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>
27 #include <linux/node.h>
30 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
31 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
32 unsigned long hugepages_treat_as_movable;
34 static int max_hstate;
35 unsigned int default_hstate_idx;
36 struct hstate hstates[HUGE_MAX_HSTATE];
38 __initdata LIST_HEAD(huge_boot_pages);
40 /* for command line parsing */
41 static struct hstate * __initdata parsed_hstate;
42 static unsigned long __initdata default_hstate_max_huge_pages;
43 static unsigned long __initdata default_hstate_size;
45 #define for_each_hstate(h) \
46 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
49 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
51 static DEFINE_SPINLOCK(hugetlb_lock);
54 * Region tracking -- allows tracking of reservations and instantiated pages
55 * across the pages in a mapping.
57 * The region data structures are protected by a combination of the mmap_sem
58 * and the hugetlb_instantion_mutex. To access or modify a region the caller
59 * must either hold the mmap_sem for write, or the mmap_sem for read and
60 * the hugetlb_instantiation mutex:
62 * down_write(&mm->mmap_sem);
64 * down_read(&mm->mmap_sem);
65 * mutex_lock(&hugetlb_instantiation_mutex);
68 struct list_head link;
73 static long region_add(struct list_head *head, long f, long t)
75 struct file_region *rg, *nrg, *trg;
77 /* Locate the region we are either in or before. */
78 list_for_each_entry(rg, head, link)
82 /* Round our left edge to the current segment if it encloses us. */
86 /* Check for and consume any regions we now overlap with. */
88 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
89 if (&rg->link == head)
94 /* If this area reaches higher then extend our area to
95 * include it completely. If this is not the first area
96 * which we intend to reuse, free it. */
109 static long region_chg(struct list_head *head, long f, long t)
111 struct file_region *rg, *nrg;
114 /* Locate the region we are before or in. */
115 list_for_each_entry(rg, head, link)
119 /* If we are below the current region then a new region is required.
120 * Subtle, allocate a new region at the position but make it zero
121 * size such that we can guarantee to record the reservation. */
122 if (&rg->link == head || t < rg->from) {
123 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
128 INIT_LIST_HEAD(&nrg->link);
129 list_add(&nrg->link, rg->link.prev);
134 /* Round our left edge to the current segment if it encloses us. */
139 /* Check for and consume any regions we now overlap with. */
140 list_for_each_entry(rg, rg->link.prev, link) {
141 if (&rg->link == head)
146 /* We overlap with this area, if it extends futher than
147 * us then we must extend ourselves. Account for its
148 * existing reservation. */
153 chg -= rg->to - rg->from;
158 static long region_truncate(struct list_head *head, long end)
160 struct file_region *rg, *trg;
163 /* Locate the region we are either in or before. */
164 list_for_each_entry(rg, head, link)
167 if (&rg->link == head)
170 /* If we are in the middle of a region then adjust it. */
171 if (end > rg->from) {
174 rg = list_entry(rg->link.next, typeof(*rg), link);
177 /* Drop any remaining regions. */
178 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
179 if (&rg->link == head)
181 chg += rg->to - rg->from;
188 static long region_count(struct list_head *head, long f, long t)
190 struct file_region *rg;
193 /* Locate each segment we overlap with, and count that overlap. */
194 list_for_each_entry(rg, head, link) {
203 seg_from = max(rg->from, f);
204 seg_to = min(rg->to, t);
206 chg += seg_to - seg_from;
213 * Convert the address within this vma to the page offset within
214 * the mapping, in pagecache page units; huge pages here.
216 static pgoff_t vma_hugecache_offset(struct hstate *h,
217 struct vm_area_struct *vma, unsigned long address)
219 return ((address - vma->vm_start) >> huge_page_shift(h)) +
220 (vma->vm_pgoff >> huge_page_order(h));
224 * Return the size of the pages allocated when backing a VMA. In the majority
225 * cases this will be same size as used by the page table entries.
227 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
229 struct hstate *hstate;
231 if (!is_vm_hugetlb_page(vma))
234 hstate = hstate_vma(vma);
236 return 1UL << (hstate->order + PAGE_SHIFT);
238 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
241 * Return the page size being used by the MMU to back a VMA. In the majority
242 * of cases, the page size used by the kernel matches the MMU size. On
243 * architectures where it differs, an architecture-specific version of this
244 * function is required.
246 #ifndef vma_mmu_pagesize
247 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
249 return vma_kernel_pagesize(vma);
254 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
255 * bits of the reservation map pointer, which are always clear due to
258 #define HPAGE_RESV_OWNER (1UL << 0)
259 #define HPAGE_RESV_UNMAPPED (1UL << 1)
260 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
263 * These helpers are used to track how many pages are reserved for
264 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
265 * is guaranteed to have their future faults succeed.
267 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
268 * the reserve counters are updated with the hugetlb_lock held. It is safe
269 * to reset the VMA at fork() time as it is not in use yet and there is no
270 * chance of the global counters getting corrupted as a result of the values.
272 * The private mapping reservation is represented in a subtly different
273 * manner to a shared mapping. A shared mapping has a region map associated
274 * with the underlying file, this region map represents the backing file
275 * pages which have ever had a reservation assigned which this persists even
276 * after the page is instantiated. A private mapping has a region map
277 * associated with the original mmap which is attached to all VMAs which
278 * reference it, this region map represents those offsets which have consumed
279 * reservation ie. where pages have been instantiated.
281 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
283 return (unsigned long)vma->vm_private_data;
286 static void set_vma_private_data(struct vm_area_struct *vma,
289 vma->vm_private_data = (void *)value;
294 struct list_head regions;
297 static struct resv_map *resv_map_alloc(void)
299 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
303 kref_init(&resv_map->refs);
304 INIT_LIST_HEAD(&resv_map->regions);
309 static void resv_map_release(struct kref *ref)
311 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
313 /* Clear out any active regions before we release the map. */
314 region_truncate(&resv_map->regions, 0);
318 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
320 VM_BUG_ON(!is_vm_hugetlb_page(vma));
321 if (!(vma->vm_flags & VM_MAYSHARE))
322 return (struct resv_map *)(get_vma_private_data(vma) &
327 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
329 VM_BUG_ON(!is_vm_hugetlb_page(vma));
330 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
332 set_vma_private_data(vma, (get_vma_private_data(vma) &
333 HPAGE_RESV_MASK) | (unsigned long)map);
336 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
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) | flags);
344 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
346 VM_BUG_ON(!is_vm_hugetlb_page(vma));
348 return (get_vma_private_data(vma) & flag) != 0;
351 /* Decrement the reserved pages in the hugepage pool by one */
352 static void decrement_hugepage_resv_vma(struct hstate *h,
353 struct vm_area_struct *vma)
355 if (vma->vm_flags & VM_NORESERVE)
358 if (vma->vm_flags & VM_MAYSHARE) {
359 /* Shared mappings always use reserves */
360 h->resv_huge_pages--;
361 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
363 * Only the process that called mmap() has reserves for
366 h->resv_huge_pages--;
370 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
371 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
373 VM_BUG_ON(!is_vm_hugetlb_page(vma));
374 if (!(vma->vm_flags & VM_MAYSHARE))
375 vma->vm_private_data = (void *)0;
378 /* Returns true if the VMA has associated reserve pages */
379 static int vma_has_reserves(struct vm_area_struct *vma)
381 if (vma->vm_flags & VM_MAYSHARE)
383 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
388 static void clear_gigantic_page(struct page *page,
389 unsigned long addr, unsigned long sz)
392 struct page *p = page;
395 for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) {
397 clear_user_highpage(p, addr + i * PAGE_SIZE);
400 static void clear_huge_page(struct page *page,
401 unsigned long addr, unsigned long sz)
405 if (unlikely(sz > MAX_ORDER_NR_PAGES)) {
406 clear_gigantic_page(page, addr, sz);
411 for (i = 0; i < sz/PAGE_SIZE; i++) {
413 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
417 static void copy_gigantic_page(struct page *dst, struct page *src,
418 unsigned long addr, struct vm_area_struct *vma)
421 struct hstate *h = hstate_vma(vma);
422 struct page *dst_base = dst;
423 struct page *src_base = src;
425 for (i = 0; i < pages_per_huge_page(h); ) {
427 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
430 dst = mem_map_next(dst, dst_base, i);
431 src = mem_map_next(src, src_base, i);
434 static void copy_huge_page(struct page *dst, struct page *src,
435 unsigned long addr, struct vm_area_struct *vma)
438 struct hstate *h = hstate_vma(vma);
440 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
441 copy_gigantic_page(dst, src, addr, vma);
446 for (i = 0; i < pages_per_huge_page(h); i++) {
448 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
452 static void enqueue_huge_page(struct hstate *h, struct page *page)
454 int nid = page_to_nid(page);
455 list_add(&page->lru, &h->hugepage_freelists[nid]);
456 h->free_huge_pages++;
457 h->free_huge_pages_node[nid]++;
460 static struct page *dequeue_huge_page_vma(struct hstate *h,
461 struct vm_area_struct *vma,
462 unsigned long address, int avoid_reserve)
465 struct page *page = NULL;
466 struct mempolicy *mpol;
467 nodemask_t *nodemask;
468 struct zonelist *zonelist = huge_zonelist(vma, address,
469 htlb_alloc_mask, &mpol, &nodemask);
474 * A child process with MAP_PRIVATE mappings created by their parent
475 * have no page reserves. This check ensures that reservations are
476 * not "stolen". The child may still get SIGKILLed
478 if (!vma_has_reserves(vma) &&
479 h->free_huge_pages - h->resv_huge_pages == 0)
482 /* If reserves cannot be used, ensure enough pages are in the pool */
483 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
486 for_each_zone_zonelist_nodemask(zone, z, zonelist,
487 MAX_NR_ZONES - 1, nodemask) {
488 nid = zone_to_nid(zone);
489 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
490 !list_empty(&h->hugepage_freelists[nid])) {
491 page = list_entry(h->hugepage_freelists[nid].next,
493 list_del(&page->lru);
494 h->free_huge_pages--;
495 h->free_huge_pages_node[nid]--;
498 decrement_hugepage_resv_vma(h, vma);
507 static void update_and_free_page(struct hstate *h, struct page *page)
511 VM_BUG_ON(h->order >= MAX_ORDER);
514 h->nr_huge_pages_node[page_to_nid(page)]--;
515 for (i = 0; i < pages_per_huge_page(h); i++) {
516 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
517 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
518 1 << PG_private | 1<< PG_writeback);
520 set_compound_page_dtor(page, NULL);
521 set_page_refcounted(page);
522 arch_release_hugepage(page);
523 __free_pages(page, huge_page_order(h));
526 struct hstate *size_to_hstate(unsigned long size)
531 if (huge_page_size(h) == size)
537 static void free_huge_page(struct page *page)
540 * Can't pass hstate in here because it is called from the
541 * compound page destructor.
543 struct hstate *h = page_hstate(page);
544 int nid = page_to_nid(page);
545 struct address_space *mapping;
547 mapping = (struct address_space *) page_private(page);
548 set_page_private(page, 0);
549 BUG_ON(page_count(page));
550 INIT_LIST_HEAD(&page->lru);
552 spin_lock(&hugetlb_lock);
553 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
554 update_and_free_page(h, page);
555 h->surplus_huge_pages--;
556 h->surplus_huge_pages_node[nid]--;
558 enqueue_huge_page(h, page);
560 spin_unlock(&hugetlb_lock);
562 hugetlb_put_quota(mapping, 1);
565 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
567 set_compound_page_dtor(page, free_huge_page);
568 spin_lock(&hugetlb_lock);
570 h->nr_huge_pages_node[nid]++;
571 spin_unlock(&hugetlb_lock);
572 put_page(page); /* free it into the hugepage allocator */
575 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
578 int nr_pages = 1 << order;
579 struct page *p = page + 1;
581 /* we rely on prep_new_huge_page to set the destructor */
582 set_compound_order(page, order);
584 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
586 p->first_page = page;
590 int PageHuge(struct page *page)
592 compound_page_dtor *dtor;
594 if (!PageCompound(page))
597 page = compound_head(page);
598 dtor = get_compound_page_dtor(page);
600 return dtor == free_huge_page;
603 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
607 if (h->order >= MAX_ORDER)
610 page = alloc_pages_exact_node(nid,
611 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
612 __GFP_REPEAT|__GFP_NOWARN,
615 if (arch_prepare_hugepage(page)) {
616 __free_pages(page, huge_page_order(h));
619 prep_new_huge_page(h, page, nid);
626 * common helper functions for hstate_next_node_to_{alloc|free}.
627 * We may have allocated or freed a huge page based on a different
628 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
629 * be outside of *nodes_allowed. Ensure that we use an allowed
630 * node for alloc or free.
632 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
634 nid = next_node(nid, *nodes_allowed);
635 if (nid == MAX_NUMNODES)
636 nid = first_node(*nodes_allowed);
637 VM_BUG_ON(nid >= MAX_NUMNODES);
642 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
644 if (!node_isset(nid, *nodes_allowed))
645 nid = next_node_allowed(nid, nodes_allowed);
650 * returns the previously saved node ["this node"] from which to
651 * allocate a persistent huge page for the pool and advance the
652 * next node from which to allocate, handling wrap at end of node
655 static int hstate_next_node_to_alloc(struct hstate *h,
656 nodemask_t *nodes_allowed)
660 VM_BUG_ON(!nodes_allowed);
662 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
663 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
668 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
675 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
676 next_nid = start_nid;
679 page = alloc_fresh_huge_page_node(h, next_nid);
684 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
685 } while (next_nid != start_nid);
688 count_vm_event(HTLB_BUDDY_PGALLOC);
690 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
696 * helper for free_pool_huge_page() - return the previously saved
697 * node ["this node"] from which to free a huge page. Advance the
698 * next node id whether or not we find a free huge page to free so
699 * that the next attempt to free addresses the next node.
701 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
705 VM_BUG_ON(!nodes_allowed);
707 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
708 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
714 * Free huge page from pool from next node to free.
715 * Attempt to keep persistent huge pages more or less
716 * balanced over allowed nodes.
717 * Called with hugetlb_lock locked.
719 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
726 start_nid = hstate_next_node_to_free(h, nodes_allowed);
727 next_nid = start_nid;
731 * If we're returning unused surplus pages, only examine
732 * nodes with surplus pages.
734 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
735 !list_empty(&h->hugepage_freelists[next_nid])) {
737 list_entry(h->hugepage_freelists[next_nid].next,
739 list_del(&page->lru);
740 h->free_huge_pages--;
741 h->free_huge_pages_node[next_nid]--;
743 h->surplus_huge_pages--;
744 h->surplus_huge_pages_node[next_nid]--;
746 update_and_free_page(h, page);
750 next_nid = hstate_next_node_to_free(h, nodes_allowed);
751 } while (next_nid != start_nid);
756 static struct page *alloc_buddy_huge_page(struct hstate *h,
757 struct vm_area_struct *vma, unsigned long address)
762 if (h->order >= MAX_ORDER)
766 * Assume we will successfully allocate the surplus page to
767 * prevent racing processes from causing the surplus to exceed
770 * This however introduces a different race, where a process B
771 * tries to grow the static hugepage pool while alloc_pages() is
772 * called by process A. B will only examine the per-node
773 * counters in determining if surplus huge pages can be
774 * converted to normal huge pages in adjust_pool_surplus(). A
775 * won't be able to increment the per-node counter, until the
776 * lock is dropped by B, but B doesn't drop hugetlb_lock until
777 * no more huge pages can be converted from surplus to normal
778 * state (and doesn't try to convert again). Thus, we have a
779 * case where a surplus huge page exists, the pool is grown, and
780 * the surplus huge page still exists after, even though it
781 * should just have been converted to a normal huge page. This
782 * does not leak memory, though, as the hugepage will be freed
783 * once it is out of use. It also does not allow the counters to
784 * go out of whack in adjust_pool_surplus() as we don't modify
785 * the node values until we've gotten the hugepage and only the
786 * per-node value is checked there.
788 spin_lock(&hugetlb_lock);
789 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
790 spin_unlock(&hugetlb_lock);
794 h->surplus_huge_pages++;
796 spin_unlock(&hugetlb_lock);
798 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
799 __GFP_REPEAT|__GFP_NOWARN,
802 if (page && arch_prepare_hugepage(page)) {
803 __free_pages(page, huge_page_order(h));
807 spin_lock(&hugetlb_lock);
810 * This page is now managed by the hugetlb allocator and has
811 * no users -- drop the buddy allocator's reference.
813 put_page_testzero(page);
814 VM_BUG_ON(page_count(page));
815 nid = page_to_nid(page);
816 set_compound_page_dtor(page, free_huge_page);
818 * We incremented the global counters already
820 h->nr_huge_pages_node[nid]++;
821 h->surplus_huge_pages_node[nid]++;
822 __count_vm_event(HTLB_BUDDY_PGALLOC);
825 h->surplus_huge_pages--;
826 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
828 spin_unlock(&hugetlb_lock);
834 * Increase the hugetlb pool such that it can accomodate a reservation
837 static int gather_surplus_pages(struct hstate *h, int delta)
839 struct list_head surplus_list;
840 struct page *page, *tmp;
842 int needed, allocated;
844 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
846 h->resv_huge_pages += delta;
851 INIT_LIST_HEAD(&surplus_list);
855 spin_unlock(&hugetlb_lock);
856 for (i = 0; i < needed; i++) {
857 page = alloc_buddy_huge_page(h, NULL, 0);
860 * We were not able to allocate enough pages to
861 * satisfy the entire reservation so we free what
862 * we've allocated so far.
864 spin_lock(&hugetlb_lock);
869 list_add(&page->lru, &surplus_list);
874 * After retaking hugetlb_lock, we need to recalculate 'needed'
875 * because either resv_huge_pages or free_huge_pages may have changed.
877 spin_lock(&hugetlb_lock);
878 needed = (h->resv_huge_pages + delta) -
879 (h->free_huge_pages + allocated);
884 * The surplus_list now contains _at_least_ the number of extra pages
885 * needed to accomodate the reservation. Add the appropriate number
886 * of pages to the hugetlb pool and free the extras back to the buddy
887 * allocator. Commit the entire reservation here to prevent another
888 * process from stealing the pages as they are added to the pool but
889 * before they are reserved.
892 h->resv_huge_pages += delta;
895 /* Free the needed pages to the hugetlb pool */
896 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
899 list_del(&page->lru);
900 enqueue_huge_page(h, page);
903 /* Free unnecessary surplus pages to the buddy allocator */
904 if (!list_empty(&surplus_list)) {
905 spin_unlock(&hugetlb_lock);
906 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
907 list_del(&page->lru);
909 * The page has a reference count of zero already, so
910 * call free_huge_page directly instead of using
911 * put_page. This must be done with hugetlb_lock
912 * unlocked which is safe because free_huge_page takes
913 * hugetlb_lock before deciding how to free the page.
915 free_huge_page(page);
917 spin_lock(&hugetlb_lock);
924 * When releasing a hugetlb pool reservation, any surplus pages that were
925 * allocated to satisfy the reservation must be explicitly freed if they were
927 * Called with hugetlb_lock held.
929 static void return_unused_surplus_pages(struct hstate *h,
930 unsigned long unused_resv_pages)
932 unsigned long nr_pages;
934 /* Uncommit the reservation */
935 h->resv_huge_pages -= unused_resv_pages;
937 /* Cannot return gigantic pages currently */
938 if (h->order >= MAX_ORDER)
941 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
944 * We want to release as many surplus pages as possible, spread
945 * evenly across all nodes. Iterate across all nodes until we
946 * can no longer free unreserved surplus pages. This occurs when
947 * the nodes with surplus pages have no free pages.
948 * free_pool_huge_page() will balance the the frees across the
949 * on-line nodes for us and will handle the hstate accounting.
952 if (!free_pool_huge_page(h, &node_online_map, 1))
958 * Determine if the huge page at addr within the vma has an associated
959 * reservation. Where it does not we will need to logically increase
960 * reservation and actually increase quota before an allocation can occur.
961 * Where any new reservation would be required the reservation change is
962 * prepared, but not committed. Once the page has been quota'd allocated
963 * an instantiated the change should be committed via vma_commit_reservation.
964 * No action is required on failure.
966 static long vma_needs_reservation(struct hstate *h,
967 struct vm_area_struct *vma, unsigned long addr)
969 struct address_space *mapping = vma->vm_file->f_mapping;
970 struct inode *inode = mapping->host;
972 if (vma->vm_flags & VM_MAYSHARE) {
973 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
974 return region_chg(&inode->i_mapping->private_list,
977 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
982 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
983 struct resv_map *reservations = vma_resv_map(vma);
985 err = region_chg(&reservations->regions, idx, idx + 1);
991 static void vma_commit_reservation(struct hstate *h,
992 struct vm_area_struct *vma, unsigned long addr)
994 struct address_space *mapping = vma->vm_file->f_mapping;
995 struct inode *inode = mapping->host;
997 if (vma->vm_flags & VM_MAYSHARE) {
998 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
999 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1001 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1002 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1003 struct resv_map *reservations = vma_resv_map(vma);
1005 /* Mark this page used in the map. */
1006 region_add(&reservations->regions, idx, idx + 1);
1010 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1011 unsigned long addr, int avoid_reserve)
1013 struct hstate *h = hstate_vma(vma);
1015 struct address_space *mapping = vma->vm_file->f_mapping;
1016 struct inode *inode = mapping->host;
1020 * Processes that did not create the mapping will have no reserves and
1021 * will not have accounted against quota. Check that the quota can be
1022 * made before satisfying the allocation
1023 * MAP_NORESERVE mappings may also need pages and quota allocated
1024 * if no reserve mapping overlaps.
1026 chg = vma_needs_reservation(h, vma, addr);
1028 return ERR_PTR(chg);
1030 if (hugetlb_get_quota(inode->i_mapping, chg))
1031 return ERR_PTR(-ENOSPC);
1033 spin_lock(&hugetlb_lock);
1034 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1035 spin_unlock(&hugetlb_lock);
1038 page = alloc_buddy_huge_page(h, vma, addr);
1040 hugetlb_put_quota(inode->i_mapping, chg);
1041 return ERR_PTR(-VM_FAULT_OOM);
1045 set_page_refcounted(page);
1046 set_page_private(page, (unsigned long) mapping);
1048 vma_commit_reservation(h, vma, addr);
1053 int __weak alloc_bootmem_huge_page(struct hstate *h)
1055 struct huge_bootmem_page *m;
1056 int nr_nodes = nodes_weight(node_online_map);
1061 addr = __alloc_bootmem_node_nopanic(
1062 NODE_DATA(hstate_next_node_to_alloc(h,
1064 huge_page_size(h), huge_page_size(h), 0);
1068 * Use the beginning of the huge page to store the
1069 * huge_bootmem_page struct (until gather_bootmem
1070 * puts them into the mem_map).
1080 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1081 /* Put them into a private list first because mem_map is not up yet */
1082 list_add(&m->list, &huge_boot_pages);
1087 static void prep_compound_huge_page(struct page *page, int order)
1089 if (unlikely(order > (MAX_ORDER - 1)))
1090 prep_compound_gigantic_page(page, order);
1092 prep_compound_page(page, order);
1095 /* Put bootmem huge pages into the standard lists after mem_map is up */
1096 static void __init gather_bootmem_prealloc(void)
1098 struct huge_bootmem_page *m;
1100 list_for_each_entry(m, &huge_boot_pages, list) {
1101 struct page *page = virt_to_page(m);
1102 struct hstate *h = m->hstate;
1103 __ClearPageReserved(page);
1104 WARN_ON(page_count(page) != 1);
1105 prep_compound_huge_page(page, h->order);
1106 prep_new_huge_page(h, page, page_to_nid(page));
1110 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1114 for (i = 0; i < h->max_huge_pages; ++i) {
1115 if (h->order >= MAX_ORDER) {
1116 if (!alloc_bootmem_huge_page(h))
1118 } else if (!alloc_fresh_huge_page(h, &node_online_map))
1121 h->max_huge_pages = i;
1124 static void __init hugetlb_init_hstates(void)
1128 for_each_hstate(h) {
1129 /* oversize hugepages were init'ed in early boot */
1130 if (h->order < MAX_ORDER)
1131 hugetlb_hstate_alloc_pages(h);
1135 static char * __init memfmt(char *buf, unsigned long n)
1137 if (n >= (1UL << 30))
1138 sprintf(buf, "%lu GB", n >> 30);
1139 else if (n >= (1UL << 20))
1140 sprintf(buf, "%lu MB", n >> 20);
1142 sprintf(buf, "%lu KB", n >> 10);
1146 static void __init report_hugepages(void)
1150 for_each_hstate(h) {
1152 printk(KERN_INFO "HugeTLB registered %s page size, "
1153 "pre-allocated %ld pages\n",
1154 memfmt(buf, huge_page_size(h)),
1155 h->free_huge_pages);
1159 #ifdef CONFIG_HIGHMEM
1160 static void try_to_free_low(struct hstate *h, unsigned long count,
1161 nodemask_t *nodes_allowed)
1165 if (h->order >= MAX_ORDER)
1168 for_each_node_mask(i, *nodes_allowed) {
1169 struct page *page, *next;
1170 struct list_head *freel = &h->hugepage_freelists[i];
1171 list_for_each_entry_safe(page, next, freel, lru) {
1172 if (count >= h->nr_huge_pages)
1174 if (PageHighMem(page))
1176 list_del(&page->lru);
1177 update_and_free_page(h, page);
1178 h->free_huge_pages--;
1179 h->free_huge_pages_node[page_to_nid(page)]--;
1184 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1185 nodemask_t *nodes_allowed)
1191 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1192 * balanced by operating on them in a round-robin fashion.
1193 * Returns 1 if an adjustment was made.
1195 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1198 int start_nid, next_nid;
1201 VM_BUG_ON(delta != -1 && delta != 1);
1204 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1206 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1207 next_nid = start_nid;
1213 * To shrink on this node, there must be a surplus page
1215 if (!h->surplus_huge_pages_node[nid]) {
1216 next_nid = hstate_next_node_to_alloc(h,
1223 * Surplus cannot exceed the total number of pages
1225 if (h->surplus_huge_pages_node[nid] >=
1226 h->nr_huge_pages_node[nid]) {
1227 next_nid = hstate_next_node_to_free(h,
1233 h->surplus_huge_pages += delta;
1234 h->surplus_huge_pages_node[nid] += delta;
1237 } while (next_nid != start_nid);
1242 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1243 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1244 nodemask_t *nodes_allowed)
1246 unsigned long min_count, ret;
1248 if (h->order >= MAX_ORDER)
1249 return h->max_huge_pages;
1252 * Increase the pool size
1253 * First take pages out of surplus state. Then make up the
1254 * remaining difference by allocating fresh huge pages.
1256 * We might race with alloc_buddy_huge_page() here and be unable
1257 * to convert a surplus huge page to a normal huge page. That is
1258 * not critical, though, it just means the overall size of the
1259 * pool might be one hugepage larger than it needs to be, but
1260 * within all the constraints specified by the sysctls.
1262 spin_lock(&hugetlb_lock);
1263 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1264 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1268 while (count > persistent_huge_pages(h)) {
1270 * If this allocation races such that we no longer need the
1271 * page, free_huge_page will handle it by freeing the page
1272 * and reducing the surplus.
1274 spin_unlock(&hugetlb_lock);
1275 ret = alloc_fresh_huge_page(h, nodes_allowed);
1276 spin_lock(&hugetlb_lock);
1283 * Decrease the pool size
1284 * First return free pages to the buddy allocator (being careful
1285 * to keep enough around to satisfy reservations). Then place
1286 * pages into surplus state as needed so the pool will shrink
1287 * to the desired size as pages become free.
1289 * By placing pages into the surplus state independent of the
1290 * overcommit value, we are allowing the surplus pool size to
1291 * exceed overcommit. There are few sane options here. Since
1292 * alloc_buddy_huge_page() is checking the global counter,
1293 * though, we'll note that we're not allowed to exceed surplus
1294 * and won't grow the pool anywhere else. Not until one of the
1295 * sysctls are changed, or the surplus pages go out of use.
1297 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1298 min_count = max(count, min_count);
1299 try_to_free_low(h, min_count, nodes_allowed);
1300 while (min_count < persistent_huge_pages(h)) {
1301 if (!free_pool_huge_page(h, nodes_allowed, 0))
1304 while (count < persistent_huge_pages(h)) {
1305 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1309 ret = persistent_huge_pages(h);
1310 spin_unlock(&hugetlb_lock);
1314 #define HSTATE_ATTR_RO(_name) \
1315 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1317 #define HSTATE_ATTR(_name) \
1318 static struct kobj_attribute _name##_attr = \
1319 __ATTR(_name, 0644, _name##_show, _name##_store)
1321 static struct kobject *hugepages_kobj;
1322 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1324 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1326 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1330 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1331 if (hstate_kobjs[i] == kobj) {
1333 *nidp = NUMA_NO_NODE;
1337 return kobj_to_node_hstate(kobj, nidp);
1340 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1341 struct kobj_attribute *attr, char *buf)
1344 unsigned long nr_huge_pages;
1347 h = kobj_to_hstate(kobj, &nid);
1348 if (nid == NUMA_NO_NODE)
1349 nr_huge_pages = h->nr_huge_pages;
1351 nr_huge_pages = h->nr_huge_pages_node[nid];
1353 return sprintf(buf, "%lu\n", nr_huge_pages);
1355 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1356 struct kobject *kobj, struct kobj_attribute *attr,
1357 const char *buf, size_t len)
1361 unsigned long count;
1363 NODEMASK_ALLOC(nodemask_t, nodes_allowed);
1365 err = strict_strtoul(buf, 10, &count);
1369 h = kobj_to_hstate(kobj, &nid);
1370 if (nid == NUMA_NO_NODE) {
1372 * global hstate attribute
1374 if (!(obey_mempolicy &&
1375 init_nodemask_of_mempolicy(nodes_allowed))) {
1376 NODEMASK_FREE(nodes_allowed);
1377 nodes_allowed = &node_states[N_HIGH_MEMORY];
1379 } else if (nodes_allowed) {
1381 * per node hstate attribute: adjust count to global,
1382 * but restrict alloc/free to the specified node.
1384 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1385 init_nodemask_of_node(nodes_allowed, nid);
1387 nodes_allowed = &node_states[N_HIGH_MEMORY];
1389 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1391 if (nodes_allowed != &node_online_map)
1392 NODEMASK_FREE(nodes_allowed);
1397 static ssize_t nr_hugepages_show(struct kobject *kobj,
1398 struct kobj_attribute *attr, char *buf)
1400 return nr_hugepages_show_common(kobj, attr, buf);
1403 static ssize_t nr_hugepages_store(struct kobject *kobj,
1404 struct kobj_attribute *attr, const char *buf, size_t len)
1406 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1408 HSTATE_ATTR(nr_hugepages);
1413 * hstate attribute for optionally mempolicy-based constraint on persistent
1414 * huge page alloc/free.
1416 static ssize_t nr_hugepages_mempolicy_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_mempolicy_store(struct kobject *kobj,
1423 struct kobj_attribute *attr, const char *buf, size_t len)
1425 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1427 HSTATE_ATTR(nr_hugepages_mempolicy);
1431 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1432 struct kobj_attribute *attr, char *buf)
1434 struct hstate *h = kobj_to_hstate(kobj, NULL);
1435 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1437 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1438 struct kobj_attribute *attr, const char *buf, size_t count)
1441 unsigned long input;
1442 struct hstate *h = kobj_to_hstate(kobj, NULL);
1444 err = strict_strtoul(buf, 10, &input);
1448 spin_lock(&hugetlb_lock);
1449 h->nr_overcommit_huge_pages = input;
1450 spin_unlock(&hugetlb_lock);
1454 HSTATE_ATTR(nr_overcommit_hugepages);
1456 static ssize_t free_hugepages_show(struct kobject *kobj,
1457 struct kobj_attribute *attr, char *buf)
1460 unsigned long free_huge_pages;
1463 h = kobj_to_hstate(kobj, &nid);
1464 if (nid == NUMA_NO_NODE)
1465 free_huge_pages = h->free_huge_pages;
1467 free_huge_pages = h->free_huge_pages_node[nid];
1469 return sprintf(buf, "%lu\n", free_huge_pages);
1471 HSTATE_ATTR_RO(free_hugepages);
1473 static ssize_t resv_hugepages_show(struct kobject *kobj,
1474 struct kobj_attribute *attr, char *buf)
1476 struct hstate *h = kobj_to_hstate(kobj, NULL);
1477 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1479 HSTATE_ATTR_RO(resv_hugepages);
1481 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1482 struct kobj_attribute *attr, char *buf)
1485 unsigned long surplus_huge_pages;
1488 h = kobj_to_hstate(kobj, &nid);
1489 if (nid == NUMA_NO_NODE)
1490 surplus_huge_pages = h->surplus_huge_pages;
1492 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1494 return sprintf(buf, "%lu\n", surplus_huge_pages);
1496 HSTATE_ATTR_RO(surplus_hugepages);
1498 static struct attribute *hstate_attrs[] = {
1499 &nr_hugepages_attr.attr,
1500 &nr_overcommit_hugepages_attr.attr,
1501 &free_hugepages_attr.attr,
1502 &resv_hugepages_attr.attr,
1503 &surplus_hugepages_attr.attr,
1505 &nr_hugepages_mempolicy_attr.attr,
1510 static struct attribute_group hstate_attr_group = {
1511 .attrs = hstate_attrs,
1514 static int __init hugetlb_sysfs_add_hstate(struct hstate *h,
1515 struct kobject *parent,
1516 struct kobject **hstate_kobjs,
1517 struct attribute_group *hstate_attr_group)
1520 int hi = h - hstates;
1522 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1523 if (!hstate_kobjs[hi])
1526 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1528 kobject_put(hstate_kobjs[hi]);
1533 static void __init hugetlb_sysfs_init(void)
1538 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1539 if (!hugepages_kobj)
1542 for_each_hstate(h) {
1543 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1544 hstate_kobjs, &hstate_attr_group);
1546 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1554 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1555 * with node sysdevs in node_devices[] using a parallel array. The array
1556 * index of a node sysdev or _hstate == node id.
1557 * This is here to avoid any static dependency of the node sysdev driver, in
1558 * the base kernel, on the hugetlb module.
1560 struct node_hstate {
1561 struct kobject *hugepages_kobj;
1562 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1564 struct node_hstate node_hstates[MAX_NUMNODES];
1567 * A subset of global hstate attributes for node sysdevs
1569 static struct attribute *per_node_hstate_attrs[] = {
1570 &nr_hugepages_attr.attr,
1571 &free_hugepages_attr.attr,
1572 &surplus_hugepages_attr.attr,
1576 static struct attribute_group per_node_hstate_attr_group = {
1577 .attrs = per_node_hstate_attrs,
1581 * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj.
1582 * Returns node id via non-NULL nidp.
1584 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1588 for (nid = 0; nid < nr_node_ids; nid++) {
1589 struct node_hstate *nhs = &node_hstates[nid];
1591 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1592 if (nhs->hstate_kobjs[i] == kobj) {
1604 * Unregister hstate attributes from a single node sysdev.
1605 * No-op if no hstate attributes attached.
1607 void hugetlb_unregister_node(struct node *node)
1610 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1612 if (!nhs->hugepages_kobj)
1616 if (nhs->hstate_kobjs[h - hstates]) {
1617 kobject_put(nhs->hstate_kobjs[h - hstates]);
1618 nhs->hstate_kobjs[h - hstates] = NULL;
1621 kobject_put(nhs->hugepages_kobj);
1622 nhs->hugepages_kobj = NULL;
1626 * hugetlb module exit: unregister hstate attributes from node sysdevs
1629 static void hugetlb_unregister_all_nodes(void)
1634 * disable node sysdev registrations.
1636 register_hugetlbfs_with_node(NULL, NULL);
1639 * remove hstate attributes from any nodes that have them.
1641 for (nid = 0; nid < nr_node_ids; nid++)
1642 hugetlb_unregister_node(&node_devices[nid]);
1646 * Register hstate attributes for a single node sysdev.
1647 * No-op if attributes already registered.
1649 void hugetlb_register_node(struct node *node)
1652 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1655 if (nhs->hugepages_kobj)
1656 return; /* already allocated */
1658 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1659 &node->sysdev.kobj);
1660 if (!nhs->hugepages_kobj)
1663 for_each_hstate(h) {
1664 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1666 &per_node_hstate_attr_group);
1668 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1670 h->name, node->sysdev.id);
1671 hugetlb_unregister_node(node);
1678 * hugetlb init time: register hstate attributes for all registered
1679 * node sysdevs. All on-line nodes should have registered their
1680 * associated sysdev by the time the hugetlb module initializes.
1682 static void hugetlb_register_all_nodes(void)
1686 for (nid = 0; nid < nr_node_ids; nid++) {
1687 struct node *node = &node_devices[nid];
1688 if (node->sysdev.id == nid)
1689 hugetlb_register_node(node);
1693 * Let the node sysdev driver know we're here so it can
1694 * [un]register hstate attributes on node hotplug.
1696 register_hugetlbfs_with_node(hugetlb_register_node,
1697 hugetlb_unregister_node);
1699 #else /* !CONFIG_NUMA */
1701 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1709 static void hugetlb_unregister_all_nodes(void) { }
1711 static void hugetlb_register_all_nodes(void) { }
1715 static void __exit hugetlb_exit(void)
1719 hugetlb_unregister_all_nodes();
1721 for_each_hstate(h) {
1722 kobject_put(hstate_kobjs[h - hstates]);
1725 kobject_put(hugepages_kobj);
1727 module_exit(hugetlb_exit);
1729 static int __init hugetlb_init(void)
1731 /* Some platform decide whether they support huge pages at boot
1732 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1733 * there is no such support
1735 if (HPAGE_SHIFT == 0)
1738 if (!size_to_hstate(default_hstate_size)) {
1739 default_hstate_size = HPAGE_SIZE;
1740 if (!size_to_hstate(default_hstate_size))
1741 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1743 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1744 if (default_hstate_max_huge_pages)
1745 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1747 hugetlb_init_hstates();
1749 gather_bootmem_prealloc();
1753 hugetlb_sysfs_init();
1755 hugetlb_register_all_nodes();
1759 module_init(hugetlb_init);
1761 /* Should be called on processing a hugepagesz=... option */
1762 void __init hugetlb_add_hstate(unsigned order)
1767 if (size_to_hstate(PAGE_SIZE << order)) {
1768 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1771 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1773 h = &hstates[max_hstate++];
1775 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1776 h->nr_huge_pages = 0;
1777 h->free_huge_pages = 0;
1778 for (i = 0; i < MAX_NUMNODES; ++i)
1779 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1780 h->next_nid_to_alloc = first_node(node_online_map);
1781 h->next_nid_to_free = first_node(node_online_map);
1782 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1783 huge_page_size(h)/1024);
1788 static int __init hugetlb_nrpages_setup(char *s)
1791 static unsigned long *last_mhp;
1794 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1795 * so this hugepages= parameter goes to the "default hstate".
1798 mhp = &default_hstate_max_huge_pages;
1800 mhp = &parsed_hstate->max_huge_pages;
1802 if (mhp == last_mhp) {
1803 printk(KERN_WARNING "hugepages= specified twice without "
1804 "interleaving hugepagesz=, ignoring\n");
1808 if (sscanf(s, "%lu", mhp) <= 0)
1812 * Global state is always initialized later in hugetlb_init.
1813 * But we need to allocate >= MAX_ORDER hstates here early to still
1814 * use the bootmem allocator.
1816 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1817 hugetlb_hstate_alloc_pages(parsed_hstate);
1823 __setup("hugepages=", hugetlb_nrpages_setup);
1825 static int __init hugetlb_default_setup(char *s)
1827 default_hstate_size = memparse(s, &s);
1830 __setup("default_hugepagesz=", hugetlb_default_setup);
1832 static unsigned int cpuset_mems_nr(unsigned int *array)
1835 unsigned int nr = 0;
1837 for_each_node_mask(node, cpuset_current_mems_allowed)
1843 #ifdef CONFIG_SYSCTL
1844 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1845 struct ctl_table *table, int write,
1846 void __user *buffer, size_t *length, loff_t *ppos)
1848 struct hstate *h = &default_hstate;
1852 tmp = h->max_huge_pages;
1855 table->maxlen = sizeof(unsigned long);
1856 proc_doulongvec_minmax(table, write, buffer, length, ppos);
1859 NODEMASK_ALLOC(nodemask_t, nodes_allowed);
1860 if (!(obey_mempolicy &&
1861 init_nodemask_of_mempolicy(nodes_allowed))) {
1862 NODEMASK_FREE(nodes_allowed);
1863 nodes_allowed = &node_states[N_HIGH_MEMORY];
1865 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1867 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1868 NODEMASK_FREE(nodes_allowed);
1874 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1875 void __user *buffer, size_t *length, loff_t *ppos)
1878 return hugetlb_sysctl_handler_common(false, table, write,
1879 buffer, length, ppos);
1883 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
1884 void __user *buffer, size_t *length, loff_t *ppos)
1886 return hugetlb_sysctl_handler_common(true, table, write,
1887 buffer, length, ppos);
1889 #endif /* CONFIG_NUMA */
1891 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1892 void __user *buffer,
1893 size_t *length, loff_t *ppos)
1895 proc_dointvec(table, write, buffer, length, ppos);
1896 if (hugepages_treat_as_movable)
1897 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1899 htlb_alloc_mask = GFP_HIGHUSER;
1903 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1904 void __user *buffer,
1905 size_t *length, loff_t *ppos)
1907 struct hstate *h = &default_hstate;
1911 tmp = h->nr_overcommit_huge_pages;
1914 table->maxlen = sizeof(unsigned long);
1915 proc_doulongvec_minmax(table, write, buffer, length, ppos);
1918 spin_lock(&hugetlb_lock);
1919 h->nr_overcommit_huge_pages = tmp;
1920 spin_unlock(&hugetlb_lock);
1926 #endif /* CONFIG_SYSCTL */
1928 void hugetlb_report_meminfo(struct seq_file *m)
1930 struct hstate *h = &default_hstate;
1932 "HugePages_Total: %5lu\n"
1933 "HugePages_Free: %5lu\n"
1934 "HugePages_Rsvd: %5lu\n"
1935 "HugePages_Surp: %5lu\n"
1936 "Hugepagesize: %8lu kB\n",
1940 h->surplus_huge_pages,
1941 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1944 int hugetlb_report_node_meminfo(int nid, char *buf)
1946 struct hstate *h = &default_hstate;
1948 "Node %d HugePages_Total: %5u\n"
1949 "Node %d HugePages_Free: %5u\n"
1950 "Node %d HugePages_Surp: %5u\n",
1951 nid, h->nr_huge_pages_node[nid],
1952 nid, h->free_huge_pages_node[nid],
1953 nid, h->surplus_huge_pages_node[nid]);
1956 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1957 unsigned long hugetlb_total_pages(void)
1959 struct hstate *h = &default_hstate;
1960 return h->nr_huge_pages * pages_per_huge_page(h);
1963 static int hugetlb_acct_memory(struct hstate *h, long delta)
1967 spin_lock(&hugetlb_lock);
1969 * When cpuset is configured, it breaks the strict hugetlb page
1970 * reservation as the accounting is done on a global variable. Such
1971 * reservation is completely rubbish in the presence of cpuset because
1972 * the reservation is not checked against page availability for the
1973 * current cpuset. Application can still potentially OOM'ed by kernel
1974 * with lack of free htlb page in cpuset that the task is in.
1975 * Attempt to enforce strict accounting with cpuset is almost
1976 * impossible (or too ugly) because cpuset is too fluid that
1977 * task or memory node can be dynamically moved between cpusets.
1979 * The change of semantics for shared hugetlb mapping with cpuset is
1980 * undesirable. However, in order to preserve some of the semantics,
1981 * we fall back to check against current free page availability as
1982 * a best attempt and hopefully to minimize the impact of changing
1983 * semantics that cpuset has.
1986 if (gather_surplus_pages(h, delta) < 0)
1989 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1990 return_unused_surplus_pages(h, delta);
1997 return_unused_surplus_pages(h, (unsigned long) -delta);
2000 spin_unlock(&hugetlb_lock);
2004 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2006 struct resv_map *reservations = vma_resv_map(vma);
2009 * This new VMA should share its siblings reservation map if present.
2010 * The VMA will only ever have a valid reservation map pointer where
2011 * it is being copied for another still existing VMA. As that VMA
2012 * has a reference to the reservation map it cannot dissappear until
2013 * after this open call completes. It is therefore safe to take a
2014 * new reference here without additional locking.
2017 kref_get(&reservations->refs);
2020 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2022 struct hstate *h = hstate_vma(vma);
2023 struct resv_map *reservations = vma_resv_map(vma);
2024 unsigned long reserve;
2025 unsigned long start;
2029 start = vma_hugecache_offset(h, vma, vma->vm_start);
2030 end = vma_hugecache_offset(h, vma, vma->vm_end);
2032 reserve = (end - start) -
2033 region_count(&reservations->regions, start, end);
2035 kref_put(&reservations->refs, resv_map_release);
2038 hugetlb_acct_memory(h, -reserve);
2039 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
2045 * We cannot handle pagefaults against hugetlb pages at all. They cause
2046 * handle_mm_fault() to try to instantiate regular-sized pages in the
2047 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2050 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2056 const struct vm_operations_struct hugetlb_vm_ops = {
2057 .fault = hugetlb_vm_op_fault,
2058 .open = hugetlb_vm_op_open,
2059 .close = hugetlb_vm_op_close,
2062 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2069 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2071 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2073 entry = pte_mkyoung(entry);
2074 entry = pte_mkhuge(entry);
2079 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2080 unsigned long address, pte_t *ptep)
2084 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2085 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
2086 update_mmu_cache(vma, address, entry);
2091 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2092 struct vm_area_struct *vma)
2094 pte_t *src_pte, *dst_pte, entry;
2095 struct page *ptepage;
2098 struct hstate *h = hstate_vma(vma);
2099 unsigned long sz = huge_page_size(h);
2101 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2103 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2104 src_pte = huge_pte_offset(src, addr);
2107 dst_pte = huge_pte_alloc(dst, addr, sz);
2111 /* If the pagetables are shared don't copy or take references */
2112 if (dst_pte == src_pte)
2115 spin_lock(&dst->page_table_lock);
2116 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2117 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2119 huge_ptep_set_wrprotect(src, addr, src_pte);
2120 entry = huge_ptep_get(src_pte);
2121 ptepage = pte_page(entry);
2123 set_huge_pte_at(dst, addr, dst_pte, entry);
2125 spin_unlock(&src->page_table_lock);
2126 spin_unlock(&dst->page_table_lock);
2134 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2135 unsigned long end, struct page *ref_page)
2137 struct mm_struct *mm = vma->vm_mm;
2138 unsigned long address;
2143 struct hstate *h = hstate_vma(vma);
2144 unsigned long sz = huge_page_size(h);
2147 * A page gathering list, protected by per file i_mmap_lock. The
2148 * lock is used to avoid list corruption from multiple unmapping
2149 * of the same page since we are using page->lru.
2151 LIST_HEAD(page_list);
2153 WARN_ON(!is_vm_hugetlb_page(vma));
2154 BUG_ON(start & ~huge_page_mask(h));
2155 BUG_ON(end & ~huge_page_mask(h));
2157 mmu_notifier_invalidate_range_start(mm, start, end);
2158 spin_lock(&mm->page_table_lock);
2159 for (address = start; address < end; address += sz) {
2160 ptep = huge_pte_offset(mm, address);
2164 if (huge_pmd_unshare(mm, &address, ptep))
2168 * If a reference page is supplied, it is because a specific
2169 * page is being unmapped, not a range. Ensure the page we
2170 * are about to unmap is the actual page of interest.
2173 pte = huge_ptep_get(ptep);
2174 if (huge_pte_none(pte))
2176 page = pte_page(pte);
2177 if (page != ref_page)
2181 * Mark the VMA as having unmapped its page so that
2182 * future faults in this VMA will fail rather than
2183 * looking like data was lost
2185 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2188 pte = huge_ptep_get_and_clear(mm, address, ptep);
2189 if (huge_pte_none(pte))
2192 page = pte_page(pte);
2194 set_page_dirty(page);
2195 list_add(&page->lru, &page_list);
2197 spin_unlock(&mm->page_table_lock);
2198 flush_tlb_range(vma, start, end);
2199 mmu_notifier_invalidate_range_end(mm, start, end);
2200 list_for_each_entry_safe(page, tmp, &page_list, lru) {
2201 list_del(&page->lru);
2206 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2207 unsigned long end, struct page *ref_page)
2209 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2210 __unmap_hugepage_range(vma, start, end, ref_page);
2211 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2215 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2216 * mappping it owns the reserve page for. The intention is to unmap the page
2217 * from other VMAs and let the children be SIGKILLed if they are faulting the
2220 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2221 struct page *page, unsigned long address)
2223 struct hstate *h = hstate_vma(vma);
2224 struct vm_area_struct *iter_vma;
2225 struct address_space *mapping;
2226 struct prio_tree_iter iter;
2230 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2231 * from page cache lookup which is in HPAGE_SIZE units.
2233 address = address & huge_page_mask(h);
2234 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
2235 + (vma->vm_pgoff >> PAGE_SHIFT);
2236 mapping = (struct address_space *)page_private(page);
2238 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2239 /* Do not unmap the current VMA */
2240 if (iter_vma == vma)
2244 * Unmap the page from other VMAs without their own reserves.
2245 * They get marked to be SIGKILLed if they fault in these
2246 * areas. This is because a future no-page fault on this VMA
2247 * could insert a zeroed page instead of the data existing
2248 * from the time of fork. This would look like data corruption
2250 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2251 unmap_hugepage_range(iter_vma,
2252 address, address + huge_page_size(h),
2259 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2260 unsigned long address, pte_t *ptep, pte_t pte,
2261 struct page *pagecache_page)
2263 struct hstate *h = hstate_vma(vma);
2264 struct page *old_page, *new_page;
2266 int outside_reserve = 0;
2268 old_page = pte_page(pte);
2271 /* If no-one else is actually using this page, avoid the copy
2272 * and just make the page writable */
2273 avoidcopy = (page_count(old_page) == 1);
2275 set_huge_ptep_writable(vma, address, ptep);
2280 * If the process that created a MAP_PRIVATE mapping is about to
2281 * perform a COW due to a shared page count, attempt to satisfy
2282 * the allocation without using the existing reserves. The pagecache
2283 * page is used to determine if the reserve at this address was
2284 * consumed or not. If reserves were used, a partial faulted mapping
2285 * at the time of fork() could consume its reserves on COW instead
2286 * of the full address range.
2288 if (!(vma->vm_flags & VM_MAYSHARE) &&
2289 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2290 old_page != pagecache_page)
2291 outside_reserve = 1;
2293 page_cache_get(old_page);
2294 new_page = alloc_huge_page(vma, address, outside_reserve);
2296 if (IS_ERR(new_page)) {
2297 page_cache_release(old_page);
2300 * If a process owning a MAP_PRIVATE mapping fails to COW,
2301 * it is due to references held by a child and an insufficient
2302 * huge page pool. To guarantee the original mappers
2303 * reliability, unmap the page from child processes. The child
2304 * may get SIGKILLed if it later faults.
2306 if (outside_reserve) {
2307 BUG_ON(huge_pte_none(pte));
2308 if (unmap_ref_private(mm, vma, old_page, address)) {
2309 BUG_ON(page_count(old_page) != 1);
2310 BUG_ON(huge_pte_none(pte));
2311 goto retry_avoidcopy;
2316 return -PTR_ERR(new_page);
2319 spin_unlock(&mm->page_table_lock);
2320 copy_huge_page(new_page, old_page, address, vma);
2321 __SetPageUptodate(new_page);
2322 spin_lock(&mm->page_table_lock);
2324 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2325 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2327 huge_ptep_clear_flush(vma, address, ptep);
2328 set_huge_pte_at(mm, address, ptep,
2329 make_huge_pte(vma, new_page, 1));
2330 /* Make the old page be freed below */
2331 new_page = old_page;
2333 page_cache_release(new_page);
2334 page_cache_release(old_page);
2338 /* Return the pagecache page at a given address within a VMA */
2339 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2340 struct vm_area_struct *vma, unsigned long address)
2342 struct address_space *mapping;
2345 mapping = vma->vm_file->f_mapping;
2346 idx = vma_hugecache_offset(h, vma, address);
2348 return find_lock_page(mapping, idx);
2352 * Return whether there is a pagecache page to back given address within VMA.
2353 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2355 static bool hugetlbfs_pagecache_present(struct hstate *h,
2356 struct vm_area_struct *vma, unsigned long address)
2358 struct address_space *mapping;
2362 mapping = vma->vm_file->f_mapping;
2363 idx = vma_hugecache_offset(h, vma, address);
2365 page = find_get_page(mapping, idx);
2368 return page != NULL;
2371 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2372 unsigned long address, pte_t *ptep, unsigned int flags)
2374 struct hstate *h = hstate_vma(vma);
2375 int ret = VM_FAULT_SIGBUS;
2379 struct address_space *mapping;
2383 * Currently, we are forced to kill the process in the event the
2384 * original mapper has unmapped pages from the child due to a failed
2385 * COW. Warn that such a situation has occured as it may not be obvious
2387 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2389 "PID %d killed due to inadequate hugepage pool\n",
2394 mapping = vma->vm_file->f_mapping;
2395 idx = vma_hugecache_offset(h, vma, address);
2398 * Use page lock to guard against racing truncation
2399 * before we get page_table_lock.
2402 page = find_lock_page(mapping, idx);
2404 size = i_size_read(mapping->host) >> huge_page_shift(h);
2407 page = alloc_huge_page(vma, address, 0);
2409 ret = -PTR_ERR(page);
2412 clear_huge_page(page, address, huge_page_size(h));
2413 __SetPageUptodate(page);
2415 if (vma->vm_flags & VM_MAYSHARE) {
2417 struct inode *inode = mapping->host;
2419 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2427 spin_lock(&inode->i_lock);
2428 inode->i_blocks += blocks_per_huge_page(h);
2429 spin_unlock(&inode->i_lock);
2435 * If we are going to COW a private mapping later, we examine the
2436 * pending reservations for this page now. This will ensure that
2437 * any allocations necessary to record that reservation occur outside
2440 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2441 if (vma_needs_reservation(h, vma, address) < 0) {
2443 goto backout_unlocked;
2446 spin_lock(&mm->page_table_lock);
2447 size = i_size_read(mapping->host) >> huge_page_shift(h);
2452 if (!huge_pte_none(huge_ptep_get(ptep)))
2455 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2456 && (vma->vm_flags & VM_SHARED)));
2457 set_huge_pte_at(mm, address, ptep, new_pte);
2459 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2460 /* Optimization, do the COW without a second fault */
2461 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2464 spin_unlock(&mm->page_table_lock);
2470 spin_unlock(&mm->page_table_lock);
2477 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2478 unsigned long address, unsigned int flags)
2483 struct page *pagecache_page = NULL;
2484 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2485 struct hstate *h = hstate_vma(vma);
2487 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2489 return VM_FAULT_OOM;
2492 * Serialize hugepage allocation and instantiation, so that we don't
2493 * get spurious allocation failures if two CPUs race to instantiate
2494 * the same page in the page cache.
2496 mutex_lock(&hugetlb_instantiation_mutex);
2497 entry = huge_ptep_get(ptep);
2498 if (huge_pte_none(entry)) {
2499 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2506 * If we are going to COW the mapping later, we examine the pending
2507 * reservations for this page now. This will ensure that any
2508 * allocations necessary to record that reservation occur outside the
2509 * spinlock. For private mappings, we also lookup the pagecache
2510 * page now as it is used to determine if a reservation has been
2513 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2514 if (vma_needs_reservation(h, vma, address) < 0) {
2519 if (!(vma->vm_flags & VM_MAYSHARE))
2520 pagecache_page = hugetlbfs_pagecache_page(h,
2524 spin_lock(&mm->page_table_lock);
2525 /* Check for a racing update before calling hugetlb_cow */
2526 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2527 goto out_page_table_lock;
2530 if (flags & FAULT_FLAG_WRITE) {
2531 if (!pte_write(entry)) {
2532 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2534 goto out_page_table_lock;
2536 entry = pte_mkdirty(entry);
2538 entry = pte_mkyoung(entry);
2539 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2540 flags & FAULT_FLAG_WRITE))
2541 update_mmu_cache(vma, address, entry);
2543 out_page_table_lock:
2544 spin_unlock(&mm->page_table_lock);
2546 if (pagecache_page) {
2547 unlock_page(pagecache_page);
2548 put_page(pagecache_page);
2552 mutex_unlock(&hugetlb_instantiation_mutex);
2557 /* Can be overriden by architectures */
2558 __attribute__((weak)) struct page *
2559 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2560 pud_t *pud, int write)
2566 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2567 struct page **pages, struct vm_area_struct **vmas,
2568 unsigned long *position, int *length, int i,
2571 unsigned long pfn_offset;
2572 unsigned long vaddr = *position;
2573 int remainder = *length;
2574 struct hstate *h = hstate_vma(vma);
2576 spin_lock(&mm->page_table_lock);
2577 while (vaddr < vma->vm_end && remainder) {
2583 * Some archs (sparc64, sh*) have multiple pte_ts to
2584 * each hugepage. We have to make sure we get the
2585 * first, for the page indexing below to work.
2587 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2588 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2591 * When coredumping, it suits get_dump_page if we just return
2592 * an error where there's an empty slot with no huge pagecache
2593 * to back it. This way, we avoid allocating a hugepage, and
2594 * the sparse dumpfile avoids allocating disk blocks, but its
2595 * huge holes still show up with zeroes where they need to be.
2597 if (absent && (flags & FOLL_DUMP) &&
2598 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2604 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2607 spin_unlock(&mm->page_table_lock);
2608 ret = hugetlb_fault(mm, vma, vaddr,
2609 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2610 spin_lock(&mm->page_table_lock);
2611 if (!(ret & VM_FAULT_ERROR))
2618 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2619 page = pte_page(huge_ptep_get(pte));
2622 pages[i] = mem_map_offset(page, pfn_offset);
2633 if (vaddr < vma->vm_end && remainder &&
2634 pfn_offset < pages_per_huge_page(h)) {
2636 * We use pfn_offset to avoid touching the pageframes
2637 * of this compound page.
2642 spin_unlock(&mm->page_table_lock);
2643 *length = remainder;
2646 return i ? i : -EFAULT;
2649 void hugetlb_change_protection(struct vm_area_struct *vma,
2650 unsigned long address, unsigned long end, pgprot_t newprot)
2652 struct mm_struct *mm = vma->vm_mm;
2653 unsigned long start = address;
2656 struct hstate *h = hstate_vma(vma);
2658 BUG_ON(address >= end);
2659 flush_cache_range(vma, address, end);
2661 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2662 spin_lock(&mm->page_table_lock);
2663 for (; address < end; address += huge_page_size(h)) {
2664 ptep = huge_pte_offset(mm, address);
2667 if (huge_pmd_unshare(mm, &address, ptep))
2669 if (!huge_pte_none(huge_ptep_get(ptep))) {
2670 pte = huge_ptep_get_and_clear(mm, address, ptep);
2671 pte = pte_mkhuge(pte_modify(pte, newprot));
2672 set_huge_pte_at(mm, address, ptep, pte);
2675 spin_unlock(&mm->page_table_lock);
2676 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2678 flush_tlb_range(vma, start, end);
2681 int hugetlb_reserve_pages(struct inode *inode,
2683 struct vm_area_struct *vma,
2687 struct hstate *h = hstate_inode(inode);
2690 * Only apply hugepage reservation if asked. At fault time, an
2691 * attempt will be made for VM_NORESERVE to allocate a page
2692 * and filesystem quota without using reserves
2694 if (acctflag & VM_NORESERVE)
2698 * Shared mappings base their reservation on the number of pages that
2699 * are already allocated on behalf of the file. Private mappings need
2700 * to reserve the full area even if read-only as mprotect() may be
2701 * called to make the mapping read-write. Assume !vma is a shm mapping
2703 if (!vma || vma->vm_flags & VM_MAYSHARE)
2704 chg = region_chg(&inode->i_mapping->private_list, from, to);
2706 struct resv_map *resv_map = resv_map_alloc();
2712 set_vma_resv_map(vma, resv_map);
2713 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2719 /* There must be enough filesystem quota for the mapping */
2720 if (hugetlb_get_quota(inode->i_mapping, chg))
2724 * Check enough hugepages are available for the reservation.
2725 * Hand back the quota if there are not
2727 ret = hugetlb_acct_memory(h, chg);
2729 hugetlb_put_quota(inode->i_mapping, chg);
2734 * Account for the reservations made. Shared mappings record regions
2735 * that have reservations as they are shared by multiple VMAs.
2736 * When the last VMA disappears, the region map says how much
2737 * the reservation was and the page cache tells how much of
2738 * the reservation was consumed. Private mappings are per-VMA and
2739 * only the consumed reservations are tracked. When the VMA
2740 * disappears, the original reservation is the VMA size and the
2741 * consumed reservations are stored in the map. Hence, nothing
2742 * else has to be done for private mappings here
2744 if (!vma || vma->vm_flags & VM_MAYSHARE)
2745 region_add(&inode->i_mapping->private_list, from, to);
2749 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2751 struct hstate *h = hstate_inode(inode);
2752 long chg = region_truncate(&inode->i_mapping->private_list, offset);
2754 spin_lock(&inode->i_lock);
2755 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
2756 spin_unlock(&inode->i_lock);
2758 hugetlb_put_quota(inode->i_mapping, (chg - freed));
2759 hugetlb_acct_memory(h, -(chg - freed));