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 |
507 1 << PG_referenced | 1 << PG_dirty |
508 1 << PG_active | 1 << PG_reserved |
509 1 << PG_private | 1 << PG_writeback);
511 set_compound_page_dtor(page, NULL);
512 set_page_refcounted(page);
513 arch_release_hugepage(page);
514 __free_pages(page, huge_page_order(h));
517 struct hstate *size_to_hstate(unsigned long size)
522 if (huge_page_size(h) == size)
528 static void free_huge_page(struct page *page)
531 * Can't pass hstate in here because it is called from the
532 * compound page destructor.
534 struct hstate *h = page_hstate(page);
535 int nid = page_to_nid(page);
536 struct address_space *mapping;
538 mapping = (struct address_space *) page_private(page);
539 set_page_private(page, 0);
540 page->mapping = NULL;
541 BUG_ON(page_count(page));
542 BUG_ON(page_mapcount(page));
543 INIT_LIST_HEAD(&page->lru);
545 spin_lock(&hugetlb_lock);
546 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
547 update_and_free_page(h, page);
548 h->surplus_huge_pages--;
549 h->surplus_huge_pages_node[nid]--;
551 enqueue_huge_page(h, page);
553 spin_unlock(&hugetlb_lock);
555 hugetlb_put_quota(mapping, 1);
558 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
560 set_compound_page_dtor(page, free_huge_page);
561 spin_lock(&hugetlb_lock);
563 h->nr_huge_pages_node[nid]++;
564 spin_unlock(&hugetlb_lock);
565 put_page(page); /* free it into the hugepage allocator */
568 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
571 int nr_pages = 1 << order;
572 struct page *p = page + 1;
574 /* we rely on prep_new_huge_page to set the destructor */
575 set_compound_order(page, order);
577 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
579 set_page_count(p, 0);
580 p->first_page = page;
584 int PageHuge(struct page *page)
586 compound_page_dtor *dtor;
588 if (!PageCompound(page))
591 page = compound_head(page);
592 dtor = get_compound_page_dtor(page);
594 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;
855 bool alloc_ok = true;
857 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
859 h->resv_huge_pages += delta;
864 INIT_LIST_HEAD(&surplus_list);
868 spin_unlock(&hugetlb_lock);
869 for (i = 0; i < needed; i++) {
870 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
875 list_add(&page->lru, &surplus_list);
880 * After retaking hugetlb_lock, we need to recalculate 'needed'
881 * because either resv_huge_pages or free_huge_pages may have changed.
883 spin_lock(&hugetlb_lock);
884 needed = (h->resv_huge_pages + delta) -
885 (h->free_huge_pages + allocated);
890 * We were not able to allocate enough pages to
891 * satisfy the entire reservation so we free what
892 * we've allocated so far.
897 * The surplus_list now contains _at_least_ the number of extra pages
898 * needed to accommodate the reservation. Add the appropriate number
899 * of pages to the hugetlb pool and free the extras back to the buddy
900 * allocator. Commit the entire reservation here to prevent another
901 * process from stealing the pages as they are added to the pool but
902 * before they are reserved.
905 h->resv_huge_pages += delta;
908 /* Free the needed pages to the hugetlb pool */
909 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
912 list_del(&page->lru);
914 * This page is now managed by the hugetlb allocator and has
915 * no users -- drop the buddy allocator's reference.
917 put_page_testzero(page);
918 VM_BUG_ON(page_count(page));
919 enqueue_huge_page(h, page);
922 spin_unlock(&hugetlb_lock);
924 /* Free unnecessary surplus pages to the buddy allocator */
925 if (!list_empty(&surplus_list)) {
926 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
927 list_del(&page->lru);
931 spin_lock(&hugetlb_lock);
937 * When releasing a hugetlb pool reservation, any surplus pages that were
938 * allocated to satisfy the reservation must be explicitly freed if they were
940 * Called with hugetlb_lock held.
942 static void return_unused_surplus_pages(struct hstate *h,
943 unsigned long unused_resv_pages)
945 unsigned long nr_pages;
947 /* Uncommit the reservation */
948 h->resv_huge_pages -= unused_resv_pages;
950 /* Cannot return gigantic pages currently */
951 if (h->order >= MAX_ORDER)
954 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
957 * We want to release as many surplus pages as possible, spread
958 * evenly across all nodes with memory. Iterate across these nodes
959 * until we can no longer free unreserved surplus pages. This occurs
960 * when the nodes with surplus pages have no free pages.
961 * free_pool_huge_page() will balance the the freed pages across the
962 * on-line nodes with memory and will handle the hstate accounting.
965 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
971 * Determine if the huge page at addr within the vma has an associated
972 * reservation. Where it does not we will need to logically increase
973 * reservation and actually increase quota before an allocation can occur.
974 * Where any new reservation would be required the reservation change is
975 * prepared, but not committed. Once the page has been quota'd allocated
976 * an instantiated the change should be committed via vma_commit_reservation.
977 * No action is required on failure.
979 static long vma_needs_reservation(struct hstate *h,
980 struct vm_area_struct *vma, unsigned long addr)
982 struct address_space *mapping = vma->vm_file->f_mapping;
983 struct inode *inode = mapping->host;
985 if (vma->vm_flags & VM_MAYSHARE) {
986 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
987 return region_chg(&inode->i_mapping->private_list,
990 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
995 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
996 struct resv_map *reservations = vma_resv_map(vma);
998 err = region_chg(&reservations->regions, idx, idx + 1);
1004 static void vma_commit_reservation(struct hstate *h,
1005 struct vm_area_struct *vma, unsigned long addr)
1007 struct address_space *mapping = vma->vm_file->f_mapping;
1008 struct inode *inode = mapping->host;
1010 if (vma->vm_flags & VM_MAYSHARE) {
1011 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1012 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1014 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1015 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1016 struct resv_map *reservations = vma_resv_map(vma);
1018 /* Mark this page used in the map. */
1019 region_add(&reservations->regions, idx, idx + 1);
1023 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1024 unsigned long addr, int avoid_reserve)
1026 struct hstate *h = hstate_vma(vma);
1028 struct address_space *mapping = vma->vm_file->f_mapping;
1029 struct inode *inode = mapping->host;
1033 * Processes that did not create the mapping will have no reserves and
1034 * will not have accounted against quota. Check that the quota can be
1035 * made before satisfying the allocation
1036 * MAP_NORESERVE mappings may also need pages and quota allocated
1037 * if no reserve mapping overlaps.
1039 chg = vma_needs_reservation(h, vma, addr);
1041 return ERR_PTR(-VM_FAULT_OOM);
1043 if (hugetlb_get_quota(inode->i_mapping, chg))
1044 return ERR_PTR(-VM_FAULT_SIGBUS);
1046 spin_lock(&hugetlb_lock);
1047 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1048 spin_unlock(&hugetlb_lock);
1051 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1053 hugetlb_put_quota(inode->i_mapping, chg);
1054 return ERR_PTR(-VM_FAULT_SIGBUS);
1058 set_page_private(page, (unsigned long) mapping);
1060 vma_commit_reservation(h, vma, addr);
1065 int __weak alloc_bootmem_huge_page(struct hstate *h)
1067 struct huge_bootmem_page *m;
1068 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1073 addr = __alloc_bootmem_node_nopanic(
1074 NODE_DATA(hstate_next_node_to_alloc(h,
1075 &node_states[N_HIGH_MEMORY])),
1076 huge_page_size(h), huge_page_size(h), 0);
1080 * Use the beginning of the huge page to store the
1081 * huge_bootmem_page struct (until gather_bootmem
1082 * puts them into the mem_map).
1092 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1093 /* Put them into a private list first because mem_map is not up yet */
1094 list_add(&m->list, &huge_boot_pages);
1099 static void prep_compound_huge_page(struct page *page, int order)
1101 if (unlikely(order > (MAX_ORDER - 1)))
1102 prep_compound_gigantic_page(page, order);
1104 prep_compound_page(page, order);
1107 /* Put bootmem huge pages into the standard lists after mem_map is up */
1108 static void __init gather_bootmem_prealloc(void)
1110 struct huge_bootmem_page *m;
1112 list_for_each_entry(m, &huge_boot_pages, list) {
1113 struct hstate *h = m->hstate;
1116 #ifdef CONFIG_HIGHMEM
1117 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1118 free_bootmem_late((unsigned long)m,
1119 sizeof(struct huge_bootmem_page));
1121 page = virt_to_page(m);
1123 __ClearPageReserved(page);
1124 WARN_ON(page_count(page) != 1);
1125 prep_compound_huge_page(page, h->order);
1126 prep_new_huge_page(h, page, page_to_nid(page));
1128 * If we had gigantic hugepages allocated at boot time, we need
1129 * to restore the 'stolen' pages to totalram_pages in order to
1130 * fix confusing memory reports from free(1) and another
1131 * side-effects, like CommitLimit going negative.
1133 if (h->order > (MAX_ORDER - 1))
1134 totalram_pages += 1 << h->order;
1138 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1142 for (i = 0; i < h->max_huge_pages; ++i) {
1143 if (h->order >= MAX_ORDER) {
1144 if (!alloc_bootmem_huge_page(h))
1146 } else if (!alloc_fresh_huge_page(h,
1147 &node_states[N_HIGH_MEMORY]))
1150 h->max_huge_pages = i;
1153 static void __init hugetlb_init_hstates(void)
1157 for_each_hstate(h) {
1158 /* oversize hugepages were init'ed in early boot */
1159 if (h->order < MAX_ORDER)
1160 hugetlb_hstate_alloc_pages(h);
1164 static char * __init memfmt(char *buf, unsigned long n)
1166 if (n >= (1UL << 30))
1167 sprintf(buf, "%lu GB", n >> 30);
1168 else if (n >= (1UL << 20))
1169 sprintf(buf, "%lu MB", n >> 20);
1171 sprintf(buf, "%lu KB", n >> 10);
1175 static void __init report_hugepages(void)
1179 for_each_hstate(h) {
1181 printk(KERN_INFO "HugeTLB registered %s page size, "
1182 "pre-allocated %ld pages\n",
1183 memfmt(buf, huge_page_size(h)),
1184 h->free_huge_pages);
1188 #ifdef CONFIG_HIGHMEM
1189 static void try_to_free_low(struct hstate *h, unsigned long count,
1190 nodemask_t *nodes_allowed)
1194 if (h->order >= MAX_ORDER)
1197 for_each_node_mask(i, *nodes_allowed) {
1198 struct page *page, *next;
1199 struct list_head *freel = &h->hugepage_freelists[i];
1200 list_for_each_entry_safe(page, next, freel, lru) {
1201 if (count >= h->nr_huge_pages)
1203 if (PageHighMem(page))
1205 list_del(&page->lru);
1206 update_and_free_page(h, page);
1207 h->free_huge_pages--;
1208 h->free_huge_pages_node[page_to_nid(page)]--;
1213 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1214 nodemask_t *nodes_allowed)
1220 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1221 * balanced by operating on them in a round-robin fashion.
1222 * Returns 1 if an adjustment was made.
1224 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1227 int start_nid, next_nid;
1230 VM_BUG_ON(delta != -1 && delta != 1);
1233 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1235 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1236 next_nid = start_nid;
1242 * To shrink on this node, there must be a surplus page
1244 if (!h->surplus_huge_pages_node[nid]) {
1245 next_nid = hstate_next_node_to_alloc(h,
1252 * Surplus cannot exceed the total number of pages
1254 if (h->surplus_huge_pages_node[nid] >=
1255 h->nr_huge_pages_node[nid]) {
1256 next_nid = hstate_next_node_to_free(h,
1262 h->surplus_huge_pages += delta;
1263 h->surplus_huge_pages_node[nid] += delta;
1266 } while (next_nid != start_nid);
1271 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1272 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1273 nodemask_t *nodes_allowed)
1275 unsigned long min_count, ret;
1277 if (h->order >= MAX_ORDER)
1278 return h->max_huge_pages;
1281 * Increase the pool size
1282 * First take pages out of surplus state. Then make up the
1283 * remaining difference by allocating fresh huge pages.
1285 * We might race with alloc_buddy_huge_page() here and be unable
1286 * to convert a surplus huge page to a normal huge page. That is
1287 * not critical, though, it just means the overall size of the
1288 * pool might be one hugepage larger than it needs to be, but
1289 * within all the constraints specified by the sysctls.
1291 spin_lock(&hugetlb_lock);
1292 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1293 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1297 while (count > persistent_huge_pages(h)) {
1299 * If this allocation races such that we no longer need the
1300 * page, free_huge_page will handle it by freeing the page
1301 * and reducing the surplus.
1303 spin_unlock(&hugetlb_lock);
1304 ret = alloc_fresh_huge_page(h, nodes_allowed);
1305 spin_lock(&hugetlb_lock);
1309 /* Bail for signals. Probably ctrl-c from user */
1310 if (signal_pending(current))
1315 * Decrease the pool size
1316 * First return free pages to the buddy allocator (being careful
1317 * to keep enough around to satisfy reservations). Then place
1318 * pages into surplus state as needed so the pool will shrink
1319 * to the desired size as pages become free.
1321 * By placing pages into the surplus state independent of the
1322 * overcommit value, we are allowing the surplus pool size to
1323 * exceed overcommit. There are few sane options here. Since
1324 * alloc_buddy_huge_page() is checking the global counter,
1325 * though, we'll note that we're not allowed to exceed surplus
1326 * and won't grow the pool anywhere else. Not until one of the
1327 * sysctls are changed, or the surplus pages go out of use.
1329 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1330 min_count = max(count, min_count);
1331 try_to_free_low(h, min_count, nodes_allowed);
1332 while (min_count < persistent_huge_pages(h)) {
1333 if (!free_pool_huge_page(h, nodes_allowed, 0))
1336 while (count < persistent_huge_pages(h)) {
1337 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1341 ret = persistent_huge_pages(h);
1342 spin_unlock(&hugetlb_lock);
1346 #define HSTATE_ATTR_RO(_name) \
1347 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1349 #define HSTATE_ATTR(_name) \
1350 static struct kobj_attribute _name##_attr = \
1351 __ATTR(_name, 0644, _name##_show, _name##_store)
1353 static struct kobject *hugepages_kobj;
1354 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1356 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1358 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1362 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1363 if (hstate_kobjs[i] == kobj) {
1365 *nidp = NUMA_NO_NODE;
1369 return kobj_to_node_hstate(kobj, nidp);
1372 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1373 struct kobj_attribute *attr, char *buf)
1376 unsigned long nr_huge_pages;
1379 h = kobj_to_hstate(kobj, &nid);
1380 if (nid == NUMA_NO_NODE)
1381 nr_huge_pages = h->nr_huge_pages;
1383 nr_huge_pages = h->nr_huge_pages_node[nid];
1385 return sprintf(buf, "%lu\n", nr_huge_pages);
1388 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1389 struct kobject *kobj, struct kobj_attribute *attr,
1390 const char *buf, size_t len)
1394 unsigned long count;
1396 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1398 err = strict_strtoul(buf, 10, &count);
1402 h = kobj_to_hstate(kobj, &nid);
1403 if (h->order >= MAX_ORDER) {
1408 if (nid == NUMA_NO_NODE) {
1410 * global hstate attribute
1412 if (!(obey_mempolicy &&
1413 init_nodemask_of_mempolicy(nodes_allowed))) {
1414 NODEMASK_FREE(nodes_allowed);
1415 nodes_allowed = &node_states[N_HIGH_MEMORY];
1417 } else if (nodes_allowed) {
1419 * per node hstate attribute: adjust count to global,
1420 * but restrict alloc/free to the specified node.
1422 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1423 init_nodemask_of_node(nodes_allowed, nid);
1425 nodes_allowed = &node_states[N_HIGH_MEMORY];
1427 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1429 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1430 NODEMASK_FREE(nodes_allowed);
1434 NODEMASK_FREE(nodes_allowed);
1438 static ssize_t nr_hugepages_show(struct kobject *kobj,
1439 struct kobj_attribute *attr, char *buf)
1441 return nr_hugepages_show_common(kobj, attr, buf);
1444 static ssize_t nr_hugepages_store(struct kobject *kobj,
1445 struct kobj_attribute *attr, const char *buf, size_t len)
1447 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1449 HSTATE_ATTR(nr_hugepages);
1454 * hstate attribute for optionally mempolicy-based constraint on persistent
1455 * huge page alloc/free.
1457 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1458 struct kobj_attribute *attr, char *buf)
1460 return nr_hugepages_show_common(kobj, attr, buf);
1463 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1464 struct kobj_attribute *attr, const char *buf, size_t len)
1466 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1468 HSTATE_ATTR(nr_hugepages_mempolicy);
1472 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1473 struct kobj_attribute *attr, char *buf)
1475 struct hstate *h = kobj_to_hstate(kobj, NULL);
1476 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1479 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1480 struct kobj_attribute *attr, const char *buf, size_t count)
1483 unsigned long input;
1484 struct hstate *h = kobj_to_hstate(kobj, NULL);
1486 if (h->order >= MAX_ORDER)
1489 err = strict_strtoul(buf, 10, &input);
1493 spin_lock(&hugetlb_lock);
1494 h->nr_overcommit_huge_pages = input;
1495 spin_unlock(&hugetlb_lock);
1499 HSTATE_ATTR(nr_overcommit_hugepages);
1501 static ssize_t free_hugepages_show(struct kobject *kobj,
1502 struct kobj_attribute *attr, char *buf)
1505 unsigned long free_huge_pages;
1508 h = kobj_to_hstate(kobj, &nid);
1509 if (nid == NUMA_NO_NODE)
1510 free_huge_pages = h->free_huge_pages;
1512 free_huge_pages = h->free_huge_pages_node[nid];
1514 return sprintf(buf, "%lu\n", free_huge_pages);
1516 HSTATE_ATTR_RO(free_hugepages);
1518 static ssize_t resv_hugepages_show(struct kobject *kobj,
1519 struct kobj_attribute *attr, char *buf)
1521 struct hstate *h = kobj_to_hstate(kobj, NULL);
1522 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1524 HSTATE_ATTR_RO(resv_hugepages);
1526 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1527 struct kobj_attribute *attr, char *buf)
1530 unsigned long surplus_huge_pages;
1533 h = kobj_to_hstate(kobj, &nid);
1534 if (nid == NUMA_NO_NODE)
1535 surplus_huge_pages = h->surplus_huge_pages;
1537 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1539 return sprintf(buf, "%lu\n", surplus_huge_pages);
1541 HSTATE_ATTR_RO(surplus_hugepages);
1543 static struct attribute *hstate_attrs[] = {
1544 &nr_hugepages_attr.attr,
1545 &nr_overcommit_hugepages_attr.attr,
1546 &free_hugepages_attr.attr,
1547 &resv_hugepages_attr.attr,
1548 &surplus_hugepages_attr.attr,
1550 &nr_hugepages_mempolicy_attr.attr,
1555 static struct attribute_group hstate_attr_group = {
1556 .attrs = hstate_attrs,
1559 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1560 struct kobject **hstate_kobjs,
1561 struct attribute_group *hstate_attr_group)
1564 int hi = h - hstates;
1566 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1567 if (!hstate_kobjs[hi])
1570 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1572 kobject_put(hstate_kobjs[hi]);
1577 static void __init hugetlb_sysfs_init(void)
1582 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1583 if (!hugepages_kobj)
1586 for_each_hstate(h) {
1587 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1588 hstate_kobjs, &hstate_attr_group);
1590 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1598 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1599 * with node devices in node_devices[] using a parallel array. The array
1600 * index of a node device or _hstate == node id.
1601 * This is here to avoid any static dependency of the node device driver, in
1602 * the base kernel, on the hugetlb module.
1604 struct node_hstate {
1605 struct kobject *hugepages_kobj;
1606 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1608 struct node_hstate node_hstates[MAX_NUMNODES];
1611 * A subset of global hstate attributes for node devices
1613 static struct attribute *per_node_hstate_attrs[] = {
1614 &nr_hugepages_attr.attr,
1615 &free_hugepages_attr.attr,
1616 &surplus_hugepages_attr.attr,
1620 static struct attribute_group per_node_hstate_attr_group = {
1621 .attrs = per_node_hstate_attrs,
1625 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1626 * Returns node id via non-NULL nidp.
1628 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1632 for (nid = 0; nid < nr_node_ids; nid++) {
1633 struct node_hstate *nhs = &node_hstates[nid];
1635 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1636 if (nhs->hstate_kobjs[i] == kobj) {
1648 * Unregister hstate attributes from a single node device.
1649 * No-op if no hstate attributes attached.
1651 void hugetlb_unregister_node(struct node *node)
1654 struct node_hstate *nhs = &node_hstates[node->dev.id];
1656 if (!nhs->hugepages_kobj)
1657 return; /* no hstate attributes */
1660 if (nhs->hstate_kobjs[h - hstates]) {
1661 kobject_put(nhs->hstate_kobjs[h - hstates]);
1662 nhs->hstate_kobjs[h - hstates] = NULL;
1665 kobject_put(nhs->hugepages_kobj);
1666 nhs->hugepages_kobj = NULL;
1670 * hugetlb module exit: unregister hstate attributes from node devices
1673 static void hugetlb_unregister_all_nodes(void)
1678 * disable node device registrations.
1680 register_hugetlbfs_with_node(NULL, NULL);
1683 * remove hstate attributes from any nodes that have them.
1685 for (nid = 0; nid < nr_node_ids; nid++)
1686 hugetlb_unregister_node(&node_devices[nid]);
1690 * Register hstate attributes for a single node device.
1691 * No-op if attributes already registered.
1693 void hugetlb_register_node(struct node *node)
1696 struct node_hstate *nhs = &node_hstates[node->dev.id];
1699 if (nhs->hugepages_kobj)
1700 return; /* already allocated */
1702 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1704 if (!nhs->hugepages_kobj)
1707 for_each_hstate(h) {
1708 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1710 &per_node_hstate_attr_group);
1712 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1714 h->name, node->dev.id);
1715 hugetlb_unregister_node(node);
1722 * hugetlb init time: register hstate attributes for all registered node
1723 * devices of nodes that have memory. All on-line nodes should have
1724 * registered their associated device by this time.
1726 static void hugetlb_register_all_nodes(void)
1730 for_each_node_state(nid, N_HIGH_MEMORY) {
1731 struct node *node = &node_devices[nid];
1732 if (node->dev.id == nid)
1733 hugetlb_register_node(node);
1737 * Let the node device driver know we're here so it can
1738 * [un]register hstate attributes on node hotplug.
1740 register_hugetlbfs_with_node(hugetlb_register_node,
1741 hugetlb_unregister_node);
1743 #else /* !CONFIG_NUMA */
1745 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1753 static void hugetlb_unregister_all_nodes(void) { }
1755 static void hugetlb_register_all_nodes(void) { }
1759 static void __exit hugetlb_exit(void)
1763 hugetlb_unregister_all_nodes();
1765 for_each_hstate(h) {
1766 kobject_put(hstate_kobjs[h - hstates]);
1769 kobject_put(hugepages_kobj);
1771 module_exit(hugetlb_exit);
1773 static int __init hugetlb_init(void)
1775 /* Some platform decide whether they support huge pages at boot
1776 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1777 * there is no such support
1779 if (HPAGE_SHIFT == 0)
1782 if (!size_to_hstate(default_hstate_size)) {
1783 default_hstate_size = HPAGE_SIZE;
1784 if (!size_to_hstate(default_hstate_size))
1785 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1787 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1788 if (default_hstate_max_huge_pages)
1789 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1791 hugetlb_init_hstates();
1793 gather_bootmem_prealloc();
1797 hugetlb_sysfs_init();
1799 hugetlb_register_all_nodes();
1803 module_init(hugetlb_init);
1805 /* Should be called on processing a hugepagesz=... option */
1806 void __init hugetlb_add_hstate(unsigned order)
1811 if (size_to_hstate(PAGE_SIZE << order)) {
1812 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1815 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1817 h = &hstates[max_hstate++];
1819 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1820 h->nr_huge_pages = 0;
1821 h->free_huge_pages = 0;
1822 for (i = 0; i < MAX_NUMNODES; ++i)
1823 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1824 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1825 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1826 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1827 huge_page_size(h)/1024);
1832 static int __init hugetlb_nrpages_setup(char *s)
1835 static unsigned long *last_mhp;
1838 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1839 * so this hugepages= parameter goes to the "default hstate".
1842 mhp = &default_hstate_max_huge_pages;
1844 mhp = &parsed_hstate->max_huge_pages;
1846 if (mhp == last_mhp) {
1847 printk(KERN_WARNING "hugepages= specified twice without "
1848 "interleaving hugepagesz=, ignoring\n");
1852 if (sscanf(s, "%lu", mhp) <= 0)
1856 * Global state is always initialized later in hugetlb_init.
1857 * But we need to allocate >= MAX_ORDER hstates here early to still
1858 * use the bootmem allocator.
1860 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1861 hugetlb_hstate_alloc_pages(parsed_hstate);
1867 __setup("hugepages=", hugetlb_nrpages_setup);
1869 static int __init hugetlb_default_setup(char *s)
1871 default_hstate_size = memparse(s, &s);
1874 __setup("default_hugepagesz=", hugetlb_default_setup);
1876 static unsigned int cpuset_mems_nr(unsigned int *array)
1879 unsigned int nr = 0;
1881 for_each_node_mask(node, cpuset_current_mems_allowed)
1887 #ifdef CONFIG_SYSCTL
1888 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1889 struct ctl_table *table, int write,
1890 void __user *buffer, size_t *length, loff_t *ppos)
1892 struct hstate *h = &default_hstate;
1896 tmp = h->max_huge_pages;
1898 if (write && h->order >= MAX_ORDER)
1902 table->maxlen = sizeof(unsigned long);
1903 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1908 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1909 GFP_KERNEL | __GFP_NORETRY);
1910 if (!(obey_mempolicy &&
1911 init_nodemask_of_mempolicy(nodes_allowed))) {
1912 NODEMASK_FREE(nodes_allowed);
1913 nodes_allowed = &node_states[N_HIGH_MEMORY];
1915 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1917 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1918 NODEMASK_FREE(nodes_allowed);
1924 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1925 void __user *buffer, size_t *length, loff_t *ppos)
1928 return hugetlb_sysctl_handler_common(false, table, write,
1929 buffer, length, ppos);
1933 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
1934 void __user *buffer, size_t *length, loff_t *ppos)
1936 return hugetlb_sysctl_handler_common(true, table, write,
1937 buffer, length, ppos);
1939 #endif /* CONFIG_NUMA */
1941 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1942 void __user *buffer,
1943 size_t *length, loff_t *ppos)
1945 proc_dointvec(table, write, buffer, length, ppos);
1946 if (hugepages_treat_as_movable)
1947 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1949 htlb_alloc_mask = GFP_HIGHUSER;
1953 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1954 void __user *buffer,
1955 size_t *length, loff_t *ppos)
1957 struct hstate *h = &default_hstate;
1961 tmp = h->nr_overcommit_huge_pages;
1963 if (write && h->order >= MAX_ORDER)
1967 table->maxlen = sizeof(unsigned long);
1968 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1973 spin_lock(&hugetlb_lock);
1974 h->nr_overcommit_huge_pages = tmp;
1975 spin_unlock(&hugetlb_lock);
1981 #endif /* CONFIG_SYSCTL */
1983 void hugetlb_report_meminfo(struct seq_file *m)
1985 struct hstate *h = &default_hstate;
1987 "HugePages_Total: %5lu\n"
1988 "HugePages_Free: %5lu\n"
1989 "HugePages_Rsvd: %5lu\n"
1990 "HugePages_Surp: %5lu\n"
1991 "Hugepagesize: %8lu kB\n",
1995 h->surplus_huge_pages,
1996 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1999 int hugetlb_report_node_meminfo(int nid, char *buf)
2001 struct hstate *h = &default_hstate;
2003 "Node %d HugePages_Total: %5u\n"
2004 "Node %d HugePages_Free: %5u\n"
2005 "Node %d HugePages_Surp: %5u\n",
2006 nid, h->nr_huge_pages_node[nid],
2007 nid, h->free_huge_pages_node[nid],
2008 nid, h->surplus_huge_pages_node[nid]);
2011 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2012 unsigned long hugetlb_total_pages(void)
2014 struct hstate *h = &default_hstate;
2015 return h->nr_huge_pages * pages_per_huge_page(h);
2018 static int hugetlb_acct_memory(struct hstate *h, long delta)
2022 spin_lock(&hugetlb_lock);
2024 * When cpuset is configured, it breaks the strict hugetlb page
2025 * reservation as the accounting is done on a global variable. Such
2026 * reservation is completely rubbish in the presence of cpuset because
2027 * the reservation is not checked against page availability for the
2028 * current cpuset. Application can still potentially OOM'ed by kernel
2029 * with lack of free htlb page in cpuset that the task is in.
2030 * Attempt to enforce strict accounting with cpuset is almost
2031 * impossible (or too ugly) because cpuset is too fluid that
2032 * task or memory node can be dynamically moved between cpusets.
2034 * The change of semantics for shared hugetlb mapping with cpuset is
2035 * undesirable. However, in order to preserve some of the semantics,
2036 * we fall back to check against current free page availability as
2037 * a best attempt and hopefully to minimize the impact of changing
2038 * semantics that cpuset has.
2041 if (gather_surplus_pages(h, delta) < 0)
2044 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2045 return_unused_surplus_pages(h, delta);
2052 return_unused_surplus_pages(h, (unsigned long) -delta);
2055 spin_unlock(&hugetlb_lock);
2059 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2061 struct resv_map *reservations = vma_resv_map(vma);
2064 * This new VMA should share its siblings reservation map if present.
2065 * The VMA will only ever have a valid reservation map pointer where
2066 * it is being copied for another still existing VMA. As that VMA
2067 * has a reference to the reservation map it cannot disappear until
2068 * after this open call completes. It is therefore safe to take a
2069 * new reference here without additional locking.
2072 kref_get(&reservations->refs);
2075 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2077 struct hstate *h = hstate_vma(vma);
2078 struct resv_map *reservations = vma_resv_map(vma);
2079 unsigned long reserve;
2080 unsigned long start;
2084 start = vma_hugecache_offset(h, vma, vma->vm_start);
2085 end = vma_hugecache_offset(h, vma, vma->vm_end);
2087 reserve = (end - start) -
2088 region_count(&reservations->regions, start, end);
2090 kref_put(&reservations->refs, resv_map_release);
2093 hugetlb_acct_memory(h, -reserve);
2094 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
2100 * We cannot handle pagefaults against hugetlb pages at all. They cause
2101 * handle_mm_fault() to try to instantiate regular-sized pages in the
2102 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2105 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2111 const struct vm_operations_struct hugetlb_vm_ops = {
2112 .fault = hugetlb_vm_op_fault,
2113 .open = hugetlb_vm_op_open,
2114 .close = hugetlb_vm_op_close,
2117 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2124 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2126 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2128 entry = pte_mkyoung(entry);
2129 entry = pte_mkhuge(entry);
2134 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2135 unsigned long address, pte_t *ptep)
2139 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2140 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2141 update_mmu_cache(vma, address, ptep);
2145 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2146 struct vm_area_struct *vma)
2148 pte_t *src_pte, *dst_pte, entry;
2149 struct page *ptepage;
2152 struct hstate *h = hstate_vma(vma);
2153 unsigned long sz = huge_page_size(h);
2155 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2157 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2158 src_pte = huge_pte_offset(src, addr);
2161 dst_pte = huge_pte_alloc(dst, addr, sz);
2165 /* If the pagetables are shared don't copy or take references */
2166 if (dst_pte == src_pte)
2169 spin_lock(&dst->page_table_lock);
2170 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2171 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2173 huge_ptep_set_wrprotect(src, addr, src_pte);
2174 entry = huge_ptep_get(src_pte);
2175 ptepage = pte_page(entry);
2177 page_dup_rmap(ptepage);
2178 set_huge_pte_at(dst, addr, dst_pte, entry);
2180 spin_unlock(&src->page_table_lock);
2181 spin_unlock(&dst->page_table_lock);
2189 static int is_hugetlb_entry_migration(pte_t pte)
2193 if (huge_pte_none(pte) || pte_present(pte))
2195 swp = pte_to_swp_entry(pte);
2196 if (non_swap_entry(swp) && is_migration_entry(swp))
2202 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2206 if (huge_pte_none(pte) || pte_present(pte))
2208 swp = pte_to_swp_entry(pte);
2209 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2215 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2216 unsigned long end, struct page *ref_page)
2218 struct mm_struct *mm = vma->vm_mm;
2219 unsigned long address;
2224 struct hstate *h = hstate_vma(vma);
2225 unsigned long sz = huge_page_size(h);
2228 * A page gathering list, protected by per file i_mmap_mutex. The
2229 * lock is used to avoid list corruption from multiple unmapping
2230 * of the same page since we are using page->lru.
2232 LIST_HEAD(page_list);
2234 WARN_ON(!is_vm_hugetlb_page(vma));
2235 BUG_ON(start & ~huge_page_mask(h));
2236 BUG_ON(end & ~huge_page_mask(h));
2238 mmu_notifier_invalidate_range_start(mm, start, end);
2239 spin_lock(&mm->page_table_lock);
2240 for (address = start; address < end; address += sz) {
2241 ptep = huge_pte_offset(mm, address);
2245 if (huge_pmd_unshare(mm, &address, ptep))
2249 * If a reference page is supplied, it is because a specific
2250 * page is being unmapped, not a range. Ensure the page we
2251 * are about to unmap is the actual page of interest.
2254 pte = huge_ptep_get(ptep);
2255 if (huge_pte_none(pte))
2257 page = pte_page(pte);
2258 if (page != ref_page)
2262 * Mark the VMA as having unmapped its page so that
2263 * future faults in this VMA will fail rather than
2264 * looking like data was lost
2266 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2269 pte = huge_ptep_get_and_clear(mm, address, ptep);
2270 if (huge_pte_none(pte))
2274 * HWPoisoned hugepage is already unmapped and dropped reference
2276 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2279 page = pte_page(pte);
2281 set_page_dirty(page);
2282 list_add(&page->lru, &page_list);
2284 flush_tlb_range(vma, start, end);
2285 spin_unlock(&mm->page_table_lock);
2286 mmu_notifier_invalidate_range_end(mm, start, end);
2287 list_for_each_entry_safe(page, tmp, &page_list, lru) {
2288 page_remove_rmap(page);
2289 list_del(&page->lru);
2294 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2295 unsigned long end, struct page *ref_page)
2297 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2298 __unmap_hugepage_range(vma, start, end, ref_page);
2299 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2303 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2304 * mappping it owns the reserve page for. The intention is to unmap the page
2305 * from other VMAs and let the children be SIGKILLed if they are faulting the
2308 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2309 struct page *page, unsigned long address)
2311 struct hstate *h = hstate_vma(vma);
2312 struct vm_area_struct *iter_vma;
2313 struct address_space *mapping;
2314 struct prio_tree_iter iter;
2318 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2319 * from page cache lookup which is in HPAGE_SIZE units.
2321 address = address & huge_page_mask(h);
2322 pgoff = vma_hugecache_offset(h, vma, address);
2323 mapping = (struct address_space *)page_private(page);
2326 * Take the mapping lock for the duration of the table walk. As
2327 * this mapping should be shared between all the VMAs,
2328 * __unmap_hugepage_range() is called as the lock is already held
2330 mutex_lock(&mapping->i_mmap_mutex);
2331 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2332 /* Do not unmap the current VMA */
2333 if (iter_vma == vma)
2337 * Unmap the page from other VMAs without their own reserves.
2338 * They get marked to be SIGKILLed if they fault in these
2339 * areas. This is because a future no-page fault on this VMA
2340 * could insert a zeroed page instead of the data existing
2341 * from the time of fork. This would look like data corruption
2343 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2344 __unmap_hugepage_range(iter_vma,
2345 address, address + huge_page_size(h),
2348 mutex_unlock(&mapping->i_mmap_mutex);
2354 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2355 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2356 * cannot race with other handlers or page migration.
2357 * Keep the pte_same checks anyway to make transition from the mutex easier.
2359 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2360 unsigned long address, pte_t *ptep, pte_t pte,
2361 struct page *pagecache_page)
2363 struct hstate *h = hstate_vma(vma);
2364 struct page *old_page, *new_page;
2366 int outside_reserve = 0;
2368 old_page = pte_page(pte);
2371 /* If no-one else is actually using this page, avoid the copy
2372 * and just make the page writable */
2373 avoidcopy = (page_mapcount(old_page) == 1);
2375 if (PageAnon(old_page))
2376 page_move_anon_rmap(old_page, vma, address);
2377 set_huge_ptep_writable(vma, address, ptep);
2382 * If the process that created a MAP_PRIVATE mapping is about to
2383 * perform a COW due to a shared page count, attempt to satisfy
2384 * the allocation without using the existing reserves. The pagecache
2385 * page is used to determine if the reserve at this address was
2386 * consumed or not. If reserves were used, a partial faulted mapping
2387 * at the time of fork() could consume its reserves on COW instead
2388 * of the full address range.
2390 if (!(vma->vm_flags & VM_MAYSHARE) &&
2391 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2392 old_page != pagecache_page)
2393 outside_reserve = 1;
2395 page_cache_get(old_page);
2397 /* Drop page_table_lock as buddy allocator may be called */
2398 spin_unlock(&mm->page_table_lock);
2399 new_page = alloc_huge_page(vma, address, outside_reserve);
2401 if (IS_ERR(new_page)) {
2402 page_cache_release(old_page);
2405 * If a process owning a MAP_PRIVATE mapping fails to COW,
2406 * it is due to references held by a child and an insufficient
2407 * huge page pool. To guarantee the original mappers
2408 * reliability, unmap the page from child processes. The child
2409 * may get SIGKILLed if it later faults.
2411 if (outside_reserve) {
2412 BUG_ON(huge_pte_none(pte));
2413 if (unmap_ref_private(mm, vma, old_page, address)) {
2414 BUG_ON(page_count(old_page) != 1);
2415 BUG_ON(huge_pte_none(pte));
2416 spin_lock(&mm->page_table_lock);
2417 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2418 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2419 goto retry_avoidcopy;
2421 * race occurs while re-acquiring page_table_lock, and
2429 /* Caller expects lock to be held */
2430 spin_lock(&mm->page_table_lock);
2431 return -PTR_ERR(new_page);
2435 * When the original hugepage is shared one, it does not have
2436 * anon_vma prepared.
2438 if (unlikely(anon_vma_prepare(vma))) {
2439 page_cache_release(new_page);
2440 page_cache_release(old_page);
2441 /* Caller expects lock to be held */
2442 spin_lock(&mm->page_table_lock);
2443 return VM_FAULT_OOM;
2446 copy_user_huge_page(new_page, old_page, address, vma,
2447 pages_per_huge_page(h));
2448 __SetPageUptodate(new_page);
2451 * Retake the page_table_lock to check for racing updates
2452 * before the page tables are altered
2454 spin_lock(&mm->page_table_lock);
2455 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2456 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2458 mmu_notifier_invalidate_range_start(mm,
2459 address & huge_page_mask(h),
2460 (address & huge_page_mask(h)) + huge_page_size(h));
2461 huge_ptep_clear_flush(vma, address, ptep);
2462 set_huge_pte_at(mm, address, ptep,
2463 make_huge_pte(vma, new_page, 1));
2464 page_remove_rmap(old_page);
2465 hugepage_add_new_anon_rmap(new_page, vma, address);
2466 /* Make the old page be freed below */
2467 new_page = old_page;
2468 mmu_notifier_invalidate_range_end(mm,
2469 address & huge_page_mask(h),
2470 (address & huge_page_mask(h)) + huge_page_size(h));
2472 page_cache_release(new_page);
2473 page_cache_release(old_page);
2477 /* Return the pagecache page at a given address within a VMA */
2478 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2479 struct vm_area_struct *vma, unsigned long address)
2481 struct address_space *mapping;
2484 mapping = vma->vm_file->f_mapping;
2485 idx = vma_hugecache_offset(h, vma, address);
2487 return find_lock_page(mapping, idx);
2491 * Return whether there is a pagecache page to back given address within VMA.
2492 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2494 static bool hugetlbfs_pagecache_present(struct hstate *h,
2495 struct vm_area_struct *vma, unsigned long address)
2497 struct address_space *mapping;
2501 mapping = vma->vm_file->f_mapping;
2502 idx = vma_hugecache_offset(h, vma, address);
2504 page = find_get_page(mapping, idx);
2507 return page != NULL;
2510 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2511 unsigned long address, pte_t *ptep, unsigned int flags)
2513 struct hstate *h = hstate_vma(vma);
2514 int ret = VM_FAULT_SIGBUS;
2519 struct address_space *mapping;
2523 * Currently, we are forced to kill the process in the event the
2524 * original mapper has unmapped pages from the child due to a failed
2525 * COW. Warn that such a situation has occurred as it may not be obvious
2527 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2529 "PID %d killed due to inadequate hugepage pool\n",
2534 mapping = vma->vm_file->f_mapping;
2535 idx = vma_hugecache_offset(h, vma, address);
2538 * Use page lock to guard against racing truncation
2539 * before we get page_table_lock.
2542 page = find_lock_page(mapping, idx);
2544 size = i_size_read(mapping->host) >> huge_page_shift(h);
2547 page = alloc_huge_page(vma, address, 0);
2549 ret = -PTR_ERR(page);
2552 clear_huge_page(page, address, pages_per_huge_page(h));
2553 __SetPageUptodate(page);
2555 if (vma->vm_flags & VM_MAYSHARE) {
2557 struct inode *inode = mapping->host;
2559 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2567 spin_lock(&inode->i_lock);
2568 inode->i_blocks += blocks_per_huge_page(h);
2569 spin_unlock(&inode->i_lock);
2572 if (unlikely(anon_vma_prepare(vma))) {
2574 goto backout_unlocked;
2580 * If memory error occurs between mmap() and fault, some process
2581 * don't have hwpoisoned swap entry for errored virtual address.
2582 * So we need to block hugepage fault by PG_hwpoison bit check.
2584 if (unlikely(PageHWPoison(page))) {
2585 ret = VM_FAULT_HWPOISON |
2586 VM_FAULT_SET_HINDEX(h - hstates);
2587 goto backout_unlocked;
2592 * If we are going to COW a private mapping later, we examine the
2593 * pending reservations for this page now. This will ensure that
2594 * any allocations necessary to record that reservation occur outside
2597 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2598 if (vma_needs_reservation(h, vma, address) < 0) {
2600 goto backout_unlocked;
2603 spin_lock(&mm->page_table_lock);
2604 size = i_size_read(mapping->host) >> huge_page_shift(h);
2609 if (!huge_pte_none(huge_ptep_get(ptep)))
2613 hugepage_add_new_anon_rmap(page, vma, address);
2615 page_dup_rmap(page);
2616 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2617 && (vma->vm_flags & VM_SHARED)));
2618 set_huge_pte_at(mm, address, ptep, new_pte);
2620 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2621 /* Optimization, do the COW without a second fault */
2622 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2625 spin_unlock(&mm->page_table_lock);
2631 spin_unlock(&mm->page_table_lock);
2638 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2639 unsigned long address, unsigned int flags)
2644 struct page *page = NULL;
2645 struct page *pagecache_page = NULL;
2646 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2647 struct hstate *h = hstate_vma(vma);
2649 address &= huge_page_mask(h);
2651 ptep = huge_pte_offset(mm, address);
2653 entry = huge_ptep_get(ptep);
2654 if (unlikely(is_hugetlb_entry_migration(entry))) {
2655 migration_entry_wait(mm, (pmd_t *)ptep, address);
2657 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2658 return VM_FAULT_HWPOISON_LARGE |
2659 VM_FAULT_SET_HINDEX(h - hstates);
2662 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2664 return VM_FAULT_OOM;
2667 * Serialize hugepage allocation and instantiation, so that we don't
2668 * get spurious allocation failures if two CPUs race to instantiate
2669 * the same page in the page cache.
2671 mutex_lock(&hugetlb_instantiation_mutex);
2672 entry = huge_ptep_get(ptep);
2673 if (huge_pte_none(entry)) {
2674 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2681 * If we are going to COW the mapping later, we examine the pending
2682 * reservations for this page now. This will ensure that any
2683 * allocations necessary to record that reservation occur outside the
2684 * spinlock. For private mappings, we also lookup the pagecache
2685 * page now as it is used to determine if a reservation has been
2688 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2689 if (vma_needs_reservation(h, vma, address) < 0) {
2694 if (!(vma->vm_flags & VM_MAYSHARE))
2695 pagecache_page = hugetlbfs_pagecache_page(h,
2700 * hugetlb_cow() requires page locks of pte_page(entry) and
2701 * pagecache_page, so here we need take the former one
2702 * when page != pagecache_page or !pagecache_page.
2703 * Note that locking order is always pagecache_page -> page,
2704 * so no worry about deadlock.
2706 page = pte_page(entry);
2707 if (page != pagecache_page)
2710 spin_lock(&mm->page_table_lock);
2711 /* Check for a racing update before calling hugetlb_cow */
2712 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2713 goto out_page_table_lock;
2716 if (flags & FAULT_FLAG_WRITE) {
2717 if (!pte_write(entry)) {
2718 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2720 goto out_page_table_lock;
2722 entry = pte_mkdirty(entry);
2724 entry = pte_mkyoung(entry);
2725 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2726 flags & FAULT_FLAG_WRITE))
2727 update_mmu_cache(vma, address, ptep);
2729 out_page_table_lock:
2730 spin_unlock(&mm->page_table_lock);
2732 if (pagecache_page) {
2733 unlock_page(pagecache_page);
2734 put_page(pagecache_page);
2736 if (page != pagecache_page)
2740 mutex_unlock(&hugetlb_instantiation_mutex);
2745 /* Can be overriden by architectures */
2746 __attribute__((weak)) struct page *
2747 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2748 pud_t *pud, int write)
2754 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2755 struct page **pages, struct vm_area_struct **vmas,
2756 unsigned long *position, int *length, int i,
2759 unsigned long pfn_offset;
2760 unsigned long vaddr = *position;
2761 int remainder = *length;
2762 struct hstate *h = hstate_vma(vma);
2764 spin_lock(&mm->page_table_lock);
2765 while (vaddr < vma->vm_end && remainder) {
2771 * Some archs (sparc64, sh*) have multiple pte_ts to
2772 * each hugepage. We have to make sure we get the
2773 * first, for the page indexing below to work.
2775 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2776 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2779 * When coredumping, it suits get_dump_page if we just return
2780 * an error where there's an empty slot with no huge pagecache
2781 * to back it. This way, we avoid allocating a hugepage, and
2782 * the sparse dumpfile avoids allocating disk blocks, but its
2783 * huge holes still show up with zeroes where they need to be.
2785 if (absent && (flags & FOLL_DUMP) &&
2786 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2792 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2795 spin_unlock(&mm->page_table_lock);
2796 ret = hugetlb_fault(mm, vma, vaddr,
2797 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2798 spin_lock(&mm->page_table_lock);
2799 if (!(ret & VM_FAULT_ERROR))
2806 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2807 page = pte_page(huge_ptep_get(pte));
2810 pages[i] = mem_map_offset(page, pfn_offset);
2821 if (vaddr < vma->vm_end && remainder &&
2822 pfn_offset < pages_per_huge_page(h)) {
2824 * We use pfn_offset to avoid touching the pageframes
2825 * of this compound page.
2830 spin_unlock(&mm->page_table_lock);
2831 *length = remainder;
2834 return i ? i : -EFAULT;
2837 void hugetlb_change_protection(struct vm_area_struct *vma,
2838 unsigned long address, unsigned long end, pgprot_t newprot)
2840 struct mm_struct *mm = vma->vm_mm;
2841 unsigned long start = address;
2844 struct hstate *h = hstate_vma(vma);
2846 BUG_ON(address >= end);
2847 flush_cache_range(vma, address, end);
2849 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2850 spin_lock(&mm->page_table_lock);
2851 for (; address < end; address += huge_page_size(h)) {
2852 ptep = huge_pte_offset(mm, address);
2855 if (huge_pmd_unshare(mm, &address, ptep))
2857 if (!huge_pte_none(huge_ptep_get(ptep))) {
2858 pte = huge_ptep_get_and_clear(mm, address, ptep);
2859 pte = pte_mkhuge(pte_modify(pte, newprot));
2860 set_huge_pte_at(mm, address, ptep, pte);
2863 spin_unlock(&mm->page_table_lock);
2864 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2866 flush_tlb_range(vma, start, end);
2869 int hugetlb_reserve_pages(struct inode *inode,
2871 struct vm_area_struct *vma,
2872 vm_flags_t vm_flags)
2875 struct hstate *h = hstate_inode(inode);
2878 * Only apply hugepage reservation if asked. At fault time, an
2879 * attempt will be made for VM_NORESERVE to allocate a page
2880 * and filesystem quota without using reserves
2882 if (vm_flags & VM_NORESERVE)
2886 * Shared mappings base their reservation on the number of pages that
2887 * are already allocated on behalf of the file. Private mappings need
2888 * to reserve the full area even if read-only as mprotect() may be
2889 * called to make the mapping read-write. Assume !vma is a shm mapping
2891 if (!vma || vma->vm_flags & VM_MAYSHARE)
2892 chg = region_chg(&inode->i_mapping->private_list, from, to);
2894 struct resv_map *resv_map = resv_map_alloc();
2900 set_vma_resv_map(vma, resv_map);
2901 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2907 /* There must be enough filesystem quota for the mapping */
2908 if (hugetlb_get_quota(inode->i_mapping, chg))
2912 * Check enough hugepages are available for the reservation.
2913 * Hand back the quota if there are not
2915 ret = hugetlb_acct_memory(h, chg);
2917 hugetlb_put_quota(inode->i_mapping, chg);
2922 * Account for the reservations made. Shared mappings record regions
2923 * that have reservations as they are shared by multiple VMAs.
2924 * When the last VMA disappears, the region map says how much
2925 * the reservation was and the page cache tells how much of
2926 * the reservation was consumed. Private mappings are per-VMA and
2927 * only the consumed reservations are tracked. When the VMA
2928 * disappears, the original reservation is the VMA size and the
2929 * consumed reservations are stored in the map. Hence, nothing
2930 * else has to be done for private mappings here
2932 if (!vma || vma->vm_flags & VM_MAYSHARE)
2933 region_add(&inode->i_mapping->private_list, from, to);
2937 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2939 struct hstate *h = hstate_inode(inode);
2940 long chg = region_truncate(&inode->i_mapping->private_list, offset);
2942 spin_lock(&inode->i_lock);
2943 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
2944 spin_unlock(&inode->i_lock);
2946 hugetlb_put_quota(inode->i_mapping, (chg - freed));
2947 hugetlb_acct_memory(h, -(chg - freed));
2950 #ifdef CONFIG_MEMORY_FAILURE
2952 /* Should be called in hugetlb_lock */
2953 static int is_hugepage_on_freelist(struct page *hpage)
2957 struct hstate *h = page_hstate(hpage);
2958 int nid = page_to_nid(hpage);
2960 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
2967 * This function is called from memory failure code.
2968 * Assume the caller holds page lock of the head page.
2970 int dequeue_hwpoisoned_huge_page(struct page *hpage)
2972 struct hstate *h = page_hstate(hpage);
2973 int nid = page_to_nid(hpage);
2976 spin_lock(&hugetlb_lock);
2977 if (is_hugepage_on_freelist(hpage)) {
2978 list_del(&hpage->lru);
2979 set_page_refcounted(hpage);
2980 h->free_huge_pages--;
2981 h->free_huge_pages_node[nid]--;
2984 spin_unlock(&hugetlb_lock);