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)
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 |
514 1 << PG_referenced | 1 << PG_dirty |
515 1 << PG_active | 1 << PG_reserved |
516 1 << PG_private | 1 << PG_writeback);
518 set_compound_page_dtor(page, NULL);
519 set_page_refcounted(page);
520 arch_release_hugepage(page);
521 __free_pages(page, huge_page_order(h));
524 struct hstate *size_to_hstate(unsigned long size)
529 if (huge_page_size(h) == size)
535 static void free_huge_page(struct page *page)
538 * Can't pass hstate in here because it is called from the
539 * compound page destructor.
541 struct hstate *h = page_hstate(page);
542 int nid = page_to_nid(page);
543 struct address_space *mapping;
545 mapping = (struct address_space *) page_private(page);
546 set_page_private(page, 0);
547 page->mapping = NULL;
548 BUG_ON(page_count(page));
549 BUG_ON(page_mapcount(page));
550 INIT_LIST_HEAD(&page->lru);
552 spin_lock(&hugetlb_lock);
553 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
554 update_and_free_page(h, page);
555 h->surplus_huge_pages--;
556 h->surplus_huge_pages_node[nid]--;
558 enqueue_huge_page(h, page);
560 spin_unlock(&hugetlb_lock);
562 hugetlb_put_quota(mapping, 1);
565 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
567 set_compound_page_dtor(page, free_huge_page);
568 spin_lock(&hugetlb_lock);
570 h->nr_huge_pages_node[nid]++;
571 spin_unlock(&hugetlb_lock);
572 put_page(page); /* free it into the hugepage allocator */
575 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
578 int nr_pages = 1 << order;
579 struct page *p = page + 1;
581 /* we rely on prep_new_huge_page to set the destructor */
582 set_compound_order(page, order);
584 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
586 set_page_count(p, 0);
587 p->first_page = page;
591 int PageHuge(struct page *page)
593 compound_page_dtor *dtor;
595 if (!PageCompound(page))
598 page = compound_head(page);
599 dtor = get_compound_page_dtor(page);
601 return dtor == free_huge_page;
603 EXPORT_SYMBOL_GPL(PageHuge);
605 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
609 if (h->order >= MAX_ORDER)
612 page = alloc_pages_exact_node(nid,
613 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
614 __GFP_REPEAT|__GFP_NOWARN,
617 if (arch_prepare_hugepage(page)) {
618 __free_pages(page, huge_page_order(h));
621 prep_new_huge_page(h, page, nid);
628 * common helper functions for hstate_next_node_to_{alloc|free}.
629 * We may have allocated or freed a huge page based on a different
630 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
631 * be outside of *nodes_allowed. Ensure that we use an allowed
632 * node for alloc or free.
634 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
636 nid = next_node(nid, *nodes_allowed);
637 if (nid == MAX_NUMNODES)
638 nid = first_node(*nodes_allowed);
639 VM_BUG_ON(nid >= MAX_NUMNODES);
644 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
646 if (!node_isset(nid, *nodes_allowed))
647 nid = next_node_allowed(nid, nodes_allowed);
652 * returns the previously saved node ["this node"] from which to
653 * allocate a persistent huge page for the pool and advance the
654 * next node from which to allocate, handling wrap at end of node
657 static int hstate_next_node_to_alloc(struct hstate *h,
658 nodemask_t *nodes_allowed)
662 VM_BUG_ON(!nodes_allowed);
664 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
665 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
670 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
677 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
678 next_nid = start_nid;
681 page = alloc_fresh_huge_page_node(h, next_nid);
686 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
687 } while (next_nid != start_nid);
690 count_vm_event(HTLB_BUDDY_PGALLOC);
692 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
698 * helper for free_pool_huge_page() - return the previously saved
699 * node ["this node"] from which to free a huge page. Advance the
700 * next node id whether or not we find a free huge page to free so
701 * that the next attempt to free addresses the next node.
703 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
707 VM_BUG_ON(!nodes_allowed);
709 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
710 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
716 * Free huge page from pool from next node to free.
717 * Attempt to keep persistent huge pages more or less
718 * balanced over allowed nodes.
719 * Called with hugetlb_lock locked.
721 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
728 start_nid = hstate_next_node_to_free(h, nodes_allowed);
729 next_nid = start_nid;
733 * If we're returning unused surplus pages, only examine
734 * nodes with surplus pages.
736 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
737 !list_empty(&h->hugepage_freelists[next_nid])) {
739 list_entry(h->hugepage_freelists[next_nid].next,
741 list_del(&page->lru);
742 h->free_huge_pages--;
743 h->free_huge_pages_node[next_nid]--;
745 h->surplus_huge_pages--;
746 h->surplus_huge_pages_node[next_nid]--;
748 update_and_free_page(h, page);
752 next_nid = hstate_next_node_to_free(h, nodes_allowed);
753 } while (next_nid != start_nid);
758 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
763 if (h->order >= MAX_ORDER)
767 * Assume we will successfully allocate the surplus page to
768 * prevent racing processes from causing the surplus to exceed
771 * This however introduces a different race, where a process B
772 * tries to grow the static hugepage pool while alloc_pages() is
773 * called by process A. B will only examine the per-node
774 * counters in determining if surplus huge pages can be
775 * converted to normal huge pages in adjust_pool_surplus(). A
776 * won't be able to increment the per-node counter, until the
777 * lock is dropped by B, but B doesn't drop hugetlb_lock until
778 * no more huge pages can be converted from surplus to normal
779 * state (and doesn't try to convert again). Thus, we have a
780 * case where a surplus huge page exists, the pool is grown, and
781 * the surplus huge page still exists after, even though it
782 * should just have been converted to a normal huge page. This
783 * does not leak memory, though, as the hugepage will be freed
784 * once it is out of use. It also does not allow the counters to
785 * go out of whack in adjust_pool_surplus() as we don't modify
786 * the node values until we've gotten the hugepage and only the
787 * per-node value is checked there.
789 spin_lock(&hugetlb_lock);
790 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
791 spin_unlock(&hugetlb_lock);
795 h->surplus_huge_pages++;
797 spin_unlock(&hugetlb_lock);
799 if (nid == NUMA_NO_NODE)
800 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
801 __GFP_REPEAT|__GFP_NOWARN,
804 page = alloc_pages_exact_node(nid,
805 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
806 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
808 if (page && arch_prepare_hugepage(page)) {
809 __free_pages(page, huge_page_order(h));
813 spin_lock(&hugetlb_lock);
815 r_nid = page_to_nid(page);
816 set_compound_page_dtor(page, free_huge_page);
818 * We incremented the global counters already
820 h->nr_huge_pages_node[r_nid]++;
821 h->surplus_huge_pages_node[r_nid]++;
822 __count_vm_event(HTLB_BUDDY_PGALLOC);
825 h->surplus_huge_pages--;
826 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
828 spin_unlock(&hugetlb_lock);
834 * This allocation function is useful in the context where vma is irrelevant.
835 * E.g. soft-offlining uses this function because it only cares physical
836 * address of error page.
838 struct page *alloc_huge_page_node(struct hstate *h, int nid)
842 spin_lock(&hugetlb_lock);
843 page = dequeue_huge_page_node(h, nid);
844 spin_unlock(&hugetlb_lock);
847 page = alloc_buddy_huge_page(h, nid);
853 * Increase the hugetlb pool such that it can accommodate a reservation
856 static int gather_surplus_pages(struct hstate *h, int delta)
858 struct list_head surplus_list;
859 struct page *page, *tmp;
861 int needed, allocated;
862 bool alloc_ok = true;
864 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
866 h->resv_huge_pages += delta;
871 INIT_LIST_HEAD(&surplus_list);
875 spin_unlock(&hugetlb_lock);
876 for (i = 0; i < needed; i++) {
877 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
882 list_add(&page->lru, &surplus_list);
887 * After retaking hugetlb_lock, we need to recalculate 'needed'
888 * because either resv_huge_pages or free_huge_pages may have changed.
890 spin_lock(&hugetlb_lock);
891 needed = (h->resv_huge_pages + delta) -
892 (h->free_huge_pages + allocated);
897 * We were not able to allocate enough pages to
898 * satisfy the entire reservation so we free what
899 * we've allocated so far.
904 * The surplus_list now contains _at_least_ the number of extra pages
905 * needed to accommodate the reservation. Add the appropriate number
906 * of pages to the hugetlb pool and free the extras back to the buddy
907 * allocator. Commit the entire reservation here to prevent another
908 * process from stealing the pages as they are added to the pool but
909 * before they are reserved.
912 h->resv_huge_pages += delta;
915 /* Free the needed pages to the hugetlb pool */
916 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
919 list_del(&page->lru);
921 * This page is now managed by the hugetlb allocator and has
922 * no users -- drop the buddy allocator's reference.
924 put_page_testzero(page);
925 VM_BUG_ON(page_count(page));
926 enqueue_huge_page(h, page);
929 spin_unlock(&hugetlb_lock);
931 /* Free unnecessary surplus pages to the buddy allocator */
932 if (!list_empty(&surplus_list)) {
933 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
934 list_del(&page->lru);
938 spin_lock(&hugetlb_lock);
944 * When releasing a hugetlb pool reservation, any surplus pages that were
945 * allocated to satisfy the reservation must be explicitly freed if they were
947 * Called with hugetlb_lock held.
949 static void return_unused_surplus_pages(struct hstate *h,
950 unsigned long unused_resv_pages)
952 unsigned long nr_pages;
954 /* Uncommit the reservation */
955 h->resv_huge_pages -= unused_resv_pages;
957 /* Cannot return gigantic pages currently */
958 if (h->order >= MAX_ORDER)
961 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
964 * We want to release as many surplus pages as possible, spread
965 * evenly across all nodes with memory. Iterate across these nodes
966 * until we can no longer free unreserved surplus pages. This occurs
967 * when the nodes with surplus pages have no free pages.
968 * free_pool_huge_page() will balance the the freed pages across the
969 * on-line nodes with memory and will handle the hstate accounting.
972 if (!free_pool_huge_page(h, &node_states[N_HIGH_MEMORY], 1))
978 * Determine if the huge page at addr within the vma has an associated
979 * reservation. Where it does not we will need to logically increase
980 * reservation and actually increase quota before an allocation can occur.
981 * Where any new reservation would be required the reservation change is
982 * prepared, but not committed. Once the page has been quota'd allocated
983 * an instantiated the change should be committed via vma_commit_reservation.
984 * No action is required on failure.
986 static long vma_needs_reservation(struct hstate *h,
987 struct vm_area_struct *vma, unsigned long addr)
989 struct address_space *mapping = vma->vm_file->f_mapping;
990 struct inode *inode = mapping->host;
992 if (vma->vm_flags & VM_MAYSHARE) {
993 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
994 return region_chg(&inode->i_mapping->private_list,
997 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1002 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1003 struct resv_map *reservations = vma_resv_map(vma);
1005 err = region_chg(&reservations->regions, idx, idx + 1);
1011 static void vma_commit_reservation(struct hstate *h,
1012 struct vm_area_struct *vma, unsigned long addr)
1014 struct address_space *mapping = vma->vm_file->f_mapping;
1015 struct inode *inode = mapping->host;
1017 if (vma->vm_flags & VM_MAYSHARE) {
1018 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1019 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1021 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1022 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1023 struct resv_map *reservations = vma_resv_map(vma);
1025 /* Mark this page used in the map. */
1026 region_add(&reservations->regions, idx, idx + 1);
1030 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1031 unsigned long addr, int avoid_reserve)
1033 struct hstate *h = hstate_vma(vma);
1035 struct address_space *mapping = vma->vm_file->f_mapping;
1036 struct inode *inode = mapping->host;
1040 * Processes that did not create the mapping will have no reserves and
1041 * will not have accounted against quota. Check that the quota can be
1042 * made before satisfying the allocation
1043 * MAP_NORESERVE mappings may also need pages and quota allocated
1044 * if no reserve mapping overlaps.
1046 chg = vma_needs_reservation(h, vma, addr);
1048 return ERR_PTR(-VM_FAULT_OOM);
1050 if (hugetlb_get_quota(inode->i_mapping, chg))
1051 return ERR_PTR(-VM_FAULT_SIGBUS);
1053 spin_lock(&hugetlb_lock);
1054 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1055 spin_unlock(&hugetlb_lock);
1058 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1060 hugetlb_put_quota(inode->i_mapping, chg);
1061 return ERR_PTR(-VM_FAULT_SIGBUS);
1065 set_page_private(page, (unsigned long) mapping);
1067 vma_commit_reservation(h, vma, addr);
1072 int __weak alloc_bootmem_huge_page(struct hstate *h)
1074 struct huge_bootmem_page *m;
1075 int nr_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
1080 addr = __alloc_bootmem_node_nopanic(
1081 NODE_DATA(hstate_next_node_to_alloc(h,
1082 &node_states[N_HIGH_MEMORY])),
1083 huge_page_size(h), huge_page_size(h), 0);
1087 * Use the beginning of the huge page to store the
1088 * huge_bootmem_page struct (until gather_bootmem
1089 * puts them into the mem_map).
1099 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1100 /* Put them into a private list first because mem_map is not up yet */
1101 list_add(&m->list, &huge_boot_pages);
1106 static void prep_compound_huge_page(struct page *page, int order)
1108 if (unlikely(order > (MAX_ORDER - 1)))
1109 prep_compound_gigantic_page(page, order);
1111 prep_compound_page(page, order);
1114 /* Put bootmem huge pages into the standard lists after mem_map is up */
1115 static void __init gather_bootmem_prealloc(void)
1117 struct huge_bootmem_page *m;
1119 list_for_each_entry(m, &huge_boot_pages, list) {
1120 struct hstate *h = m->hstate;
1123 #ifdef CONFIG_HIGHMEM
1124 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1125 free_bootmem_late((unsigned long)m,
1126 sizeof(struct huge_bootmem_page));
1128 page = virt_to_page(m);
1130 __ClearPageReserved(page);
1131 WARN_ON(page_count(page) != 1);
1132 prep_compound_huge_page(page, h->order);
1133 prep_new_huge_page(h, page, page_to_nid(page));
1135 * If we had gigantic hugepages allocated at boot time, we need
1136 * to restore the 'stolen' pages to totalram_pages in order to
1137 * fix confusing memory reports from free(1) and another
1138 * side-effects, like CommitLimit going negative.
1140 if (h->order > (MAX_ORDER - 1))
1141 totalram_pages += 1 << h->order;
1145 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1149 for (i = 0; i < h->max_huge_pages; ++i) {
1150 if (h->order >= MAX_ORDER) {
1151 if (!alloc_bootmem_huge_page(h))
1153 } else if (!alloc_fresh_huge_page(h,
1154 &node_states[N_HIGH_MEMORY]))
1157 h->max_huge_pages = i;
1160 static void __init hugetlb_init_hstates(void)
1164 for_each_hstate(h) {
1165 /* oversize hugepages were init'ed in early boot */
1166 if (h->order < MAX_ORDER)
1167 hugetlb_hstate_alloc_pages(h);
1171 static char * __init memfmt(char *buf, unsigned long n)
1173 if (n >= (1UL << 30))
1174 sprintf(buf, "%lu GB", n >> 30);
1175 else if (n >= (1UL << 20))
1176 sprintf(buf, "%lu MB", n >> 20);
1178 sprintf(buf, "%lu KB", n >> 10);
1182 static void __init report_hugepages(void)
1186 for_each_hstate(h) {
1188 printk(KERN_INFO "HugeTLB registered %s page size, "
1189 "pre-allocated %ld pages\n",
1190 memfmt(buf, huge_page_size(h)),
1191 h->free_huge_pages);
1195 #ifdef CONFIG_HIGHMEM
1196 static void try_to_free_low(struct hstate *h, unsigned long count,
1197 nodemask_t *nodes_allowed)
1201 if (h->order >= MAX_ORDER)
1204 for_each_node_mask(i, *nodes_allowed) {
1205 struct page *page, *next;
1206 struct list_head *freel = &h->hugepage_freelists[i];
1207 list_for_each_entry_safe(page, next, freel, lru) {
1208 if (count >= h->nr_huge_pages)
1210 if (PageHighMem(page))
1212 list_del(&page->lru);
1213 update_and_free_page(h, page);
1214 h->free_huge_pages--;
1215 h->free_huge_pages_node[page_to_nid(page)]--;
1220 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1221 nodemask_t *nodes_allowed)
1227 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1228 * balanced by operating on them in a round-robin fashion.
1229 * Returns 1 if an adjustment was made.
1231 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1234 int start_nid, next_nid;
1237 VM_BUG_ON(delta != -1 && delta != 1);
1240 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1242 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1243 next_nid = start_nid;
1249 * To shrink on this node, there must be a surplus page
1251 if (!h->surplus_huge_pages_node[nid]) {
1252 next_nid = hstate_next_node_to_alloc(h,
1259 * Surplus cannot exceed the total number of pages
1261 if (h->surplus_huge_pages_node[nid] >=
1262 h->nr_huge_pages_node[nid]) {
1263 next_nid = hstate_next_node_to_free(h,
1269 h->surplus_huge_pages += delta;
1270 h->surplus_huge_pages_node[nid] += delta;
1273 } while (next_nid != start_nid);
1278 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1279 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1280 nodemask_t *nodes_allowed)
1282 unsigned long min_count, ret;
1284 if (h->order >= MAX_ORDER)
1285 return h->max_huge_pages;
1288 * Increase the pool size
1289 * First take pages out of surplus state. Then make up the
1290 * remaining difference by allocating fresh huge pages.
1292 * We might race with alloc_buddy_huge_page() here and be unable
1293 * to convert a surplus huge page to a normal huge page. That is
1294 * not critical, though, it just means the overall size of the
1295 * pool might be one hugepage larger than it needs to be, but
1296 * within all the constraints specified by the sysctls.
1298 spin_lock(&hugetlb_lock);
1299 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1300 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1304 while (count > persistent_huge_pages(h)) {
1306 * If this allocation races such that we no longer need the
1307 * page, free_huge_page will handle it by freeing the page
1308 * and reducing the surplus.
1310 spin_unlock(&hugetlb_lock);
1311 ret = alloc_fresh_huge_page(h, nodes_allowed);
1312 spin_lock(&hugetlb_lock);
1316 /* Bail for signals. Probably ctrl-c from user */
1317 if (signal_pending(current))
1322 * Decrease the pool size
1323 * First return free pages to the buddy allocator (being careful
1324 * to keep enough around to satisfy reservations). Then place
1325 * pages into surplus state as needed so the pool will shrink
1326 * to the desired size as pages become free.
1328 * By placing pages into the surplus state independent of the
1329 * overcommit value, we are allowing the surplus pool size to
1330 * exceed overcommit. There are few sane options here. Since
1331 * alloc_buddy_huge_page() is checking the global counter,
1332 * though, we'll note that we're not allowed to exceed surplus
1333 * and won't grow the pool anywhere else. Not until one of the
1334 * sysctls are changed, or the surplus pages go out of use.
1336 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1337 min_count = max(count, min_count);
1338 try_to_free_low(h, min_count, nodes_allowed);
1339 while (min_count < persistent_huge_pages(h)) {
1340 if (!free_pool_huge_page(h, nodes_allowed, 0))
1343 while (count < persistent_huge_pages(h)) {
1344 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1348 ret = persistent_huge_pages(h);
1349 spin_unlock(&hugetlb_lock);
1353 #define HSTATE_ATTR_RO(_name) \
1354 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1356 #define HSTATE_ATTR(_name) \
1357 static struct kobj_attribute _name##_attr = \
1358 __ATTR(_name, 0644, _name##_show, _name##_store)
1360 static struct kobject *hugepages_kobj;
1361 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1363 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1365 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1369 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1370 if (hstate_kobjs[i] == kobj) {
1372 *nidp = NUMA_NO_NODE;
1376 return kobj_to_node_hstate(kobj, nidp);
1379 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1380 struct kobj_attribute *attr, char *buf)
1383 unsigned long nr_huge_pages;
1386 h = kobj_to_hstate(kobj, &nid);
1387 if (nid == NUMA_NO_NODE)
1388 nr_huge_pages = h->nr_huge_pages;
1390 nr_huge_pages = h->nr_huge_pages_node[nid];
1392 return sprintf(buf, "%lu\n", nr_huge_pages);
1395 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1396 struct kobject *kobj, struct kobj_attribute *attr,
1397 const char *buf, size_t len)
1401 unsigned long count;
1403 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1405 err = strict_strtoul(buf, 10, &count);
1409 h = kobj_to_hstate(kobj, &nid);
1410 if (h->order >= MAX_ORDER) {
1415 if (nid == NUMA_NO_NODE) {
1417 * global hstate attribute
1419 if (!(obey_mempolicy &&
1420 init_nodemask_of_mempolicy(nodes_allowed))) {
1421 NODEMASK_FREE(nodes_allowed);
1422 nodes_allowed = &node_states[N_HIGH_MEMORY];
1424 } else if (nodes_allowed) {
1426 * per node hstate attribute: adjust count to global,
1427 * but restrict alloc/free to the specified node.
1429 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1430 init_nodemask_of_node(nodes_allowed, nid);
1432 nodes_allowed = &node_states[N_HIGH_MEMORY];
1434 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1436 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1437 NODEMASK_FREE(nodes_allowed);
1441 NODEMASK_FREE(nodes_allowed);
1445 static ssize_t nr_hugepages_show(struct kobject *kobj,
1446 struct kobj_attribute *attr, char *buf)
1448 return nr_hugepages_show_common(kobj, attr, buf);
1451 static ssize_t nr_hugepages_store(struct kobject *kobj,
1452 struct kobj_attribute *attr, const char *buf, size_t len)
1454 return nr_hugepages_store_common(false, kobj, attr, buf, len);
1456 HSTATE_ATTR(nr_hugepages);
1461 * hstate attribute for optionally mempolicy-based constraint on persistent
1462 * huge page alloc/free.
1464 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1465 struct kobj_attribute *attr, char *buf)
1467 return nr_hugepages_show_common(kobj, attr, buf);
1470 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1471 struct kobj_attribute *attr, const char *buf, size_t len)
1473 return nr_hugepages_store_common(true, kobj, attr, buf, len);
1475 HSTATE_ATTR(nr_hugepages_mempolicy);
1479 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1480 struct kobj_attribute *attr, char *buf)
1482 struct hstate *h = kobj_to_hstate(kobj, NULL);
1483 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1486 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1487 struct kobj_attribute *attr, const char *buf, size_t count)
1490 unsigned long input;
1491 struct hstate *h = kobj_to_hstate(kobj, NULL);
1493 if (h->order >= MAX_ORDER)
1496 err = strict_strtoul(buf, 10, &input);
1500 spin_lock(&hugetlb_lock);
1501 h->nr_overcommit_huge_pages = input;
1502 spin_unlock(&hugetlb_lock);
1506 HSTATE_ATTR(nr_overcommit_hugepages);
1508 static ssize_t free_hugepages_show(struct kobject *kobj,
1509 struct kobj_attribute *attr, char *buf)
1512 unsigned long free_huge_pages;
1515 h = kobj_to_hstate(kobj, &nid);
1516 if (nid == NUMA_NO_NODE)
1517 free_huge_pages = h->free_huge_pages;
1519 free_huge_pages = h->free_huge_pages_node[nid];
1521 return sprintf(buf, "%lu\n", free_huge_pages);
1523 HSTATE_ATTR_RO(free_hugepages);
1525 static ssize_t resv_hugepages_show(struct kobject *kobj,
1526 struct kobj_attribute *attr, char *buf)
1528 struct hstate *h = kobj_to_hstate(kobj, NULL);
1529 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1531 HSTATE_ATTR_RO(resv_hugepages);
1533 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1534 struct kobj_attribute *attr, char *buf)
1537 unsigned long surplus_huge_pages;
1540 h = kobj_to_hstate(kobj, &nid);
1541 if (nid == NUMA_NO_NODE)
1542 surplus_huge_pages = h->surplus_huge_pages;
1544 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1546 return sprintf(buf, "%lu\n", surplus_huge_pages);
1548 HSTATE_ATTR_RO(surplus_hugepages);
1550 static struct attribute *hstate_attrs[] = {
1551 &nr_hugepages_attr.attr,
1552 &nr_overcommit_hugepages_attr.attr,
1553 &free_hugepages_attr.attr,
1554 &resv_hugepages_attr.attr,
1555 &surplus_hugepages_attr.attr,
1557 &nr_hugepages_mempolicy_attr.attr,
1562 static struct attribute_group hstate_attr_group = {
1563 .attrs = hstate_attrs,
1566 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1567 struct kobject **hstate_kobjs,
1568 struct attribute_group *hstate_attr_group)
1571 int hi = h - hstates;
1573 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1574 if (!hstate_kobjs[hi])
1577 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1579 kobject_put(hstate_kobjs[hi]);
1584 static void __init hugetlb_sysfs_init(void)
1589 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1590 if (!hugepages_kobj)
1593 for_each_hstate(h) {
1594 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1595 hstate_kobjs, &hstate_attr_group);
1597 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1605 * node_hstate/s - associate per node hstate attributes, via their kobjects,
1606 * with node devices in node_devices[] using a parallel array. The array
1607 * index of a node device or _hstate == node id.
1608 * This is here to avoid any static dependency of the node device driver, in
1609 * the base kernel, on the hugetlb module.
1611 struct node_hstate {
1612 struct kobject *hugepages_kobj;
1613 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1615 struct node_hstate node_hstates[MAX_NUMNODES];
1618 * A subset of global hstate attributes for node devices
1620 static struct attribute *per_node_hstate_attrs[] = {
1621 &nr_hugepages_attr.attr,
1622 &free_hugepages_attr.attr,
1623 &surplus_hugepages_attr.attr,
1627 static struct attribute_group per_node_hstate_attr_group = {
1628 .attrs = per_node_hstate_attrs,
1632 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1633 * Returns node id via non-NULL nidp.
1635 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1639 for (nid = 0; nid < nr_node_ids; nid++) {
1640 struct node_hstate *nhs = &node_hstates[nid];
1642 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1643 if (nhs->hstate_kobjs[i] == kobj) {
1655 * Unregister hstate attributes from a single node device.
1656 * No-op if no hstate attributes attached.
1658 void hugetlb_unregister_node(struct node *node)
1661 struct node_hstate *nhs = &node_hstates[node->dev.id];
1663 if (!nhs->hugepages_kobj)
1664 return; /* no hstate attributes */
1667 if (nhs->hstate_kobjs[h - hstates]) {
1668 kobject_put(nhs->hstate_kobjs[h - hstates]);
1669 nhs->hstate_kobjs[h - hstates] = NULL;
1672 kobject_put(nhs->hugepages_kobj);
1673 nhs->hugepages_kobj = NULL;
1677 * hugetlb module exit: unregister hstate attributes from node devices
1680 static void hugetlb_unregister_all_nodes(void)
1685 * disable node device registrations.
1687 register_hugetlbfs_with_node(NULL, NULL);
1690 * remove hstate attributes from any nodes that have them.
1692 for (nid = 0; nid < nr_node_ids; nid++)
1693 hugetlb_unregister_node(&node_devices[nid]);
1697 * Register hstate attributes for a single node device.
1698 * No-op if attributes already registered.
1700 void hugetlb_register_node(struct node *node)
1703 struct node_hstate *nhs = &node_hstates[node->dev.id];
1706 if (nhs->hugepages_kobj)
1707 return; /* already allocated */
1709 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1711 if (!nhs->hugepages_kobj)
1714 for_each_hstate(h) {
1715 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1717 &per_node_hstate_attr_group);
1719 printk(KERN_ERR "Hugetlb: Unable to add hstate %s"
1721 h->name, node->dev.id);
1722 hugetlb_unregister_node(node);
1729 * hugetlb init time: register hstate attributes for all registered node
1730 * devices of nodes that have memory. All on-line nodes should have
1731 * registered their associated device by this time.
1733 static void hugetlb_register_all_nodes(void)
1737 for_each_node_state(nid, N_HIGH_MEMORY) {
1738 struct node *node = &node_devices[nid];
1739 if (node->dev.id == nid)
1740 hugetlb_register_node(node);
1744 * Let the node device driver know we're here so it can
1745 * [un]register hstate attributes on node hotplug.
1747 register_hugetlbfs_with_node(hugetlb_register_node,
1748 hugetlb_unregister_node);
1750 #else /* !CONFIG_NUMA */
1752 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1760 static void hugetlb_unregister_all_nodes(void) { }
1762 static void hugetlb_register_all_nodes(void) { }
1766 static void __exit hugetlb_exit(void)
1770 hugetlb_unregister_all_nodes();
1772 for_each_hstate(h) {
1773 kobject_put(hstate_kobjs[h - hstates]);
1776 kobject_put(hugepages_kobj);
1778 module_exit(hugetlb_exit);
1780 static int __init hugetlb_init(void)
1782 /* Some platform decide whether they support huge pages at boot
1783 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1784 * there is no such support
1786 if (HPAGE_SHIFT == 0)
1789 if (!size_to_hstate(default_hstate_size)) {
1790 default_hstate_size = HPAGE_SIZE;
1791 if (!size_to_hstate(default_hstate_size))
1792 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1794 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1795 if (default_hstate_max_huge_pages)
1796 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1798 hugetlb_init_hstates();
1800 gather_bootmem_prealloc();
1804 hugetlb_sysfs_init();
1806 hugetlb_register_all_nodes();
1810 module_init(hugetlb_init);
1812 /* Should be called on processing a hugepagesz=... option */
1813 void __init hugetlb_add_hstate(unsigned order)
1818 if (size_to_hstate(PAGE_SIZE << order)) {
1819 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1822 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1824 h = &hstates[max_hstate++];
1826 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1827 h->nr_huge_pages = 0;
1828 h->free_huge_pages = 0;
1829 for (i = 0; i < MAX_NUMNODES; ++i)
1830 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1831 h->next_nid_to_alloc = first_node(node_states[N_HIGH_MEMORY]);
1832 h->next_nid_to_free = first_node(node_states[N_HIGH_MEMORY]);
1833 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1834 huge_page_size(h)/1024);
1839 static int __init hugetlb_nrpages_setup(char *s)
1842 static unsigned long *last_mhp;
1845 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1846 * so this hugepages= parameter goes to the "default hstate".
1849 mhp = &default_hstate_max_huge_pages;
1851 mhp = &parsed_hstate->max_huge_pages;
1853 if (mhp == last_mhp) {
1854 printk(KERN_WARNING "hugepages= specified twice without "
1855 "interleaving hugepagesz=, ignoring\n");
1859 if (sscanf(s, "%lu", mhp) <= 0)
1863 * Global state is always initialized later in hugetlb_init.
1864 * But we need to allocate >= MAX_ORDER hstates here early to still
1865 * use the bootmem allocator.
1867 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1868 hugetlb_hstate_alloc_pages(parsed_hstate);
1874 __setup("hugepages=", hugetlb_nrpages_setup);
1876 static int __init hugetlb_default_setup(char *s)
1878 default_hstate_size = memparse(s, &s);
1881 __setup("default_hugepagesz=", hugetlb_default_setup);
1883 static unsigned int cpuset_mems_nr(unsigned int *array)
1886 unsigned int nr = 0;
1888 for_each_node_mask(node, cpuset_current_mems_allowed)
1894 #ifdef CONFIG_SYSCTL
1895 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
1896 struct ctl_table *table, int write,
1897 void __user *buffer, size_t *length, loff_t *ppos)
1899 struct hstate *h = &default_hstate;
1903 tmp = h->max_huge_pages;
1905 if (write && h->order >= MAX_ORDER)
1909 table->maxlen = sizeof(unsigned long);
1910 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1915 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
1916 GFP_KERNEL | __GFP_NORETRY);
1917 if (!(obey_mempolicy &&
1918 init_nodemask_of_mempolicy(nodes_allowed))) {
1919 NODEMASK_FREE(nodes_allowed);
1920 nodes_allowed = &node_states[N_HIGH_MEMORY];
1922 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
1924 if (nodes_allowed != &node_states[N_HIGH_MEMORY])
1925 NODEMASK_FREE(nodes_allowed);
1931 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1932 void __user *buffer, size_t *length, loff_t *ppos)
1935 return hugetlb_sysctl_handler_common(false, table, write,
1936 buffer, length, ppos);
1940 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
1941 void __user *buffer, size_t *length, loff_t *ppos)
1943 return hugetlb_sysctl_handler_common(true, table, write,
1944 buffer, length, ppos);
1946 #endif /* CONFIG_NUMA */
1948 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1949 void __user *buffer,
1950 size_t *length, loff_t *ppos)
1952 proc_dointvec(table, write, buffer, length, ppos);
1953 if (hugepages_treat_as_movable)
1954 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1956 htlb_alloc_mask = GFP_HIGHUSER;
1960 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1961 void __user *buffer,
1962 size_t *length, loff_t *ppos)
1964 struct hstate *h = &default_hstate;
1968 tmp = h->nr_overcommit_huge_pages;
1970 if (write && h->order >= MAX_ORDER)
1974 table->maxlen = sizeof(unsigned long);
1975 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
1980 spin_lock(&hugetlb_lock);
1981 h->nr_overcommit_huge_pages = tmp;
1982 spin_unlock(&hugetlb_lock);
1988 #endif /* CONFIG_SYSCTL */
1990 void hugetlb_report_meminfo(struct seq_file *m)
1992 struct hstate *h = &default_hstate;
1994 "HugePages_Total: %5lu\n"
1995 "HugePages_Free: %5lu\n"
1996 "HugePages_Rsvd: %5lu\n"
1997 "HugePages_Surp: %5lu\n"
1998 "Hugepagesize: %8lu kB\n",
2002 h->surplus_huge_pages,
2003 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2006 int hugetlb_report_node_meminfo(int nid, char *buf)
2008 struct hstate *h = &default_hstate;
2010 "Node %d HugePages_Total: %5u\n"
2011 "Node %d HugePages_Free: %5u\n"
2012 "Node %d HugePages_Surp: %5u\n",
2013 nid, h->nr_huge_pages_node[nid],
2014 nid, h->free_huge_pages_node[nid],
2015 nid, h->surplus_huge_pages_node[nid]);
2018 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2019 unsigned long hugetlb_total_pages(void)
2021 struct hstate *h = &default_hstate;
2022 return h->nr_huge_pages * pages_per_huge_page(h);
2025 static int hugetlb_acct_memory(struct hstate *h, long delta)
2029 spin_lock(&hugetlb_lock);
2031 * When cpuset is configured, it breaks the strict hugetlb page
2032 * reservation as the accounting is done on a global variable. Such
2033 * reservation is completely rubbish in the presence of cpuset because
2034 * the reservation is not checked against page availability for the
2035 * current cpuset. Application can still potentially OOM'ed by kernel
2036 * with lack of free htlb page in cpuset that the task is in.
2037 * Attempt to enforce strict accounting with cpuset is almost
2038 * impossible (or too ugly) because cpuset is too fluid that
2039 * task or memory node can be dynamically moved between cpusets.
2041 * The change of semantics for shared hugetlb mapping with cpuset is
2042 * undesirable. However, in order to preserve some of the semantics,
2043 * we fall back to check against current free page availability as
2044 * a best attempt and hopefully to minimize the impact of changing
2045 * semantics that cpuset has.
2048 if (gather_surplus_pages(h, delta) < 0)
2051 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2052 return_unused_surplus_pages(h, delta);
2059 return_unused_surplus_pages(h, (unsigned long) -delta);
2062 spin_unlock(&hugetlb_lock);
2066 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2068 struct resv_map *reservations = vma_resv_map(vma);
2071 * This new VMA should share its siblings reservation map if present.
2072 * The VMA will only ever have a valid reservation map pointer where
2073 * it is being copied for another still existing VMA. As that VMA
2074 * has a reference to the reservation map it cannot disappear until
2075 * after this open call completes. It is therefore safe to take a
2076 * new reference here without additional locking.
2079 kref_get(&reservations->refs);
2082 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2084 struct hstate *h = hstate_vma(vma);
2085 struct resv_map *reservations = vma_resv_map(vma);
2086 unsigned long reserve;
2087 unsigned long start;
2091 start = vma_hugecache_offset(h, vma, vma->vm_start);
2092 end = vma_hugecache_offset(h, vma, vma->vm_end);
2094 reserve = (end - start) -
2095 region_count(&reservations->regions, start, end);
2097 kref_put(&reservations->refs, resv_map_release);
2100 hugetlb_acct_memory(h, -reserve);
2101 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
2107 * We cannot handle pagefaults against hugetlb pages at all. They cause
2108 * handle_mm_fault() to try to instantiate regular-sized pages in the
2109 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2112 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2118 const struct vm_operations_struct hugetlb_vm_ops = {
2119 .fault = hugetlb_vm_op_fault,
2120 .open = hugetlb_vm_op_open,
2121 .close = hugetlb_vm_op_close,
2124 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2131 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
2133 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
2135 entry = pte_mkyoung(entry);
2136 entry = pte_mkhuge(entry);
2141 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2142 unsigned long address, pte_t *ptep)
2146 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
2147 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2148 update_mmu_cache(vma, address, ptep);
2152 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2153 struct vm_area_struct *vma)
2155 pte_t *src_pte, *dst_pte, entry;
2156 struct page *ptepage;
2159 struct hstate *h = hstate_vma(vma);
2160 unsigned long sz = huge_page_size(h);
2162 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2164 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2165 src_pte = huge_pte_offset(src, addr);
2168 dst_pte = huge_pte_alloc(dst, addr, sz);
2172 /* If the pagetables are shared don't copy or take references */
2173 if (dst_pte == src_pte)
2176 spin_lock(&dst->page_table_lock);
2177 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2178 if (!huge_pte_none(huge_ptep_get(src_pte))) {
2180 huge_ptep_set_wrprotect(src, addr, src_pte);
2181 entry = huge_ptep_get(src_pte);
2182 ptepage = pte_page(entry);
2184 page_dup_rmap(ptepage);
2185 set_huge_pte_at(dst, addr, dst_pte, entry);
2187 spin_unlock(&src->page_table_lock);
2188 spin_unlock(&dst->page_table_lock);
2196 static int is_hugetlb_entry_migration(pte_t pte)
2200 if (huge_pte_none(pte) || pte_present(pte))
2202 swp = pte_to_swp_entry(pte);
2203 if (non_swap_entry(swp) && is_migration_entry(swp))
2209 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2213 if (huge_pte_none(pte) || pte_present(pte))
2215 swp = pte_to_swp_entry(pte);
2216 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2222 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2223 unsigned long end, struct page *ref_page)
2225 struct mm_struct *mm = vma->vm_mm;
2226 unsigned long address;
2231 struct hstate *h = hstate_vma(vma);
2232 unsigned long sz = huge_page_size(h);
2235 * A page gathering list, protected by per file i_mmap_mutex. The
2236 * lock is used to avoid list corruption from multiple unmapping
2237 * of the same page since we are using page->lru.
2239 LIST_HEAD(page_list);
2241 WARN_ON(!is_vm_hugetlb_page(vma));
2242 BUG_ON(start & ~huge_page_mask(h));
2243 BUG_ON(end & ~huge_page_mask(h));
2245 mmu_notifier_invalidate_range_start(mm, start, end);
2246 spin_lock(&mm->page_table_lock);
2247 for (address = start; address < end; address += sz) {
2248 ptep = huge_pte_offset(mm, address);
2252 if (huge_pmd_unshare(mm, &address, ptep))
2256 * If a reference page is supplied, it is because a specific
2257 * page is being unmapped, not a range. Ensure the page we
2258 * are about to unmap is the actual page of interest.
2261 pte = huge_ptep_get(ptep);
2262 if (huge_pte_none(pte))
2264 page = pte_page(pte);
2265 if (page != ref_page)
2269 * Mark the VMA as having unmapped its page so that
2270 * future faults in this VMA will fail rather than
2271 * looking like data was lost
2273 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2276 pte = huge_ptep_get_and_clear(mm, address, ptep);
2277 if (huge_pte_none(pte))
2281 * HWPoisoned hugepage is already unmapped and dropped reference
2283 if (unlikely(is_hugetlb_entry_hwpoisoned(pte)))
2286 page = pte_page(pte);
2288 set_page_dirty(page);
2289 list_add(&page->lru, &page_list);
2291 /* Bail out after unmapping reference page if supplied */
2295 flush_tlb_range(vma, start, end);
2296 spin_unlock(&mm->page_table_lock);
2297 mmu_notifier_invalidate_range_end(mm, start, end);
2298 list_for_each_entry_safe(page, tmp, &page_list, lru) {
2299 page_remove_rmap(page);
2300 list_del(&page->lru);
2305 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2306 unsigned long end, struct page *ref_page)
2308 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2309 __unmap_hugepage_range(vma, start, end, ref_page);
2310 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2314 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2315 * mappping it owns the reserve page for. The intention is to unmap the page
2316 * from other VMAs and let the children be SIGKILLed if they are faulting the
2319 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2320 struct page *page, unsigned long address)
2322 struct hstate *h = hstate_vma(vma);
2323 struct vm_area_struct *iter_vma;
2324 struct address_space *mapping;
2325 struct prio_tree_iter iter;
2329 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2330 * from page cache lookup which is in HPAGE_SIZE units.
2332 address = address & huge_page_mask(h);
2333 pgoff = vma_hugecache_offset(h, vma, address);
2334 mapping = (struct address_space *)page_private(page);
2337 * Take the mapping lock for the duration of the table walk. As
2338 * this mapping should be shared between all the VMAs,
2339 * __unmap_hugepage_range() is called as the lock is already held
2341 mutex_lock(&mapping->i_mmap_mutex);
2342 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
2343 /* Do not unmap the current VMA */
2344 if (iter_vma == vma)
2348 * Unmap the page from other VMAs without their own reserves.
2349 * They get marked to be SIGKILLed if they fault in these
2350 * areas. This is because a future no-page fault on this VMA
2351 * could insert a zeroed page instead of the data existing
2352 * from the time of fork. This would look like data corruption
2354 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2355 __unmap_hugepage_range(iter_vma,
2356 address, address + huge_page_size(h),
2359 mutex_unlock(&mapping->i_mmap_mutex);
2365 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2366 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2367 * cannot race with other handlers or page migration.
2368 * Keep the pte_same checks anyway to make transition from the mutex easier.
2370 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2371 unsigned long address, pte_t *ptep, pte_t pte,
2372 struct page *pagecache_page)
2374 struct hstate *h = hstate_vma(vma);
2375 struct page *old_page, *new_page;
2377 int outside_reserve = 0;
2379 old_page = pte_page(pte);
2382 /* If no-one else is actually using this page, avoid the copy
2383 * and just make the page writable */
2384 avoidcopy = (page_mapcount(old_page) == 1);
2386 if (PageAnon(old_page))
2387 page_move_anon_rmap(old_page, vma, address);
2388 set_huge_ptep_writable(vma, address, ptep);
2393 * If the process that created a MAP_PRIVATE mapping is about to
2394 * perform a COW due to a shared page count, attempt to satisfy
2395 * the allocation without using the existing reserves. The pagecache
2396 * page is used to determine if the reserve at this address was
2397 * consumed or not. If reserves were used, a partial faulted mapping
2398 * at the time of fork() could consume its reserves on COW instead
2399 * of the full address range.
2401 if (!(vma->vm_flags & VM_MAYSHARE) &&
2402 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2403 old_page != pagecache_page)
2404 outside_reserve = 1;
2406 page_cache_get(old_page);
2408 /* Drop page_table_lock as buddy allocator may be called */
2409 spin_unlock(&mm->page_table_lock);
2410 new_page = alloc_huge_page(vma, address, outside_reserve);
2412 if (IS_ERR(new_page)) {
2413 page_cache_release(old_page);
2416 * If a process owning a MAP_PRIVATE mapping fails to COW,
2417 * it is due to references held by a child and an insufficient
2418 * huge page pool. To guarantee the original mappers
2419 * reliability, unmap the page from child processes. The child
2420 * may get SIGKILLed if it later faults.
2422 if (outside_reserve) {
2423 BUG_ON(huge_pte_none(pte));
2424 if (unmap_ref_private(mm, vma, old_page, address)) {
2425 BUG_ON(page_count(old_page) != 1);
2426 BUG_ON(huge_pte_none(pte));
2427 spin_lock(&mm->page_table_lock);
2428 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2429 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2430 goto retry_avoidcopy;
2432 * race occurs while re-acquiring page_table_lock, and
2440 /* Caller expects lock to be held */
2441 spin_lock(&mm->page_table_lock);
2442 return -PTR_ERR(new_page);
2446 * When the original hugepage is shared one, it does not have
2447 * anon_vma prepared.
2449 if (unlikely(anon_vma_prepare(vma))) {
2450 page_cache_release(new_page);
2451 page_cache_release(old_page);
2452 /* Caller expects lock to be held */
2453 spin_lock(&mm->page_table_lock);
2454 return VM_FAULT_OOM;
2457 copy_user_huge_page(new_page, old_page, address, vma,
2458 pages_per_huge_page(h));
2459 __SetPageUptodate(new_page);
2462 * Retake the page_table_lock to check for racing updates
2463 * before the page tables are altered
2465 spin_lock(&mm->page_table_lock);
2466 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2467 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2469 mmu_notifier_invalidate_range_start(mm,
2470 address & huge_page_mask(h),
2471 (address & huge_page_mask(h)) + huge_page_size(h));
2472 huge_ptep_clear_flush(vma, address, ptep);
2473 set_huge_pte_at(mm, address, ptep,
2474 make_huge_pte(vma, new_page, 1));
2475 page_remove_rmap(old_page);
2476 hugepage_add_new_anon_rmap(new_page, vma, address);
2477 /* Make the old page be freed below */
2478 new_page = old_page;
2479 mmu_notifier_invalidate_range_end(mm,
2480 address & huge_page_mask(h),
2481 (address & huge_page_mask(h)) + huge_page_size(h));
2483 page_cache_release(new_page);
2484 page_cache_release(old_page);
2488 /* Return the pagecache page at a given address within a VMA */
2489 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2490 struct vm_area_struct *vma, unsigned long address)
2492 struct address_space *mapping;
2495 mapping = vma->vm_file->f_mapping;
2496 idx = vma_hugecache_offset(h, vma, address);
2498 return find_lock_page(mapping, idx);
2502 * Return whether there is a pagecache page to back given address within VMA.
2503 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2505 static bool hugetlbfs_pagecache_present(struct hstate *h,
2506 struct vm_area_struct *vma, unsigned long address)
2508 struct address_space *mapping;
2512 mapping = vma->vm_file->f_mapping;
2513 idx = vma_hugecache_offset(h, vma, address);
2515 page = find_get_page(mapping, idx);
2518 return page != NULL;
2521 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2522 unsigned long address, pte_t *ptep, unsigned int flags)
2524 struct hstate *h = hstate_vma(vma);
2525 int ret = VM_FAULT_SIGBUS;
2530 struct address_space *mapping;
2534 * Currently, we are forced to kill the process in the event the
2535 * original mapper has unmapped pages from the child due to a failed
2536 * COW. Warn that such a situation has occurred as it may not be obvious
2538 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2540 "PID %d killed due to inadequate hugepage pool\n",
2545 mapping = vma->vm_file->f_mapping;
2546 idx = vma_hugecache_offset(h, vma, address);
2549 * Use page lock to guard against racing truncation
2550 * before we get page_table_lock.
2553 page = find_lock_page(mapping, idx);
2555 size = i_size_read(mapping->host) >> huge_page_shift(h);
2558 page = alloc_huge_page(vma, address, 0);
2560 ret = -PTR_ERR(page);
2563 clear_huge_page(page, address, pages_per_huge_page(h));
2564 __SetPageUptodate(page);
2566 if (vma->vm_flags & VM_MAYSHARE) {
2568 struct inode *inode = mapping->host;
2570 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2578 spin_lock(&inode->i_lock);
2579 inode->i_blocks += blocks_per_huge_page(h);
2580 spin_unlock(&inode->i_lock);
2583 if (unlikely(anon_vma_prepare(vma))) {
2585 goto backout_unlocked;
2591 * If memory error occurs between mmap() and fault, some process
2592 * don't have hwpoisoned swap entry for errored virtual address.
2593 * So we need to block hugepage fault by PG_hwpoison bit check.
2595 if (unlikely(PageHWPoison(page))) {
2596 ret = VM_FAULT_HWPOISON |
2597 VM_FAULT_SET_HINDEX(h - hstates);
2598 goto backout_unlocked;
2603 * If we are going to COW a private mapping later, we examine the
2604 * pending reservations for this page now. This will ensure that
2605 * any allocations necessary to record that reservation occur outside
2608 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2609 if (vma_needs_reservation(h, vma, address) < 0) {
2611 goto backout_unlocked;
2614 spin_lock(&mm->page_table_lock);
2615 size = i_size_read(mapping->host) >> huge_page_shift(h);
2620 if (!huge_pte_none(huge_ptep_get(ptep)))
2624 hugepage_add_new_anon_rmap(page, vma, address);
2626 page_dup_rmap(page);
2627 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2628 && (vma->vm_flags & VM_SHARED)));
2629 set_huge_pte_at(mm, address, ptep, new_pte);
2631 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2632 /* Optimization, do the COW without a second fault */
2633 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2636 spin_unlock(&mm->page_table_lock);
2642 spin_unlock(&mm->page_table_lock);
2649 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2650 unsigned long address, unsigned int flags)
2655 struct page *page = NULL;
2656 struct page *pagecache_page = NULL;
2657 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2658 struct hstate *h = hstate_vma(vma);
2660 address &= huge_page_mask(h);
2662 ptep = huge_pte_offset(mm, address);
2664 entry = huge_ptep_get(ptep);
2665 if (unlikely(is_hugetlb_entry_migration(entry))) {
2666 migration_entry_wait(mm, (pmd_t *)ptep, address);
2668 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2669 return VM_FAULT_HWPOISON_LARGE |
2670 VM_FAULT_SET_HINDEX(h - hstates);
2673 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2675 return VM_FAULT_OOM;
2678 * Serialize hugepage allocation and instantiation, so that we don't
2679 * get spurious allocation failures if two CPUs race to instantiate
2680 * the same page in the page cache.
2682 mutex_lock(&hugetlb_instantiation_mutex);
2683 entry = huge_ptep_get(ptep);
2684 if (huge_pte_none(entry)) {
2685 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2692 * If we are going to COW the mapping later, we examine the pending
2693 * reservations for this page now. This will ensure that any
2694 * allocations necessary to record that reservation occur outside the
2695 * spinlock. For private mappings, we also lookup the pagecache
2696 * page now as it is used to determine if a reservation has been
2699 if ((flags & FAULT_FLAG_WRITE) && !pte_write(entry)) {
2700 if (vma_needs_reservation(h, vma, address) < 0) {
2705 if (!(vma->vm_flags & VM_MAYSHARE))
2706 pagecache_page = hugetlbfs_pagecache_page(h,
2711 * hugetlb_cow() requires page locks of pte_page(entry) and
2712 * pagecache_page, so here we need take the former one
2713 * when page != pagecache_page or !pagecache_page.
2714 * Note that locking order is always pagecache_page -> page,
2715 * so no worry about deadlock.
2717 page = pte_page(entry);
2718 if (page != pagecache_page)
2721 spin_lock(&mm->page_table_lock);
2722 /* Check for a racing update before calling hugetlb_cow */
2723 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2724 goto out_page_table_lock;
2727 if (flags & FAULT_FLAG_WRITE) {
2728 if (!pte_write(entry)) {
2729 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2731 goto out_page_table_lock;
2733 entry = pte_mkdirty(entry);
2735 entry = pte_mkyoung(entry);
2736 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
2737 flags & FAULT_FLAG_WRITE))
2738 update_mmu_cache(vma, address, ptep);
2740 out_page_table_lock:
2741 spin_unlock(&mm->page_table_lock);
2743 if (pagecache_page) {
2744 unlock_page(pagecache_page);
2745 put_page(pagecache_page);
2747 if (page != pagecache_page)
2751 mutex_unlock(&hugetlb_instantiation_mutex);
2756 /* Can be overriden by architectures */
2757 __attribute__((weak)) struct page *
2758 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2759 pud_t *pud, int write)
2765 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2766 struct page **pages, struct vm_area_struct **vmas,
2767 unsigned long *position, int *length, int i,
2770 unsigned long pfn_offset;
2771 unsigned long vaddr = *position;
2772 int remainder = *length;
2773 struct hstate *h = hstate_vma(vma);
2775 spin_lock(&mm->page_table_lock);
2776 while (vaddr < vma->vm_end && remainder) {
2782 * Some archs (sparc64, sh*) have multiple pte_ts to
2783 * each hugepage. We have to make sure we get the
2784 * first, for the page indexing below to work.
2786 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2787 absent = !pte || huge_pte_none(huge_ptep_get(pte));
2790 * When coredumping, it suits get_dump_page if we just return
2791 * an error where there's an empty slot with no huge pagecache
2792 * to back it. This way, we avoid allocating a hugepage, and
2793 * the sparse dumpfile avoids allocating disk blocks, but its
2794 * huge holes still show up with zeroes where they need to be.
2796 if (absent && (flags & FOLL_DUMP) &&
2797 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
2803 ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
2806 spin_unlock(&mm->page_table_lock);
2807 ret = hugetlb_fault(mm, vma, vaddr,
2808 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
2809 spin_lock(&mm->page_table_lock);
2810 if (!(ret & VM_FAULT_ERROR))
2817 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2818 page = pte_page(huge_ptep_get(pte));
2821 pages[i] = mem_map_offset(page, pfn_offset);
2832 if (vaddr < vma->vm_end && remainder &&
2833 pfn_offset < pages_per_huge_page(h)) {
2835 * We use pfn_offset to avoid touching the pageframes
2836 * of this compound page.
2841 spin_unlock(&mm->page_table_lock);
2842 *length = remainder;
2845 return i ? i : -EFAULT;
2848 void hugetlb_change_protection(struct vm_area_struct *vma,
2849 unsigned long address, unsigned long end, pgprot_t newprot)
2851 struct mm_struct *mm = vma->vm_mm;
2852 unsigned long start = address;
2855 struct hstate *h = hstate_vma(vma);
2857 BUG_ON(address >= end);
2858 flush_cache_range(vma, address, end);
2860 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
2861 spin_lock(&mm->page_table_lock);
2862 for (; address < end; address += huge_page_size(h)) {
2863 ptep = huge_pte_offset(mm, address);
2866 if (huge_pmd_unshare(mm, &address, ptep))
2868 if (!huge_pte_none(huge_ptep_get(ptep))) {
2869 pte = huge_ptep_get_and_clear(mm, address, ptep);
2870 pte = pte_mkhuge(pte_modify(pte, newprot));
2871 set_huge_pte_at(mm, address, ptep, pte);
2874 spin_unlock(&mm->page_table_lock);
2875 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
2877 flush_tlb_range(vma, start, end);
2880 int hugetlb_reserve_pages(struct inode *inode,
2882 struct vm_area_struct *vma,
2883 vm_flags_t vm_flags)
2886 struct hstate *h = hstate_inode(inode);
2889 * Only apply hugepage reservation if asked. At fault time, an
2890 * attempt will be made for VM_NORESERVE to allocate a page
2891 * and filesystem quota without using reserves
2893 if (vm_flags & VM_NORESERVE)
2897 * Shared mappings base their reservation on the number of pages that
2898 * are already allocated on behalf of the file. Private mappings need
2899 * to reserve the full area even if read-only as mprotect() may be
2900 * called to make the mapping read-write. Assume !vma is a shm mapping
2902 if (!vma || vma->vm_flags & VM_MAYSHARE)
2903 chg = region_chg(&inode->i_mapping->private_list, from, to);
2905 struct resv_map *resv_map = resv_map_alloc();
2911 set_vma_resv_map(vma, resv_map);
2912 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2918 /* There must be enough filesystem quota for the mapping */
2919 if (hugetlb_get_quota(inode->i_mapping, chg))
2923 * Check enough hugepages are available for the reservation.
2924 * Hand back the quota if there are not
2926 ret = hugetlb_acct_memory(h, chg);
2928 hugetlb_put_quota(inode->i_mapping, chg);
2933 * Account for the reservations made. Shared mappings record regions
2934 * that have reservations as they are shared by multiple VMAs.
2935 * When the last VMA disappears, the region map says how much
2936 * the reservation was and the page cache tells how much of
2937 * the reservation was consumed. Private mappings are per-VMA and
2938 * only the consumed reservations are tracked. When the VMA
2939 * disappears, the original reservation is the VMA size and the
2940 * consumed reservations are stored in the map. Hence, nothing
2941 * else has to be done for private mappings here
2943 if (!vma || vma->vm_flags & VM_MAYSHARE)
2944 region_add(&inode->i_mapping->private_list, from, to);
2948 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2950 struct hstate *h = hstate_inode(inode);
2951 long chg = region_truncate(&inode->i_mapping->private_list, offset);
2953 spin_lock(&inode->i_lock);
2954 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
2955 spin_unlock(&inode->i_lock);
2957 hugetlb_put_quota(inode->i_mapping, (chg - freed));
2958 hugetlb_acct_memory(h, -(chg - freed));
2961 #ifdef CONFIG_MEMORY_FAILURE
2963 /* Should be called in hugetlb_lock */
2964 static int is_hugepage_on_freelist(struct page *hpage)
2968 struct hstate *h = page_hstate(hpage);
2969 int nid = page_to_nid(hpage);
2971 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
2978 * This function is called from memory failure code.
2979 * Assume the caller holds page lock of the head page.
2981 int dequeue_hwpoisoned_huge_page(struct page *hpage)
2983 struct hstate *h = page_hstate(hpage);
2984 int nid = page_to_nid(hpage);
2987 spin_lock(&hugetlb_lock);
2988 if (is_hugepage_on_freelist(hpage)) {
2989 list_del(&hpage->lru);
2990 set_page_refcounted(hpage);
2991 h->free_huge_pages--;
2992 h->free_huge_pages_node[nid]--;
2995 spin_unlock(&hugetlb_lock);