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;
463 unsigned int cpuset_mems_cookie;
466 cpuset_mems_cookie = get_mems_allowed();
467 zonelist = huge_zonelist(vma, address,
468 htlb_alloc_mask, &mpol, &nodemask);
470 * A child process with MAP_PRIVATE mappings created by their parent
471 * have no page reserves. This check ensures that reservations are
472 * not "stolen". The child may still get SIGKILLed
474 if (!vma_has_reserves(vma) &&
475 h->free_huge_pages - h->resv_huge_pages == 0)
478 /* If reserves cannot be used, ensure enough pages are in the pool */
479 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
482 for_each_zone_zonelist_nodemask(zone, z, zonelist,
483 MAX_NR_ZONES - 1, nodemask) {
484 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
485 page = dequeue_huge_page_node(h, zone_to_nid(zone));
488 decrement_hugepage_resv_vma(h, vma);
495 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
504 static void update_and_free_page(struct hstate *h, struct page *page)
508 VM_BUG_ON(h->order >= MAX_ORDER);
511 h->nr_huge_pages_node[page_to_nid(page)]--;
512 for (i = 0; i < pages_per_huge_page(h); i++) {
513 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
514 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
515 1 << PG_private | 1<< PG_writeback);
517 set_compound_page_dtor(page, NULL);
518 set_page_refcounted(page);
519 arch_release_hugepage(page);
520 __free_pages(page, huge_page_order(h));
523 struct hstate *size_to_hstate(unsigned long size)
528 if (huge_page_size(h) == size)
534 static void free_huge_page(struct page *page)
537 * Can't pass hstate in here because it is called from the
538 * compound page destructor.
540 struct hstate *h = page_hstate(page);
541 int nid = page_to_nid(page);
542 struct address_space *mapping;
544 mapping = (struct address_space *) page_private(page);
545 set_page_private(page, 0);
546 page->mapping = NULL;
547 BUG_ON(page_count(page));
548 BUG_ON(page_mapcount(page));
549 INIT_LIST_HEAD(&page->lru);
551 spin_lock(&hugetlb_lock);
552 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
553 update_and_free_page(h, page);
554 h->surplus_huge_pages--;
555 h->surplus_huge_pages_node[nid]--;
557 enqueue_huge_page(h, page);
559 spin_unlock(&hugetlb_lock);
561 hugetlb_put_quota(mapping, 1);
564 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
566 set_compound_page_dtor(page, free_huge_page);
567 spin_lock(&hugetlb_lock);
569 h->nr_huge_pages_node[nid]++;
570 spin_unlock(&hugetlb_lock);
571 put_page(page); /* free it into the hugepage allocator */
574 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
577 int nr_pages = 1 << order;
578 struct page *p = page + 1;
580 /* we rely on prep_new_huge_page to set the destructor */
581 set_compound_order(page, order);
583 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
585 set_page_count(p, 0);
586 p->first_page = page;
590 int PageHuge(struct page *page)
592 compound_page_dtor *dtor;
594 if (!PageCompound(page))
597 page = compound_head(page);
598 dtor = get_compound_page_dtor(page);
600 return dtor == free_huge_page;
603 EXPORT_SYMBOL_GPL(PageHuge);
605 pgoff_t __basepage_index(struct page *page)
607 struct page *page_head = compound_head(page);
608 pgoff_t index = page_index(page_head);
609 unsigned long compound_idx;
611 if (!PageHuge(page_head))
612 return page_index(page);
614 if (compound_order(page_head) >= MAX_ORDER)
615 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
617 compound_idx = page - page_head;
619 return (index << compound_order(page_head)) + compound_idx;
622 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
626 if (h->order >= MAX_ORDER)
629 page = alloc_pages_exact_node(nid,
630 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
631 __GFP_REPEAT|__GFP_NOWARN,
634 if (arch_prepare_hugepage(page)) {
635 __free_pages(page, huge_page_order(h));
638 prep_new_huge_page(h, page, nid);
645 * common helper functions for hstate_next_node_to_{alloc|free}.
646 * We may have allocated or freed a huge page based on a different
647 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
648 * be outside of *nodes_allowed. Ensure that we use an allowed
649 * node for alloc or free.
651 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
653 nid = next_node(nid, *nodes_allowed);
654 if (nid == MAX_NUMNODES)
655 nid = first_node(*nodes_allowed);
656 VM_BUG_ON(nid >= MAX_NUMNODES);
661 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
663 if (!node_isset(nid, *nodes_allowed))
664 nid = next_node_allowed(nid, nodes_allowed);
669 * returns the previously saved node ["this node"] from which to
670 * allocate a persistent huge page for the pool and advance the
671 * next node from which to allocate, handling wrap at end of node
674 static int hstate_next_node_to_alloc(struct hstate *h,
675 nodemask_t *nodes_allowed)
679 VM_BUG_ON(!nodes_allowed);
681 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
682 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
687 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
694 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
695 next_nid = start_nid;
698 page = alloc_fresh_huge_page_node(h, next_nid);
703 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
704 } while (next_nid != start_nid);
707 count_vm_event(HTLB_BUDDY_PGALLOC);
709 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
715 * helper for free_pool_huge_page() - return the previously saved
716 * node ["this node"] from which to free a huge page. Advance the
717 * next node id whether or not we find a free huge page to free so
718 * that the next attempt to free addresses the next node.
720 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
724 VM_BUG_ON(!nodes_allowed);
726 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
727 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
733 * Free huge page from pool from next node to free.
734 * Attempt to keep persistent huge pages more or less
735 * balanced over allowed nodes.
736 * Called with hugetlb_lock locked.
738 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
745 start_nid = hstate_next_node_to_free(h, nodes_allowed);
746 next_nid = start_nid;
750 * If we're returning unused surplus pages, only examine
751 * nodes with surplus pages.
753 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
754 !list_empty(&h->hugepage_freelists[next_nid])) {
756 list_entry(h->hugepage_freelists[next_nid].next,
758 list_del(&page->lru);
759 h->free_huge_pages--;
760 h->free_huge_pages_node[next_nid]--;
762 h->surplus_huge_pages--;
763 h->surplus_huge_pages_node[next_nid]--;
765 update_and_free_page(h, page);
769 next_nid = hstate_next_node_to_free(h, nodes_allowed);
770 } while (next_nid != start_nid);
775 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
780 if (h->order >= MAX_ORDER)
784 * Assume we will successfully allocate the surplus page to
785 * prevent racing processes from causing the surplus to exceed
788 * This however introduces a different race, where a process B
789 * tries to grow the static hugepage pool while alloc_pages() is
790 * called by process A. B will only examine the per-node
791 * counters in determining if surplus huge pages can be
792 * converted to normal huge pages in adjust_pool_surplus(). A
793 * won't be able to increment the per-node counter, until the
794 * lock is dropped by B, but B doesn't drop hugetlb_lock until
795 * no more huge pages can be converted from surplus to normal
796 * state (and doesn't try to convert again). Thus, we have a
797 * case where a surplus huge page exists, the pool is grown, and
798 * the surplus huge page still exists after, even though it
799 * should just have been converted to a normal huge page. This
800 * does not leak memory, though, as the hugepage will be freed
801 * once it is out of use. It also does not allow the counters to
802 * go out of whack in adjust_pool_surplus() as we don't modify
803 * the node values until we've gotten the hugepage and only the
804 * per-node value is checked there.
806 spin_lock(&hugetlb_lock);
807 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
808 spin_unlock(&hugetlb_lock);
812 h->surplus_huge_pages++;
814 spin_unlock(&hugetlb_lock);
816 if (nid == NUMA_NO_NODE)
817 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
818 __GFP_REPEAT|__GFP_NOWARN,
821 page = alloc_pages_exact_node(nid,
822 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
823 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
825 if (page && arch_prepare_hugepage(page)) {
826 __free_pages(page, huge_page_order(h));
830 spin_lock(&hugetlb_lock);
832 r_nid = page_to_nid(page);
833 set_compound_page_dtor(page, free_huge_page);
835 * We incremented the global counters already
837 h->nr_huge_pages_node[r_nid]++;
838 h->surplus_huge_pages_node[r_nid]++;
839 __count_vm_event(HTLB_BUDDY_PGALLOC);
842 h->surplus_huge_pages--;
843 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
845 spin_unlock(&hugetlb_lock);
851 * This allocation function is useful in the context where vma is irrelevant.
852 * E.g. soft-offlining uses this function because it only cares physical
853 * address of error page.
855 struct page *alloc_huge_page_node(struct hstate *h, int nid)
859 spin_lock(&hugetlb_lock);
860 page = dequeue_huge_page_node(h, nid);
861 spin_unlock(&hugetlb_lock);
864 page = alloc_buddy_huge_page(h, nid);
870 * Increase the hugetlb pool such that it can accommodate a reservation
873 static int gather_surplus_pages(struct hstate *h, int delta)
875 struct list_head surplus_list;
876 struct page *page, *tmp;
878 int needed, allocated;
880 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
882 h->resv_huge_pages += delta;
887 INIT_LIST_HEAD(&surplus_list);
891 spin_unlock(&hugetlb_lock);
892 for (i = 0; i < needed; i++) {
893 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
896 * We were not able to allocate enough pages to
897 * satisfy the entire reservation so we free what
898 * we've allocated so far.
902 list_add(&page->lru, &surplus_list);
907 * After retaking hugetlb_lock, we need to recalculate 'needed'
908 * because either resv_huge_pages or free_huge_pages may have changed.
910 spin_lock(&hugetlb_lock);
911 needed = (h->resv_huge_pages + delta) -
912 (h->free_huge_pages + allocated);
917 * The surplus_list now contains _at_least_ the number of extra pages
918 * needed to accommodate the reservation. Add the appropriate number
919 * of pages to the hugetlb pool and free the extras back to the buddy
920 * allocator. Commit the entire reservation here to prevent another
921 * process from stealing the pages as they are added to the pool but
922 * before they are reserved.
925 h->resv_huge_pages += delta;
928 /* Free the needed pages to the hugetlb pool */
929 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
932 list_del(&page->lru);
934 * This page is now managed by the hugetlb allocator and has
935 * no users -- drop the buddy allocator's reference.
937 put_page_testzero(page);
938 VM_BUG_ON(page_count(page));
939 enqueue_huge_page(h, page);
941 spin_unlock(&hugetlb_lock);
943 /* Free unnecessary surplus pages to the buddy allocator */
945 if (!list_empty(&surplus_list)) {
946 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
947 list_del(&page->lru);
951 spin_lock(&hugetlb_lock);
957 * When releasing a hugetlb pool reservation, any surplus pages that were
958 * allocated to satisfy the reservation must be explicitly freed if they were
960 * Called with hugetlb_lock held.
962 static void return_unused_surplus_pages(struct hstate *h,
963 unsigned long unused_resv_pages)
965 unsigned long nr_pages;
967 /* Uncommit the reservation */
968 h->resv_huge_pages -= unused_resv_pages;
970 /* Cannot return gigantic pages currently */
971 if (h->order >= MAX_ORDER)
974 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
977 * We want to release as many surplus pages as possible, spread
978 * evenly across all nodes with memory. Iterate across these nodes
979 * until we can no longer free unreserved surplus pages. This occurs
980 * when the nodes with surplus pages have no free pages.
981 * free_pool_huge_page() will balance the the freed pages across the
982 * on-line nodes with memory and will handle the hstate accounting.
985 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
991 * Determine if the huge page at addr within the vma has an associated
992 * reservation. Where it does not we will need to logically increase
993 * reservation and actually increase quota before an allocation can occur.
994 * Where any new reservation would be required the reservation change is
995 * prepared, but not committed. Once the page has been quota'd allocated
996 * an instantiated the change should be committed via vma_commit_reservation.
997 * No action is required on failure.
999 static long vma_needs_reservation(struct hstate *h,
1000 struct vm_area_struct *vma, unsigned long addr)
1002 struct address_space *mapping = vma->vm_file->f_mapping;
1003 struct inode *inode = mapping->host;
1005 if (vma->vm_flags & VM_MAYSHARE) {
1006 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1007 return region_chg(&inode->i_mapping->private_list,
1010 } 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 err = region_chg(&reservations->regions, idx, idx + 1);
1024 static void vma_commit_reservation(struct hstate *h,
1025 struct vm_area_struct *vma, unsigned long addr)
1027 struct address_space *mapping = vma->vm_file->f_mapping;
1028 struct inode *inode = mapping->host;
1030 if (vma->vm_flags & VM_MAYSHARE) {
1031 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1032 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1034 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1035 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1036 struct resv_map *reservations = vma_resv_map(vma);
1038 /* Mark this page used in the map. */
1039 region_add(&reservations->regions, idx, idx + 1);
1043 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1044 unsigned long addr, int avoid_reserve)
1046 struct hstate *h = hstate_vma(vma);
1048 struct address_space *mapping = vma->vm_file->f_mapping;
1049 struct inode *inode = mapping->host;
1053 * Processes that did not create the mapping will have no reserves and
1054 * will not have accounted against quota. Check that the quota can be
1055 * made before satisfying the allocation
1056 * MAP_NORESERVE mappings may also need pages and quota allocated
1057 * if no reserve mapping overlaps.
1059 chg = vma_needs_reservation(h, vma, addr);
1061 return ERR_PTR(-VM_FAULT_OOM);
1063 if (hugetlb_get_quota(inode->i_mapping, chg))
1064 return ERR_PTR(-VM_FAULT_SIGBUS);
1066 spin_lock(&hugetlb_lock);
1067 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1068 spin_unlock(&hugetlb_lock);
1071 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1073 hugetlb_put_quota(inode->i_mapping, chg);
1074 return ERR_PTR(-VM_FAULT_SIGBUS);
1078 set_page_private(page, (unsigned long) mapping);
1080 vma_commit_reservation(h, vma, addr);
1085 int __weak alloc_bootmem_huge_page(struct hstate *h)
1087 struct huge_bootmem_page *m;
1088 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1093 addr = __alloc_bootmem_node_nopanic(
1094 NODE_DATA(hstate_next_node_to_alloc(h,
1095 &node_states[N_HIGH_MEMORY])),
1096 huge_page_size(h), huge_page_size(h), 0);
1100 * Use the beginning of the huge page to store the
1101 * huge_bootmem_page struct (until gather_bootmem
1102 * puts them into the mem_map).
1112 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1113 /* Put them into a private list first because mem_map is not up yet */
1114 list_add(&m->list, &huge_boot_pages);
1119 static void prep_compound_huge_page(struct page *page, int order)
1121 if (unlikely(order > (MAX_ORDER - 1)))
1122 prep_compound_gigantic_page(page, order);
1124 prep_compound_page(page, order);
1127 /* Put bootmem huge pages into the standard lists after mem_map is up */
1128 static void __init gather_bootmem_prealloc(void)
1130 struct huge_bootmem_page *m;
1132 list_for_each_entry(m, &huge_boot_pages, list) {
1133 struct page *page = virt_to_page(m);
1134 struct hstate *h = m->hstate;
1135 __ClearPageReserved(page);
1136 WARN_ON(page_count(page) != 1);
1137 prep_compound_huge_page(page, h->order);
1138 prep_new_huge_page(h, page, page_to_nid(page));
1140 * If we had gigantic hugepages allocated at boot time, we need
1141 * to restore the 'stolen' pages to totalram_pages in order to
1142 * fix confusing memory reports from free(1) and another
1143 * side-effects, like CommitLimit going negative.
1145 if (h->order > (MAX_ORDER - 1))
1146 totalram_pages += 1 << h->order;
1150 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1154 for (i = 0; i < h->max_huge_pages; ++i) {
1155 if (h->order >= MAX_ORDER) {
1156 if (!alloc_bootmem_huge_page(h))
1158 } else if (!alloc_fresh_huge_page(h,
1159 &node_states[N_HIGH_MEMORY]))
1162 h->max_huge_pages = i;
1165 static void __init hugetlb_init_hstates(void)
1169 for_each_hstate(h) {
1170 /* oversize hugepages were init'ed in early boot */
1171 if (h->order < MAX_ORDER)
1172 hugetlb_hstate_alloc_pages(h);
1176 static char * __init memfmt(char *buf, unsigned long n)
1178 if (n >= (1UL << 30))
1179 sprintf(buf, "%lu GB", n >> 30);
1180 else if (n >= (1UL << 20))
1181 sprintf(buf, "%lu MB", n >> 20);
1183 sprintf(buf, "%lu KB", n >> 10);
1187 static void __init report_hugepages(void)
1191 for_each_hstate(h) {
1193 printk(KERN_INFO "HugeTLB registered %s page size, "
1194 "pre-allocated %ld pages\n",
1195 memfmt(buf, huge_page_size(h)),
1196 h->free_huge_pages);
1200 #ifdef CONFIG_HIGHMEM
1201 static void try_to_free_low(struct hstate *h, unsigned long count,
1202 nodemask_t *nodes_allowed)
1206 if (h->order >= MAX_ORDER)
1209 for_each_node_mask(i, *nodes_allowed) {
1210 struct page *page, *next;
1211 struct list_head *freel = &h->hugepage_freelists[i];
1212 list_for_each_entry_safe(page, next, freel, lru) {
1213 if (count >= h->nr_huge_pages)
1215 if (PageHighMem(page))
1217 list_del(&page->lru);
1218 update_and_free_page(h, page);
1219 h->free_huge_pages--;
1220 h->free_huge_pages_node[page_to_nid(page)]--;
1225 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1226 nodemask_t *nodes_allowed)
1232 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1233 * balanced by operating on them in a round-robin fashion.
1234 * Returns 1 if an adjustment was made.
1236 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1239 int start_nid, next_nid;
1242 VM_BUG_ON(delta != -1 && delta != 1);
1245 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1247 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1248 next_nid = start_nid;
1254 * To shrink on this node, there must be a surplus page
1256 if (!h->surplus_huge_pages_node[nid]) {
1257 next_nid = hstate_next_node_to_alloc(h,
1264 * Surplus cannot exceed the total number of pages
1266 if (h->surplus_huge_pages_node[nid] >=
1267 h->nr_huge_pages_node[nid]) {
1268 next_nid = hstate_next_node_to_free(h,
1274 h->surplus_huge_pages += delta;
1275 h->surplus_huge_pages_node[nid] += delta;
1278 } while (next_nid != start_nid);
1283 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1284 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1285 nodemask_t *nodes_allowed)
1287 unsigned long min_count, ret;
1289 if (h->order >= MAX_ORDER)
1290 return h->max_huge_pages;
1293 * Increase the pool size
1294 * First take pages out of surplus state. Then make up the
1295 * remaining difference by allocating fresh huge pages.
1297 * We might race with alloc_buddy_huge_page() here and be unable
1298 * to convert a surplus huge page to a normal huge page. That is
1299 * not critical, though, it just means the overall size of the
1300 * pool might be one hugepage larger than it needs to be, but
1301 * within all the constraints specified by the sysctls.
1303 spin_lock(&hugetlb_lock);
1304 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1305 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1309 while (count > persistent_huge_pages(h)) {
1311 * If this allocation races such that we no longer need the
1312 * page, free_huge_page will handle it by freeing the page
1313 * and reducing the surplus.
1315 spin_unlock(&hugetlb_lock);
1316 ret = alloc_fresh_huge_page(h, nodes_allowed);
1317 spin_lock(&hugetlb_lock);
1321 /* Bail for signals. Probably ctrl-c from user */
1322 if (signal_pending(current))
1327 * Decrease the pool size
1328 * First return free pages to the buddy allocator (being careful
1329 * to keep enough around to satisfy reservations). Then place
1330 * pages into surplus state as needed so the pool will shrink
1331 * to the desired size as pages become free.
1333 * By placing pages into the surplus state independent of the
1334 * overcommit value, we are allowing the surplus pool size to
1335 * exceed overcommit. There are few sane options here. Since
1336 * alloc_buddy_huge_page() is checking the global counter,
1337 * though, we'll note that we're not allowed to exceed surplus
1338 * and won't grow the pool anywhere else. Not until one of the
1339 * sysctls are changed, or the surplus pages go out of use.
1341 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1342 min_count = max(count, min_count);
1343 try_to_free_low(h, min_count, nodes_allowed);
1344 while (min_count < persistent_huge_pages(h)) {
1345 if (!free_pool_huge_page(h, nodes_allowed, 0))
1348 while (count < persistent_huge_pages(h)) {
1349 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1353 ret = persistent_huge_pages(h);
1354 spin_unlock(&hugetlb_lock);
1358 #define HSTATE_ATTR_RO(_name) \
1359 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1361 #define HSTATE_ATTR(_name) \
1362 static struct kobj_attribute _name##_attr = \
1363 __ATTR(_name, 0644, _name##_show, _name##_store)
1365 static struct kobject *hugepages_kobj;
1366 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1368 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1370 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1374 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1375 if (hstate_kobjs[i] == kobj) {
1377 *nidp = NUMA_NO_NODE;
1381 return kobj_to_node_hstate(kobj, nidp);
1384 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1385 struct kobj_attribute *attr, char *buf)
1388 unsigned long nr_huge_pages;
1391 h = kobj_to_hstate(kobj, &nid);
1392 if (nid == NUMA_NO_NODE)
1393 nr_huge_pages = h->nr_huge_pages;
1395 nr_huge_pages = h->nr_huge_pages_node[nid];
1397 return sprintf(buf, "%lu\n", nr_huge_pages);
1400 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1401 struct kobject *kobj, struct kobj_attribute *attr,
1402 const char *buf, size_t len)
1406 unsigned long count;
1408 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1410 err = strict_strtoul(buf, 10, &count);
1414 h = kobj_to_hstate(kobj, &nid);
1415 if (h->order >= MAX_ORDER) {
1420 if (nid == NUMA_NO_NODE) {
1422 * global hstate attribute
1424 if (!(obey_mempolicy &&
1425 init_nodemask_of_mempolicy(nodes_allowed))) {
1426 NODEMASK_FREE(nodes_allowed);
1427 nodes_allowed = &node_states[N_HIGH_MEMORY];
1429 } else if (nodes_allowed) {
1431 * per node hstate attribute: adjust count to global,
1432 * but restrict alloc/free to the specified node.
1434 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1435 init_nodemask_of_node(nodes_allowed, nid);
1437 nodes_allowed = &node_states[N_HIGH_MEMORY];
1439 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1441 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1442 NODEMASK_FREE(nodes_allowed);
1446 NODEMASK_FREE(nodes_allowed);
1450 static ssize_t nr_hugepages_show(struct kobject *kobj,
1451 struct kobj_attribute *attr, char *buf)
1453 return nr_hugepages_show_common(kobj, attr, buf);
1456 static ssize_t nr_hugepages_store(struct kobject *kobj,
1457 struct kobj_attribute *attr, const char *buf, size_t len)
1459 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1461 HSTATE_ATTR(nr_hugepages);
1466 * hstate attribute for optionally mempolicy-based constraint on persistent
1467 * huge page alloc/free.
1469 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1470 struct kobj_attribute *attr, char *buf)
1472 return nr_hugepages_show_common(kobj, attr, buf);
1475 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1476 struct kobj_attribute *attr, const char *buf, size_t len)
1478 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1480 HSTATE_ATTR(nr_hugepages_mempolicy);
1484 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1485 struct kobj_attribute *attr, char *buf)
1487 struct hstate *h = kobj_to_hstate(kobj, NULL);
1488 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1491 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1492 struct kobj_attribute *attr, const char *buf, size_t count)
1495 unsigned long input;
1496 struct hstate *h = kobj_to_hstate(kobj, NULL);
1498 if (h->order >= MAX_ORDER)
1501 err = strict_strtoul(buf, 10, &input);
1505 spin_lock(&hugetlb_lock);
1506 h->nr_overcommit_huge_pages = input;
1507 spin_unlock(&hugetlb_lock);
1511 HSTATE_ATTR(nr_overcommit_hugepages);
1513 static ssize_t free_hugepages_show(struct kobject *kobj,
1514 struct kobj_attribute *attr, char *buf)
1517 unsigned long free_huge_pages;
1520 h = kobj_to_hstate(kobj, &nid);
1521 if (nid == NUMA_NO_NODE)
1522 free_huge_pages = h->free_huge_pages;
1524 free_huge_pages = h->free_huge_pages_node[nid];
1526 return sprintf(buf, "%lu\n", free_huge_pages);
1528 HSTATE_ATTR_RO(free_hugepages);
1530 static ssize_t resv_hugepages_show(struct kobject *kobj,
1531 struct kobj_attribute *attr, char *buf)
1533 struct hstate *h = kobj_to_hstate(kobj, NULL);
1534 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1536 HSTATE_ATTR_RO(resv_hugepages);
1538 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1539 struct kobj_attribute *attr, char *buf)
1542 unsigned long surplus_huge_pages;
1545 h = kobj_to_hstate(kobj, &nid);
1546 if (nid == NUMA_NO_NODE)
1547 surplus_huge_pages = h->surplus_huge_pages;
1549 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1551 return sprintf(buf, "%lu\n", surplus_huge_pages);
1553 HSTATE_ATTR_RO(surplus_hugepages);
1555 static struct attribute *hstate_attrs[] = {
1556 &nr_hugepages_attr.attr,
1557 &nr_overcommit_hugepages_attr.attr,
1558 &free_hugepages_attr.attr,
1559 &resv_hugepages_attr.attr,
1560 &surplus_hugepages_attr.attr,
1562 &nr_hugepages_mempolicy_attr.attr,
1567 static struct attribute_group hstate_attr_group = {
1568 .attrs = hstate_attrs,
1571 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1572 struct kobject **hstate_kobjs,
1573 struct attribute_group *hstate_attr_group)
1576 int hi = h - hstates;
1578 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1579 if (!hstate_kobjs[hi])
1582 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1584 kobject_put(hstate_kobjs[hi]);
1589 static void __init hugetlb_sysfs_init(void)
1594 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1595 if (!hugepages_kobj)
1598 for_each_hstate(h) {
1599 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1600 hstate_kobjs, &hstate_attr_group);
1602 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1610 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1611 * with node sysdevs in node_devices[] using a parallel array. The array
1612 * index of a node sysdev or _hstate == node id.
1613 * This is here to avoid any static dependency of the node sysdev driver, in
1614 * the base kernel, on the hugetlb module.
1616 struct node_hstate {
1617 struct kobject *hugepages_kobj;
1618 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1620 struct node_hstate node_hstates[MAX_NUMNODES];
1623 * A subset of global hstate attributes for node sysdevs
1625 static struct attribute *per_node_hstate_attrs[] = {
1626 &nr_hugepages_attr.attr,
1627 &free_hugepages_attr.attr,
1628 &surplus_hugepages_attr.attr,
1632 static struct attribute_group per_node_hstate_attr_group = {
1633 .attrs = per_node_hstate_attrs,
1637 * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj.
1638 * Returns node id via non-NULL nidp.
1640 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1644 for (nid = 0; nid < nr_node_ids; nid++) {
1645 struct node_hstate *nhs = &node_hstates[nid];
1647 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1648 if (nhs->hstate_kobjs[i] == kobj) {
1660 * Unregister hstate attributes from a single node sysdev.
1661 * No-op if no hstate attributes attached.
1663 void hugetlb_unregister_node(struct node *node)
1666 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1668 if (!nhs->hugepages_kobj)
1669 return; /* no hstate attributes */
1672 if (nhs->hstate_kobjs[h - hstates]) {
1673 kobject_put(nhs->hstate_kobjs[h - hstates]);
1674 nhs->hstate_kobjs[h - hstates] = NULL;
1677 kobject_put(nhs->hugepages_kobj);
1678 nhs->hugepages_kobj = NULL;
1682 * hugetlb module exit: unregister hstate attributes from node sysdevs
1685 static void hugetlb_unregister_all_nodes(void)
1690 * disable node sysdev registrations.
1692 register_hugetlbfs_with_node(NULL, NULL);
1695 * remove hstate attributes from any nodes that have them.
1697 for (nid = 0; nid < nr_node_ids; nid++)
1698 hugetlb_unregister_node(&node_devices[nid]);
1702 * Register hstate attributes for a single node sysdev.
1703 * No-op if attributes already registered.
1705 void hugetlb_register_node(struct node *node)
1708 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1711 if (nhs->hugepages_kobj)
1712 return; /* already allocated */
1714 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1715 &node->sysdev.kobj);
1716 if (!nhs->hugepages_kobj)
1719 for_each_hstate(h) {
1720 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1722 &per_node_hstate_attr_group);
1724 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1726 h->name, node->sysdev.id);
1727 hugetlb_unregister_node(node);
1734 * hugetlb init time: register hstate attributes for all registered node
1735 * sysdevs of nodes that have memory. All on-line nodes should have
1736 * registered their associated sysdev by this time.
1738 static void hugetlb_register_all_nodes(void)
1742 for_each_node_state(nid, N_HIGH_MEMORY) {
1743 struct node *node = &node_devices[nid];
1744 if (node->sysdev.id == nid)
1745 hugetlb_register_node(node);
1749 * Let the node sysdev driver know we're here so it can
1750 * [un]register hstate attributes on node hotplug.
1752 register_hugetlbfs_with_node(hugetlb_register_node,
1753 hugetlb_unregister_node);
1755 #else /* !CONFIG_NUMA */
1757 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1765 static void hugetlb_unregister_all_nodes(void) { }
1767 static void hugetlb_register_all_nodes(void) { }
1771 static void __exit hugetlb_exit(void)
1775 hugetlb_unregister_all_nodes();
1777 for_each_hstate(h) {
1778 kobject_put(hstate_kobjs[h - hstates]);
1781 kobject_put(hugepages_kobj);
1783 module_exit(hugetlb_exit);
1785 static int __init hugetlb_init(void)
1787 /* Some platform decide whether they support huge pages at boot
1788 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1789 * there is no such support
1791 if (HPAGE_SHIFT == 0)
1794 if (!size_to_hstate(default_hstate_size)) {
1795 default_hstate_size = HPAGE_SIZE;
1796 if (!size_to_hstate(default_hstate_size))
1797 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1799 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1800 if (default_hstate_max_huge_pages)
1801 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1803 hugetlb_init_hstates();
1805 gather_bootmem_prealloc();
1809 hugetlb_sysfs_init();
1811 hugetlb_register_all_nodes();
1815 module_init(hugetlb_init);
1817 /* Should be called on processing a hugepagesz=... option */
1818 void __init hugetlb_add_hstate(unsigned order)
1823 if (size_to_hstate(PAGE_SIZE << order)) {
1824 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1827 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1829 h = &hstates[max_hstate++];
1831 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1832 h->nr_huge_pages = 0;
1833 h->free_huge_pages = 0;
1834 for (i = 0; i < MAX_NUMNODES; ++i)
1835 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1836 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1837 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1838 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1839 huge_page_size(h)/1024);
1844 static int __init hugetlb_nrpages_setup(char *s)
1847 static unsigned long *last_mhp;
1850 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1851 * so this hugepages= parameter goes to the "default hstate".
1854 mhp = &default_hstate_max_huge_pages;
1856 mhp = &parsed_hstate->max_huge_pages;
1858 if (mhp == last_mhp) {
1859 printk(KERN_WARNING "hugepages= specified twice without "
1860 "interleaving hugepagesz=, ignoring\n");
1864 if (sscanf(s, "%lu", mhp) <= 0)
1868 * Global state is always initialized later in hugetlb_init.
1869 * But we need to allocate >= MAX_ORDER hstates here early to still
1870 * use the bootmem allocator.
1872 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1873 hugetlb_hstate_alloc_pages(parsed_hstate);
1879 __setup("hugepages=", hugetlb_nrpages_setup);
1881 static int __init hugetlb_default_setup(char *s)
1883 default_hstate_size = memparse(s, &s);
1886 __setup("default_hugepagesz=", hugetlb_default_setup);
1888 static unsigned int cpuset_mems_nr(unsigned int *array)
1891 unsigned int nr = 0;
1893 for_each_node_mask(node, cpuset_current_mems_allowed)
1899 #ifdef CONFIG_SYSCTL
1900 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1901 struct ctl_table *table, int write,
1902 void __user *buffer, size_t *length, loff_t *ppos)
1904 struct hstate *h = &default_hstate;
1908 tmp = h->max_huge_pages;
1910 if (write && h->order >= MAX_ORDER)
1914 table->maxlen = sizeof(unsigned long);
1915 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1920 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1921 GFP_KERNEL | __GFP_NORETRY);
1922 if (!(obey_mempolicy &&
1923 init_nodemask_of_mempolicy(nodes_allowed))) {
1924 NODEMASK_FREE(nodes_allowed);
1925 nodes_allowed = &node_states[N_HIGH_MEMORY];
1927 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1929 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1930 NODEMASK_FREE(nodes_allowed);
1936 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1937 void __user *buffer, size_t *length, loff_t *ppos)
1940 return hugetlb_sysctl_handler_common(false, table, write,
1941 buffer, length, ppos);
1945 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
1946 void __user *buffer, size_t *length, loff_t *ppos)
1948 return hugetlb_sysctl_handler_common(true, table, write,
1949 buffer, length, ppos);
1951 #endif /* CONFIG_NUMA */
1953 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1954 void __user *buffer,
1955 size_t *length, loff_t *ppos)
1957 proc_dointvec(table, write, buffer, length, ppos);
1958 if (hugepages_treat_as_movable)
1959 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1961 htlb_alloc_mask = GFP_HIGHUSER;
1965 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1966 void __user *buffer,
1967 size_t *length, loff_t *ppos)
1969 struct hstate *h = &default_hstate;
1973 tmp = h->nr_overcommit_huge_pages;
1975 if (write && h->order >= MAX_ORDER)
1979 table->maxlen = sizeof(unsigned long);
1980 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1985 spin_lock(&hugetlb_lock);
1986 h->nr_overcommit_huge_pages = tmp;
1987 spin_unlock(&hugetlb_lock);
1993 #endif /* CONFIG_SYSCTL */
1995 void hugetlb_report_meminfo(struct seq_file *m)
1997 struct hstate *h = &default_hstate;
1999 "HugePages_Total: %5lu\n"
2000 "HugePages_Free: %5lu\n"
2001 "HugePages_Rsvd: %5lu\n"
2002 "HugePages_Surp: %5lu\n"
2003 "Hugepagesize: %8lu kB\n",
2007 h->surplus_huge_pages,
2008 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2011 int hugetlb_report_node_meminfo(int nid, char *buf)
2013 struct hstate *h = &default_hstate;
2015 "Node %d HugePages_Total: %5u\n"
2016 "Node %d HugePages_Free: %5u\n"
2017 "Node %d HugePages_Surp: %5u\n",
2018 nid, h->nr_huge_pages_node[nid],
2019 nid, h->free_huge_pages_node[nid],
2020 nid, h->surplus_huge_pages_node[nid]);
2023 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2024 unsigned long hugetlb_total_pages(void)
2027 unsigned long nr_total_pages = 0;
2030 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2031 return nr_total_pages;
2034 static int hugetlb_acct_memory(struct hstate *h, long delta)
2038 spin_lock(&hugetlb_lock);
2040 * When cpuset is configured, it breaks the strict hugetlb page
2041 * reservation as the accounting is done on a global variable. Such
2042 * reservation is completely rubbish in the presence of cpuset because
2043 * the reservation is not checked against page availability for the
2044 * current cpuset. Application can still potentially OOM'ed by kernel
2045 * with lack of free htlb page in cpuset that the task is in.
2046 * Attempt to enforce strict accounting with cpuset is almost
2047 * impossible (or too ugly) because cpuset is too fluid that
2048 * task or memory node can be dynamically moved between cpusets.
2050 * The change of semantics for shared hugetlb mapping with cpuset is
2051 * undesirable. However, in order to preserve some of the semantics,
2052 * we fall back to check against current free page availability as
2053 * a best attempt and hopefully to minimize the impact of changing
2054 * semantics that cpuset has.
2057 if (gather_surplus_pages(h, delta) < 0)
2060 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2061 return_unused_surplus_pages(h, delta);
2068 return_unused_surplus_pages(h, (unsigned long) -delta);
2071 spin_unlock(&hugetlb_lock);
2075 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2077 struct resv_map *reservations = vma_resv_map(vma);
2080 * This new VMA should share its siblings reservation map if present.
2081 * The VMA will only ever have a valid reservation map pointer where
2082 * it is being copied for another still existing VMA. As that VMA
2083 * has a reference to the reservation map it cannot disappear until
2084 * after this open call completes. It is therefore safe to take a
2085 * new reference here without additional locking.
2088 kref_get(&reservations->refs);
2091 static void resv_map_put(struct vm_area_struct *vma)
2093 struct resv_map *reservations = vma_resv_map(vma);
2097 kref_put(&reservations->refs, resv_map_release);
2100 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2102 struct hstate *h = hstate_vma(vma);
2103 struct resv_map *reservations = vma_resv_map(vma);
2104 unsigned long reserve;
2105 unsigned long start;
2109 start = vma_hugecache_offset(h, vma, vma->vm_start);
2110 end = vma_hugecache_offset(h, vma, vma->vm_end);
2112 reserve = (end - start) -
2113 region_count(&reservations->regions, start, end);
2118 hugetlb_acct_memory(h, -reserve);
2119 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
2125 * We cannot handle pagefaults against hugetlb pages at all. They cause
2126 * handle_mm_fault() to try to instantiate regular-sized pages in the
2127 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2130 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2136 const struct vm_operations_struct hugetlb_vm_ops = {
2137 .fault = hugetlb_vm_op_fault,
2138 .open = hugetlb_vm_op_open,
2139 .close = hugetlb_vm_op_close,
2142 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2149 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2151 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2153 entry = pte_mkyoung(entry);
2154 entry = pte_mkhuge(entry);
2159 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2160 unsigned long address, pte_t *ptep)
2164 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2165 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
2166 update_mmu_cache(vma, address, ptep);
2171 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2172 struct vm_area_struct *vma)
2174 pte_t *src_pte, *dst_pte, entry;
2175 struct page *ptepage;
2178 struct hstate *h = hstate_vma(vma);
2179 unsigned long sz = huge_page_size(h);
2181 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2183 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2184 src_pte = huge_pte_offset(src, addr);
2187 dst_pte = huge_pte_alloc(dst, addr, sz);
2191 /* If the pagetables are shared don't copy or take references */
2192 if (dst_pte == src_pte)
2195 spin_lock(&dst->page_table_lock);
2196 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2197 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2199 huge_ptep_set_wrprotect(src, addr, src_pte);
2200 entry = huge_ptep_get(src_pte);
2201 ptepage = pte_page(entry);
2203 page_dup_rmap(ptepage);
2204 set_huge_pte_at(dst, addr, dst_pte, entry);
2206 spin_unlock(&src->page_table_lock);
2207 spin_unlock(&dst->page_table_lock);
2215 static int is_hugetlb_entry_migration(pte_t pte)
2219 if (huge_pte_none(pte) || pte_present(pte))
2221 swp = pte_to_swp_entry(pte);
2222 if (non_swap_entry(swp) && is_migration_entry(swp)) {
2228 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2232 if (huge_pte_none(pte) || pte_present(pte))
2234 swp = pte_to_swp_entry(pte);
2235 if (non_swap_entry(swp) && is_hwpoison_entry(swp)) {
2241 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2242 unsigned long end, struct page *ref_page)
2244 struct mm_struct *mm = vma->vm_mm;
2245 unsigned long address;
2250 struct hstate *h = hstate_vma(vma);
2251 unsigned long sz = huge_page_size(h);
2254 * A page gathering list, protected by per file i_mmap_mutex. The
2255 * lock is used to avoid list corruption from multiple unmapping
2256 * of the same page since we are using page->lru.
2258 LIST_HEAD(page_list);
2260 WARN_ON(!is_vm_hugetlb_page(vma));
2261 BUG_ON(start & ~huge_page_mask(h));
2262 BUG_ON(end & ~huge_page_mask(h));
2264 mmu_notifier_invalidate_range_start(mm, start, end);
2265 spin_lock(&mm->page_table_lock);
2266 for (address = start; address < end; address += sz) {
2267 ptep = huge_pte_offset(mm, address);
2271 if (huge_pmd_unshare(mm, &address, ptep))
2275 * If a reference page is supplied, it is because a specific
2276 * page is being unmapped, not a range. Ensure the page we
2277 * are about to unmap is the actual page of interest.
2280 pte = huge_ptep_get(ptep);
2281 if (huge_pte_none(pte))
2283 page = pte_page(pte);
2284 if (page != ref_page)
2288 * Mark the VMA as having unmapped its page so that
2289 * future faults in this VMA will fail rather than
2290 * looking like data was lost
2292 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2295 pte = huge_ptep_get_and_clear(mm, address, ptep);
2296 if (huge_pte_none(pte))
2300 * HWPoisoned hugepage is already unmapped and dropped reference
2302 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2305 page = pte_page(pte);
2307 set_page_dirty(page);
2308 list_add(&page->lru, &page_list);
2310 spin_unlock(&mm->page_table_lock);
2311 flush_tlb_range(vma, start, end);
2312 mmu_notifier_invalidate_range_end(mm, start, end);
2313 list_for_each_entry_safe(page, tmp, &page_list, lru) {
2314 page_remove_rmap(page);
2315 list_del(&page->lru);
2320 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2321 unsigned long end, struct page *ref_page)
2323 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2324 __unmap_hugepage_range(vma, start, end, ref_page);
2326 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2327 * test will fail on a vma being torn down, and not grab a page table
2328 * on its way out. We're lucky that the flag has such an appropriate
2329 * name, and can in fact be safely cleared here. We could clear it
2330 * before the __unmap_hugepage_range above, but all that's necessary
2331 * is to clear it before releasing the i_mmap_mutex below.
2333 * This works because in the contexts this is called, the VMA is
2334 * going to be destroyed. It is not vunerable to madvise(DONTNEED)
2335 * because madvise is not supported on hugetlbfs. The same applies
2336 * for direct IO. unmap_hugepage_range() is only being called just
2337 * before free_pgtables() so clearing VM_MAYSHARE will not cause
2340 vma->vm_flags &= ~VM_MAYSHARE;
2341 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2345 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2346 * mappping it owns the reserve page for. The intention is to unmap the page
2347 * from other VMAs and let the children be SIGKILLed if they are faulting the
2350 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2351 struct page *page, unsigned long address)
2353 struct hstate *h = hstate_vma(vma);
2354 struct vm_area_struct *iter_vma;
2355 struct address_space *mapping;
2356 struct prio_tree_iter iter;
2360 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2361 * from page cache lookup which is in HPAGE_SIZE units.
2363 address = address & huge_page_mask(h);
2364 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
2365 + (vma->vm_pgoff >> PAGE_SHIFT);
2366 mapping = (struct address_space *)page_private(page);
2369 * Take the mapping lock for the duration of the table walk. As
2370 * this mapping should be shared between all the VMAs,
2371 * __unmap_hugepage_range() is called as the lock is already held
2373 mutex_lock(&mapping->i_mmap_mutex);
2374 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2375 /* Do not unmap the current VMA */
2376 if (iter_vma == vma)
2380 * Unmap the page from other VMAs without their own reserves.
2381 * They get marked to be SIGKILLed if they fault in these
2382 * areas. This is because a future no-page fault on this VMA
2383 * could insert a zeroed page instead of the data existing
2384 * from the time of fork. This would look like data corruption
2386 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2387 __unmap_hugepage_range(iter_vma,
2388 address, address + huge_page_size(h),
2391 mutex_unlock(&mapping->i_mmap_mutex);
2397 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2399 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2400 unsigned long address, pte_t *ptep, pte_t pte,
2401 struct page *pagecache_page)
2403 struct hstate *h = hstate_vma(vma);
2404 struct page *old_page, *new_page;
2406 int outside_reserve = 0;
2408 old_page = pte_page(pte);
2411 /* If no-one else is actually using this page, avoid the copy
2412 * and just make the page writable */
2413 avoidcopy = (page_mapcount(old_page) == 1);
2415 if (PageAnon(old_page))
2416 page_move_anon_rmap(old_page, vma, address);
2417 set_huge_ptep_writable(vma, address, ptep);
2422 * If the process that created a MAP_PRIVATE mapping is about to
2423 * perform a COW due to a shared page count, attempt to satisfy
2424 * the allocation without using the existing reserves. The pagecache
2425 * page is used to determine if the reserve at this address was
2426 * consumed or not. If reserves were used, a partial faulted mapping
2427 * at the time of fork() could consume its reserves on COW instead
2428 * of the full address range.
2430 if (!(vma->vm_flags & VM_MAYSHARE) &&
2431 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2432 old_page != pagecache_page)
2433 outside_reserve = 1;
2435 page_cache_get(old_page);
2437 /* Drop page_table_lock as buddy allocator may be called */
2438 spin_unlock(&mm->page_table_lock);
2439 new_page = alloc_huge_page(vma, address, outside_reserve);
2441 if (IS_ERR(new_page)) {
2442 page_cache_release(old_page);
2445 * If a process owning a MAP_PRIVATE mapping fails to COW,
2446 * it is due to references held by a child and an insufficient
2447 * huge page pool. To guarantee the original mappers
2448 * reliability, unmap the page from child processes. The child
2449 * may get SIGKILLed if it later faults.
2451 if (outside_reserve) {
2452 BUG_ON(huge_pte_none(pte));
2453 if (unmap_ref_private(mm, vma, old_page, address)) {
2454 BUG_ON(huge_pte_none(pte));
2455 spin_lock(&mm->page_table_lock);
2456 goto retry_avoidcopy;
2461 /* Caller expects lock to be held */
2462 spin_lock(&mm->page_table_lock);
2463 return -PTR_ERR(new_page);
2467 * When the original hugepage is shared one, it does not have
2468 * anon_vma prepared.
2470 if (unlikely(anon_vma_prepare(vma))) {
2471 page_cache_release(new_page);
2472 page_cache_release(old_page);
2473 /* Caller expects lock to be held */
2474 spin_lock(&mm->page_table_lock);
2475 return VM_FAULT_OOM;
2478 copy_user_huge_page(new_page, old_page, address, vma,
2479 pages_per_huge_page(h));
2480 __SetPageUptodate(new_page);
2483 * Retake the page_table_lock to check for racing updates
2484 * before the page tables are altered
2486 spin_lock(&mm->page_table_lock);
2487 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2488 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2490 mmu_notifier_invalidate_range_start(mm,
2491 address & huge_page_mask(h),
2492 (address & huge_page_mask(h)) + huge_page_size(h));
2493 huge_ptep_clear_flush(vma, address, ptep);
2494 set_huge_pte_at(mm, address, ptep,
2495 make_huge_pte(vma, new_page, 1));
2496 page_remove_rmap(old_page);
2497 hugepage_add_new_anon_rmap(new_page, vma, address);
2498 /* Make the old page be freed below */
2499 new_page = old_page;
2500 mmu_notifier_invalidate_range_end(mm,
2501 address & huge_page_mask(h),
2502 (address & huge_page_mask(h)) + huge_page_size(h));
2504 page_cache_release(new_page);
2505 page_cache_release(old_page);
2509 /* Return the pagecache page at a given address within a VMA */
2510 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2511 struct vm_area_struct *vma, unsigned long address)
2513 struct address_space *mapping;
2516 mapping = vma->vm_file->f_mapping;
2517 idx = vma_hugecache_offset(h, vma, address);
2519 return find_lock_page(mapping, idx);
2523 * Return whether there is a pagecache page to back given address within VMA.
2524 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2526 static bool hugetlbfs_pagecache_present(struct hstate *h,
2527 struct vm_area_struct *vma, unsigned long address)
2529 struct address_space *mapping;
2533 mapping = vma->vm_file->f_mapping;
2534 idx = vma_hugecache_offset(h, vma, address);
2536 page = find_get_page(mapping, idx);
2539 return page != NULL;
2542 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2543 unsigned long address, pte_t *ptep, unsigned int flags)
2545 struct hstate *h = hstate_vma(vma);
2546 int ret = VM_FAULT_SIGBUS;
2550 struct address_space *mapping;
2554 * Currently, we are forced to kill the process in the event the
2555 * original mapper has unmapped pages from the child due to a failed
2556 * COW. Warn that such a situation has occurred as it may not be obvious
2558 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2560 "PID %d killed due to inadequate hugepage pool\n",
2565 mapping = vma->vm_file->f_mapping;
2566 idx = vma_hugecache_offset(h, vma, address);
2569 * Use page lock to guard against racing truncation
2570 * before we get page_table_lock.
2573 page = find_lock_page(mapping, idx);
2575 size = i_size_read(mapping->host) >> huge_page_shift(h);
2578 page = alloc_huge_page(vma, address, 0);
2580 ret = -PTR_ERR(page);
2583 clear_huge_page(page, address, pages_per_huge_page(h));
2584 __SetPageUptodate(page);
2586 if (vma->vm_flags & VM_MAYSHARE) {
2588 struct inode *inode = mapping->host;
2590 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2598 spin_lock(&inode->i_lock);
2599 inode->i_blocks += blocks_per_huge_page(h);
2600 spin_unlock(&inode->i_lock);
2601 page_dup_rmap(page);
2604 if (unlikely(anon_vma_prepare(vma))) {
2606 goto backout_unlocked;
2608 hugepage_add_new_anon_rmap(page, vma, address);
2612 * If memory error occurs between mmap() and fault, some process
2613 * don't have hwpoisoned swap entry for errored virtual address.
2614 * So we need to block hugepage fault by PG_hwpoison bit check.
2616 if (unlikely(PageHWPoison(page))) {
2617 ret = VM_FAULT_HWPOISON |
2618 VM_FAULT_SET_HINDEX(h - hstates);
2619 goto backout_unlocked;
2621 page_dup_rmap(page);
2625 * If we are going to COW a private mapping later, we examine the
2626 * pending reservations for this page now. This will ensure that
2627 * any allocations necessary to record that reservation occur outside
2630 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2631 if (vma_needs_reservation(h, vma, address) < 0) {
2633 goto backout_unlocked;
2636 spin_lock(&mm->page_table_lock);
2637 size = i_size_read(mapping->host) >> huge_page_shift(h);
2642 if (!huge_pte_none(huge_ptep_get(ptep)))
2645 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2646 && (vma->vm_flags & VM_SHARED)));
2647 set_huge_pte_at(mm, address, ptep, new_pte);
2649 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2650 /* Optimization, do the COW without a second fault */
2651 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2654 spin_unlock(&mm->page_table_lock);
2660 spin_unlock(&mm->page_table_lock);
2667 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2668 unsigned long address, unsigned int flags)
2673 struct page *page = NULL;
2674 struct page *pagecache_page = NULL;
2675 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2676 struct hstate *h = hstate_vma(vma);
2678 ptep = huge_pte_offset(mm, address);
2680 entry = huge_ptep_get(ptep);
2681 if (unlikely(is_hugetlb_entry_migration(entry))) {
2682 migration_entry_wait_huge(mm, ptep);
2684 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2685 return VM_FAULT_HWPOISON_LARGE |
2686 VM_FAULT_SET_HINDEX(h - hstates);
2689 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2691 return VM_FAULT_OOM;
2694 * Serialize hugepage allocation and instantiation, so that we don't
2695 * get spurious allocation failures if two CPUs race to instantiate
2696 * the same page in the page cache.
2698 mutex_lock(&hugetlb_instantiation_mutex);
2699 entry = huge_ptep_get(ptep);
2700 if (huge_pte_none(entry)) {
2701 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2708 * If we are going to COW the mapping later, we examine the pending
2709 * reservations for this page now. This will ensure that any
2710 * allocations necessary to record that reservation occur outside the
2711 * spinlock. For private mappings, we also lookup the pagecache
2712 * page now as it is used to determine if a reservation has been
2715 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2716 if (vma_needs_reservation(h, vma, address) < 0) {
2721 if (!(vma->vm_flags & VM_MAYSHARE))
2722 pagecache_page = hugetlbfs_pagecache_page(h,
2727 * hugetlb_cow() requires page locks of pte_page(entry) and
2728 * pagecache_page, so here we need take the former one
2729 * when page != pagecache_page or !pagecache_page.
2730 * Note that locking order is always pagecache_page -> page,
2731 * so no worry about deadlock.
2733 page = pte_page(entry);
2735 if (page != pagecache_page)
2738 spin_lock(&mm->page_table_lock);
2739 /* Check for a racing update before calling hugetlb_cow */
2740 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2741 goto out_page_table_lock;
2744 if (flags & FAULT_FLAG_WRITE) {
2745 if (!pte_write(entry)) {
2746 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2748 goto out_page_table_lock;
2750 entry = pte_mkdirty(entry);
2752 entry = pte_mkyoung(entry);
2753 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2754 flags & FAULT_FLAG_WRITE))
2755 update_mmu_cache(vma, address, ptep);
2757 out_page_table_lock:
2758 spin_unlock(&mm->page_table_lock);
2760 if (pagecache_page) {
2761 unlock_page(pagecache_page);
2762 put_page(pagecache_page);
2764 if (page != pagecache_page)
2769 mutex_unlock(&hugetlb_instantiation_mutex);
2774 /* Can be overriden by architectures */
2775 __attribute__((weak)) struct page *
2776 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2777 pud_t *pud, int write)
2783 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2784 struct page **pages, struct vm_area_struct **vmas,
2785 unsigned long *position, int *length, int i,
2788 unsigned long pfn_offset;
2789 unsigned long vaddr = *position;
2790 int remainder = *length;
2791 struct hstate *h = hstate_vma(vma);
2793 spin_lock(&mm->page_table_lock);
2794 while (vaddr < vma->vm_end && remainder) {
2800 * Some archs (sparc64, sh*) have multiple pte_ts to
2801 * each hugepage. We have to make sure we get the
2802 * first, for the page indexing below to work.
2804 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2805 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2808 * When coredumping, it suits get_dump_page if we just return
2809 * an error where there's an empty slot with no huge pagecache
2810 * to back it. This way, we avoid allocating a hugepage, and
2811 * the sparse dumpfile avoids allocating disk blocks, but its
2812 * huge holes still show up with zeroes where they need to be.
2814 if (absent && (flags & FOLL_DUMP) &&
2815 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2821 * We need call hugetlb_fault for both hugepages under migration
2822 * (in which case hugetlb_fault waits for the migration,) and
2823 * hwpoisoned hugepages (in which case we need to prevent the
2824 * caller from accessing to them.) In order to do this, we use
2825 * here is_swap_pte instead of is_hugetlb_entry_migration and
2826 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
2827 * both cases, and because we can't follow correct pages
2828 * directly from any kind of swap entries.
2830 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
2831 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2834 spin_unlock(&mm->page_table_lock);
2835 ret = hugetlb_fault(mm, vma, vaddr,
2836 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2837 spin_lock(&mm->page_table_lock);
2838 if (!(ret & VM_FAULT_ERROR))
2845 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2846 page = pte_page(huge_ptep_get(pte));
2849 pages[i] = mem_map_offset(page, pfn_offset);
2860 if (vaddr < vma->vm_end && remainder &&
2861 pfn_offset < pages_per_huge_page(h)) {
2863 * We use pfn_offset to avoid touching the pageframes
2864 * of this compound page.
2869 spin_unlock(&mm->page_table_lock);
2870 *length = remainder;
2873 return i ? i : -EFAULT;
2876 void hugetlb_change_protection(struct vm_area_struct *vma,
2877 unsigned long address, unsigned long end, pgprot_t newprot)
2879 struct mm_struct *mm = vma->vm_mm;
2880 unsigned long start = address;
2883 struct hstate *h = hstate_vma(vma);
2885 BUG_ON(address >= end);
2886 flush_cache_range(vma, address, end);
2888 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2889 spin_lock(&mm->page_table_lock);
2890 for (; address < end; address += huge_page_size(h)) {
2891 ptep = huge_pte_offset(mm, address);
2894 if (huge_pmd_unshare(mm, &address, ptep))
2896 if (!huge_pte_none(huge_ptep_get(ptep))) {
2897 pte = huge_ptep_get_and_clear(mm, address, ptep);
2898 pte = pte_mkhuge(pte_modify(pte, newprot));
2899 set_huge_pte_at(mm, address, ptep, pte);
2902 spin_unlock(&mm->page_table_lock);
2904 * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
2905 * may have cleared our pud entry and done put_page on the page table:
2906 * once we release i_mmap_mutex, another task can do the final put_page
2907 * and that page table be reused and filled with junk.
2909 flush_tlb_range(vma, start, end);
2910 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2913 int hugetlb_reserve_pages(struct inode *inode,
2915 struct vm_area_struct *vma,
2916 vm_flags_t vm_flags)
2919 struct hstate *h = hstate_inode(inode);
2922 * Only apply hugepage reservation if asked. At fault time, an
2923 * attempt will be made for VM_NORESERVE to allocate a page
2924 * and filesystem quota without using reserves
2926 if (vm_flags & VM_NORESERVE)
2930 * Shared mappings base their reservation on the number of pages that
2931 * are already allocated on behalf of the file. Private mappings need
2932 * to reserve the full area even if read-only as mprotect() may be
2933 * called to make the mapping read-write. Assume !vma is a shm mapping
2935 if (!vma || vma->vm_flags & VM_MAYSHARE)
2936 chg = region_chg(&inode->i_mapping->private_list, from, to);
2938 struct resv_map *resv_map = resv_map_alloc();
2944 set_vma_resv_map(vma, resv_map);
2945 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2953 /* There must be enough filesystem quota for the mapping */
2954 if (hugetlb_get_quota(inode->i_mapping, chg)) {
2960 * Check enough hugepages are available for the reservation.
2961 * Hand back the quota if there are not
2963 ret = hugetlb_acct_memory(h, chg);
2965 hugetlb_put_quota(inode->i_mapping, chg);
2970 * Account for the reservations made. Shared mappings record regions
2971 * that have reservations as they are shared by multiple VMAs.
2972 * When the last VMA disappears, the region map says how much
2973 * the reservation was and the page cache tells how much of
2974 * the reservation was consumed. Private mappings are per-VMA and
2975 * only the consumed reservations are tracked. When the VMA
2976 * disappears, the original reservation is the VMA size and the
2977 * consumed reservations are stored in the map. Hence, nothing
2978 * else has to be done for private mappings here
2980 if (!vma || vma->vm_flags & VM_MAYSHARE)
2981 region_add(&inode->i_mapping->private_list, from, to);
2989 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2991 struct hstate *h = hstate_inode(inode);
2992 long chg = region_truncate(&inode->i_mapping->private_list, offset);
2994 spin_lock(&inode->i_lock);
2995 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
2996 spin_unlock(&inode->i_lock);
2998 hugetlb_put_quota(inode->i_mapping, (chg - freed));
2999 hugetlb_acct_memory(h, -(chg - freed));
3002 #ifdef CONFIG_MEMORY_FAILURE
3004 /* Should be called in hugetlb_lock */
3005 static int is_hugepage_on_freelist(struct page *hpage)
3009 struct hstate *h = page_hstate(hpage);
3010 int nid = page_to_nid(hpage);
3012 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3019 * This function is called from memory failure code.
3020 * Assume the caller holds page lock of the head page.
3022 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3024 struct hstate *h = page_hstate(hpage);
3025 int nid = page_to_nid(hpage);
3028 spin_lock(&hugetlb_lock);
3029 if (is_hugepage_on_freelist(hpage)) {
3030 list_del(&hpage->lru);
3031 set_page_refcounted(hpage);
3032 h->free_huge_pages--;
3033 h->free_huge_pages_node[nid]--;
3036 spin_unlock(&hugetlb_lock);