2 * Generic hugetlb support.
3 * (C) William Irwin, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/module.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/rmap.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
26 #include <asm/pgtable.h>
29 #include <linux/hugetlb.h>
30 #include <linux/node.h>
33 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
34 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
35 unsigned long hugepages_treat_as_movable;
37 static int max_hstate;
38 unsigned int default_hstate_idx;
39 struct hstate hstates[HUGE_MAX_HSTATE];
41 __initdata LIST_HEAD(huge_boot_pages);
43 /* for command line parsing */
44 static struct hstate * __initdata parsed_hstate;
45 static unsigned long __initdata default_hstate_max_huge_pages;
46 static unsigned long __initdata default_hstate_size;
48 #define for_each_hstate(h) \
49 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
52 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
54 static DEFINE_SPINLOCK(hugetlb_lock);
57 * Region tracking -- allows tracking of reservations and instantiated pages
58 * across the pages in a mapping.
60 * The region data structures are protected by a combination of the mmap_sem
61 * and the hugetlb_instantion_mutex. To access or modify a region the caller
62 * must either hold the mmap_sem for write, or the mmap_sem for read and
63 * the hugetlb_instantiation mutex:
65 * down_write(&mm->mmap_sem);
67 * down_read(&mm->mmap_sem);
68 * mutex_lock(&hugetlb_instantiation_mutex);
71 struct list_head link;
76 static long region_add(struct list_head *head, long f, long t)
78 struct file_region *rg, *nrg, *trg;
80 /* Locate the region we are either in or before. */
81 list_for_each_entry(rg, head, link)
85 /* Round our left edge to the current segment if it encloses us. */
89 /* Check for and consume any regions we now overlap with. */
91 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
92 if (&rg->link == head)
97 /* If this area reaches higher then extend our area to
98 * include it completely. If this is not the first area
99 * which we intend to reuse, free it. */
112 static long region_chg(struct list_head *head, long f, long t)
114 struct file_region *rg, *nrg;
117 /* Locate the region we are before or in. */
118 list_for_each_entry(rg, head, link)
122 /* If we are below the current region then a new region is required.
123 * Subtle, allocate a new region at the position but make it zero
124 * size such that we can guarantee to record the reservation. */
125 if (&rg->link == head || t < rg->from) {
126 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
131 INIT_LIST_HEAD(&nrg->link);
132 list_add(&nrg->link, rg->link.prev);
137 /* Round our left edge to the current segment if it encloses us. */
142 /* Check for and consume any regions we now overlap with. */
143 list_for_each_entry(rg, rg->link.prev, link) {
144 if (&rg->link == head)
149 /* We overlap with this area, if it extends futher than
150 * us then we must extend ourselves. Account for its
151 * existing reservation. */
156 chg -= rg->to - rg->from;
161 static long region_truncate(struct list_head *head, long end)
163 struct file_region *rg, *trg;
166 /* Locate the region we are either in or before. */
167 list_for_each_entry(rg, head, link)
170 if (&rg->link == head)
173 /* If we are in the middle of a region then adjust it. */
174 if (end > rg->from) {
177 rg = list_entry(rg->link.next, typeof(*rg), link);
180 /* Drop any remaining regions. */
181 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
182 if (&rg->link == head)
184 chg += rg->to - rg->from;
191 static long region_count(struct list_head *head, long f, long t)
193 struct file_region *rg;
196 /* Locate each segment we overlap with, and count that overlap. */
197 list_for_each_entry(rg, head, link) {
206 seg_from = max(rg->from, f);
207 seg_to = min(rg->to, t);
209 chg += seg_to - seg_from;
216 * Convert the address within this vma to the page offset within
217 * the mapping, in pagecache page units; huge pages here.
219 static pgoff_t vma_hugecache_offset(struct hstate *h,
220 struct vm_area_struct *vma, unsigned long address)
222 return ((address - vma->vm_start) >> huge_page_shift(h)) +
223 (vma->vm_pgoff >> huge_page_order(h));
226 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
227 unsigned long address)
229 return vma_hugecache_offset(hstate_vma(vma), vma, address);
233 * Return the size of the pages allocated when backing a VMA. In the majority
234 * cases this will be same size as used by the page table entries.
236 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
238 struct hstate *hstate;
240 if (!is_vm_hugetlb_page(vma))
243 hstate = hstate_vma(vma);
245 return 1UL << (hstate->order + PAGE_SHIFT);
247 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
250 * Return the page size being used by the MMU to back a VMA. In the majority
251 * of cases, the page size used by the kernel matches the MMU size. On
252 * architectures where it differs, an architecture-specific version of this
253 * function is required.
255 #ifndef vma_mmu_pagesize
256 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
258 return vma_kernel_pagesize(vma);
263 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
264 * bits of the reservation map pointer, which are always clear due to
267 #define HPAGE_RESV_OWNER (1UL << 0)
268 #define HPAGE_RESV_UNMAPPED (1UL << 1)
269 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
272 * These helpers are used to track how many pages are reserved for
273 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
274 * is guaranteed to have their future faults succeed.
276 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
277 * the reserve counters are updated with the hugetlb_lock held. It is safe
278 * to reset the VMA at fork() time as it is not in use yet and there is no
279 * chance of the global counters getting corrupted as a result of the values.
281 * The private mapping reservation is represented in a subtly different
282 * manner to a shared mapping. A shared mapping has a region map associated
283 * with the underlying file, this region map represents the backing file
284 * pages which have ever had a reservation assigned which this persists even
285 * after the page is instantiated. A private mapping has a region map
286 * associated with the original mmap which is attached to all VMAs which
287 * reference it, this region map represents those offsets which have consumed
288 * reservation ie. where pages have been instantiated.
290 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
292 return (unsigned long)vma->vm_private_data;
295 static void set_vma_private_data(struct vm_area_struct *vma,
298 vma->vm_private_data = (void *)value;
303 struct list_head regions;
306 static struct resv_map *resv_map_alloc(void)
308 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
312 kref_init(&resv_map->refs);
313 INIT_LIST_HEAD(&resv_map->regions);
318 static void resv_map_release(struct kref *ref)
320 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
322 /* Clear out any active regions before we release the map. */
323 region_truncate(&resv_map->regions, 0);
327 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
329 VM_BUG_ON(!is_vm_hugetlb_page(vma));
330 if (!(vma->vm_flags & VM_MAYSHARE))
331 return (struct resv_map *)(get_vma_private_data(vma) &
336 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
338 VM_BUG_ON(!is_vm_hugetlb_page(vma));
339 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
341 set_vma_private_data(vma, (get_vma_private_data(vma) &
342 HPAGE_RESV_MASK) | (unsigned long)map);
345 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
347 VM_BUG_ON(!is_vm_hugetlb_page(vma));
348 VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
350 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
353 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
355 VM_BUG_ON(!is_vm_hugetlb_page(vma));
357 return (get_vma_private_data(vma) & flag) != 0;
360 /* Decrement the reserved pages in the hugepage pool by one */
361 static void decrement_hugepage_resv_vma(struct hstate *h,
362 struct vm_area_struct *vma)
364 if (vma->vm_flags & VM_NORESERVE)
367 if (vma->vm_flags & VM_MAYSHARE) {
368 /* Shared mappings always use reserves */
369 h->resv_huge_pages--;
370 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
372 * Only the process that called mmap() has reserves for
375 h->resv_huge_pages--;
379 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
380 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
382 VM_BUG_ON(!is_vm_hugetlb_page(vma));
383 if (!(vma->vm_flags & VM_MAYSHARE))
384 vma->vm_private_data = (void *)0;
387 /* Returns true if the VMA has associated reserve pages */
388 static int vma_has_reserves(struct vm_area_struct *vma)
390 if (vma->vm_flags & VM_MAYSHARE)
392 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
397 static void clear_gigantic_page(struct page *page,
398 unsigned long addr, unsigned long sz)
401 struct page *p = page;
404 for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) {
406 clear_user_highpage(p, addr + i * PAGE_SIZE);
409 static void clear_huge_page(struct page *page,
410 unsigned long addr, unsigned long sz)
414 if (unlikely(sz/PAGE_SIZE > MAX_ORDER_NR_PAGES)) {
415 clear_gigantic_page(page, addr, sz);
420 for (i = 0; i < sz/PAGE_SIZE; i++) {
422 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
426 static void copy_gigantic_page(struct page *dst, struct page *src,
427 unsigned long addr, struct vm_area_struct *vma)
430 struct hstate *h = hstate_vma(vma);
431 struct page *dst_base = dst;
432 struct page *src_base = src;
434 for (i = 0; i < pages_per_huge_page(h); ) {
436 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
439 dst = mem_map_next(dst, dst_base, i);
440 src = mem_map_next(src, src_base, i);
443 static void copy_huge_page(struct page *dst, struct page *src,
444 unsigned long addr, struct vm_area_struct *vma)
447 struct hstate *h = hstate_vma(vma);
449 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
450 copy_gigantic_page(dst, src, addr, vma);
455 for (i = 0; i < pages_per_huge_page(h); i++) {
457 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
461 static void enqueue_huge_page(struct hstate *h, struct page *page)
463 int nid = page_to_nid(page);
464 list_add(&page->lru, &h->hugepage_freelists[nid]);
465 h->free_huge_pages++;
466 h->free_huge_pages_node[nid]++;
469 static struct page *dequeue_huge_page_vma(struct hstate *h,
470 struct vm_area_struct *vma,
471 unsigned long address, int avoid_reserve)
474 struct page *page = NULL;
475 struct mempolicy *mpol;
476 nodemask_t *nodemask;
477 struct zonelist *zonelist;
482 zonelist = huge_zonelist(vma, address,
483 htlb_alloc_mask, &mpol, &nodemask);
485 * A child process with MAP_PRIVATE mappings created by their parent
486 * have no page reserves. This check ensures that reservations are
487 * not "stolen". The child may still get SIGKILLed
489 if (!vma_has_reserves(vma) &&
490 h->free_huge_pages - h->resv_huge_pages == 0)
493 /* If reserves cannot be used, ensure enough pages are in the pool */
494 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
497 for_each_zone_zonelist_nodemask(zone, z, zonelist,
498 MAX_NR_ZONES - 1, nodemask) {
499 nid = zone_to_nid(zone);
500 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
501 !list_empty(&h->hugepage_freelists[nid])) {
502 page = list_entry(h->hugepage_freelists[nid].next,
504 list_del(&page->lru);
505 h->free_huge_pages--;
506 h->free_huge_pages_node[nid]--;
509 decrement_hugepage_resv_vma(h, vma);
520 static void update_and_free_page(struct hstate *h, struct page *page)
524 VM_BUG_ON(h->order >= MAX_ORDER);
527 h->nr_huge_pages_node[page_to_nid(page)]--;
528 for (i = 0; i < pages_per_huge_page(h); i++) {
529 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
530 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
531 1 << PG_private | 1<< PG_writeback);
533 set_compound_page_dtor(page, NULL);
534 set_page_refcounted(page);
535 arch_release_hugepage(page);
536 __free_pages(page, huge_page_order(h));
539 struct hstate *size_to_hstate(unsigned long size)
544 if (huge_page_size(h) == size)
550 static void free_huge_page(struct page *page)
553 * Can't pass hstate in here because it is called from the
554 * compound page destructor.
556 struct hstate *h = page_hstate(page);
557 int nid = page_to_nid(page);
558 struct address_space *mapping;
560 mapping = (struct address_space *) page_private(page);
561 set_page_private(page, 0);
562 page->mapping = NULL;
563 BUG_ON(page_count(page));
564 BUG_ON(page_mapcount(page));
565 INIT_LIST_HEAD(&page->lru);
567 spin_lock(&hugetlb_lock);
568 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
569 update_and_free_page(h, page);
570 h->surplus_huge_pages--;
571 h->surplus_huge_pages_node[nid]--;
573 enqueue_huge_page(h, page);
575 spin_unlock(&hugetlb_lock);
577 hugetlb_put_quota(mapping, 1);
580 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
582 set_compound_page_dtor(page, free_huge_page);
583 spin_lock(&hugetlb_lock);
585 h->nr_huge_pages_node[nid]++;
586 spin_unlock(&hugetlb_lock);
587 put_page(page); /* free it into the hugepage allocator */
590 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
593 int nr_pages = 1 << order;
594 struct page *p = page + 1;
596 /* we rely on prep_new_huge_page to set the destructor */
597 set_compound_order(page, order);
599 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
601 p->first_page = page;
605 int PageHuge(struct page *page)
607 compound_page_dtor *dtor;
609 if (!PageCompound(page))
612 page = compound_head(page);
613 dtor = get_compound_page_dtor(page);
615 return dtor == free_huge_page;
618 EXPORT_SYMBOL_GPL(PageHuge);
620 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
624 if (h->order >= MAX_ORDER)
627 page = alloc_pages_exact_node(nid,
628 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
629 __GFP_REPEAT|__GFP_NOWARN,
632 if (arch_prepare_hugepage(page)) {
633 __free_pages(page, huge_page_order(h));
636 prep_new_huge_page(h, page, nid);
643 * common helper functions for hstate_next_node_to_{alloc|free}.
644 * We may have allocated or freed a huge page based on a different
645 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
646 * be outside of *nodes_allowed. Ensure that we use an allowed
647 * node for alloc or free.
649 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
651 nid = next_node(nid, *nodes_allowed);
652 if (nid == MAX_NUMNODES)
653 nid = first_node(*nodes_allowed);
654 VM_BUG_ON(nid >= MAX_NUMNODES);
659 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
661 if (!node_isset(nid, *nodes_allowed))
662 nid = next_node_allowed(nid, nodes_allowed);
667 * returns the previously saved node ["this node"] from which to
668 * allocate a persistent huge page for the pool and advance the
669 * next node from which to allocate, handling wrap at end of node
672 static int hstate_next_node_to_alloc(struct hstate *h,
673 nodemask_t *nodes_allowed)
677 VM_BUG_ON(!nodes_allowed);
679 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
680 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
685 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
692 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
693 next_nid = start_nid;
696 page = alloc_fresh_huge_page_node(h, next_nid);
701 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
702 } while (next_nid != start_nid);
705 count_vm_event(HTLB_BUDDY_PGALLOC);
707 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
713 * helper for free_pool_huge_page() - return the previously saved
714 * node ["this node"] from which to free a huge page. Advance the
715 * next node id whether or not we find a free huge page to free so
716 * that the next attempt to free addresses the next node.
718 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
722 VM_BUG_ON(!nodes_allowed);
724 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
725 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
731 * Free huge page from pool from next node to free.
732 * Attempt to keep persistent huge pages more or less
733 * balanced over allowed nodes.
734 * Called with hugetlb_lock locked.
736 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
743 start_nid = hstate_next_node_to_free(h, nodes_allowed);
744 next_nid = start_nid;
748 * If we're returning unused surplus pages, only examine
749 * nodes with surplus pages.
751 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
752 !list_empty(&h->hugepage_freelists[next_nid])) {
754 list_entry(h->hugepage_freelists[next_nid].next,
756 list_del(&page->lru);
757 h->free_huge_pages--;
758 h->free_huge_pages_node[next_nid]--;
760 h->surplus_huge_pages--;
761 h->surplus_huge_pages_node[next_nid]--;
763 update_and_free_page(h, page);
767 next_nid = hstate_next_node_to_free(h, nodes_allowed);
768 } while (next_nid != start_nid);
773 static struct page *alloc_buddy_huge_page(struct hstate *h,
774 struct vm_area_struct *vma, unsigned long address)
779 if (h->order >= MAX_ORDER)
783 * Assume we will successfully allocate the surplus page to
784 * prevent racing processes from causing the surplus to exceed
787 * This however introduces a different race, where a process B
788 * tries to grow the static hugepage pool while alloc_pages() is
789 * called by process A. B will only examine the per-node
790 * counters in determining if surplus huge pages can be
791 * converted to normal huge pages in adjust_pool_surplus(). A
792 * won't be able to increment the per-node counter, until the
793 * lock is dropped by B, but B doesn't drop hugetlb_lock until
794 * no more huge pages can be converted from surplus to normal
795 * state (and doesn't try to convert again). Thus, we have a
796 * case where a surplus huge page exists, the pool is grown, and
797 * the surplus huge page still exists after, even though it
798 * should just have been converted to a normal huge page. This
799 * does not leak memory, though, as the hugepage will be freed
800 * once it is out of use. It also does not allow the counters to
801 * go out of whack in adjust_pool_surplus() as we don't modify
802 * the node values until we've gotten the hugepage and only the
803 * per-node value is checked there.
805 spin_lock(&hugetlb_lock);
806 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
807 spin_unlock(&hugetlb_lock);
811 h->surplus_huge_pages++;
813 spin_unlock(&hugetlb_lock);
815 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
816 __GFP_REPEAT|__GFP_NOWARN,
819 if (page && arch_prepare_hugepage(page)) {
820 __free_pages(page, huge_page_order(h));
824 spin_lock(&hugetlb_lock);
827 * This page is now managed by the hugetlb allocator and has
828 * no users -- drop the buddy allocator's reference.
830 put_page_testzero(page);
831 VM_BUG_ON(page_count(page));
832 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[nid]++;
838 h->surplus_huge_pages_node[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 * Increase the hugetlb pool such that it can accomodate a reservation
854 static int gather_surplus_pages(struct hstate *h, int delta)
856 struct list_head surplus_list;
857 struct page *page, *tmp;
859 int needed, allocated;
861 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
863 h->resv_huge_pages += delta;
868 INIT_LIST_HEAD(&surplus_list);
872 spin_unlock(&hugetlb_lock);
873 for (i = 0; i < needed; i++) {
874 page = alloc_buddy_huge_page(h, NULL, 0);
877 * We were not able to allocate enough pages to
878 * satisfy the entire reservation so we free what
879 * we've allocated so far.
881 spin_lock(&hugetlb_lock);
886 list_add(&page->lru, &surplus_list);
891 * After retaking hugetlb_lock, we need to recalculate 'needed'
892 * because either resv_huge_pages or free_huge_pages may have changed.
894 spin_lock(&hugetlb_lock);
895 needed = (h->resv_huge_pages + delta) -
896 (h->free_huge_pages + allocated);
901 * The surplus_list now contains _at_least_ the number of extra pages
902 * needed to accomodate the reservation. Add the appropriate number
903 * of pages to the hugetlb pool and free the extras back to the buddy
904 * allocator. Commit the entire reservation here to prevent another
905 * process from stealing the pages as they are added to the pool but
906 * before they are reserved.
909 h->resv_huge_pages += delta;
912 /* Free the needed pages to the hugetlb pool */
913 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
916 list_del(&page->lru);
917 enqueue_huge_page(h, page);
920 /* Free unnecessary surplus pages to the buddy allocator */
921 if (!list_empty(&surplus_list)) {
922 spin_unlock(&hugetlb_lock);
923 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
924 list_del(&page->lru);
926 * The page has a reference count of zero already, so
927 * call free_huge_page directly instead of using
928 * put_page. This must be done with hugetlb_lock
929 * unlocked which is safe because free_huge_page takes
930 * hugetlb_lock before deciding how to free the page.
932 free_huge_page(page);
934 spin_lock(&hugetlb_lock);
941 * When releasing a hugetlb pool reservation, any surplus pages that were
942 * allocated to satisfy the reservation must be explicitly freed if they were
944 * Called with hugetlb_lock held.
946 static void return_unused_surplus_pages(struct hstate *h,
947 unsigned long unused_resv_pages)
949 unsigned long nr_pages;
951 /* Uncommit the reservation */
952 h->resv_huge_pages -= unused_resv_pages;
954 /* Cannot return gigantic pages currently */
955 if (h->order >= MAX_ORDER)
958 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
961 * We want to release as many surplus pages as possible, spread
962 * evenly across all nodes with memory. Iterate across these nodes
963 * until we can no longer free unreserved surplus pages. This occurs
964 * when the nodes with surplus pages have no free pages.
965 * free_pool_huge_page() will balance the the freed pages across the
966 * on-line nodes with memory and will handle the hstate accounting.
969 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
975 * Determine if the huge page at addr within the vma has an associated
976 * reservation. Where it does not we will need to logically increase
977 * reservation and actually increase quota before an allocation can occur.
978 * Where any new reservation would be required the reservation change is
979 * prepared, but not committed. Once the page has been quota'd allocated
980 * an instantiated the change should be committed via vma_commit_reservation.
981 * No action is required on failure.
983 static long vma_needs_reservation(struct hstate *h,
984 struct vm_area_struct *vma, unsigned long addr)
986 struct address_space *mapping = vma->vm_file->f_mapping;
987 struct inode *inode = mapping->host;
989 if (vma->vm_flags & VM_MAYSHARE) {
990 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
991 return region_chg(&inode->i_mapping->private_list,
994 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
999 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1000 struct resv_map *reservations = vma_resv_map(vma);
1002 err = region_chg(&reservations->regions, idx, idx + 1);
1008 static void vma_commit_reservation(struct hstate *h,
1009 struct vm_area_struct *vma, unsigned long addr)
1011 struct address_space *mapping = vma->vm_file->f_mapping;
1012 struct inode *inode = mapping->host;
1014 if (vma->vm_flags & VM_MAYSHARE) {
1015 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1016 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1018 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1019 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1020 struct resv_map *reservations = vma_resv_map(vma);
1022 /* Mark this page used in the map. */
1023 region_add(&reservations->regions, idx, idx + 1);
1027 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1028 unsigned long addr, int avoid_reserve)
1030 struct hstate *h = hstate_vma(vma);
1032 struct address_space *mapping = vma->vm_file->f_mapping;
1033 struct inode *inode = mapping->host;
1037 * Processes that did not create the mapping will have no reserves and
1038 * will not have accounted against quota. Check that the quota can be
1039 * made before satisfying the allocation
1040 * MAP_NORESERVE mappings may also need pages and quota allocated
1041 * if no reserve mapping overlaps.
1043 chg = vma_needs_reservation(h, vma, addr);
1045 return ERR_PTR(chg);
1047 if (hugetlb_get_quota(inode->i_mapping, chg))
1048 return ERR_PTR(-ENOSPC);
1050 spin_lock(&hugetlb_lock);
1051 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1052 spin_unlock(&hugetlb_lock);
1055 page = alloc_buddy_huge_page(h, vma, addr);
1057 hugetlb_put_quota(inode->i_mapping, chg);
1058 return ERR_PTR(-VM_FAULT_SIGBUS);
1062 set_page_refcounted(page);
1063 set_page_private(page, (unsigned long) mapping);
1065 vma_commit_reservation(h, vma, addr);
1070 int __weak alloc_bootmem_huge_page(struct hstate *h)
1072 struct huge_bootmem_page *m;
1073 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1078 addr = __alloc_bootmem_node_nopanic(
1079 NODE_DATA(hstate_next_node_to_alloc(h,
1080 &node_states[N_HIGH_MEMORY])),
1081 huge_page_size(h), huge_page_size(h), 0);
1085 * Use the beginning of the huge page to store the
1086 * huge_bootmem_page struct (until gather_bootmem
1087 * puts them into the mem_map).
1097 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1098 /* Put them into a private list first because mem_map is not up yet */
1099 list_add(&m->list, &huge_boot_pages);
1104 static void prep_compound_huge_page(struct page *page, int order)
1106 if (unlikely(order > (MAX_ORDER - 1)))
1107 prep_compound_gigantic_page(page, order);
1109 prep_compound_page(page, order);
1112 /* Put bootmem huge pages into the standard lists after mem_map is up */
1113 static void __init gather_bootmem_prealloc(void)
1115 struct huge_bootmem_page *m;
1117 list_for_each_entry(m, &huge_boot_pages, list) {
1118 struct page *page = virt_to_page(m);
1119 struct hstate *h = m->hstate;
1120 __ClearPageReserved(page);
1121 WARN_ON(page_count(page) != 1);
1122 prep_compound_huge_page(page, h->order);
1123 prep_new_huge_page(h, page, page_to_nid(page));
1127 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1131 for (i = 0; i < h->max_huge_pages; ++i) {
1132 if (h->order >= MAX_ORDER) {
1133 if (!alloc_bootmem_huge_page(h))
1135 } else if (!alloc_fresh_huge_page(h,
1136 &node_states[N_HIGH_MEMORY]))
1139 h->max_huge_pages = i;
1142 static void __init hugetlb_init_hstates(void)
1146 for_each_hstate(h) {
1147 /* oversize hugepages were init'ed in early boot */
1148 if (h->order < MAX_ORDER)
1149 hugetlb_hstate_alloc_pages(h);
1153 static char * __init memfmt(char *buf, unsigned long n)
1155 if (n >= (1UL << 30))
1156 sprintf(buf, "%lu GB", n >> 30);
1157 else if (n >= (1UL << 20))
1158 sprintf(buf, "%lu MB", n >> 20);
1160 sprintf(buf, "%lu KB", n >> 10);
1164 static void __init report_hugepages(void)
1168 for_each_hstate(h) {
1170 printk(KERN_INFO "HugeTLB registered %s page size, "
1171 "pre-allocated %ld pages\n",
1172 memfmt(buf, huge_page_size(h)),
1173 h->free_huge_pages);
1177 #ifdef CONFIG_HIGHMEM
1178 static void try_to_free_low(struct hstate *h, unsigned long count,
1179 nodemask_t *nodes_allowed)
1183 if (h->order >= MAX_ORDER)
1186 for_each_node_mask(i, *nodes_allowed) {
1187 struct page *page, *next;
1188 struct list_head *freel = &h->hugepage_freelists[i];
1189 list_for_each_entry_safe(page, next, freel, lru) {
1190 if (count >= h->nr_huge_pages)
1192 if (PageHighMem(page))
1194 list_del(&page->lru);
1195 update_and_free_page(h, page);
1196 h->free_huge_pages--;
1197 h->free_huge_pages_node[page_to_nid(page)]--;
1202 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1203 nodemask_t *nodes_allowed)
1209 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1210 * balanced by operating on them in a round-robin fashion.
1211 * Returns 1 if an adjustment was made.
1213 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1216 int start_nid, next_nid;
1219 VM_BUG_ON(delta != -1 && delta != 1);
1222 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1224 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1225 next_nid = start_nid;
1231 * To shrink on this node, there must be a surplus page
1233 if (!h->surplus_huge_pages_node[nid]) {
1234 next_nid = hstate_next_node_to_alloc(h,
1241 * Surplus cannot exceed the total number of pages
1243 if (h->surplus_huge_pages_node[nid] >=
1244 h->nr_huge_pages_node[nid]) {
1245 next_nid = hstate_next_node_to_free(h,
1251 h->surplus_huge_pages += delta;
1252 h->surplus_huge_pages_node[nid] += delta;
1255 } while (next_nid != start_nid);
1260 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1261 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1262 nodemask_t *nodes_allowed)
1264 unsigned long min_count, ret;
1266 if (h->order >= MAX_ORDER)
1267 return h->max_huge_pages;
1270 * Increase the pool size
1271 * First take pages out of surplus state. Then make up the
1272 * remaining difference by allocating fresh huge pages.
1274 * We might race with alloc_buddy_huge_page() here and be unable
1275 * to convert a surplus huge page to a normal huge page. That is
1276 * not critical, though, it just means the overall size of the
1277 * pool might be one hugepage larger than it needs to be, but
1278 * within all the constraints specified by the sysctls.
1280 spin_lock(&hugetlb_lock);
1281 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1282 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1286 while (count > persistent_huge_pages(h)) {
1288 * If this allocation races such that we no longer need the
1289 * page, free_huge_page will handle it by freeing the page
1290 * and reducing the surplus.
1292 spin_unlock(&hugetlb_lock);
1293 ret = alloc_fresh_huge_page(h, nodes_allowed);
1294 spin_lock(&hugetlb_lock);
1298 /* Bail for signals. Probably ctrl-c from user */
1299 if (signal_pending(current))
1304 * Decrease the pool size
1305 * First return free pages to the buddy allocator (being careful
1306 * to keep enough around to satisfy reservations). Then place
1307 * pages into surplus state as needed so the pool will shrink
1308 * to the desired size as pages become free.
1310 * By placing pages into the surplus state independent of the
1311 * overcommit value, we are allowing the surplus pool size to
1312 * exceed overcommit. There are few sane options here. Since
1313 * alloc_buddy_huge_page() is checking the global counter,
1314 * though, we'll note that we're not allowed to exceed surplus
1315 * and won't grow the pool anywhere else. Not until one of the
1316 * sysctls are changed, or the surplus pages go out of use.
1318 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1319 min_count = max(count, min_count);
1320 try_to_free_low(h, min_count, nodes_allowed);
1321 while (min_count < persistent_huge_pages(h)) {
1322 if (!free_pool_huge_page(h, nodes_allowed, 0))
1325 while (count < persistent_huge_pages(h)) {
1326 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1330 ret = persistent_huge_pages(h);
1331 spin_unlock(&hugetlb_lock);
1335 #define HSTATE_ATTR_RO(_name) \
1336 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1338 #define HSTATE_ATTR(_name) \
1339 static struct kobj_attribute _name##_attr = \
1340 __ATTR(_name, 0644, _name##_show, _name##_store)
1342 static struct kobject *hugepages_kobj;
1343 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1345 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1347 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1351 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1352 if (hstate_kobjs[i] == kobj) {
1354 *nidp = NUMA_NO_NODE;
1358 return kobj_to_node_hstate(kobj, nidp);
1361 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1362 struct kobj_attribute *attr, char *buf)
1365 unsigned long nr_huge_pages;
1368 h = kobj_to_hstate(kobj, &nid);
1369 if (nid == NUMA_NO_NODE)
1370 nr_huge_pages = h->nr_huge_pages;
1372 nr_huge_pages = h->nr_huge_pages_node[nid];
1374 return sprintf(buf, "%lu\n", nr_huge_pages);
1376 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1377 struct kobject *kobj, struct kobj_attribute *attr,
1378 const char *buf, size_t len)
1382 unsigned long count;
1384 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1386 err = strict_strtoul(buf, 10, &count);
1390 h = kobj_to_hstate(kobj, &nid);
1391 if (nid == NUMA_NO_NODE) {
1393 * global hstate attribute
1395 if (!(obey_mempolicy &&
1396 init_nodemask_of_mempolicy(nodes_allowed))) {
1397 NODEMASK_FREE(nodes_allowed);
1398 nodes_allowed = &node_states[N_HIGH_MEMORY];
1400 } else if (nodes_allowed) {
1402 * per node hstate attribute: adjust count to global,
1403 * but restrict alloc/free to the specified node.
1405 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1406 init_nodemask_of_node(nodes_allowed, nid);
1408 nodes_allowed = &node_states[N_HIGH_MEMORY];
1410 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1412 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1413 NODEMASK_FREE(nodes_allowed);
1418 static ssize_t nr_hugepages_show(struct kobject *kobj,
1419 struct kobj_attribute *attr, char *buf)
1421 return nr_hugepages_show_common(kobj, attr, buf);
1424 static ssize_t nr_hugepages_store(struct kobject *kobj,
1425 struct kobj_attribute *attr, const char *buf, size_t len)
1427 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1429 HSTATE_ATTR(nr_hugepages);
1434 * hstate attribute for optionally mempolicy-based constraint on persistent
1435 * huge page alloc/free.
1437 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1438 struct kobj_attribute *attr, char *buf)
1440 return nr_hugepages_show_common(kobj, attr, buf);
1443 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1444 struct kobj_attribute *attr, const char *buf, size_t len)
1446 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1448 HSTATE_ATTR(nr_hugepages_mempolicy);
1452 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1453 struct kobj_attribute *attr, char *buf)
1455 struct hstate *h = kobj_to_hstate(kobj, NULL);
1456 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1458 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1459 struct kobj_attribute *attr, const char *buf, size_t count)
1462 unsigned long input;
1463 struct hstate *h = kobj_to_hstate(kobj, NULL);
1465 err = strict_strtoul(buf, 10, &input);
1469 spin_lock(&hugetlb_lock);
1470 h->nr_overcommit_huge_pages = input;
1471 spin_unlock(&hugetlb_lock);
1475 HSTATE_ATTR(nr_overcommit_hugepages);
1477 static ssize_t free_hugepages_show(struct kobject *kobj,
1478 struct kobj_attribute *attr, char *buf)
1481 unsigned long free_huge_pages;
1484 h = kobj_to_hstate(kobj, &nid);
1485 if (nid == NUMA_NO_NODE)
1486 free_huge_pages = h->free_huge_pages;
1488 free_huge_pages = h->free_huge_pages_node[nid];
1490 return sprintf(buf, "%lu\n", free_huge_pages);
1492 HSTATE_ATTR_RO(free_hugepages);
1494 static ssize_t resv_hugepages_show(struct kobject *kobj,
1495 struct kobj_attribute *attr, char *buf)
1497 struct hstate *h = kobj_to_hstate(kobj, NULL);
1498 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1500 HSTATE_ATTR_RO(resv_hugepages);
1502 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1503 struct kobj_attribute *attr, char *buf)
1506 unsigned long surplus_huge_pages;
1509 h = kobj_to_hstate(kobj, &nid);
1510 if (nid == NUMA_NO_NODE)
1511 surplus_huge_pages = h->surplus_huge_pages;
1513 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1515 return sprintf(buf, "%lu\n", surplus_huge_pages);
1517 HSTATE_ATTR_RO(surplus_hugepages);
1519 static struct attribute *hstate_attrs[] = {
1520 &nr_hugepages_attr.attr,
1521 &nr_overcommit_hugepages_attr.attr,
1522 &free_hugepages_attr.attr,
1523 &resv_hugepages_attr.attr,
1524 &surplus_hugepages_attr.attr,
1526 &nr_hugepages_mempolicy_attr.attr,
1531 static struct attribute_group hstate_attr_group = {
1532 .attrs = hstate_attrs,
1535 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1536 struct kobject **hstate_kobjs,
1537 struct attribute_group *hstate_attr_group)
1540 int hi = h - hstates;
1542 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1543 if (!hstate_kobjs[hi])
1546 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1548 kobject_put(hstate_kobjs[hi]);
1553 static void __init hugetlb_sysfs_init(void)
1558 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1559 if (!hugepages_kobj)
1562 for_each_hstate(h) {
1563 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1564 hstate_kobjs, &hstate_attr_group);
1566 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1574 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1575 * with node sysdevs in node_devices[] using a parallel array. The array
1576 * index of a node sysdev or _hstate == node id.
1577 * This is here to avoid any static dependency of the node sysdev driver, in
1578 * the base kernel, on the hugetlb module.
1580 struct node_hstate {
1581 struct kobject *hugepages_kobj;
1582 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1584 struct node_hstate node_hstates[MAX_NUMNODES];
1587 * A subset of global hstate attributes for node sysdevs
1589 static struct attribute *per_node_hstate_attrs[] = {
1590 &nr_hugepages_attr.attr,
1591 &free_hugepages_attr.attr,
1592 &surplus_hugepages_attr.attr,
1596 static struct attribute_group per_node_hstate_attr_group = {
1597 .attrs = per_node_hstate_attrs,
1601 * kobj_to_node_hstate - lookup global hstate for node sysdev hstate attr kobj.
1602 * Returns node id via non-NULL nidp.
1604 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1608 for (nid = 0; nid < nr_node_ids; nid++) {
1609 struct node_hstate *nhs = &node_hstates[nid];
1611 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1612 if (nhs->hstate_kobjs[i] == kobj) {
1624 * Unregister hstate attributes from a single node sysdev.
1625 * No-op if no hstate attributes attached.
1627 void hugetlb_unregister_node(struct node *node)
1630 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1632 if (!nhs->hugepages_kobj)
1633 return; /* no hstate attributes */
1636 if (nhs->hstate_kobjs[h - hstates]) {
1637 kobject_put(nhs->hstate_kobjs[h - hstates]);
1638 nhs->hstate_kobjs[h - hstates] = NULL;
1641 kobject_put(nhs->hugepages_kobj);
1642 nhs->hugepages_kobj = NULL;
1646 * hugetlb module exit: unregister hstate attributes from node sysdevs
1649 static void hugetlb_unregister_all_nodes(void)
1654 * disable node sysdev registrations.
1656 register_hugetlbfs_with_node(NULL, NULL);
1659 * remove hstate attributes from any nodes that have them.
1661 for (nid = 0; nid < nr_node_ids; nid++)
1662 hugetlb_unregister_node(&node_devices[nid]);
1666 * Register hstate attributes for a single node sysdev.
1667 * No-op if attributes already registered.
1669 void hugetlb_register_node(struct node *node)
1672 struct node_hstate *nhs = &node_hstates[node->sysdev.id];
1675 if (nhs->hugepages_kobj)
1676 return; /* already allocated */
1678 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1679 &node->sysdev.kobj);
1680 if (!nhs->hugepages_kobj)
1683 for_each_hstate(h) {
1684 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1686 &per_node_hstate_attr_group);
1688 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1690 h->name, node->sysdev.id);
1691 hugetlb_unregister_node(node);
1698 * hugetlb init time: register hstate attributes for all registered node
1699 * sysdevs of nodes that have memory. All on-line nodes should have
1700 * registered their associated sysdev by this time.
1702 static void hugetlb_register_all_nodes(void)
1706 for_each_node_state(nid, N_HIGH_MEMORY) {
1707 struct node *node = &node_devices[nid];
1708 if (node->sysdev.id == nid)
1709 hugetlb_register_node(node);
1713 * Let the node sysdev driver know we're here so it can
1714 * [un]register hstate attributes on node hotplug.
1716 register_hugetlbfs_with_node(hugetlb_register_node,
1717 hugetlb_unregister_node);
1719 #else /* !CONFIG_NUMA */
1721 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1729 static void hugetlb_unregister_all_nodes(void) { }
1731 static void hugetlb_register_all_nodes(void) { }
1735 static void __exit hugetlb_exit(void)
1739 hugetlb_unregister_all_nodes();
1741 for_each_hstate(h) {
1742 kobject_put(hstate_kobjs[h - hstates]);
1745 kobject_put(hugepages_kobj);
1747 module_exit(hugetlb_exit);
1749 static int __init hugetlb_init(void)
1751 /* Some platform decide whether they support huge pages at boot
1752 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1753 * there is no such support
1755 if (HPAGE_SHIFT == 0)
1758 if (!size_to_hstate(default_hstate_size)) {
1759 default_hstate_size = HPAGE_SIZE;
1760 if (!size_to_hstate(default_hstate_size))
1761 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1763 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1764 if (default_hstate_max_huge_pages)
1765 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1767 hugetlb_init_hstates();
1769 gather_bootmem_prealloc();
1773 hugetlb_sysfs_init();
1775 hugetlb_register_all_nodes();
1779 module_init(hugetlb_init);
1781 /* Should be called on processing a hugepagesz=... option */
1782 void __init hugetlb_add_hstate(unsigned order)
1787 if (size_to_hstate(PAGE_SIZE << order)) {
1788 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1791 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1793 h = &hstates[max_hstate++];
1795 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1796 h->nr_huge_pages = 0;
1797 h->free_huge_pages = 0;
1798 for (i = 0; i < MAX_NUMNODES; ++i)
1799 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1800 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1801 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1802 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1803 huge_page_size(h)/1024);
1808 static int __init hugetlb_nrpages_setup(char *s)
1811 static unsigned long *last_mhp;
1814 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1815 * so this hugepages= parameter goes to the "default hstate".
1818 mhp = &default_hstate_max_huge_pages;
1820 mhp = &parsed_hstate->max_huge_pages;
1822 if (mhp == last_mhp) {
1823 printk(KERN_WARNING "hugepages= specified twice without "
1824 "interleaving hugepagesz=, ignoring\n");
1828 if (sscanf(s, "%lu", mhp) <= 0)
1832 * Global state is always initialized later in hugetlb_init.
1833 * But we need to allocate >= MAX_ORDER hstates here early to still
1834 * use the bootmem allocator.
1836 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1837 hugetlb_hstate_alloc_pages(parsed_hstate);
1843 __setup("hugepages=", hugetlb_nrpages_setup);
1845 static int __init hugetlb_default_setup(char *s)
1847 default_hstate_size = memparse(s, &s);
1850 __setup("default_hugepagesz=", hugetlb_default_setup);
1852 static unsigned int cpuset_mems_nr(unsigned int *array)
1855 unsigned int nr = 0;
1857 for_each_node_mask(node, cpuset_current_mems_allowed)
1863 #ifdef CONFIG_SYSCTL
1864 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1865 struct ctl_table *table, int write,
1866 void __user *buffer, size_t *length, loff_t *ppos)
1868 struct hstate *h = &default_hstate;
1872 tmp = h->max_huge_pages;
1875 table->maxlen = sizeof(unsigned long);
1876 proc_doulongvec_minmax(table, write, buffer, length, ppos);
1879 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1880 GFP_KERNEL | __GFP_NORETRY);
1881 if (!(obey_mempolicy &&
1882 init_nodemask_of_mempolicy(nodes_allowed))) {
1883 NODEMASK_FREE(nodes_allowed);
1884 nodes_allowed = &node_states[N_HIGH_MEMORY];
1886 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1888 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1889 NODEMASK_FREE(nodes_allowed);
1895 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1896 void __user *buffer, size_t *length, loff_t *ppos)
1899 return hugetlb_sysctl_handler_common(false, table, write,
1900 buffer, length, ppos);
1904 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
1905 void __user *buffer, size_t *length, loff_t *ppos)
1907 return hugetlb_sysctl_handler_common(true, table, write,
1908 buffer, length, ppos);
1910 #endif /* CONFIG_NUMA */
1912 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1913 void __user *buffer,
1914 size_t *length, loff_t *ppos)
1916 proc_dointvec(table, write, buffer, length, ppos);
1917 if (hugepages_treat_as_movable)
1918 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1920 htlb_alloc_mask = GFP_HIGHUSER;
1924 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1925 void __user *buffer,
1926 size_t *length, loff_t *ppos)
1928 struct hstate *h = &default_hstate;
1932 tmp = h->nr_overcommit_huge_pages;
1935 table->maxlen = sizeof(unsigned long);
1936 proc_doulongvec_minmax(table, write, buffer, length, ppos);
1939 spin_lock(&hugetlb_lock);
1940 h->nr_overcommit_huge_pages = tmp;
1941 spin_unlock(&hugetlb_lock);
1947 #endif /* CONFIG_SYSCTL */
1949 void hugetlb_report_meminfo(struct seq_file *m)
1951 struct hstate *h = &default_hstate;
1953 "HugePages_Total: %5lu\n"
1954 "HugePages_Free: %5lu\n"
1955 "HugePages_Rsvd: %5lu\n"
1956 "HugePages_Surp: %5lu\n"
1957 "Hugepagesize: %8lu kB\n",
1961 h->surplus_huge_pages,
1962 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1965 int hugetlb_report_node_meminfo(int nid, char *buf)
1967 struct hstate *h = &default_hstate;
1969 "Node %d HugePages_Total: %5u\n"
1970 "Node %d HugePages_Free: %5u\n"
1971 "Node %d HugePages_Surp: %5u\n",
1972 nid, h->nr_huge_pages_node[nid],
1973 nid, h->free_huge_pages_node[nid],
1974 nid, h->surplus_huge_pages_node[nid]);
1977 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1978 unsigned long hugetlb_total_pages(void)
1980 struct hstate *h = &default_hstate;
1981 return h->nr_huge_pages * pages_per_huge_page(h);
1984 static int hugetlb_acct_memory(struct hstate *h, long delta)
1988 spin_lock(&hugetlb_lock);
1990 * When cpuset is configured, it breaks the strict hugetlb page
1991 * reservation as the accounting is done on a global variable. Such
1992 * reservation is completely rubbish in the presence of cpuset because
1993 * the reservation is not checked against page availability for the
1994 * current cpuset. Application can still potentially OOM'ed by kernel
1995 * with lack of free htlb page in cpuset that the task is in.
1996 * Attempt to enforce strict accounting with cpuset is almost
1997 * impossible (or too ugly) because cpuset is too fluid that
1998 * task or memory node can be dynamically moved between cpusets.
2000 * The change of semantics for shared hugetlb mapping with cpuset is
2001 * undesirable. However, in order to preserve some of the semantics,
2002 * we fall back to check against current free page availability as
2003 * a best attempt and hopefully to minimize the impact of changing
2004 * semantics that cpuset has.
2007 if (gather_surplus_pages(h, delta) < 0)
2010 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2011 return_unused_surplus_pages(h, delta);
2018 return_unused_surplus_pages(h, (unsigned long) -delta);
2021 spin_unlock(&hugetlb_lock);
2025 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2027 struct resv_map *reservations = vma_resv_map(vma);
2030 * This new VMA should share its siblings reservation map if present.
2031 * The VMA will only ever have a valid reservation map pointer where
2032 * it is being copied for another still existing VMA. As that VMA
2033 * has a reference to the reservation map it cannot dissappear until
2034 * after this open call completes. It is therefore safe to take a
2035 * new reference here without additional locking.
2038 kref_get(&reservations->refs);
2041 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2043 struct hstate *h = hstate_vma(vma);
2044 struct resv_map *reservations = vma_resv_map(vma);
2045 unsigned long reserve;
2046 unsigned long start;
2050 start = vma_hugecache_offset(h, vma, vma->vm_start);
2051 end = vma_hugecache_offset(h, vma, vma->vm_end);
2053 reserve = (end - start) -
2054 region_count(&reservations->regions, start, end);
2056 kref_put(&reservations->refs, resv_map_release);
2059 hugetlb_acct_memory(h, -reserve);
2060 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
2066 * We cannot handle pagefaults against hugetlb pages at all. They cause
2067 * handle_mm_fault() to try to instantiate regular-sized pages in the
2068 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2071 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2077 const struct vm_operations_struct hugetlb_vm_ops = {
2078 .fault = hugetlb_vm_op_fault,
2079 .open = hugetlb_vm_op_open,
2080 .close = hugetlb_vm_op_close,
2083 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2090 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2092 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2094 entry = pte_mkyoung(entry);
2095 entry = pte_mkhuge(entry);
2100 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2101 unsigned long address, pte_t *ptep)
2105 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2106 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
2107 update_mmu_cache(vma, address, ptep);
2112 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2113 struct vm_area_struct *vma)
2115 pte_t *src_pte, *dst_pte, entry;
2116 struct page *ptepage;
2119 struct hstate *h = hstate_vma(vma);
2120 unsigned long sz = huge_page_size(h);
2122 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2124 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2125 src_pte = huge_pte_offset(src, addr);
2128 dst_pte = huge_pte_alloc(dst, addr, sz);
2132 /* If the pagetables are shared don't copy or take references */
2133 if (dst_pte == src_pte)
2136 spin_lock(&dst->page_table_lock);
2137 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2138 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2140 huge_ptep_set_wrprotect(src, addr, src_pte);
2141 entry = huge_ptep_get(src_pte);
2142 ptepage = pte_page(entry);
2144 page_dup_rmap(ptepage);
2145 set_huge_pte_at(dst, addr, dst_pte, entry);
2147 spin_unlock(&src->page_table_lock);
2148 spin_unlock(&dst->page_table_lock);
2156 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2160 if (huge_pte_none(pte) || pte_present(pte))
2162 swp = pte_to_swp_entry(pte);
2163 if (non_swap_entry(swp) && is_hwpoison_entry(swp)) {
2169 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2170 unsigned long end, struct page *ref_page)
2172 struct mm_struct *mm = vma->vm_mm;
2173 unsigned long address;
2178 struct hstate *h = hstate_vma(vma);
2179 unsigned long sz = huge_page_size(h);
2182 * A page gathering list, protected by per file i_mmap_lock. The
2183 * lock is used to avoid list corruption from multiple unmapping
2184 * of the same page since we are using page->lru.
2186 LIST_HEAD(page_list);
2188 WARN_ON(!is_vm_hugetlb_page(vma));
2189 BUG_ON(start & ~huge_page_mask(h));
2190 BUG_ON(end & ~huge_page_mask(h));
2192 mmu_notifier_invalidate_range_start(mm, start, end);
2193 spin_lock(&mm->page_table_lock);
2194 for (address = start; address < end; address += sz) {
2195 ptep = huge_pte_offset(mm, address);
2199 if (huge_pmd_unshare(mm, &address, ptep))
2203 * If a reference page is supplied, it is because a specific
2204 * page is being unmapped, not a range. Ensure the page we
2205 * are about to unmap is the actual page of interest.
2208 pte = huge_ptep_get(ptep);
2209 if (huge_pte_none(pte))
2211 page = pte_page(pte);
2212 if (page != ref_page)
2216 * Mark the VMA as having unmapped its page so that
2217 * future faults in this VMA will fail rather than
2218 * looking like data was lost
2220 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2223 pte = huge_ptep_get_and_clear(mm, address, ptep);
2224 if (huge_pte_none(pte))
2228 * HWPoisoned hugepage is already unmapped and dropped reference
2230 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2233 page = pte_page(pte);
2235 set_page_dirty(page);
2236 list_add(&page->lru, &page_list);
2238 spin_unlock(&mm->page_table_lock);
2239 flush_tlb_range(vma, start, end);
2240 mmu_notifier_invalidate_range_end(mm, start, end);
2241 list_for_each_entry_safe(page, tmp, &page_list, lru) {
2242 page_remove_rmap(page);
2243 list_del(&page->lru);
2248 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2249 unsigned long end, struct page *ref_page)
2251 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2252 __unmap_hugepage_range(vma, start, end, ref_page);
2253 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2257 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2258 * mappping it owns the reserve page for. The intention is to unmap the page
2259 * from other VMAs and let the children be SIGKILLed if they are faulting the
2262 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2263 struct page *page, unsigned long address)
2265 struct hstate *h = hstate_vma(vma);
2266 struct vm_area_struct *iter_vma;
2267 struct address_space *mapping;
2268 struct prio_tree_iter iter;
2272 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2273 * from page cache lookup which is in HPAGE_SIZE units.
2275 address = address & huge_page_mask(h);
2276 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
2277 + (vma->vm_pgoff >> PAGE_SHIFT);
2278 mapping = (struct address_space *)page_private(page);
2281 * Take the mapping lock for the duration of the table walk. As
2282 * this mapping should be shared between all the VMAs,
2283 * __unmap_hugepage_range() is called as the lock is already held
2285 spin_lock(&mapping->i_mmap_lock);
2286 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2287 /* Do not unmap the current VMA */
2288 if (iter_vma == vma)
2292 * Unmap the page from other VMAs without their own reserves.
2293 * They get marked to be SIGKILLed if they fault in these
2294 * areas. This is because a future no-page fault on this VMA
2295 * could insert a zeroed page instead of the data existing
2296 * from the time of fork. This would look like data corruption
2298 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2299 __unmap_hugepage_range(iter_vma,
2300 address, address + huge_page_size(h),
2303 spin_unlock(&mapping->i_mmap_lock);
2309 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2311 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2312 unsigned long address, pte_t *ptep, pte_t pte,
2313 struct page *pagecache_page)
2315 struct hstate *h = hstate_vma(vma);
2316 struct page *old_page, *new_page;
2318 int outside_reserve = 0;
2320 old_page = pte_page(pte);
2323 /* If no-one else is actually using this page, avoid the copy
2324 * and just make the page writable */
2325 avoidcopy = (page_mapcount(old_page) == 1);
2327 if (PageAnon(old_page))
2328 page_move_anon_rmap(old_page, vma, address);
2329 set_huge_ptep_writable(vma, address, ptep);
2334 * If the process that created a MAP_PRIVATE mapping is about to
2335 * perform a COW due to a shared page count, attempt to satisfy
2336 * the allocation without using the existing reserves. The pagecache
2337 * page is used to determine if the reserve at this address was
2338 * consumed or not. If reserves were used, a partial faulted mapping
2339 * at the time of fork() could consume its reserves on COW instead
2340 * of the full address range.
2342 if (!(vma->vm_flags & VM_MAYSHARE) &&
2343 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2344 old_page != pagecache_page)
2345 outside_reserve = 1;
2347 page_cache_get(old_page);
2349 /* Drop page_table_lock as buddy allocator may be called */
2350 spin_unlock(&mm->page_table_lock);
2351 new_page = alloc_huge_page(vma, address, outside_reserve);
2353 if (IS_ERR(new_page)) {
2354 page_cache_release(old_page);
2357 * If a process owning a MAP_PRIVATE mapping fails to COW,
2358 * it is due to references held by a child and an insufficient
2359 * huge page pool. To guarantee the original mappers
2360 * reliability, unmap the page from child processes. The child
2361 * may get SIGKILLed if it later faults.
2363 if (outside_reserve) {
2364 BUG_ON(huge_pte_none(pte));
2365 if (unmap_ref_private(mm, vma, old_page, address)) {
2366 BUG_ON(page_count(old_page) != 1);
2367 BUG_ON(huge_pte_none(pte));
2368 spin_lock(&mm->page_table_lock);
2369 goto retry_avoidcopy;
2374 /* Caller expects lock to be held */
2375 spin_lock(&mm->page_table_lock);
2376 return -PTR_ERR(new_page);
2380 * When the original hugepage is shared one, it does not have
2381 * anon_vma prepared.
2383 if (unlikely(anon_vma_prepare(vma)))
2384 return VM_FAULT_OOM;
2386 copy_huge_page(new_page, old_page, address, vma);
2387 __SetPageUptodate(new_page);
2390 * Retake the page_table_lock to check for racing updates
2391 * before the page tables are altered
2393 spin_lock(&mm->page_table_lock);
2394 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2395 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2397 mmu_notifier_invalidate_range_start(mm,
2398 address & huge_page_mask(h),
2399 (address & huge_page_mask(h)) + huge_page_size(h));
2400 huge_ptep_clear_flush(vma, address, ptep);
2401 set_huge_pte_at(mm, address, ptep,
2402 make_huge_pte(vma, new_page, 1));
2403 page_remove_rmap(old_page);
2404 hugepage_add_new_anon_rmap(new_page, vma, address);
2405 /* Make the old page be freed below */
2406 new_page = old_page;
2407 mmu_notifier_invalidate_range_end(mm,
2408 address & huge_page_mask(h),
2409 (address & huge_page_mask(h)) + huge_page_size(h));
2411 page_cache_release(new_page);
2412 page_cache_release(old_page);
2416 /* Return the pagecache page at a given address within a VMA */
2417 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2418 struct vm_area_struct *vma, unsigned long address)
2420 struct address_space *mapping;
2423 mapping = vma->vm_file->f_mapping;
2424 idx = vma_hugecache_offset(h, vma, address);
2426 return find_lock_page(mapping, idx);
2430 * Return whether there is a pagecache page to back given address within VMA.
2431 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2433 static bool hugetlbfs_pagecache_present(struct hstate *h,
2434 struct vm_area_struct *vma, unsigned long address)
2436 struct address_space *mapping;
2440 mapping = vma->vm_file->f_mapping;
2441 idx = vma_hugecache_offset(h, vma, address);
2443 page = find_get_page(mapping, idx);
2446 return page != NULL;
2449 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2450 unsigned long address, pte_t *ptep, unsigned int flags)
2452 struct hstate *h = hstate_vma(vma);
2453 int ret = VM_FAULT_SIGBUS;
2457 struct address_space *mapping;
2461 * Currently, we are forced to kill the process in the event the
2462 * original mapper has unmapped pages from the child due to a failed
2463 * COW. Warn that such a situation has occured as it may not be obvious
2465 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2467 "PID %d killed due to inadequate hugepage pool\n",
2472 mapping = vma->vm_file->f_mapping;
2473 idx = vma_hugecache_offset(h, vma, address);
2476 * Use page lock to guard against racing truncation
2477 * before we get page_table_lock.
2480 page = find_lock_page(mapping, idx);
2482 size = i_size_read(mapping->host) >> huge_page_shift(h);
2485 page = alloc_huge_page(vma, address, 0);
2487 ret = -PTR_ERR(page);
2490 clear_huge_page(page, address, huge_page_size(h));
2491 __SetPageUptodate(page);
2493 if (vma->vm_flags & VM_MAYSHARE) {
2495 struct inode *inode = mapping->host;
2497 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2505 spin_lock(&inode->i_lock);
2506 inode->i_blocks += blocks_per_huge_page(h);
2507 spin_unlock(&inode->i_lock);
2508 page_dup_rmap(page);
2511 if (unlikely(anon_vma_prepare(vma))) {
2513 goto backout_unlocked;
2515 hugepage_add_new_anon_rmap(page, vma, address);
2518 page_dup_rmap(page);
2522 * Since memory error handler replaces pte into hwpoison swap entry
2523 * at the time of error handling, a process which reserved but not have
2524 * the mapping to the error hugepage does not have hwpoison swap entry.
2525 * So we need to block accesses from such a process by checking
2526 * PG_hwpoison bit here.
2528 if (unlikely(PageHWPoison(page))) {
2529 ret = VM_FAULT_HWPOISON;
2530 goto backout_unlocked;
2534 * If we are going to COW a private mapping later, we examine the
2535 * pending reservations for this page now. This will ensure that
2536 * any allocations necessary to record that reservation occur outside
2539 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2540 if (vma_needs_reservation(h, vma, address) < 0) {
2542 goto backout_unlocked;
2545 spin_lock(&mm->page_table_lock);
2546 size = i_size_read(mapping->host) >> huge_page_shift(h);
2551 if (!huge_pte_none(huge_ptep_get(ptep)))
2554 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2555 && (vma->vm_flags & VM_SHARED)));
2556 set_huge_pte_at(mm, address, ptep, new_pte);
2558 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2559 /* Optimization, do the COW without a second fault */
2560 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2563 spin_unlock(&mm->page_table_lock);
2569 spin_unlock(&mm->page_table_lock);
2576 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2577 unsigned long address, unsigned int flags)
2582 struct page *page = NULL;
2583 struct page *pagecache_page = NULL;
2584 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2585 struct hstate *h = hstate_vma(vma);
2587 ptep = huge_pte_offset(mm, address);
2589 entry = huge_ptep_get(ptep);
2590 if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2591 return VM_FAULT_HWPOISON;
2594 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2596 return VM_FAULT_OOM;
2599 * Serialize hugepage allocation and instantiation, so that we don't
2600 * get spurious allocation failures if two CPUs race to instantiate
2601 * the same page in the page cache.
2603 mutex_lock(&hugetlb_instantiation_mutex);
2604 entry = huge_ptep_get(ptep);
2605 if (huge_pte_none(entry)) {
2606 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2613 * If we are going to COW the mapping later, we examine the pending
2614 * reservations for this page now. This will ensure that any
2615 * allocations necessary to record that reservation occur outside the
2616 * spinlock. For private mappings, we also lookup the pagecache
2617 * page now as it is used to determine if a reservation has been
2620 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2621 if (vma_needs_reservation(h, vma, address) < 0) {
2626 if (!(vma->vm_flags & VM_MAYSHARE))
2627 pagecache_page = hugetlbfs_pagecache_page(h,
2632 * hugetlb_cow() requires page locks of pte_page(entry) and
2633 * pagecache_page, so here we need take the former one
2634 * when page != pagecache_page or !pagecache_page.
2635 * Note that locking order is always pagecache_page -> page,
2636 * so no worry about deadlock.
2638 page = pte_page(entry);
2639 if (page != pagecache_page)
2642 spin_lock(&mm->page_table_lock);
2643 /* Check for a racing update before calling hugetlb_cow */
2644 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2645 goto out_page_table_lock;
2648 if (flags & FAULT_FLAG_WRITE) {
2649 if (!pte_write(entry)) {
2650 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2652 goto out_page_table_lock;
2654 entry = pte_mkdirty(entry);
2656 entry = pte_mkyoung(entry);
2657 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2658 flags & FAULT_FLAG_WRITE))
2659 update_mmu_cache(vma, address, ptep);
2661 out_page_table_lock:
2662 spin_unlock(&mm->page_table_lock);
2664 if (pagecache_page) {
2665 unlock_page(pagecache_page);
2666 put_page(pagecache_page);
2671 mutex_unlock(&hugetlb_instantiation_mutex);
2676 /* Can be overriden by architectures */
2677 __attribute__((weak)) struct page *
2678 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2679 pud_t *pud, int write)
2685 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2686 struct page **pages, struct vm_area_struct **vmas,
2687 unsigned long *position, int *length, int i,
2690 unsigned long pfn_offset;
2691 unsigned long vaddr = *position;
2692 int remainder = *length;
2693 struct hstate *h = hstate_vma(vma);
2695 spin_lock(&mm->page_table_lock);
2696 while (vaddr < vma->vm_end && remainder) {
2702 * Some archs (sparc64, sh*) have multiple pte_ts to
2703 * each hugepage. We have to make sure we get the
2704 * first, for the page indexing below to work.
2706 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2707 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2710 * When coredumping, it suits get_dump_page if we just return
2711 * an error where there's an empty slot with no huge pagecache
2712 * to back it. This way, we avoid allocating a hugepage, and
2713 * the sparse dumpfile avoids allocating disk blocks, but its
2714 * huge holes still show up with zeroes where they need to be.
2716 if (absent && (flags & FOLL_DUMP) &&
2717 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2723 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2726 spin_unlock(&mm->page_table_lock);
2727 ret = hugetlb_fault(mm, vma, vaddr,
2728 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2729 spin_lock(&mm->page_table_lock);
2730 if (!(ret & VM_FAULT_ERROR))
2737 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2738 page = pte_page(huge_ptep_get(pte));
2741 pages[i] = mem_map_offset(page, pfn_offset);
2752 if (vaddr < vma->vm_end && remainder &&
2753 pfn_offset < pages_per_huge_page(h)) {
2755 * We use pfn_offset to avoid touching the pageframes
2756 * of this compound page.
2761 spin_unlock(&mm->page_table_lock);
2762 *length = remainder;
2765 return i ? i : -EFAULT;
2768 void hugetlb_change_protection(struct vm_area_struct *vma,
2769 unsigned long address, unsigned long end, pgprot_t newprot)
2771 struct mm_struct *mm = vma->vm_mm;
2772 unsigned long start = address;
2775 struct hstate *h = hstate_vma(vma);
2777 BUG_ON(address >= end);
2778 flush_cache_range(vma, address, end);
2780 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2781 spin_lock(&mm->page_table_lock);
2782 for (; address < end; address += huge_page_size(h)) {
2783 ptep = huge_pte_offset(mm, address);
2786 if (huge_pmd_unshare(mm, &address, ptep))
2788 if (!huge_pte_none(huge_ptep_get(ptep))) {
2789 pte = huge_ptep_get_and_clear(mm, address, ptep);
2790 pte = pte_mkhuge(pte_modify(pte, newprot));
2791 set_huge_pte_at(mm, address, ptep, pte);
2794 spin_unlock(&mm->page_table_lock);
2795 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2797 flush_tlb_range(vma, start, end);
2800 int hugetlb_reserve_pages(struct inode *inode,
2802 struct vm_area_struct *vma,
2806 struct hstate *h = hstate_inode(inode);
2809 * Only apply hugepage reservation if asked. At fault time, an
2810 * attempt will be made for VM_NORESERVE to allocate a page
2811 * and filesystem quota without using reserves
2813 if (acctflag & VM_NORESERVE)
2817 * Shared mappings base their reservation on the number of pages that
2818 * are already allocated on behalf of the file. Private mappings need
2819 * to reserve the full area even if read-only as mprotect() may be
2820 * called to make the mapping read-write. Assume !vma is a shm mapping
2822 if (!vma || vma->vm_flags & VM_MAYSHARE)
2823 chg = region_chg(&inode->i_mapping->private_list, from, to);
2825 struct resv_map *resv_map = resv_map_alloc();
2831 set_vma_resv_map(vma, resv_map);
2832 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2838 /* There must be enough filesystem quota for the mapping */
2839 if (hugetlb_get_quota(inode->i_mapping, chg))
2843 * Check enough hugepages are available for the reservation.
2844 * Hand back the quota if there are not
2846 ret = hugetlb_acct_memory(h, chg);
2848 hugetlb_put_quota(inode->i_mapping, chg);
2853 * Account for the reservations made. Shared mappings record regions
2854 * that have reservations as they are shared by multiple VMAs.
2855 * When the last VMA disappears, the region map says how much
2856 * the reservation was and the page cache tells how much of
2857 * the reservation was consumed. Private mappings are per-VMA and
2858 * only the consumed reservations are tracked. When the VMA
2859 * disappears, the original reservation is the VMA size and the
2860 * consumed reservations are stored in the map. Hence, nothing
2861 * else has to be done for private mappings here
2863 if (!vma || vma->vm_flags & VM_MAYSHARE)
2864 region_add(&inode->i_mapping->private_list, from, to);
2868 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2870 struct hstate *h = hstate_inode(inode);
2871 long chg = region_truncate(&inode->i_mapping->private_list, offset);
2873 spin_lock(&inode->i_lock);
2874 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
2875 spin_unlock(&inode->i_lock);
2877 hugetlb_put_quota(inode->i_mapping, (chg - freed));
2878 hugetlb_acct_memory(h, -(chg - freed));
2882 * This function is called from memory failure code.
2883 * Assume the caller holds page lock of the head page.
2885 void __isolate_hwpoisoned_huge_page(struct page *hpage)
2887 struct hstate *h = page_hstate(hpage);
2888 int nid = page_to_nid(hpage);
2890 spin_lock(&hugetlb_lock);
2891 list_del(&hpage->lru);
2892 h->free_huge_pages--;
2893 h->free_huge_pages_node[nid]--;
2894 spin_unlock(&hugetlb_lock);