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 further 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 copy_gigantic_page(struct page *dst, struct page *src)
400 struct hstate *h = page_hstate(src);
401 struct page *dst_base = dst;
402 struct page *src_base = src;
404 for (i = 0; i < pages_per_huge_page(h); ) {
406 copy_highpage(dst, src);
409 dst = mem_map_next(dst, dst_base, i);
410 src = mem_map_next(src, src_base, i);
414 void copy_huge_page(struct page *dst, struct page *src)
417 struct hstate *h = page_hstate(src);
419 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
420 copy_gigantic_page(dst, src);
425 for (i = 0; i < pages_per_huge_page(h); i++) {
427 copy_highpage(dst + i, src + i);
431 static void enqueue_huge_page(struct hstate *h, struct page *page)
433 int nid = page_to_nid(page);
434 list_add(&page->lru, &h->hugepage_freelists[nid]);
435 h->free_huge_pages++;
436 h->free_huge_pages_node[nid]++;
439 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
443 if (list_empty(&h->hugepage_freelists[nid]))
445 page = list_entry(h->hugepage_freelists[nid].next, struct page, lru);
446 list_del(&page->lru);
447 set_page_refcounted(page);
448 h->free_huge_pages--;
449 h->free_huge_pages_node[nid]--;
453 static struct page *dequeue_huge_page_vma(struct hstate *h,
454 struct vm_area_struct *vma,
455 unsigned long address, int avoid_reserve)
457 struct page *page = NULL;
458 struct mempolicy *mpol;
459 nodemask_t *nodemask;
460 struct zonelist *zonelist;
465 zonelist = huge_zonelist(vma, address,
466 htlb_alloc_mask, &mpol, &nodemask);
468 * A child process with MAP_PRIVATE mappings created by their parent
469 * have no page reserves. This check ensures that reservations are
470 * not "stolen". The child may still get SIGKILLed
472 if (!vma_has_reserves(vma) &&
473 h->free_huge_pages - h->resv_huge_pages == 0)
476 /* If reserves cannot be used, ensure enough pages are in the pool */
477 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
480 for_each_zone_zonelist_nodemask(zone, z, zonelist,
481 MAX_NR_ZONES - 1, nodemask) {
482 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
483 page = dequeue_huge_page_node(h, zone_to_nid(zone));
486 decrement_hugepage_resv_vma(h, vma);
497 static void update_and_free_page(struct hstate *h, struct page *page)
501 VM_BUG_ON(h->order >= MAX_ORDER);
504 h->nr_huge_pages_node[page_to_nid(page)]--;
505 for (i = 0; i < pages_per_huge_page(h); i++) {
506 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
507 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
508 1 << PG_private | 1<< PG_writeback);
510 set_compound_page_dtor(page, NULL);
511 set_page_refcounted(page);
512 arch_release_hugepage(page);
513 __free_pages(page, huge_page_order(h));
516 struct hstate *size_to_hstate(unsigned long size)
521 if (huge_page_size(h) == size)
527 static void free_huge_page(struct page *page)
530 * Can't pass hstate in here because it is called from the
531 * compound page destructor.
533 struct hstate *h = page_hstate(page);
534 int nid = page_to_nid(page);
535 struct address_space *mapping;
537 mapping = (struct address_space *) page_private(page);
538 set_page_private(page, 0);
539 page->mapping = NULL;
540 BUG_ON(page_count(page));
541 BUG_ON(page_mapcount(page));
542 INIT_LIST_HEAD(&page->lru);
544 spin_lock(&hugetlb_lock);
545 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
546 update_and_free_page(h, page);
547 h->surplus_huge_pages--;
548 h->surplus_huge_pages_node[nid]--;
550 enqueue_huge_page(h, page);
552 spin_unlock(&hugetlb_lock);
554 hugetlb_put_quota(mapping, 1);
557 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
559 set_compound_page_dtor(page, free_huge_page);
560 spin_lock(&hugetlb_lock);
562 h->nr_huge_pages_node[nid]++;
563 spin_unlock(&hugetlb_lock);
564 put_page(page); /* free it into the hugepage allocator */
567 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
570 int nr_pages = 1 << order;
571 struct page *p = page + 1;
573 /* we rely on prep_new_huge_page to set the destructor */
574 set_compound_order(page, order);
576 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
578 set_page_count(p, 0);
579 p->first_page = page;
583 int PageHuge(struct page *page)
585 compound_page_dtor *dtor;
587 if (!PageCompound(page))
590 page = compound_head(page);
591 dtor = get_compound_page_dtor(page);
593 return dtor == free_huge_page;
596 EXPORT_SYMBOL_GPL(PageHuge);
598 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
602 if (h->order >= MAX_ORDER)
605 page = alloc_pages_exact_node(nid,
606 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
607 __GFP_REPEAT|__GFP_NOWARN,
610 if (arch_prepare_hugepage(page)) {
611 __free_pages(page, huge_page_order(h));
614 prep_new_huge_page(h, page, nid);
621 * common helper functions for hstate_next_node_to_{alloc|free}.
622 * We may have allocated or freed a huge page based on a different
623 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
624 * be outside of *nodes_allowed. Ensure that we use an allowed
625 * node for alloc or free.
627 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
629 nid = next_node(nid, *nodes_allowed);
630 if (nid == MAX_NUMNODES)
631 nid = first_node(*nodes_allowed);
632 VM_BUG_ON(nid >= MAX_NUMNODES);
637 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
639 if (!node_isset(nid, *nodes_allowed))
640 nid = next_node_allowed(nid, nodes_allowed);
645 * returns the previously saved node ["this node"] from which to
646 * allocate a persistent huge page for the pool and advance the
647 * next node from which to allocate, handling wrap at end of node
650 static int hstate_next_node_to_alloc(struct hstate *h,
651 nodemask_t *nodes_allowed)
655 VM_BUG_ON(!nodes_allowed);
657 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
658 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
663 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
670 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
671 next_nid = start_nid;
674 page = alloc_fresh_huge_page_node(h, next_nid);
679 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
680 } while (next_nid != start_nid);
683 count_vm_event(HTLB_BUDDY_PGALLOC);
685 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
691 * helper for free_pool_huge_page() - return the previously saved
692 * node ["this node"] from which to free a huge page. Advance the
693 * next node id whether or not we find a free huge page to free so
694 * that the next attempt to free addresses the next node.
696 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
700 VM_BUG_ON(!nodes_allowed);
702 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
703 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
709 * Free huge page from pool from next node to free.
710 * Attempt to keep persistent huge pages more or less
711 * balanced over allowed nodes.
712 * Called with hugetlb_lock locked.
714 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
721 start_nid = hstate_next_node_to_free(h, nodes_allowed);
722 next_nid = start_nid;
726 * If we're returning unused surplus pages, only examine
727 * nodes with surplus pages.
729 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
730 !list_empty(&h->hugepage_freelists[next_nid])) {
732 list_entry(h->hugepage_freelists[next_nid].next,
734 list_del(&page->lru);
735 h->free_huge_pages--;
736 h->free_huge_pages_node[next_nid]--;
738 h->surplus_huge_pages--;
739 h->surplus_huge_pages_node[next_nid]--;
741 update_and_free_page(h, page);
745 next_nid = hstate_next_node_to_free(h, nodes_allowed);
746 } while (next_nid != start_nid);
751 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
756 if (h->order >= MAX_ORDER)
760 * Assume we will successfully allocate the surplus page to
761 * prevent racing processes from causing the surplus to exceed
764 * This however introduces a different race, where a process B
765 * tries to grow the static hugepage pool while alloc_pages() is
766 * called by process A. B will only examine the per-node
767 * counters in determining if surplus huge pages can be
768 * converted to normal huge pages in adjust_pool_surplus(). A
769 * won't be able to increment the per-node counter, until the
770 * lock is dropped by B, but B doesn't drop hugetlb_lock until
771 * no more huge pages can be converted from surplus to normal
772 * state (and doesn't try to convert again). Thus, we have a
773 * case where a surplus huge page exists, the pool is grown, and
774 * the surplus huge page still exists after, even though it
775 * should just have been converted to a normal huge page. This
776 * does not leak memory, though, as the hugepage will be freed
777 * once it is out of use. It also does not allow the counters to
778 * go out of whack in adjust_pool_surplus() as we don't modify
779 * the node values until we've gotten the hugepage and only the
780 * per-node value is checked there.
782 spin_lock(&hugetlb_lock);
783 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
784 spin_unlock(&hugetlb_lock);
788 h->surplus_huge_pages++;
790 spin_unlock(&hugetlb_lock);
792 if (nid == NUMA_NO_NODE)
793 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
794 __GFP_REPEAT|__GFP_NOWARN,
797 page = alloc_pages_exact_node(nid,
798 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
799 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
801 if (page && arch_prepare_hugepage(page)) {
802 __free_pages(page, huge_page_order(h));
806 spin_lock(&hugetlb_lock);
808 r_nid = page_to_nid(page);
809 set_compound_page_dtor(page, free_huge_page);
811 * We incremented the global counters already
813 h->nr_huge_pages_node[r_nid]++;
814 h->surplus_huge_pages_node[r_nid]++;
815 __count_vm_event(HTLB_BUDDY_PGALLOC);
818 h->surplus_huge_pages--;
819 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
821 spin_unlock(&hugetlb_lock);
827 * This allocation function is useful in the context where vma is irrelevant.
828 * E.g. soft-offlining uses this function because it only cares physical
829 * address of error page.
831 struct page *alloc_huge_page_node(struct hstate *h, int nid)
835 spin_lock(&hugetlb_lock);
836 page = dequeue_huge_page_node(h, nid);
837 spin_unlock(&hugetlb_lock);
840 page = alloc_buddy_huge_page(h, nid);
846 * Increase the hugetlb pool such that it can accommodate a reservation
849 static int gather_surplus_pages(struct hstate *h, int delta)
851 struct list_head surplus_list;
852 struct page *page, *tmp;
854 int needed, allocated;
856 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
858 h->resv_huge_pages += delta;
863 INIT_LIST_HEAD(&surplus_list);
867 spin_unlock(&hugetlb_lock);
868 for (i = 0; i < needed; i++) {
869 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
872 * We were not able to allocate enough pages to
873 * satisfy the entire reservation so we free what
874 * we've allocated so far.
878 list_add(&page->lru, &surplus_list);
883 * After retaking hugetlb_lock, we need to recalculate 'needed'
884 * because either resv_huge_pages or free_huge_pages may have changed.
886 spin_lock(&hugetlb_lock);
887 needed = (h->resv_huge_pages + delta) -
888 (h->free_huge_pages + allocated);
893 * The surplus_list now contains _at_least_ the number of extra pages
894 * needed to accommodate the reservation. Add the appropriate number
895 * of pages to the hugetlb pool and free the extras back to the buddy
896 * allocator. Commit the entire reservation here to prevent another
897 * process from stealing the pages as they are added to the pool but
898 * before they are reserved.
901 h->resv_huge_pages += delta;
904 /* Free the needed pages to the hugetlb pool */
905 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
908 list_del(&page->lru);
910 * This page is now managed by the hugetlb allocator and has
911 * no users -- drop the buddy allocator's reference.
913 put_page_testzero(page);
914 VM_BUG_ON(page_count(page));
915 enqueue_huge_page(h, page);
917 spin_unlock(&hugetlb_lock);
919 /* Free unnecessary surplus pages to the buddy allocator */
921 if (!list_empty(&surplus_list)) {
922 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
923 list_del(&page->lru);
927 spin_lock(&hugetlb_lock);
933 * When releasing a hugetlb pool reservation, any surplus pages that were
934 * allocated to satisfy the reservation must be explicitly freed if they were
936 * Called with hugetlb_lock held.
938 static void return_unused_surplus_pages(struct hstate *h,
939 unsigned long unused_resv_pages)
941 unsigned long nr_pages;
943 /* Uncommit the reservation */
944 h->resv_huge_pages -= unused_resv_pages;
946 /* Cannot return gigantic pages currently */
947 if (h->order >= MAX_ORDER)
950 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
953 * We want to release as many surplus pages as possible, spread
954 * evenly across all nodes with memory. Iterate across these nodes
955 * until we can no longer free unreserved surplus pages. This occurs
956 * when the nodes with surplus pages have no free pages.
957 * free_pool_huge_page() will balance the the freed pages across the
958 * on-line nodes with memory and will handle the hstate accounting.
961 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
967 * Determine if the huge page at addr within the vma has an associated
968 * reservation. Where it does not we will need to logically increase
969 * reservation and actually increase quota before an allocation can occur.
970 * Where any new reservation would be required the reservation change is
971 * prepared, but not committed. Once the page has been quota'd allocated
972 * an instantiated the change should be committed via vma_commit_reservation.
973 * No action is required on failure.
975 static long vma_needs_reservation(struct hstate *h,
976 struct vm_area_struct *vma, unsigned long addr)
978 struct address_space *mapping = vma->vm_file->f_mapping;
979 struct inode *inode = mapping->host;
981 if (vma->vm_flags & VM_MAYSHARE) {
982 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
983 return region_chg(&inode->i_mapping->private_list,
986 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
991 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
992 struct resv_map *reservations = vma_resv_map(vma);
994 err = region_chg(&reservations->regions, idx, idx + 1);
1000 static void vma_commit_reservation(struct hstate *h,
1001 struct vm_area_struct *vma, unsigned long addr)
1003 struct address_space *mapping = vma->vm_file->f_mapping;
1004 struct inode *inode = mapping->host;
1006 if (vma->vm_flags & VM_MAYSHARE) {
1007 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1008 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1010 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1011 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1012 struct resv_map *reservations = vma_resv_map(vma);
1014 /* Mark this page used in the map. */
1015 region_add(&reservations->regions, idx, idx + 1);
1019 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1020 unsigned long addr, int avoid_reserve)
1022 struct hstate *h = hstate_vma(vma);
1024 struct address_space *mapping = vma->vm_file->f_mapping;
1025 struct inode *inode = mapping->host;
1029 * Processes that did not create the mapping will have no reserves and
1030 * will not have accounted against quota. Check that the quota can be
1031 * made before satisfying the allocation
1032 * MAP_NORESERVE mappings may also need pages and quota allocated
1033 * if no reserve mapping overlaps.
1035 chg = vma_needs_reservation(h, vma, addr);
1037 return ERR_PTR(-VM_FAULT_OOM);
1039 if (hugetlb_get_quota(inode->i_mapping, chg))
1040 return ERR_PTR(-VM_FAULT_SIGBUS);
1042 spin_lock(&hugetlb_lock);
1043 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1044 spin_unlock(&hugetlb_lock);
1047 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1049 hugetlb_put_quota(inode->i_mapping, chg);
1050 return ERR_PTR(-VM_FAULT_SIGBUS);
1054 set_page_private(page, (unsigned long) mapping);
1056 vma_commit_reservation(h, vma, addr);
1061 int __weak alloc_bootmem_huge_page(struct hstate *h)
1063 struct huge_bootmem_page *m;
1064 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1069 addr = __alloc_bootmem_node_nopanic(
1070 NODE_DATA(hstate_next_node_to_alloc(h,
1071 &node_states[N_HIGH_MEMORY])),
1072 huge_page_size(h), huge_page_size(h), 0);
1076 * Use the beginning of the huge page to store the
1077 * huge_bootmem_page struct (until gather_bootmem
1078 * puts them into the mem_map).
1088 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1089 /* Put them into a private list first because mem_map is not up yet */
1090 list_add(&m->list, &huge_boot_pages);
1095 static void prep_compound_huge_page(struct page *page, int order)
1097 if (unlikely(order > (MAX_ORDER - 1)))
1098 prep_compound_gigantic_page(page, order);
1100 prep_compound_page(page, order);
1103 /* Put bootmem huge pages into the standard lists after mem_map is up */
1104 static void __init gather_bootmem_prealloc(void)
1106 struct huge_bootmem_page *m;
1108 list_for_each_entry(m, &huge_boot_pages, list) {
1109 struct page *page = virt_to_page(m);
1110 struct hstate *h = m->hstate;
1111 __ClearPageReserved(page);
1112 WARN_ON(page_count(page) != 1);
1113 prep_compound_huge_page(page, h->order);
1114 prep_new_huge_page(h, page, page_to_nid(page));
1116 * If we had gigantic hugepages allocated at boot time, we need
1117 * to restore the 'stolen' pages to totalram_pages in order to
1118 * fix confusing memory reports from free(1) and another
1119 * side-effects, like CommitLimit going negative.
1121 if (h->order > (MAX_ORDER - 1))
1122 totalram_pages += 1 << h->order;
1126 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1130 for (i = 0; i < h->max_huge_pages; ++i) {
1131 if (h->order >= MAX_ORDER) {
1132 if (!alloc_bootmem_huge_page(h))
1134 } else if (!alloc_fresh_huge_page(h,
1135 &node_states[N_HIGH_MEMORY]))
1138 h->max_huge_pages = i;
1141 static void __init hugetlb_init_hstates(void)
1145 for_each_hstate(h) {
1146 /* oversize hugepages were init'ed in early boot */
1147 if (h->order < MAX_ORDER)
1148 hugetlb_hstate_alloc_pages(h);
1152 static char * __init memfmt(char *buf, unsigned long n)
1154 if (n >= (1UL << 30))
1155 sprintf(buf, "%lu GB", n >> 30);
1156 else if (n >= (1UL << 20))
1157 sprintf(buf, "%lu MB", n >> 20);
1159 sprintf(buf, "%lu KB", n >> 10);
1163 static void __init report_hugepages(void)
1167 for_each_hstate(h) {
1169 printk(KERN_INFO "HugeTLB registered %s page size, "
1170 "pre-allocated %ld pages\n",
1171 memfmt(buf, huge_page_size(h)),
1172 h->free_huge_pages);
1176 #ifdef CONFIG_HIGHMEM
1177 static void try_to_free_low(struct hstate *h, unsigned long count,
1178 nodemask_t *nodes_allowed)
1182 if (h->order >= MAX_ORDER)
1185 for_each_node_mask(i, *nodes_allowed) {
1186 struct page *page, *next;
1187 struct list_head *freel = &h->hugepage_freelists[i];
1188 list_for_each_entry_safe(page, next, freel, lru) {
1189 if (count >= h->nr_huge_pages)
1191 if (PageHighMem(page))
1193 list_del(&page->lru);
1194 update_and_free_page(h, page);
1195 h->free_huge_pages--;
1196 h->free_huge_pages_node[page_to_nid(page)]--;
1201 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1202 nodemask_t *nodes_allowed)
1208 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1209 * balanced by operating on them in a round-robin fashion.
1210 * Returns 1 if an adjustment was made.
1212 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1215 int start_nid, next_nid;
1218 VM_BUG_ON(delta != -1 && delta != 1);
1221 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1223 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1224 next_nid = start_nid;
1230 * To shrink on this node, there must be a surplus page
1232 if (!h->surplus_huge_pages_node[nid]) {
1233 next_nid = hstate_next_node_to_alloc(h,
1240 * Surplus cannot exceed the total number of pages
1242 if (h->surplus_huge_pages_node[nid] >=
1243 h->nr_huge_pages_node[nid]) {
1244 next_nid = hstate_next_node_to_free(h,
1250 h->surplus_huge_pages += delta;
1251 h->surplus_huge_pages_node[nid] += delta;
1254 } while (next_nid != start_nid);
1259 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1260 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1261 nodemask_t *nodes_allowed)
1263 unsigned long min_count, ret;
1265 if (h->order >= MAX_ORDER)
1266 return h->max_huge_pages;
1269 * Increase the pool size
1270 * First take pages out of surplus state. Then make up the
1271 * remaining difference by allocating fresh huge pages.
1273 * We might race with alloc_buddy_huge_page() here and be unable
1274 * to convert a surplus huge page to a normal huge page. That is
1275 * not critical, though, it just means the overall size of the
1276 * pool might be one hugepage larger than it needs to be, but
1277 * within all the constraints specified by the sysctls.
1279 spin_lock(&hugetlb_lock);
1280 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1281 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1285 while (count > persistent_huge_pages(h)) {
1287 * If this allocation races such that we no longer need the
1288 * page, free_huge_page will handle it by freeing the page
1289 * and reducing the surplus.
1291 spin_unlock(&hugetlb_lock);
1292 ret = alloc_fresh_huge_page(h, nodes_allowed);
1293 spin_lock(&hugetlb_lock);
1297 /* Bail for signals. Probably ctrl-c from user */
1298 if (signal_pending(current))
1303 * Decrease the pool size
1304 * First return free pages to the buddy allocator (being careful
1305 * to keep enough around to satisfy reservations). Then place
1306 * pages into surplus state as needed so the pool will shrink
1307 * to the desired size as pages become free.
1309 * By placing pages into the surplus state independent of the
1310 * overcommit value, we are allowing the surplus pool size to
1311 * exceed overcommit. There are few sane options here. Since
1312 * alloc_buddy_huge_page() is checking the global counter,
1313 * though, we'll note that we're not allowed to exceed surplus
1314 * and won't grow the pool anywhere else. Not until one of the
1315 * sysctls are changed, or the surplus pages go out of use.
1317 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1318 min_count = max(count, min_count);
1319 try_to_free_low(h, min_count, nodes_allowed);
1320 while (min_count < persistent_huge_pages(h)) {
1321 if (!free_pool_huge_page(h, nodes_allowed, 0))
1324 while (count < persistent_huge_pages(h)) {
1325 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1329 ret = persistent_huge_pages(h);
1330 spin_unlock(&hugetlb_lock);
1334 #define HSTATE_ATTR_RO(_name) \
1335 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1337 #define HSTATE_ATTR(_name) \
1338 static struct kobj_attribute _name##_attr = \
1339 __ATTR(_name, 0644, _name##_show, _name##_store)
1341 static struct kobject *hugepages_kobj;
1342 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1344 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1346 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1350 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1351 if (hstate_kobjs[i] == kobj) {
1353 *nidp = NUMA_NO_NODE;
1357 return kobj_to_node_hstate(kobj, nidp);
1360 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1361 struct kobj_attribute *attr, char *buf)
1364 unsigned long nr_huge_pages;
1367 h = kobj_to_hstate(kobj, &nid);
1368 if (nid == NUMA_NO_NODE)
1369 nr_huge_pages = h->nr_huge_pages;
1371 nr_huge_pages = h->nr_huge_pages_node[nid];
1373 return sprintf(buf, "%lu\n", nr_huge_pages);
1376 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1377 struct kobject *kobj, struct kobj_attribute *attr,
1378 const char *buf, size_t len)
1382 unsigned long count;
1384 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1386 err = strict_strtoul(buf, 10, &count);
1390 h = kobj_to_hstate(kobj, &nid);
1391 if (h->order >= MAX_ORDER) {
1396 if (nid == NUMA_NO_NODE) {
1398 * global hstate attribute
1400 if (!(obey_mempolicy &&
1401 init_nodemask_of_mempolicy(nodes_allowed))) {
1402 NODEMASK_FREE(nodes_allowed);
1403 nodes_allowed = &node_states[N_HIGH_MEMORY];
1405 } else if (nodes_allowed) {
1407 * per node hstate attribute: adjust count to global,
1408 * but restrict alloc/free to the specified node.
1410 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1411 init_nodemask_of_node(nodes_allowed, nid);
1413 nodes_allowed = &node_states[N_HIGH_MEMORY];
1415 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1417 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1418 NODEMASK_FREE(nodes_allowed);
1422 NODEMASK_FREE(nodes_allowed);
1426 static ssize_t nr_hugepages_show(struct kobject *kobj,
1427 struct kobj_attribute *attr, char *buf)
1429 return nr_hugepages_show_common(kobj, attr, buf);
1432 static ssize_t nr_hugepages_store(struct kobject *kobj,
1433 struct kobj_attribute *attr, const char *buf, size_t len)
1435 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1437 HSTATE_ATTR(nr_hugepages);
1442 * hstate attribute for optionally mempolicy-based constraint on persistent
1443 * huge page alloc/free.
1445 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1446 struct kobj_attribute *attr, char *buf)
1448 return nr_hugepages_show_common(kobj, attr, buf);
1451 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1452 struct kobj_attribute *attr, const char *buf, size_t len)
1454 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1456 HSTATE_ATTR(nr_hugepages_mempolicy);
1460 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1461 struct kobj_attribute *attr, char *buf)
1463 struct hstate *h = kobj_to_hstate(kobj, NULL);
1464 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1467 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1468 struct kobj_attribute *attr, const char *buf, size_t count)
1471 unsigned long input;
1472 struct hstate *h = kobj_to_hstate(kobj, NULL);
1474 if (h->order >= MAX_ORDER)
1477 err = strict_strtoul(buf, 10, &input);
1481 spin_lock(&hugetlb_lock);
1482 h->nr_overcommit_huge_pages = input;
1483 spin_unlock(&hugetlb_lock);
1487 HSTATE_ATTR(nr_overcommit_hugepages);
1489 static ssize_t free_hugepages_show(struct kobject *kobj,
1490 struct kobj_attribute *attr, char *buf)
1493 unsigned long free_huge_pages;
1496 h = kobj_to_hstate(kobj, &nid);
1497 if (nid == NUMA_NO_NODE)
1498 free_huge_pages = h->free_huge_pages;
1500 free_huge_pages = h->free_huge_pages_node[nid];
1502 return sprintf(buf, "%lu\n", free_huge_pages);
1504 HSTATE_ATTR_RO(free_hugepages);
1506 static ssize_t resv_hugepages_show(struct kobject *kobj,
1507 struct kobj_attribute *attr, char *buf)
1509 struct hstate *h = kobj_to_hstate(kobj, NULL);
1510 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1512 HSTATE_ATTR_RO(resv_hugepages);
1514 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1515 struct kobj_attribute *attr, char *buf)
1518 unsigned long surplus_huge_pages;
1521 h = kobj_to_hstate(kobj, &nid);
1522 if (nid == NUMA_NO_NODE)
1523 surplus_huge_pages = h->surplus_huge_pages;
1525 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1527 return sprintf(buf, "%lu\n", surplus_huge_pages);
1529 HSTATE_ATTR_RO(surplus_hugepages);
1531 static struct attribute *hstate_attrs[] = {
1532 &nr_hugepages_attr.attr,
1533 &nr_overcommit_hugepages_attr.attr,
1534 &free_hugepages_attr.attr,
1535 &resv_hugepages_attr.attr,
1536 &surplus_hugepages_attr.attr,
1538 &nr_hugepages_mempolicy_attr.attr,
1543 static struct attribute_group hstate_attr_group = {
1544 .attrs = hstate_attrs,
1547 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1548 struct kobject **hstate_kobjs,
1549 struct attribute_group *hstate_attr_group)
1552 int hi = h - hstates;
1554 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1555 if (!hstate_kobjs[hi])
1558 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1560 kobject_put(hstate_kobjs[hi]);
1565 static void __init hugetlb_sysfs_init(void)
1570 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1571 if (!hugepages_kobj)
1574 for_each_hstate(h) {
1575 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1576 hstate_kobjs, &hstate_attr_group);
1578 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1586 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1587 * with node sysdevs in node_devices[] using a parallel array. The array
1588 * index of a node sysdev or _hstate == node id.
1589 * This is here to avoid any static dependency of the node sysdev driver, in
1590 * the base kernel, on the hugetlb module.
1592 struct node_hstate {
1593 struct kobject *hugepages_kobj;
1594 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1596 struct node_hstate node_hstates[MAX_NUMNODES];
1599 * A subset of global hstate attributes for node sysdevs
1601 static struct attribute *per_node_hstate_attrs[] = {
1602 &nr_hugepages_attr.attr,
1603 &free_hugepages_attr.attr,
1604 &surplus_hugepages_attr.attr,
1608 static struct attribute_group per_node_hstate_attr_group = {
1609 .attrs = per_node_hstate_attrs,
1613 * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj.
1614 * Returns node id via non-NULL nidp.
1616 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1620 for (nid = 0; nid < nr_node_ids; nid++) {
1621 struct node_hstate *nhs = &node_hstates[nid];
1623 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1624 if (nhs->hstate_kobjs[i] == kobj) {
1636 * Unregister hstate attributes from a single node sysdev.
1637 * No-op if no hstate attributes attached.
1639 void hugetlb_unregister_node(struct node *node)
1642 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1644 if (!nhs->hugepages_kobj)
1645 return; /* no hstate attributes */
1648 if (nhs->hstate_kobjs[h - hstates]) {
1649 kobject_put(nhs->hstate_kobjs[h - hstates]);
1650 nhs->hstate_kobjs[h - hstates] = NULL;
1653 kobject_put(nhs->hugepages_kobj);
1654 nhs->hugepages_kobj = NULL;
1658 * hugetlb module exit: unregister hstate attributes from node sysdevs
1661 static void hugetlb_unregister_all_nodes(void)
1666 * disable node sysdev registrations.
1668 register_hugetlbfs_with_node(NULL, NULL);
1671 * remove hstate attributes from any nodes that have them.
1673 for (nid = 0; nid < nr_node_ids; nid++)
1674 hugetlb_unregister_node(&node_devices[nid]);
1678 * Register hstate attributes for a single node sysdev.
1679 * No-op if attributes already registered.
1681 void hugetlb_register_node(struct node *node)
1684 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1687 if (nhs->hugepages_kobj)
1688 return; /* already allocated */
1690 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1691 &node->sysdev.kobj);
1692 if (!nhs->hugepages_kobj)
1695 for_each_hstate(h) {
1696 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1698 &per_node_hstate_attr_group);
1700 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1702 h->name, node->sysdev.id);
1703 hugetlb_unregister_node(node);
1710 * hugetlb init time: register hstate attributes for all registered node
1711 * sysdevs of nodes that have memory. All on-line nodes should have
1712 * registered their associated sysdev by this time.
1714 static void hugetlb_register_all_nodes(void)
1718 for_each_node_state(nid, N_HIGH_MEMORY) {
1719 struct node *node = &node_devices[nid];
1720 if (node->sysdev.id == nid)
1721 hugetlb_register_node(node);
1725 * Let the node sysdev driver know we're here so it can
1726 * [un]register hstate attributes on node hotplug.
1728 register_hugetlbfs_with_node(hugetlb_register_node,
1729 hugetlb_unregister_node);
1731 #else /* !CONFIG_NUMA */
1733 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1741 static void hugetlb_unregister_all_nodes(void) { }
1743 static void hugetlb_register_all_nodes(void) { }
1747 static void __exit hugetlb_exit(void)
1751 hugetlb_unregister_all_nodes();
1753 for_each_hstate(h) {
1754 kobject_put(hstate_kobjs[h - hstates]);
1757 kobject_put(hugepages_kobj);
1759 module_exit(hugetlb_exit);
1761 static int __init hugetlb_init(void)
1763 /* Some platform decide whether they support huge pages at boot
1764 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1765 * there is no such support
1767 if (HPAGE_SHIFT == 0)
1770 if (!size_to_hstate(default_hstate_size)) {
1771 default_hstate_size = HPAGE_SIZE;
1772 if (!size_to_hstate(default_hstate_size))
1773 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1775 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1776 if (default_hstate_max_huge_pages)
1777 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1779 hugetlb_init_hstates();
1781 gather_bootmem_prealloc();
1785 hugetlb_sysfs_init();
1787 hugetlb_register_all_nodes();
1791 module_init(hugetlb_init);
1793 /* Should be called on processing a hugepagesz=... option */
1794 void __init hugetlb_add_hstate(unsigned order)
1799 if (size_to_hstate(PAGE_SIZE << order)) {
1800 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1803 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1805 h = &hstates[max_hstate++];
1807 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1808 h->nr_huge_pages = 0;
1809 h->free_huge_pages = 0;
1810 for (i = 0; i < MAX_NUMNODES; ++i)
1811 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1812 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1813 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1814 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1815 huge_page_size(h)/1024);
1820 static int __init hugetlb_nrpages_setup(char *s)
1823 static unsigned long *last_mhp;
1826 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1827 * so this hugepages= parameter goes to the "default hstate".
1830 mhp = &default_hstate_max_huge_pages;
1832 mhp = &parsed_hstate->max_huge_pages;
1834 if (mhp == last_mhp) {
1835 printk(KERN_WARNING "hugepages= specified twice without "
1836 "interleaving hugepagesz=, ignoring\n");
1840 if (sscanf(s, "%lu", mhp) <= 0)
1844 * Global state is always initialized later in hugetlb_init.
1845 * But we need to allocate >= MAX_ORDER hstates here early to still
1846 * use the bootmem allocator.
1848 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1849 hugetlb_hstate_alloc_pages(parsed_hstate);
1855 __setup("hugepages=", hugetlb_nrpages_setup);
1857 static int __init hugetlb_default_setup(char *s)
1859 default_hstate_size = memparse(s, &s);
1862 __setup("default_hugepagesz=", hugetlb_default_setup);
1864 static unsigned int cpuset_mems_nr(unsigned int *array)
1867 unsigned int nr = 0;
1869 for_each_node_mask(node, cpuset_current_mems_allowed)
1875 #ifdef CONFIG_SYSCTL
1876 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1877 struct ctl_table *table, int write,
1878 void __user *buffer, size_t *length, loff_t *ppos)
1880 struct hstate *h = &default_hstate;
1884 tmp = h->max_huge_pages;
1886 if (write && h->order >= MAX_ORDER)
1890 table->maxlen = sizeof(unsigned long);
1891 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1896 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1897 GFP_KERNEL | __GFP_NORETRY);
1898 if (!(obey_mempolicy &&
1899 init_nodemask_of_mempolicy(nodes_allowed))) {
1900 NODEMASK_FREE(nodes_allowed);
1901 nodes_allowed = &node_states[N_HIGH_MEMORY];
1903 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1905 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1906 NODEMASK_FREE(nodes_allowed);
1912 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1913 void __user *buffer, size_t *length, loff_t *ppos)
1916 return hugetlb_sysctl_handler_common(false, table, write,
1917 buffer, length, ppos);
1921 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
1922 void __user *buffer, size_t *length, loff_t *ppos)
1924 return hugetlb_sysctl_handler_common(true, table, write,
1925 buffer, length, ppos);
1927 #endif /* CONFIG_NUMA */
1929 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1930 void __user *buffer,
1931 size_t *length, loff_t *ppos)
1933 proc_dointvec(table, write, buffer, length, ppos);
1934 if (hugepages_treat_as_movable)
1935 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1937 htlb_alloc_mask = GFP_HIGHUSER;
1941 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1942 void __user *buffer,
1943 size_t *length, loff_t *ppos)
1945 struct hstate *h = &default_hstate;
1949 tmp = h->nr_overcommit_huge_pages;
1951 if (write && h->order >= MAX_ORDER)
1955 table->maxlen = sizeof(unsigned long);
1956 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1961 spin_lock(&hugetlb_lock);
1962 h->nr_overcommit_huge_pages = tmp;
1963 spin_unlock(&hugetlb_lock);
1969 #endif /* CONFIG_SYSCTL */
1971 void hugetlb_report_meminfo(struct seq_file *m)
1973 struct hstate *h = &default_hstate;
1975 "HugePages_Total: %5lu\n"
1976 "HugePages_Free: %5lu\n"
1977 "HugePages_Rsvd: %5lu\n"
1978 "HugePages_Surp: %5lu\n"
1979 "Hugepagesize: %8lu kB\n",
1983 h->surplus_huge_pages,
1984 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1987 int hugetlb_report_node_meminfo(int nid, char *buf)
1989 struct hstate *h = &default_hstate;
1991 "Node %d HugePages_Total: %5u\n"
1992 "Node %d HugePages_Free: %5u\n"
1993 "Node %d HugePages_Surp: %5u\n",
1994 nid, h->nr_huge_pages_node[nid],
1995 nid, h->free_huge_pages_node[nid],
1996 nid, h->surplus_huge_pages_node[nid]);
1999 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2000 unsigned long hugetlb_total_pages(void)
2002 struct hstate *h = &default_hstate;
2003 return h->nr_huge_pages * pages_per_huge_page(h);
2006 static int hugetlb_acct_memory(struct hstate *h, long delta)
2010 spin_lock(&hugetlb_lock);
2012 * When cpuset is configured, it breaks the strict hugetlb page
2013 * reservation as the accounting is done on a global variable. Such
2014 * reservation is completely rubbish in the presence of cpuset because
2015 * the reservation is not checked against page availability for the
2016 * current cpuset. Application can still potentially OOM'ed by kernel
2017 * with lack of free htlb page in cpuset that the task is in.
2018 * Attempt to enforce strict accounting with cpuset is almost
2019 * impossible (or too ugly) because cpuset is too fluid that
2020 * task or memory node can be dynamically moved between cpusets.
2022 * The change of semantics for shared hugetlb mapping with cpuset is
2023 * undesirable. However, in order to preserve some of the semantics,
2024 * we fall back to check against current free page availability as
2025 * a best attempt and hopefully to minimize the impact of changing
2026 * semantics that cpuset has.
2029 if (gather_surplus_pages(h, delta) < 0)
2032 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2033 return_unused_surplus_pages(h, delta);
2040 return_unused_surplus_pages(h, (unsigned long) -delta);
2043 spin_unlock(&hugetlb_lock);
2047 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2049 struct resv_map *reservations = vma_resv_map(vma);
2052 * This new VMA should share its siblings reservation map if present.
2053 * The VMA will only ever have a valid reservation map pointer where
2054 * it is being copied for another still existing VMA. As that VMA
2055 * has a reference to the reservation map it cannot disappear until
2056 * after this open call completes. It is therefore safe to take a
2057 * new reference here without additional locking.
2060 kref_get(&reservations->refs);
2063 static void resv_map_put(struct vm_area_struct *vma)
2065 struct resv_map *reservations = vma_resv_map(vma);
2069 kref_put(&reservations->refs, resv_map_release);
2072 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2074 struct hstate *h = hstate_vma(vma);
2075 struct resv_map *reservations = vma_resv_map(vma);
2076 unsigned long reserve;
2077 unsigned long start;
2081 start = vma_hugecache_offset(h, vma, vma->vm_start);
2082 end = vma_hugecache_offset(h, vma, vma->vm_end);
2084 reserve = (end - start) -
2085 region_count(&reservations->regions, start, end);
2090 hugetlb_acct_memory(h, -reserve);
2091 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
2097 * We cannot handle pagefaults against hugetlb pages at all. They cause
2098 * handle_mm_fault() to try to instantiate regular-sized pages in the
2099 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2102 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2108 const struct vm_operations_struct hugetlb_vm_ops = {
2109 .fault = hugetlb_vm_op_fault,
2110 .open = hugetlb_vm_op_open,
2111 .close = hugetlb_vm_op_close,
2114 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2121 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2123 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2125 entry = pte_mkyoung(entry);
2126 entry = pte_mkhuge(entry);
2131 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2132 unsigned long address, pte_t *ptep)
2136 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2137 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
2138 update_mmu_cache(vma, address, ptep);
2143 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2144 struct vm_area_struct *vma)
2146 pte_t *src_pte, *dst_pte, entry;
2147 struct page *ptepage;
2150 struct hstate *h = hstate_vma(vma);
2151 unsigned long sz = huge_page_size(h);
2153 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2155 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2156 src_pte = huge_pte_offset(src, addr);
2159 dst_pte = huge_pte_alloc(dst, addr, sz);
2163 /* If the pagetables are shared don't copy or take references */
2164 if (dst_pte == src_pte)
2167 spin_lock(&dst->page_table_lock);
2168 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2169 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2171 huge_ptep_set_wrprotect(src, addr, src_pte);
2172 entry = huge_ptep_get(src_pte);
2173 ptepage = pte_page(entry);
2175 page_dup_rmap(ptepage);
2176 set_huge_pte_at(dst, addr, dst_pte, entry);
2178 spin_unlock(&src->page_table_lock);
2179 spin_unlock(&dst->page_table_lock);
2187 static int is_hugetlb_entry_migration(pte_t pte)
2191 if (huge_pte_none(pte) || pte_present(pte))
2193 swp = pte_to_swp_entry(pte);
2194 if (non_swap_entry(swp) && is_migration_entry(swp)) {
2200 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2204 if (huge_pte_none(pte) || pte_present(pte))
2206 swp = pte_to_swp_entry(pte);
2207 if (non_swap_entry(swp) && is_hwpoison_entry(swp)) {
2213 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2214 unsigned long end, struct page *ref_page)
2216 struct mm_struct *mm = vma->vm_mm;
2217 unsigned long address;
2222 struct hstate *h = hstate_vma(vma);
2223 unsigned long sz = huge_page_size(h);
2226 * A page gathering list, protected by per file i_mmap_mutex. The
2227 * lock is used to avoid list corruption from multiple unmapping
2228 * of the same page since we are using page->lru.
2230 LIST_HEAD(page_list);
2232 WARN_ON(!is_vm_hugetlb_page(vma));
2233 BUG_ON(start & ~huge_page_mask(h));
2234 BUG_ON(end & ~huge_page_mask(h));
2236 mmu_notifier_invalidate_range_start(mm, start, end);
2237 spin_lock(&mm->page_table_lock);
2238 for (address = start; address < end; address += sz) {
2239 ptep = huge_pte_offset(mm, address);
2243 if (huge_pmd_unshare(mm, &address, ptep))
2247 * If a reference page is supplied, it is because a specific
2248 * page is being unmapped, not a range. Ensure the page we
2249 * are about to unmap is the actual page of interest.
2252 pte = huge_ptep_get(ptep);
2253 if (huge_pte_none(pte))
2255 page = pte_page(pte);
2256 if (page != ref_page)
2260 * Mark the VMA as having unmapped its page so that
2261 * future faults in this VMA will fail rather than
2262 * looking like data was lost
2264 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2267 pte = huge_ptep_get_and_clear(mm, address, ptep);
2268 if (huge_pte_none(pte))
2272 * HWPoisoned hugepage is already unmapped and dropped reference
2274 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2277 page = pte_page(pte);
2279 set_page_dirty(page);
2280 list_add(&page->lru, &page_list);
2282 spin_unlock(&mm->page_table_lock);
2283 flush_tlb_range(vma, start, end);
2284 mmu_notifier_invalidate_range_end(mm, start, end);
2285 list_for_each_entry_safe(page, tmp, &page_list, lru) {
2286 page_remove_rmap(page);
2287 list_del(&page->lru);
2292 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2293 unsigned long end, struct page *ref_page)
2295 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2296 __unmap_hugepage_range(vma, start, end, ref_page);
2297 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2301 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2302 * mappping it owns the reserve page for. The intention is to unmap the page
2303 * from other VMAs and let the children be SIGKILLed if they are faulting the
2306 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2307 struct page *page, unsigned long address)
2309 struct hstate *h = hstate_vma(vma);
2310 struct vm_area_struct *iter_vma;
2311 struct address_space *mapping;
2312 struct prio_tree_iter iter;
2316 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2317 * from page cache lookup which is in HPAGE_SIZE units.
2319 address = address & huge_page_mask(h);
2320 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
2321 + (vma->vm_pgoff >> PAGE_SHIFT);
2322 mapping = (struct address_space *)page_private(page);
2325 * Take the mapping lock for the duration of the table walk. As
2326 * this mapping should be shared between all the VMAs,
2327 * __unmap_hugepage_range() is called as the lock is already held
2329 mutex_lock(&mapping->i_mmap_mutex);
2330 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2331 /* Do not unmap the current VMA */
2332 if (iter_vma == vma)
2336 * Unmap the page from other VMAs without their own reserves.
2337 * They get marked to be SIGKILLed if they fault in these
2338 * areas. This is because a future no-page fault on this VMA
2339 * could insert a zeroed page instead of the data existing
2340 * from the time of fork. This would look like data corruption
2342 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2343 __unmap_hugepage_range(iter_vma,
2344 address, address + huge_page_size(h),
2347 mutex_unlock(&mapping->i_mmap_mutex);
2353 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2355 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2356 unsigned long address, pte_t *ptep, pte_t pte,
2357 struct page *pagecache_page)
2359 struct hstate *h = hstate_vma(vma);
2360 struct page *old_page, *new_page;
2362 int outside_reserve = 0;
2364 old_page = pte_page(pte);
2367 /* If no-one else is actually using this page, avoid the copy
2368 * and just make the page writable */
2369 avoidcopy = (page_mapcount(old_page) == 1);
2371 if (PageAnon(old_page))
2372 page_move_anon_rmap(old_page, vma, address);
2373 set_huge_ptep_writable(vma, address, ptep);
2378 * If the process that created a MAP_PRIVATE mapping is about to
2379 * perform a COW due to a shared page count, attempt to satisfy
2380 * the allocation without using the existing reserves. The pagecache
2381 * page is used to determine if the reserve at this address was
2382 * consumed or not. If reserves were used, a partial faulted mapping
2383 * at the time of fork() could consume its reserves on COW instead
2384 * of the full address range.
2386 if (!(vma->vm_flags & VM_MAYSHARE) &&
2387 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2388 old_page != pagecache_page)
2389 outside_reserve = 1;
2391 page_cache_get(old_page);
2393 /* Drop page_table_lock as buddy allocator may be called */
2394 spin_unlock(&mm->page_table_lock);
2395 new_page = alloc_huge_page(vma, address, outside_reserve);
2397 if (IS_ERR(new_page)) {
2398 page_cache_release(old_page);
2401 * If a process owning a MAP_PRIVATE mapping fails to COW,
2402 * it is due to references held by a child and an insufficient
2403 * huge page pool. To guarantee the original mappers
2404 * reliability, unmap the page from child processes. The child
2405 * may get SIGKILLed if it later faults.
2407 if (outside_reserve) {
2408 BUG_ON(huge_pte_none(pte));
2409 if (unmap_ref_private(mm, vma, old_page, address)) {
2410 BUG_ON(huge_pte_none(pte));
2411 spin_lock(&mm->page_table_lock);
2412 goto retry_avoidcopy;
2417 /* Caller expects lock to be held */
2418 spin_lock(&mm->page_table_lock);
2419 return -PTR_ERR(new_page);
2423 * When the original hugepage is shared one, it does not have
2424 * anon_vma prepared.
2426 if (unlikely(anon_vma_prepare(vma))) {
2427 page_cache_release(new_page);
2428 page_cache_release(old_page);
2429 /* Caller expects lock to be held */
2430 spin_lock(&mm->page_table_lock);
2431 return VM_FAULT_OOM;
2434 copy_user_huge_page(new_page, old_page, address, vma,
2435 pages_per_huge_page(h));
2436 __SetPageUptodate(new_page);
2439 * Retake the page_table_lock to check for racing updates
2440 * before the page tables are altered
2442 spin_lock(&mm->page_table_lock);
2443 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2444 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2446 mmu_notifier_invalidate_range_start(mm,
2447 address & huge_page_mask(h),
2448 (address & huge_page_mask(h)) + huge_page_size(h));
2449 huge_ptep_clear_flush(vma, address, ptep);
2450 set_huge_pte_at(mm, address, ptep,
2451 make_huge_pte(vma, new_page, 1));
2452 page_remove_rmap(old_page);
2453 hugepage_add_new_anon_rmap(new_page, vma, address);
2454 /* Make the old page be freed below */
2455 new_page = old_page;
2456 mmu_notifier_invalidate_range_end(mm,
2457 address & huge_page_mask(h),
2458 (address & huge_page_mask(h)) + huge_page_size(h));
2460 page_cache_release(new_page);
2461 page_cache_release(old_page);
2465 /* Return the pagecache page at a given address within a VMA */
2466 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2467 struct vm_area_struct *vma, unsigned long address)
2469 struct address_space *mapping;
2472 mapping = vma->vm_file->f_mapping;
2473 idx = vma_hugecache_offset(h, vma, address);
2475 return find_lock_page(mapping, idx);
2479 * Return whether there is a pagecache page to back given address within VMA.
2480 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2482 static bool hugetlbfs_pagecache_present(struct hstate *h,
2483 struct vm_area_struct *vma, unsigned long address)
2485 struct address_space *mapping;
2489 mapping = vma->vm_file->f_mapping;
2490 idx = vma_hugecache_offset(h, vma, address);
2492 page = find_get_page(mapping, idx);
2495 return page != NULL;
2498 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2499 unsigned long address, pte_t *ptep, unsigned int flags)
2501 struct hstate *h = hstate_vma(vma);
2502 int ret = VM_FAULT_SIGBUS;
2506 struct address_space *mapping;
2510 * Currently, we are forced to kill the process in the event the
2511 * original mapper has unmapped pages from the child due to a failed
2512 * COW. Warn that such a situation has occurred as it may not be obvious
2514 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2516 "PID %d killed due to inadequate hugepage pool\n",
2521 mapping = vma->vm_file->f_mapping;
2522 idx = vma_hugecache_offset(h, vma, address);
2525 * Use page lock to guard against racing truncation
2526 * before we get page_table_lock.
2529 page = find_lock_page(mapping, idx);
2531 size = i_size_read(mapping->host) >> huge_page_shift(h);
2534 page = alloc_huge_page(vma, address, 0);
2536 ret = -PTR_ERR(page);
2539 clear_huge_page(page, address, pages_per_huge_page(h));
2540 __SetPageUptodate(page);
2542 if (vma->vm_flags & VM_MAYSHARE) {
2544 struct inode *inode = mapping->host;
2546 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2554 spin_lock(&inode->i_lock);
2555 inode->i_blocks += blocks_per_huge_page(h);
2556 spin_unlock(&inode->i_lock);
2557 page_dup_rmap(page);
2560 if (unlikely(anon_vma_prepare(vma))) {
2562 goto backout_unlocked;
2564 hugepage_add_new_anon_rmap(page, vma, address);
2568 * If memory error occurs between mmap() and fault, some process
2569 * don't have hwpoisoned swap entry for errored virtual address.
2570 * So we need to block hugepage fault by PG_hwpoison bit check.
2572 if (unlikely(PageHWPoison(page))) {
2573 ret = VM_FAULT_HWPOISON |
2574 VM_FAULT_SET_HINDEX(h - hstates);
2575 goto backout_unlocked;
2577 page_dup_rmap(page);
2581 * If we are going to COW a private mapping later, we examine the
2582 * pending reservations for this page now. This will ensure that
2583 * any allocations necessary to record that reservation occur outside
2586 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2587 if (vma_needs_reservation(h, vma, address) < 0) {
2589 goto backout_unlocked;
2592 spin_lock(&mm->page_table_lock);
2593 size = i_size_read(mapping->host) >> huge_page_shift(h);
2598 if (!huge_pte_none(huge_ptep_get(ptep)))
2601 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2602 && (vma->vm_flags & VM_SHARED)));
2603 set_huge_pte_at(mm, address, ptep, new_pte);
2605 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2606 /* Optimization, do the COW without a second fault */
2607 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2610 spin_unlock(&mm->page_table_lock);
2616 spin_unlock(&mm->page_table_lock);
2623 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2624 unsigned long address, unsigned int flags)
2629 struct page *page = NULL;
2630 struct page *pagecache_page = NULL;
2631 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2632 struct hstate *h = hstate_vma(vma);
2634 ptep = huge_pte_offset(mm, address);
2636 entry = huge_ptep_get(ptep);
2637 if (unlikely(is_hugetlb_entry_migration(entry))) {
2638 migration_entry_wait(mm, (pmd_t *)ptep, address);
2640 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2641 return VM_FAULT_HWPOISON_LARGE |
2642 VM_FAULT_SET_HINDEX(h - hstates);
2645 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2647 return VM_FAULT_OOM;
2650 * Serialize hugepage allocation and instantiation, so that we don't
2651 * get spurious allocation failures if two CPUs race to instantiate
2652 * the same page in the page cache.
2654 mutex_lock(&hugetlb_instantiation_mutex);
2655 entry = huge_ptep_get(ptep);
2656 if (huge_pte_none(entry)) {
2657 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2664 * If we are going to COW the mapping later, we examine the pending
2665 * reservations for this page now. This will ensure that any
2666 * allocations necessary to record that reservation occur outside the
2667 * spinlock. For private mappings, we also lookup the pagecache
2668 * page now as it is used to determine if a reservation has been
2671 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2672 if (vma_needs_reservation(h, vma, address) < 0) {
2677 if (!(vma->vm_flags & VM_MAYSHARE))
2678 pagecache_page = hugetlbfs_pagecache_page(h,
2683 * hugetlb_cow() requires page locks of pte_page(entry) and
2684 * pagecache_page, so here we need take the former one
2685 * when page != pagecache_page or !pagecache_page.
2686 * Note that locking order is always pagecache_page -> page,
2687 * so no worry about deadlock.
2689 page = pte_page(entry);
2691 if (page != pagecache_page)
2694 spin_lock(&mm->page_table_lock);
2695 /* Check for a racing update before calling hugetlb_cow */
2696 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2697 goto out_page_table_lock;
2700 if (flags & FAULT_FLAG_WRITE) {
2701 if (!pte_write(entry)) {
2702 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2704 goto out_page_table_lock;
2706 entry = pte_mkdirty(entry);
2708 entry = pte_mkyoung(entry);
2709 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2710 flags & FAULT_FLAG_WRITE))
2711 update_mmu_cache(vma, address, ptep);
2713 out_page_table_lock:
2714 spin_unlock(&mm->page_table_lock);
2716 if (pagecache_page) {
2717 unlock_page(pagecache_page);
2718 put_page(pagecache_page);
2720 if (page != pagecache_page)
2725 mutex_unlock(&hugetlb_instantiation_mutex);
2730 /* Can be overriden by architectures */
2731 __attribute__((weak)) struct page *
2732 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2733 pud_t *pud, int write)
2739 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2740 struct page **pages, struct vm_area_struct **vmas,
2741 unsigned long *position, int *length, int i,
2744 unsigned long pfn_offset;
2745 unsigned long vaddr = *position;
2746 int remainder = *length;
2747 struct hstate *h = hstate_vma(vma);
2749 spin_lock(&mm->page_table_lock);
2750 while (vaddr < vma->vm_end && remainder) {
2756 * Some archs (sparc64, sh*) have multiple pte_ts to
2757 * each hugepage. We have to make sure we get the
2758 * first, for the page indexing below to work.
2760 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2761 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2764 * When coredumping, it suits get_dump_page if we just return
2765 * an error where there's an empty slot with no huge pagecache
2766 * to back it. This way, we avoid allocating a hugepage, and
2767 * the sparse dumpfile avoids allocating disk blocks, but its
2768 * huge holes still show up with zeroes where they need to be.
2770 if (absent && (flags & FOLL_DUMP) &&
2771 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2777 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2780 spin_unlock(&mm->page_table_lock);
2781 ret = hugetlb_fault(mm, vma, vaddr,
2782 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2783 spin_lock(&mm->page_table_lock);
2784 if (!(ret & VM_FAULT_ERROR))
2791 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2792 page = pte_page(huge_ptep_get(pte));
2795 pages[i] = mem_map_offset(page, pfn_offset);
2806 if (vaddr < vma->vm_end && remainder &&
2807 pfn_offset < pages_per_huge_page(h)) {
2809 * We use pfn_offset to avoid touching the pageframes
2810 * of this compound page.
2815 spin_unlock(&mm->page_table_lock);
2816 *length = remainder;
2819 return i ? i : -EFAULT;
2822 void hugetlb_change_protection(struct vm_area_struct *vma,
2823 unsigned long address, unsigned long end, pgprot_t newprot)
2825 struct mm_struct *mm = vma->vm_mm;
2826 unsigned long start = address;
2829 struct hstate *h = hstate_vma(vma);
2831 BUG_ON(address >= end);
2832 flush_cache_range(vma, address, end);
2834 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2835 spin_lock(&mm->page_table_lock);
2836 for (; address < end; address += huge_page_size(h)) {
2837 ptep = huge_pte_offset(mm, address);
2840 if (huge_pmd_unshare(mm, &address, ptep))
2842 if (!huge_pte_none(huge_ptep_get(ptep))) {
2843 pte = huge_ptep_get_and_clear(mm, address, ptep);
2844 pte = pte_mkhuge(pte_modify(pte, newprot));
2845 set_huge_pte_at(mm, address, ptep, pte);
2848 spin_unlock(&mm->page_table_lock);
2849 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2851 flush_tlb_range(vma, start, end);
2854 int hugetlb_reserve_pages(struct inode *inode,
2856 struct vm_area_struct *vma,
2857 vm_flags_t vm_flags)
2860 struct hstate *h = hstate_inode(inode);
2863 * Only apply hugepage reservation if asked. At fault time, an
2864 * attempt will be made for VM_NORESERVE to allocate a page
2865 * and filesystem quota without using reserves
2867 if (vm_flags & VM_NORESERVE)
2871 * Shared mappings base their reservation on the number of pages that
2872 * are already allocated on behalf of the file. Private mappings need
2873 * to reserve the full area even if read-only as mprotect() may be
2874 * called to make the mapping read-write. Assume !vma is a shm mapping
2876 if (!vma || vma->vm_flags & VM_MAYSHARE)
2877 chg = region_chg(&inode->i_mapping->private_list, from, to);
2879 struct resv_map *resv_map = resv_map_alloc();
2885 set_vma_resv_map(vma, resv_map);
2886 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2894 /* There must be enough filesystem quota for the mapping */
2895 if (hugetlb_get_quota(inode->i_mapping, chg)) {
2901 * Check enough hugepages are available for the reservation.
2902 * Hand back the quota if there are not
2904 ret = hugetlb_acct_memory(h, chg);
2906 hugetlb_put_quota(inode->i_mapping, chg);
2911 * Account for the reservations made. Shared mappings record regions
2912 * that have reservations as they are shared by multiple VMAs.
2913 * When the last VMA disappears, the region map says how much
2914 * the reservation was and the page cache tells how much of
2915 * the reservation was consumed. Private mappings are per-VMA and
2916 * only the consumed reservations are tracked. When the VMA
2917 * disappears, the original reservation is the VMA size and the
2918 * consumed reservations are stored in the map. Hence, nothing
2919 * else has to be done for private mappings here
2921 if (!vma || vma->vm_flags & VM_MAYSHARE)
2922 region_add(&inode->i_mapping->private_list, from, to);
2930 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2932 struct hstate *h = hstate_inode(inode);
2933 long chg = region_truncate(&inode->i_mapping->private_list, offset);
2935 spin_lock(&inode->i_lock);
2936 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
2937 spin_unlock(&inode->i_lock);
2939 hugetlb_put_quota(inode->i_mapping, (chg - freed));
2940 hugetlb_acct_memory(h, -(chg - freed));
2943 #ifdef CONFIG_MEMORY_FAILURE
2945 /* Should be called in hugetlb_lock */
2946 static int is_hugepage_on_freelist(struct page *hpage)
2950 struct hstate *h = page_hstate(hpage);
2951 int nid = page_to_nid(hpage);
2953 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
2960 * This function is called from memory failure code.
2961 * Assume the caller holds page lock of the head page.
2963 int dequeue_hwpoisoned_huge_page(struct page *hpage)
2965 struct hstate *h = page_hstate(hpage);
2966 int nid = page_to_nid(hpage);
2969 spin_lock(&hugetlb_lock);
2970 if (is_hugepage_on_freelist(hpage)) {
2971 list_del(&hpage->lru);
2972 set_page_refcounted(hpage);
2973 h->free_huge_pages--;
2974 h->free_huge_pages_node[nid]--;
2977 spin_unlock(&hugetlb_lock);