO_NODE)
1368 nr_huge_pages = h->nr_huge_pages;
1370 nr_huge_pages = h->nr_huge_pages_node[nid];
1372 return sprintf(buf, "%lu\n", nr_huge_pages);
1374 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1375 struct kobject *kobj, struct kobj_attribute *attr,
1376 const char *buf, size_t len)
1380 unsigned long count;
1382 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1384 err = strict_strtoul(buf, 10, &count);
1388 h = kobj_to_hstate(kobj, &nid);
1389 if (nid == NUMA_NO_NODE) {
1391 * global hstate attribute
1393 if (!(obey_mempolicy &&
1394 init_nodemask_of_mempolicy(nodes_allowed))) {
1395 NODEMASK_FREE(nodes_allowed);
1396 nodes_allowed = &node_states[N_HIGH_MEMORY];
1398 } else if (nodes_allowed) {
1400 * per node hstate attribute: adjust count to global,
1401 * but restrict alloc/free to the specified node.
1403 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1404 init_nodemask_of_node(nodes_allowed, nid);
1406 nodes_allowed = &node_states[N_HIGH_MEMORY];
1408 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1410 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1411 NODEMASK_FREE(nodes_allowed);
1416 static ssize_t nr_hugepages_show(struct kobject *kobj,
1417 struct kobj_attribute *attr, char *buf)
1419 return nr_hugepages_show_common(kobj, attr, buf);
1422 static ssize_t nr_hugepages_store(struct kobject *kobj,
1423 struct kobj_attribute *attr, const char *buf, size_t len)
1425 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1427 HSTATE_ATTR(nr_hugepages);
1432 * hstate attribute for optionally mempolicy-based constraint on persistent
1433 * huge page alloc/free.
1435 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1436 struct kobj_attribute *attr, char *buf)
1438 return nr_hugepages_show_common(kobj, attr, buf);
1441 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1442 struct kobj_attribute *attr, const char *buf, size_t len)
1444 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1446 HSTATE_ATTR(nr_hugepages_mempolicy);
1450 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1451 struct kobj_attribute *attr, char *buf)
1453 struct hstate *h = kobj_to_hstate(kobj, NULL);
1454 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1456 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1457 struct kobj_attribute *attr, const char *buf, size_t count)
1460 unsigned long input;
1461 struct hstate *h = kobj_to_hstate(kobj, NULL);
1463 err = strict_strtoul(buf, 10, &input);
1467 spin_lock(&hugetlb_lock);
1468 h->nr_overcommit_huge_pages = input;
1469 spin_unlock(&hugetlb_lock);
1473 HSTATE_ATTR(nr_overcommit_hugepages);
1475 static ssize_t free_hugepages_show(struct kobject *kobj,
1476 struct kobj_attribute *attr, char *buf)
1479 unsigned long free_huge_pages;
1482 h = kobj_to_hstate(kobj, &nid);
1483 if (nid == NUMA_NO_NODE)
1484 free_huge_pages = h->free_huge_pages;
1486 free_huge_pages = h->free_huge_pages_node[nid];
1488 return sprintf(buf, "%lu\n", free_huge_pages);
1490 HSTATE_ATTR_RO(free_hugepages);
1492 static ssize_t resv_hugepages_show(struct kobject *kobj,
1493 struct kobj_attribute *attr, char *buf)
1495 struct hstate *h = kobj_to_hstate(kobj, NULL);
1496 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1498 HSTATE_ATTR_RO(resv_hugepages);
1500 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1501 struct kobj_attribute *attr, char *buf)
1504 unsigned long surplus_huge_pages;
1507 h = kobj_to_hstate(kobj, &nid);
1508 if (nid == NUMA_NO_NODE)
1509 surplus_huge_pages = h->surplus_huge_pages;
1511 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1513 return sprintf(buf, "%lu\n", surplus_huge_pages);
1515 HSTATE_ATTR_RO(surplus_hugepages);
1517 static struct attribute *hstate_attrs[] = {
1518 &nr_hugepages_attr.attr,
1519 &nr_overcommit_hugepages_attr.attr,
1520 &free_hugepages_attr.attr,
1521 &resv_hugepages_attr.attr,
1522 &surplus_hugepages_attr.attr,
1524 &nr_hugepages_mempolicy_attr.attr,
1529 static struct attribute_group hstate_attr_group = {
1530 .attrs = hstate_attrs,
1533 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1534 struct kobject **hstate_kobjs,
1535 struct attribute_group *hstate_attr_group)
1538 int hi = h - hstates;
1540 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1541 if (!hstate_kobjs[hi])
1544 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1546 kobject_put(hstate_kobjs[hi]);
1551 static void __init hugetlb_sysfs_init(void)
1556 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1557 if (!hugepages_kobj)
1560 for_each_hstate(h) {
1561 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1562 hstate_kobjs, &hstate_attr_group);
1564 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1572 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1573 * with node sysdevs in node_devices[] using a parallel array. The array
1574 * index of a node sysdev or _hstate == node id.
1575 * This is here to avoid any static dependency of the node sysdev driver, in
1576 * the base kernel, on the hugetlb module.
1578 struct node_hstate {
1579 struct kobject *hugepages_kobj;
1580 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1582 struct node_hstate node_hstates[MAX_NUMNODES];
1585 * A subset of global hstate attributes for node sysdevs
1587 static struct attribute *per_node_hstate_attrs[] = {
1588 &nr_hugepages_attr.attr,
1589 &free_hugepages_attr.attr,
1590 &surplus_hugepages_attr.attr,
1594 static struct attribute_group per_node_hstate_attr_group = {
1595 .attrs = per_node_hstate_attrs,
1599 * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj.
1600 * Returns node id via non-NULL nidp.
1602 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1606 for (nid = 0; nid < nr_node_ids; nid++) {
1607 struct node_hstate *nhs = &node_hstates[nid];
1609 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1610 if (nhs->hstate_kobjs[i] == kobj) {
1622 * Unregister hstate attributes from a single node sysdev.
1623 * No-op if no hstate attributes attached.
1625 void hugetlb_unregister_node(struct node *node)
1628 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1630 if (!nhs->hugepages_kobj)
1631 return; /* no hstate attributes */
1634 if (nhs->hstate_kobjs[h - hstates]) {
1635 kobject_put(nhs->hstate_kobjs[h - hstates]);
1636 nhs->hstate_kobjs[h - hstates] = NULL;
1639 kobject_put(nhs->hugepages_kobj);
1640 nhs->hugepages_kobj = NULL;
1644 * hugetlb module exit: unregister hstate attributes from node sysdevs
1647 static void hugetlb_unregister_all_nodes(void)
1652 * disable node sysdev registrations.
1654 register_hugetlbfs_with_node(NULL, NULL);
1657 * remove hstate attributes from any nodes that have them.
1659 for (nid = 0; nid < nr_node_ids; nid++)
1660 hugetlb_unregister_node(&node_devices[nid]);
1664 * Register hstate attributes for a single node sysdev.
1665 * No-op if attributes already registered.
1667 void hugetlb_register_node(struct node *node)
1670 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1673 if (nhs->hugepages_kobj)
1674 return; /* already allocated */
1676 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1677 &node->sysdev.kobj);
1678 if (!nhs->hugepages_kobj)
1681 for_each_hstate(h) {
1682 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1684 &per_node_hstate_attr_group);
1686 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1688 h->name, node->sysdev.id);
1689 hugetlb_unregister_node(node);
1696 * hugetlb init time: register hstate attributes for all registered node
1697 * sysdevs of nodes that have memory. All on-line nodes should have
1698 * registered their associated sysdev by this time.
1700 static void hugetlb_register_all_nodes(void)
1704 for_each_node_state(nid, N_HIGH_MEMORY) {
1705 struct node *node = &node_devices[nid];
1706 if (node->sysdev.id == nid)
1707 hugetlb_register_node(node);
1711 * Let the node sysdev driver know we're here so it can
1712 * [un]register hstate attributes on node hotplug.
1714 register_hugetlbfs_with_node(hugetlb_register_node,
1715 hugetlb_unregister_node);
1717 #else /* !CONFIG_NUMA */
1719 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1727 static void hugetlb_unregister_all_nodes(void) { }
1729 static void hugetlb_register_all_nodes(void) { }
1733 static void __exit hugetlb_exit(void)
1737 hugetlb_unregister_all_nodes();
1739 for_each_hstate(h) {
1740 kobject_put(hstate_kobjs[h - hstates]);
1743 kobject_put(hugepages_kobj);
1745 module_exit(hugetlb_exit);
1747 static int __init hugetlb_init(void)
1749 /* Some platform decide whether they support huge pages at boot
1750 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1751 * there is no such support
1753 if (HPAGE_SHIFT == 0)
1756 if (!size_to_hstate(default_hstate_size)) {
1757 default_hstate_size = HPAGE_SIZE;
1758 if (!size_to_hstate(default_hstate_size))
1759 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1761 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1762 if (default_hstate_max_huge_pages)
1763 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1765 hugetlb_init_hstates();
1767 gather_bootmem_prealloc();
1771 hugetlb_sysfs_init();
1773 hugetlb_register_all_nodes();
1777 module_init(hugetlb_init);
1779 /* Should be called on processing a hugepagesz=... option */
1780 void __init hugetlb_add_hstate(unsigned order)
1785 if (size_to_hstate(PAGE_SIZE << order)) {
1786 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1789 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1791 h = &hstates[max_hstate++];
1793 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1794 h->nr_huge_pages = 0;
1795 h->free_huge_pages = 0;
1796 for (i = 0; i < MAX_NUMNODES; ++i)
1797 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1798 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1799 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1800 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1801 huge_page_size(h)/1024);
1806 static int __init hugetlb_nrpages_setup(char *s)
1809 static unsigned long *last_mhp;
1812 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1813 * so this hugepages= parameter goes to the "default hstate".
1816 mhp = &default_hstate_max_huge_pages;
1818 mhp = &parsed_hstate->max_huge_pages;
1820 if (mhp == last_mhp) {
1821 printk(KERN_WARNING "hugepages= specified twice without "
1822 "interleaving hugepagesz=, ignoring\n");
1826 if (sscanf(s, "%lu", mhp) <= 0)
1830 * Global state is always initialized later in hugetlb_init.
1831 * But we need to allocate >= MAX_ORDER hstates here early to still
1832 * use the bootmem allocator.
1834 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1835 hugetlb_hstate_alloc_pages(parsed_hstate);
1841 __setup("hugepages=", hugetlb_nrpages_setup);
1843 static int __init hugetlb_default_setup(char *s)
1845 default_hstate_size = memparse(s, &s);
1848 __setup("default_hugepagesz=", hugetlb_default_setup);
1850 static unsigned int cpuset_mems_nr(unsigned int *array)
1853 unsigned int nr = 0;
1855 for_each_node_mask(node, cpuset_current_mems_allowed)
1861 #ifdef CONFIG_SYSCTL
1862 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1863 struct ctl_table *table, int write,
1864 void __user *buffer, size_t *length, loff_t *ppos)
1866 struct hstate *h = &default_hstate;
1870 tmp = h->max_huge_pages;
1873 table->maxlen = sizeof(unsigned long);
1874 proc_doulongvec_minmax(table, write, buffer, length, ppos);
1877 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1878 GFP_KERNEL | __GFP_NORETRY);
1879 if (!(obey_mempolicy &&
1880 init_nodemask_of_mempolicy(nodes_allowed))) {
1881 NODEMASK_FREE(nodes_allowed);
1882 nodes_allowed = &node_states[N_HIGH_MEMORY];
1884 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1886 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1887 NODEMASK_FREE(nodes_allowed);
1893 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1894 void __user *buffer, size_t *length, loff_t *ppos)
1897 return hugetlb_sysctl_handler_common(false, table, write,
1898 buffer, length, ppos);
1902 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
1903 void __user *buffer, size_t *length, loff_t *ppos)
1905 return hugetlb_sysctl_handler_common(true, table, write,
1906 buffer, length, ppos);
1908 #endif /* CONFIG_NUMA */
1910 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1911 void __user *buffer,
1912 size_t *length, loff_t *ppos)
1914 proc_dointvec(table, write, buffer, length, ppos);
1915 if (hugepages_treat_as_movable)
1916 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1918 htlb_alloc_mask = GFP_HIGHUSER;
1922 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1923 void __user *buffer,
1924 size_t *length, loff_t *ppos)
1926 struct hstate *h = &default_hstate;
1930 tmp = h->nr_overcommit_huge_pages;
1933 table->maxlen = sizeof(unsigned long);
1934 proc_doulongvec_minmax(table, write, buffer, length, ppos);
1937 spin_lock(&hugetlb_lock);
1938 h->nr_overcommit_huge_pages = tmp;
1939 spin_unlock(&hugetlb_lock);
1945 #endif /* CONFIG_SYSCTL */
1947 void hugetlb_report_meminfo(struct seq_file *m)
1949 struct hstate *h = &default_hstate;
1951 "HugePages_Total: %5lu\n"
1952 "HugePages_Free: %5lu\n"
1953 "HugePages_Rsvd: %5lu\n"
1954 "HugePages_Surp: %5lu\n"
1955 "Hugepagesize: %8lu kB\n",
1959 h->surplus_huge_pages,
1960 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1963 int hugetlb_report_node_meminfo(int nid, char *buf)
1965 struct hstate *h = &default_hstate;
1967 "Node %d HugePages_Total: %5u\n"
1968 "Node %d HugePages_Free: %5u\n"
1969 "Node %d HugePages_Surp: %5u\n",
1970 nid, h->nr_huge_pages_node[nid],
1971 nid, h->free_huge_pages_node[nid],
1972 nid, h->surplus_huge_pages_node[nid]);
1975 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1976 unsigned long hugetlb_total_pages(void)
1978 struct hstate *h = &default_hstate;
1979 return h->nr_huge_pages * pages_per_huge_page(h);
1982 static int hugetlb_acct_memory(struct hstate *h, long delta)
1986 spin_lock(&hugetlb_lock);
1988 * When cpuset is configured, it breaks the strict hugetlb page
1989 * reservation as the accounting is done on a global variable. Such
1990 * reservation is completely rubbish in the presence of cpuset because
1991 * the reservation is not checked against page availability for the
1992 * current cpuset. Application can still potentially OOM'ed by kernel
1993 * with lack of free htlb page in cpuset that the task is in.
1994 * Attempt to enforce strict accounting with cpuset is almost
1995 * impossible (or too ugly) because cpuset is too fluid that
1996 * task or memory node can be dynamically moved between cpusets.
1998 * The change of semantics for shared hugetlb mapping with cpuset is
1999 * undesirable. However, in order to preserve some of the semantics,
2000 * we fall back to check against current free page availability as
2001 * a best attempt and hopefully to minimize the impact of changing
2002 * semantics that cpuset has.
2005 if (gather_surplus_pages(h, delta) < 0)
2008 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2009 return_unused_surplus_pages(h, delta);
2016 return_unused_surplus_pages(h, (unsigned long) -delta);
2019 spin_unlock(&hugetlb_lock);
2023 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2025 struct resv_map *reservations = vma_resv_map(vma);
2028 * This new VMA should share its siblings reservation map if present.
2029 * The VMA will only ever have a valid reservation map pointer where
2030 * it is being copied for another still existing VMA. As that VMA
2031 * has a reference to the reservation map it cannot dissappear until
2032 * after this open call completes. It is therefore safe to take a
2033 * new reference here without additional locking.
2036 kref_get(&reservations->refs);
2039 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2041 struct hstate *h = hstate_vma(vma);
2042 struct resv_map *reservations = vma_resv_map(vma);
2043 unsigned long reserve;
2044 unsigned long start;
2048 start = vma_hugecache_offset(h, vma, vma->vm_start);
2049 end = vma_hugecache_offset(h, vma, vma->vm_end);
2051 reserve = (end - start) -
2052 region_count(&reservations->regions, start, end);
2054 kref_put(&reservations->refs, resv_map_release);
2057 hugetlb_acct_memory(h, -reserve);
2058 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
2064 * We cannot handle pagefaults against hugetlb pages at all. They cause
2065 * handle_mm_fault() to try to instantiate regular-sized pages in the
2066 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2069 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2075 const struct vm_operations_struct hugetlb_vm_ops = {
2076 .fault = hugetlb_vm_op_fault,
2077 .open = hugetlb_vm_op_open,
2078 .close = hugetlb_vm_op_close,
2081 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2088 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2090 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2092 entry = pte_mkyoung(entry);
2093 entry = pte_mkhuge(entry);
2098 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2099 unsigned long address, pte_t *ptep)
2103 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2104 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
2105 update_mmu_cache(vma, address, ptep);
2110 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2111 struct vm_area_struct *vma)
2113 pte_t *src_pte, *dst_pte, entry;
2114 struct page *ptepage;
2117 struct hstate *h = hstate_vma(vma);
2118 unsigned long sz = huge_page_size(h);
2120 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2122 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2123 src_pte = huge_pte_offset(src, addr);
2126 dst_pte = huge_pte_alloc(dst, addr, sz);
2130 /* If the pagetables are shared don't copy or take references */
2131 if (dst_pte == src_pte)
2134 spin_lock(&dst->page_table_lock);
2135 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2136 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2138 huge_ptep_set_wrprotect(src, addr, src_pte);
2139 entry = huge_ptep_get(src_pte);
2140 ptepage = pte_page(entry);
2142 page_dup_rmap(ptepage);
2143 set_huge_pte_at(dst, addr, dst_pte, entry);
2145 spin_unlock(&src->page_table_lock);
2146 spin_unlock(&dst->page_table_lock);
2154 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2158 if (huge_pte_none(pte) || pte_present(pte))
2160 swp = pte_to_swp_entry(pte);
2161 if (non_swap_entry(swp) && is_hwpoison_entry(swp)) {
2167 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2168 unsigned long end, struct page *ref_page)
2170 struct mm_struct *mm = vma->vm_mm;
2171 unsigned long address;
2176 struct hstate *h = hstate_vma(vma);
2177 unsigned long sz = huge_page_size(h);
2180 * A page gathering list, protected by per file i_mmap_lock. The
2181 * lock is used to avoid list corruption from multiple unmapping
2182 * of the same page since we are using page->lru.
2184 LIST_HEAD(page_list);
2186 WARN_ON(!is_vm_hugetlb_page(vma));
2187 BUG_ON(start & ~huge_page_mask(h));
2188 BUG_ON(end & ~huge_page_mask(h));
2190 mmu_notifier_invalidate_range_start(mm, start, end);
2191 spin_lock(&mm->page_table_lock);
2192 for (address = start; address < end; address += sz) {
2193 ptep = huge_pte_offset(mm, address);
2197 if (huge_pmd_unshare(mm, &address, ptep))
2201 * If a reference page is supplied, it is because a specific
2202 * page is being unmapped, not a range. Ensure the page we
2203 * are about to unmap is the actual page of interest.
2206 pte = huge_ptep_get(ptep);
2207 if (huge_pte_none(pte))
2209 page = pte_page(pte);
2210 if (page != ref_page)
2214 * Mark the VMA as having unmapped its page so that
2215 * future faults in this VMA will fail rather than
2216 * looking like data was lost
2218 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2221 pte = huge_ptep_get_and_clear(mm, address, ptep);
2222 if (huge_pte_none(pte))
2226 * HWPoisoned hugepage is already unmapped and dropped reference
2228 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2231 page = pte_page(pte);
2233 set_page_dirty(page);
2234 list_add(&page->lru, &page_list);
2236 spin_unlock(&mm->page_table_lock);
2237 flush_tlb_range(vma, start, end);
2238 mmu_notifier_invalidate_range_end(mm, start, end);
2239 list_for_each_entry_safe(page, tmp, &page_list, lru) {
2240 page_remove_rmap(page);
2241 list_del(&page->lru);
2246 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2247 unsigned long end, struct page *ref_page)
2249 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2250 __unmap_hugepage_range(vma, start, end, ref_page);
2251 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2255 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2256 * mappping it owns the reserve page for. The intention is to unmap the page
2257 * from other VMAs and let the children be SIGKILLed if they are faulting the
2260 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2261 struct page *page, unsigned long address)
2263 struct hstate *h = hstate_vma(vma);
2264 struct vm_area_struct *iter_vma;
2265 struct address_space *mapping;
2266 struct prio_tree_iter iter;
2270 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2271 * from page cache lookup which is in HPAGE_SIZE units.
2273 address = address & huge_page_mask(h);
2274 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
2275 + (vma->vm_pgoff >> PAGE_SHIFT);
2276 mapping = (struct address_space *)page_private(page);
2279 * Take the mapping lock for the duration of the table walk. As
2280 * this mapping should be shared between all the VMAs,
2281 * __unmap_hugepage_range() is called as the lock is already held
2283 spin_lock(&mapping->i_mmap_lock);
2284 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2285 /* Do not unmap the current VMA */
2286 if (iter_vma == vma)
2290 * Unmap the page from other VMAs without their own reserves.
2291 * They get marked to be SIGKILLed if they fault in these
2292 * areas. This is because a future no-page fault on this VMA
2293 * could insert a zeroed page instead of the data existing
2294 * from the time of fork. This would look like data corruption
2296 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2297 __unmap_hugepage_range(iter_vma,
2298 address, address + huge_page_size(h),
2301 spin_unlock(&mapping->i_mmap_lock);
2307 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2309 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2310 unsigned long address, pte_t *ptep, pte_t pte,
2311 struct page *pagecache_page)
2313 struct hstate *h = hstate_vma(vma);
2314 struct page *old_page, *new_page;
2316 int outside_reserve = 0;
2318 old_page = pte_page(pte);
2321 /* If no-one else is actually using this page, avoid the copy
2322 * and just make the page writable */
2323 avoidcopy = (page_mapcount(old_page) == 1);
2325 if (!trylock_page(old_page))
2326 if (PageAnon(old_page))
2327 page_move_anon_rmap(old_page, vma, address);
2328 set_huge_ptep_writable(vma, address, ptep);
2333 * If the process that created a MAP_PRIVATE mapping is about to
2334 * perform a COW due to a shared page count, attempt to satisfy
2335 * the allocation without using the existing reserves. The pagecache
2336 * page is used to determine if the reserve at this address was
2337 * consumed or not. If reserves were used, a partial faulted mapping
2338 * at the time of fork() could consume its reserves on COW instead
2339 * of the full address range.
2341 if (!(vma->vm_flags & VM_MAYSHARE) &&
2342 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2343 old_page != pagecache_page)
2344 outside_reserve = 1;
2346 page_cache_get(old_page);
2348 /* Drop page_table_lock as buddy allocator may be called */
2349 spin_unlock(&mm->page_table_lock);
2350 new_page = alloc_huge_page(vma, address, outside_reserve);
2352 if (IS_ERR(new_page)) {
2353 page_cache_release(old_page);
2356 * If a process owning a MAP_PRIVATE mapping fails to COW,
2357 * it is due to references held by a child and an insufficient
2358 * huge page pool. To guarantee the original mappers
2359 * reliability, unmap the page from child processes. The child
2360 * may get SIGKILLed if it later faults.
2362 if (outside_reserve) {
2363 BUG_ON(huge_pte_none(pte));
2364 if (unmap_ref_private(mm, vma, old_page, address)) {
2365 BUG_ON(page_count(old_page) != 1);
2366 BUG_ON(huge_pte_none(pte));
2367 spin_lock(&mm->page_table_lock);
2368 goto retry_avoidcopy;
2373 /* Caller expects lock to be held */
2374 spin_lock(&mm->page_table_lock);
2375 return -PTR_ERR(new_page);
2379 * When the original hugepage is shared one, it does not have
2380 * anon_vma prepared.
2382 if (unlikely(anon_vma_prepare(vma)))
2383 return VM_FAULT_OOM;
2385 copy_huge_page(new_page, old_page, address, vma);
2386 __SetPageUptodate(new_page);
2389 * Retake the page_table_lock to check for racing updates
2390 * before the page tables are altered
2392 spin_lock(&mm->page_table_lock);
2393 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2394 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2396 huge_ptep_clear_flush(vma, address, ptep);
2397 set_huge_pte_at(mm, address, ptep,
2398 make_huge_pte(vma, new_page, 1));
2399 page_remove_rmap(old_page);
2400 hugepage_add_anon_rmap(new_page, vma, address);
2401 /* Make the old page be freed below */
2402 new_page = old_page;
2404 page_cache_release(new_page);
2405 page_cache_release(old_page);
2409 /* Return the pagecache page at a given address within a VMA */
2410 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2411 struct vm_area_struct *vma, unsigned long address)
2413 struct address_space *mapping;
2416 mapping = vma->vm_file->f_mapping;
2417 idx = vma_hugecache_offset(h, vma, address);
2419 return find_lock_page(mapping, idx);
2423 * Return whether there is a pagecache page to back given address within VMA.
2424 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2426 static bool hugetlbfs_pagecache_present(struct hstate *h,
2427 struct vm_area_struct *vma, unsigned long address)
2429 struct address_space *mapping;
2433 mapping = vma->vm_file->f_mapping;
2434 idx = vma_hugecache_offset(h, vma, address);
2436 page = find_get_page(mapping, idx);
2439 return page != NULL;
2442 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2443 unsigned long address, pte_t *ptep, unsigned int flags)
2445 struct hstate *h = hstate_vma(vma);
2446 int ret = VM_FAULT_SIGBUS;
2450 struct address_space *mapping;
2454 * Currently, we are forced to kill the process in the event the
2455 * original mapper has unmapped pages from the child due to a failed
2456 * COW. Warn that such a situation has occured as it may not be obvious
2458 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2460 "PID %d killed due to inadequate hugepage pool\n",
2465 mapping = vma->vm_file->f_mapping;
2466 idx = vma_hugecache_offset(h, vma, address);
2469 * Use page lock to guard against racing truncation
2470 * before we get page_table_lock.
2473 page = find_lock_page(mapping, idx);
2475 size = i_size_read(mapping->host) >> huge_page_shift(h);
2478 page = alloc_huge_page(vma, address, 0);
2480 ret = -PTR_ERR(page);
2483 clear_huge_page(page, address, huge_page_size(h));
2484 __SetPageUptodate(page);
2486 if (vma->vm_flags & VM_MAYSHARE) {
2488 struct inode *inode = mapping->host;
2490 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2498 spin_lock(&inode->i_lock);
2499 inode->i_blocks += blocks_per_huge_page(h);
2500 spin_unlock(&inode->i_lock);
2501 page_dup_rmap(page);
2504 if (unlikely(anon_vma_prepare(vma))) {
2506 goto backout_unlocked;
2508 hugepage_add_new_anon_rmap(page, vma, address);
2511 page_dup_rmap(page);
2515 * Since memory error handler replaces pte into hwpoison swap entry
2516 * at the time of error handling, a process which reserved but not have
2517 * the mapping to the error hugepage does not have hwpoison swap entry.
2518 * So we need to block accesses from such a process by checking
2519 * PG_hwpoison bit here.
2521 if (unlikely(PageHWPoison(page))) {
2522 ret = VM_FAULT_HWPOISON;
2523 goto backout_unlocked;
2527 * If we are going to COW a private mapping later, we examine the
2528 * pending reservations for this page now. This will ensure that
2529 * any allocations necessary to record that reservation occur outside
2532 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2533 if (vma_needs_reservation(h, vma, address) < 0) {
2535 goto backout_unlocked;
2538 spin_lock(&mm->page_table_lock);
2539 size = i_size_read(mapping->host) >> huge_page_shift(h);
2544 if (!huge_pte_none(huge_ptep_get(ptep)))
2547 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2548 && (vma->vm_flags & VM_SHARED)));
2549 set_huge_pte_at(mm, address, ptep, new_pte);
2551 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2552 /* Optimization, do the COW without a second fault */
2553 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2556 spin_unlock(&mm->page_table_lock);
2562 spin_unlock(&mm->page_table_lock);
2569 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2570 unsigned long address, unsigned int flags)
2575 struct page *page = NULL;
2576 struct page *pagecache_page = NULL;
2577 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2578 struct hstate *h = hstate_vma(vma);
2580 ptep = huge_pte_offset(mm, address);
2582 entry = huge_ptep_get(ptep);
2583 if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2584 return VM_FAULT_HWPOISON;
2587 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2589 return VM_FAULT_OOM;
2592 * Serialize hugepage allocation and instantiation, so that we don't
2593 * get spurious allocation failures if two CPUs race to instantiate
2594 * the same page in the page cache.
2596 mutex_lock(&hugetlb_instantiation_mutex);
2597 entry = huge_ptep_get(ptep);
2598 if (huge_pte_none(entry)) {
2599 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2606 * If we are going to COW the mapping later, we examine the pending
2607 * reservations for this page now. This will ensure that any
2608 * allocations necessary to record that reservation occur outside the
2609 * spinlock. For private mappings, we also lookup the pagecache
2610 * page now as it is used to determine if a reservation has been
2613 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2614 if (vma_needs_reservation(h, vma, address) < 0) {
2619 if (!(vma->vm_flags & VM_MAYSHARE))
2620 pagecache_page = hugetlbfs_pagecache_page(h,
2624 if (!pagecache_page) {
2625 page = pte_page(entry);
2629 spin_lock(&mm->page_table_lock);
2630 /* Check for a racing update before calling hugetlb_cow */
2631 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2632 goto out_page_table_lock;
2635 if (flags & FAULT_FLAG_WRITE) {
2636 if (!pte_write(entry)) {
2637 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2639 goto out_page_table_lock;
2641 entry = pte_mkdirty(entry);
2643 entry = pte_mkyoung(entry);
2644 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2645 flags & FAULT_FLAG_WRITE))
2646 update_mmu_cache(vma, address, ptep);
2648 out_page_table_lock:
2649 spin_unlock(&mm->page_table_lock);
2651 if (pagecache_page) {
2652 unlock_page(pagecache_page);
2653 put_page(pagecache_page);
2659 mutex_unlock(&hugetlb_instantiation_mutex);
2664 /* Can be overriden by architectures */
2665 __attribute__((weak)) struct page *
2666 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2667 pud_t *pud, int write)
2673 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2674 struct page **pages, struct vm_area_struct **vmas,
2675 unsigned long *position, int *length, int i,
2678 unsigned long pfn_offset;
2679 unsigned long vaddr = *position;
2680 int remainder = *length;
2681 struct hstate *h = hstate_vma(vma);
2683 spin_lock(&mm->page_table_lock);
2684 while (vaddr < vma->vm_end && remainder) {
2690 * Some archs (sparc64, sh*) have multiple pte_ts to
2691 * each hugepage. We have to make sure we get the
2692 * first, for the page indexing below to work.
2694 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2695 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2698 * When coredumping, it suits get_dump_page if we just return
2699 * an error where there's an empty slot with no huge pagecache
2700 * to back it. This way, we avoid allocating a hugepage, and
2701 * the sparse dumpfile avoids allocating disk blocks, but its
2702 * huge holes still show up with zeroes where they need to be.
2704 if (absent && (flags & FOLL_DUMP) &&
2705 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2711 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2714 spin_unlock(&mm->page_table_lock);
2715 ret = hugetlb_fault(mm, vma, vaddr,
2716 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2717 spin_lock(&mm->page_table_lock);
2718 if (!(ret & VM_FAULT_ERROR))
2725 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2726 page = pte_page(huge_ptep_get(pte));
2729 pages[i] = mem_map_offset(page, pfn_offset);
2740 if (vaddr < vma->vm_end && remainder &&
2741 pfn_offset < pages_per_huge_page(h)) {
2743 * We use pfn_offset to avoid touching the pageframes
2744 * of this compound page.
2749 spin_unlock(&mm->page_table_lock);
2750 *length = remainder;
2753 return i ? i : -EFAULT;
2756 void hugetlb_change_protection(struct vm_area_struct *vma,
2757 unsigned long address, unsigned long end, pgprot_t newprot)
2759 struct mm_struct *mm = vma->vm_mm;
2760 unsigned long start = address;
2763 struct hstate *h = hstate_vma(vma);
2765 BUG_ON(address >= end);
2766 flush_cache_range(vma, address, end);
2768 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2769 spin_lock(&mm->page_table_lock);
2770 for (; address < end; address += huge_page_size(h)) {
2771 ptep = huge_pte_offset(mm, address);
2774 if (huge_pmd_unshare(mm, &address, ptep))
2776 if (!huge_pte_none(huge_ptep_get(ptep))) {
2777 pte = huge_ptep_get_and_clear(mm, address, ptep);
2778 pte = pte_mkhuge(pte_modify(pte, newprot));
2779 set_huge_pte_at(mm, address, ptep, pte);
2782 spin_unlock(&mm->page_table_lock);
2783 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2785 flush_tlb_range(vma, start, end);
2788 int hugetlb_reserve_pages(struct inode *inode,
2790 struct vm_area_struct *vma,
2794 struct hstate *h = hstate_inode(inode);
2797 * Only apply hugepage reservation if asked. At fault time, an
2798 * attempt will be made for VM_NORESERVE to allocate a page
2799 * and filesystem quota without using reserves
2801 if (acctflag & VM_NORESERVE)
2805 * Shared mappings base their reservation on the number of pages that
2806 * are already allocated on behalf of the file. Private mappings need
2807 * to reserve the full area even if read-only as mprotect() may be
2808 * called to make the mapping read-write. Assume !vma is a shm mapping
2810 if (!vma || vma->vm_flags & VM_MAYSHARE)
2811 chg = region_chg(&inode->i_mapping->private_list, from, to);
2813 struct resv_map *resv_map = resv_map_alloc();
2819 set_vma_resv_map(vma, resv_map);
2820 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2826 /* There must be enough filesystem quota for the mapping */
2827 if (hugetlb_get_quota(inode->i_mapping, chg))
2831 * Check enough hugepages are available for the reservation.
2832 * Hand back the quota if there are not
2834 ret = hugetlb_acct_memory(h, chg);
2836 hugetlb_put_quota(inode->i_mapping, chg);
2841 * Account for the reservations made. Shared mappings record regions
2842 * that have reservations as they are shared by multiple VMAs.
2843 * When the last VMA disappears, the region map says how much
2844 * the reservation was and the page cache tells how much of
2845 * the reservation was consumed. Private mappings are per-VMA and
2846 * only the consumed reservations are tracked. When the VMA
2847 * disappears, the original reservation is the VMA size and the
2848 * consumed reservations are stored in the map. Hence, nothing
2849 * else has to be done for private mappings here
2851 if (!vma || vma->vm_flags & VM_MAYSHARE)
2852 region_add(&inode->i_mapping->private_list, from, to);
2856 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2858 struct hstate *h = hstate_inode(inode);
2859 long chg = region_truncate(&inode->i_mapping->private_list, offset);
2861 spin_lock(&inode->i_lock);
2862 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
2863 spin_unlock(&inode->i_lock);
2865 hugetlb_put_quota(inode->i_mapping, (chg - freed));
2866 hugetlb_acct_memory(h, -(chg - freed));
2870 * This function is called from memory failure code.
2871 * Assume the caller holds page lock of the head page.
2873 void __isolate_hwpoisoned_huge_page(struct page *hpage)
2875 struct hstate *h = page_hstate(hpage);
2876 int nid = page_to_nid(hpage);
2878 spin_lock(&hugetlb_lock);
2879 list_del(&hpage->lru);
2880 h->free_huge_pages--;
2881 h->free_huge_pages_node[nid]--;
2882 spin_unlock(&hugetlb_lock);