2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, 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/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/bootmem.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/rmap.h>
23 #include <linux/swap.h>
24 #include <linux/swapops.h>
25 #include <linux/page-isolation.h>
26 #include <linux/jhash.h>
29 #include <asm/pgtable.h>
33 #include <linux/hugetlb.h>
34 #include <linux/hugetlb_cgroup.h>
35 #include <linux/node.h>
38 int hugepages_treat_as_movable;
40 int hugetlb_max_hstate __read_mostly;
41 unsigned int default_hstate_idx;
42 struct hstate hstates[HUGE_MAX_HSTATE];
44 * Minimum page order among possible hugepage sizes, set to a proper value
47 static unsigned int minimum_order __read_mostly = UINT_MAX;
49 __initdata LIST_HEAD(huge_boot_pages);
51 /* for command line parsing */
52 static struct hstate * __initdata parsed_hstate;
53 static unsigned long __initdata default_hstate_max_huge_pages;
54 static unsigned long __initdata default_hstate_size;
57 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
58 * free_huge_pages, and surplus_huge_pages.
60 DEFINE_SPINLOCK(hugetlb_lock);
63 * Serializes faults on the same logical page. This is used to
64 * prevent spurious OOMs when the hugepage pool is fully utilized.
66 static int num_fault_mutexes;
67 static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp;
69 /* Forward declaration */
70 static int hugetlb_acct_memory(struct hstate *h, long delta);
72 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
74 bool free = (spool->count == 0) && (spool->used_hpages == 0);
76 spin_unlock(&spool->lock);
78 /* If no pages are used, and no other handles to the subpool
79 * remain, give up any reservations mased on minimum size and
82 if (spool->min_hpages != -1)
83 hugetlb_acct_memory(spool->hstate,
89 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
92 struct hugepage_subpool *spool;
94 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
98 spin_lock_init(&spool->lock);
100 spool->max_hpages = max_hpages;
102 spool->min_hpages = min_hpages;
104 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
108 spool->rsv_hpages = min_hpages;
113 void hugepage_put_subpool(struct hugepage_subpool *spool)
115 spin_lock(&spool->lock);
116 BUG_ON(!spool->count);
118 unlock_or_release_subpool(spool);
122 * Subpool accounting for allocating and reserving pages.
123 * Return -ENOMEM if there are not enough resources to satisfy the
124 * the request. Otherwise, return the number of pages by which the
125 * global pools must be adjusted (upward). The returned value may
126 * only be different than the passed value (delta) in the case where
127 * a subpool minimum size must be manitained.
129 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
137 spin_lock(&spool->lock);
139 if (spool->max_hpages != -1) { /* maximum size accounting */
140 if ((spool->used_hpages + delta) <= spool->max_hpages)
141 spool->used_hpages += delta;
148 if (spool->min_hpages != -1) { /* minimum size accounting */
149 if (delta > spool->rsv_hpages) {
151 * Asking for more reserves than those already taken on
152 * behalf of subpool. Return difference.
154 ret = delta - spool->rsv_hpages;
155 spool->rsv_hpages = 0;
157 ret = 0; /* reserves already accounted for */
158 spool->rsv_hpages -= delta;
163 spin_unlock(&spool->lock);
168 * Subpool accounting for freeing and unreserving pages.
169 * Return the number of global page reservations that must be dropped.
170 * The return value may only be different than the passed value (delta)
171 * in the case where a subpool minimum size must be maintained.
173 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
181 spin_lock(&spool->lock);
183 if (spool->max_hpages != -1) /* maximum size accounting */
184 spool->used_hpages -= delta;
186 if (spool->min_hpages != -1) { /* minimum size accounting */
187 if (spool->rsv_hpages + delta <= spool->min_hpages)
190 ret = spool->rsv_hpages + delta - spool->min_hpages;
192 spool->rsv_hpages += delta;
193 if (spool->rsv_hpages > spool->min_hpages)
194 spool->rsv_hpages = spool->min_hpages;
198 * If hugetlbfs_put_super couldn't free spool due to an outstanding
199 * quota reference, free it now.
201 unlock_or_release_subpool(spool);
206 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
208 return HUGETLBFS_SB(inode->i_sb)->spool;
211 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
213 return subpool_inode(file_inode(vma->vm_file));
217 * Region tracking -- allows tracking of reservations and instantiated pages
218 * across the pages in a mapping.
220 * The region data structures are embedded into a resv_map and protected
221 * by a resv_map's lock. The set of regions within the resv_map represent
222 * reservations for huge pages, or huge pages that have already been
223 * instantiated within the map. The from and to elements are huge page
224 * indicies into the associated mapping. from indicates the starting index
225 * of the region. to represents the first index past the end of the region.
227 * For example, a file region structure with from == 0 and to == 4 represents
228 * four huge pages in a mapping. It is important to note that the to element
229 * represents the first element past the end of the region. This is used in
230 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
232 * Interval notation of the form [from, to) will be used to indicate that
233 * the endpoint from is inclusive and to is exclusive.
236 struct list_head link;
242 * Add the huge page range represented by [f, t) to the reserve
243 * map. Existing regions will be expanded to accommodate the
244 * specified range. We know only existing regions need to be
245 * expanded, because region_add is only called after region_chg
246 * with the same range. If a new file_region structure must
247 * be allocated, it is done in region_chg.
249 * Return the number of new huge pages added to the map. This
250 * number is greater than or equal to zero.
252 static long region_add(struct resv_map *resv, long f, long t)
254 struct list_head *head = &resv->regions;
255 struct file_region *rg, *nrg, *trg;
258 spin_lock(&resv->lock);
259 /* Locate the region we are either in or before. */
260 list_for_each_entry(rg, head, link)
264 /* Round our left edge to the current segment if it encloses us. */
268 /* Check for and consume any regions we now overlap with. */
270 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
271 if (&rg->link == head)
276 /* If this area reaches higher then extend our area to
277 * include it completely. If this is not the first area
278 * which we intend to reuse, free it. */
282 /* Decrement return value by the deleted range.
283 * Another range will span this area so that by
284 * end of routine add will be >= zero
286 add -= (rg->to - rg->from);
292 add += (nrg->from - f); /* Added to beginning of region */
294 add += t - nrg->to; /* Added to end of region */
297 spin_unlock(&resv->lock);
303 * Examine the existing reserve map and determine how many
304 * huge pages in the specified range [f, t) are NOT currently
305 * represented. This routine is called before a subsequent
306 * call to region_add that will actually modify the reserve
307 * map to add the specified range [f, t). region_chg does
308 * not change the number of huge pages represented by the
309 * map. However, if the existing regions in the map can not
310 * be expanded to represent the new range, a new file_region
311 * structure is added to the map as a placeholder. This is
312 * so that the subsequent region_add call will have all the
313 * regions it needs and will not fail.
315 * Returns the number of huge pages that need to be added
316 * to the existing reservation map for the range [f, t).
317 * This number is greater or equal to zero. -ENOMEM is
318 * returned if a new file_region structure is needed and can
321 static long region_chg(struct resv_map *resv, long f, long t)
323 struct list_head *head = &resv->regions;
324 struct file_region *rg, *nrg = NULL;
328 spin_lock(&resv->lock);
329 /* Locate the region we are before or in. */
330 list_for_each_entry(rg, head, link)
334 /* If we are below the current region then a new region is required.
335 * Subtle, allocate a new region at the position but make it zero
336 * size such that we can guarantee to record the reservation. */
337 if (&rg->link == head || t < rg->from) {
339 spin_unlock(&resv->lock);
340 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
346 INIT_LIST_HEAD(&nrg->link);
350 list_add(&nrg->link, rg->link.prev);
355 /* Round our left edge to the current segment if it encloses us. */
360 /* Check for and consume any regions we now overlap with. */
361 list_for_each_entry(rg, rg->link.prev, link) {
362 if (&rg->link == head)
367 /* We overlap with this area, if it extends further than
368 * us then we must extend ourselves. Account for its
369 * existing reservation. */
374 chg -= rg->to - rg->from;
378 spin_unlock(&resv->lock);
379 /* We already know we raced and no longer need the new region */
383 spin_unlock(&resv->lock);
388 * Truncate the reserve map at index 'end'. Modify/truncate any
389 * region which contains end. Delete any regions past end.
390 * Return the number of huge pages removed from the map.
392 static long region_truncate(struct resv_map *resv, long end)
394 struct list_head *head = &resv->regions;
395 struct file_region *rg, *trg;
398 spin_lock(&resv->lock);
399 /* Locate the region we are either in or before. */
400 list_for_each_entry(rg, head, link)
403 if (&rg->link == head)
406 /* If we are in the middle of a region then adjust it. */
407 if (end > rg->from) {
410 rg = list_entry(rg->link.next, typeof(*rg), link);
413 /* Drop any remaining regions. */
414 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
415 if (&rg->link == head)
417 chg += rg->to - rg->from;
423 spin_unlock(&resv->lock);
428 * Count and return the number of huge pages in the reserve map
429 * that intersect with the range [f, t).
431 static long region_count(struct resv_map *resv, long f, long t)
433 struct list_head *head = &resv->regions;
434 struct file_region *rg;
437 spin_lock(&resv->lock);
438 /* Locate each segment we overlap with, and count that overlap. */
439 list_for_each_entry(rg, head, link) {
448 seg_from = max(rg->from, f);
449 seg_to = min(rg->to, t);
451 chg += seg_to - seg_from;
453 spin_unlock(&resv->lock);
459 * Convert the address within this vma to the page offset within
460 * the mapping, in pagecache page units; huge pages here.
462 static pgoff_t vma_hugecache_offset(struct hstate *h,
463 struct vm_area_struct *vma, unsigned long address)
465 return ((address - vma->vm_start) >> huge_page_shift(h)) +
466 (vma->vm_pgoff >> huge_page_order(h));
469 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
470 unsigned long address)
472 return vma_hugecache_offset(hstate_vma(vma), vma, address);
476 * Return the size of the pages allocated when backing a VMA. In the majority
477 * cases this will be same size as used by the page table entries.
479 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
481 struct hstate *hstate;
483 if (!is_vm_hugetlb_page(vma))
486 hstate = hstate_vma(vma);
488 return 1UL << huge_page_shift(hstate);
490 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
493 * Return the page size being used by the MMU to back a VMA. In the majority
494 * of cases, the page size used by the kernel matches the MMU size. On
495 * architectures where it differs, an architecture-specific version of this
496 * function is required.
498 #ifndef vma_mmu_pagesize
499 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
501 return vma_kernel_pagesize(vma);
506 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
507 * bits of the reservation map pointer, which are always clear due to
510 #define HPAGE_RESV_OWNER (1UL << 0)
511 #define HPAGE_RESV_UNMAPPED (1UL << 1)
512 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
515 * These helpers are used to track how many pages are reserved for
516 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
517 * is guaranteed to have their future faults succeed.
519 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
520 * the reserve counters are updated with the hugetlb_lock held. It is safe
521 * to reset the VMA at fork() time as it is not in use yet and there is no
522 * chance of the global counters getting corrupted as a result of the values.
524 * The private mapping reservation is represented in a subtly different
525 * manner to a shared mapping. A shared mapping has a region map associated
526 * with the underlying file, this region map represents the backing file
527 * pages which have ever had a reservation assigned which this persists even
528 * after the page is instantiated. A private mapping has a region map
529 * associated with the original mmap which is attached to all VMAs which
530 * reference it, this region map represents those offsets which have consumed
531 * reservation ie. where pages have been instantiated.
533 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
535 return (unsigned long)vma->vm_private_data;
538 static void set_vma_private_data(struct vm_area_struct *vma,
541 vma->vm_private_data = (void *)value;
544 struct resv_map *resv_map_alloc(void)
546 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
550 kref_init(&resv_map->refs);
551 spin_lock_init(&resv_map->lock);
552 INIT_LIST_HEAD(&resv_map->regions);
557 void resv_map_release(struct kref *ref)
559 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
561 /* Clear out any active regions before we release the map. */
562 region_truncate(resv_map, 0);
566 static inline struct resv_map *inode_resv_map(struct inode *inode)
568 return inode->i_mapping->private_data;
571 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
573 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
574 if (vma->vm_flags & VM_MAYSHARE) {
575 struct address_space *mapping = vma->vm_file->f_mapping;
576 struct inode *inode = mapping->host;
578 return inode_resv_map(inode);
581 return (struct resv_map *)(get_vma_private_data(vma) &
586 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
588 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
589 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
591 set_vma_private_data(vma, (get_vma_private_data(vma) &
592 HPAGE_RESV_MASK) | (unsigned long)map);
595 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
597 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
598 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
600 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
603 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
605 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
607 return (get_vma_private_data(vma) & flag) != 0;
610 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
611 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
613 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
614 if (!(vma->vm_flags & VM_MAYSHARE))
615 vma->vm_private_data = (void *)0;
618 /* Returns true if the VMA has associated reserve pages */
619 static int vma_has_reserves(struct vm_area_struct *vma, long chg)
621 if (vma->vm_flags & VM_NORESERVE) {
623 * This address is already reserved by other process(chg == 0),
624 * so, we should decrement reserved count. Without decrementing,
625 * reserve count remains after releasing inode, because this
626 * allocated page will go into page cache and is regarded as
627 * coming from reserved pool in releasing step. Currently, we
628 * don't have any other solution to deal with this situation
629 * properly, so add work-around here.
631 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
637 /* Shared mappings always use reserves */
638 if (vma->vm_flags & VM_MAYSHARE)
642 * Only the process that called mmap() has reserves for
645 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
651 static void enqueue_huge_page(struct hstate *h, struct page *page)
653 int nid = page_to_nid(page);
654 list_move(&page->lru, &h->hugepage_freelists[nid]);
655 h->free_huge_pages++;
656 h->free_huge_pages_node[nid]++;
659 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
663 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
664 if (!is_migrate_isolate_page(page))
667 * if 'non-isolated free hugepage' not found on the list,
668 * the allocation fails.
670 if (&h->hugepage_freelists[nid] == &page->lru)
672 list_move(&page->lru, &h->hugepage_activelist);
673 set_page_refcounted(page);
674 h->free_huge_pages--;
675 h->free_huge_pages_node[nid]--;
679 /* Movability of hugepages depends on migration support. */
680 static inline gfp_t htlb_alloc_mask(struct hstate *h)
682 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
683 return GFP_HIGHUSER_MOVABLE;
688 static struct page *dequeue_huge_page_vma(struct hstate *h,
689 struct vm_area_struct *vma,
690 unsigned long address, int avoid_reserve,
693 struct page *page = NULL;
694 struct mempolicy *mpol;
695 nodemask_t *nodemask;
696 struct zonelist *zonelist;
699 unsigned int cpuset_mems_cookie;
702 * A child process with MAP_PRIVATE mappings created by their parent
703 * have no page reserves. This check ensures that reservations are
704 * not "stolen". The child may still get SIGKILLed
706 if (!vma_has_reserves(vma, chg) &&
707 h->free_huge_pages - h->resv_huge_pages == 0)
710 /* If reserves cannot be used, ensure enough pages are in the pool */
711 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
715 cpuset_mems_cookie = read_mems_allowed_begin();
716 zonelist = huge_zonelist(vma, address,
717 htlb_alloc_mask(h), &mpol, &nodemask);
719 for_each_zone_zonelist_nodemask(zone, z, zonelist,
720 MAX_NR_ZONES - 1, nodemask) {
721 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
722 page = dequeue_huge_page_node(h, zone_to_nid(zone));
726 if (!vma_has_reserves(vma, chg))
729 SetPagePrivate(page);
730 h->resv_huge_pages--;
737 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
746 * common helper functions for hstate_next_node_to_{alloc|free}.
747 * We may have allocated or freed a huge page based on a different
748 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
749 * be outside of *nodes_allowed. Ensure that we use an allowed
750 * node for alloc or free.
752 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
754 nid = next_node(nid, *nodes_allowed);
755 if (nid == MAX_NUMNODES)
756 nid = first_node(*nodes_allowed);
757 VM_BUG_ON(nid >= MAX_NUMNODES);
762 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
764 if (!node_isset(nid, *nodes_allowed))
765 nid = next_node_allowed(nid, nodes_allowed);
770 * returns the previously saved node ["this node"] from which to
771 * allocate a persistent huge page for the pool and advance the
772 * next node from which to allocate, handling wrap at end of node
775 static int hstate_next_node_to_alloc(struct hstate *h,
776 nodemask_t *nodes_allowed)
780 VM_BUG_ON(!nodes_allowed);
782 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
783 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
789 * helper for free_pool_huge_page() - return the previously saved
790 * node ["this node"] from which to free a huge page. Advance the
791 * next node id whether or not we find a free huge page to free so
792 * that the next attempt to free addresses the next node.
794 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
798 VM_BUG_ON(!nodes_allowed);
800 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
801 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
806 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
807 for (nr_nodes = nodes_weight(*mask); \
809 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
812 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
813 for (nr_nodes = nodes_weight(*mask); \
815 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
818 #if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
819 static void destroy_compound_gigantic_page(struct page *page,
823 int nr_pages = 1 << order;
824 struct page *p = page + 1;
826 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
828 set_page_refcounted(p);
829 p->first_page = NULL;
832 set_compound_order(page, 0);
833 __ClearPageHead(page);
836 static void free_gigantic_page(struct page *page, unsigned order)
838 free_contig_range(page_to_pfn(page), 1 << order);
841 static int __alloc_gigantic_page(unsigned long start_pfn,
842 unsigned long nr_pages)
844 unsigned long end_pfn = start_pfn + nr_pages;
845 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
848 static bool pfn_range_valid_gigantic(unsigned long start_pfn,
849 unsigned long nr_pages)
851 unsigned long i, end_pfn = start_pfn + nr_pages;
854 for (i = start_pfn; i < end_pfn; i++) {
858 page = pfn_to_page(i);
860 if (PageReserved(page))
863 if (page_count(page) > 0)
873 static bool zone_spans_last_pfn(const struct zone *zone,
874 unsigned long start_pfn, unsigned long nr_pages)
876 unsigned long last_pfn = start_pfn + nr_pages - 1;
877 return zone_spans_pfn(zone, last_pfn);
880 static struct page *alloc_gigantic_page(int nid, unsigned order)
882 unsigned long nr_pages = 1 << order;
883 unsigned long ret, pfn, flags;
886 z = NODE_DATA(nid)->node_zones;
887 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
888 spin_lock_irqsave(&z->lock, flags);
890 pfn = ALIGN(z->zone_start_pfn, nr_pages);
891 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
892 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
894 * We release the zone lock here because
895 * alloc_contig_range() will also lock the zone
896 * at some point. If there's an allocation
897 * spinning on this lock, it may win the race
898 * and cause alloc_contig_range() to fail...
900 spin_unlock_irqrestore(&z->lock, flags);
901 ret = __alloc_gigantic_page(pfn, nr_pages);
903 return pfn_to_page(pfn);
904 spin_lock_irqsave(&z->lock, flags);
909 spin_unlock_irqrestore(&z->lock, flags);
915 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
916 static void prep_compound_gigantic_page(struct page *page, unsigned long order);
918 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
922 page = alloc_gigantic_page(nid, huge_page_order(h));
924 prep_compound_gigantic_page(page, huge_page_order(h));
925 prep_new_huge_page(h, page, nid);
931 static int alloc_fresh_gigantic_page(struct hstate *h,
932 nodemask_t *nodes_allowed)
934 struct page *page = NULL;
937 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
938 page = alloc_fresh_gigantic_page_node(h, node);
946 static inline bool gigantic_page_supported(void) { return true; }
948 static inline bool gigantic_page_supported(void) { return false; }
949 static inline void free_gigantic_page(struct page *page, unsigned order) { }
950 static inline void destroy_compound_gigantic_page(struct page *page,
951 unsigned long order) { }
952 static inline int alloc_fresh_gigantic_page(struct hstate *h,
953 nodemask_t *nodes_allowed) { return 0; }
956 static void update_and_free_page(struct hstate *h, struct page *page)
960 if (hstate_is_gigantic(h) && !gigantic_page_supported())
964 h->nr_huge_pages_node[page_to_nid(page)]--;
965 for (i = 0; i < pages_per_huge_page(h); i++) {
966 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
967 1 << PG_referenced | 1 << PG_dirty |
968 1 << PG_active | 1 << PG_private |
971 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
972 set_compound_page_dtor(page, NULL);
973 set_page_refcounted(page);
974 if (hstate_is_gigantic(h)) {
975 destroy_compound_gigantic_page(page, huge_page_order(h));
976 free_gigantic_page(page, huge_page_order(h));
978 arch_release_hugepage(page);
979 __free_pages(page, huge_page_order(h));
983 struct hstate *size_to_hstate(unsigned long size)
988 if (huge_page_size(h) == size)
995 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
996 * to hstate->hugepage_activelist.)
998 * This function can be called for tail pages, but never returns true for them.
1000 bool page_huge_active(struct page *page)
1002 VM_BUG_ON_PAGE(!PageHuge(page), page);
1003 return PageHead(page) && PagePrivate(&page[1]);
1006 /* never called for tail page */
1007 static void set_page_huge_active(struct page *page)
1009 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1010 SetPagePrivate(&page[1]);
1013 static void clear_page_huge_active(struct page *page)
1015 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1016 ClearPagePrivate(&page[1]);
1019 void free_huge_page(struct page *page)
1022 * Can't pass hstate in here because it is called from the
1023 * compound page destructor.
1025 struct hstate *h = page_hstate(page);
1026 int nid = page_to_nid(page);
1027 struct hugepage_subpool *spool =
1028 (struct hugepage_subpool *)page_private(page);
1029 bool restore_reserve;
1031 set_page_private(page, 0);
1032 page->mapping = NULL;
1033 BUG_ON(page_count(page));
1034 BUG_ON(page_mapcount(page));
1035 restore_reserve = PagePrivate(page);
1036 ClearPagePrivate(page);
1039 * A return code of zero implies that the subpool will be under its
1040 * minimum size if the reservation is not restored after page is free.
1041 * Therefore, force restore_reserve operation.
1043 if (hugepage_subpool_put_pages(spool, 1) == 0)
1044 restore_reserve = true;
1046 spin_lock(&hugetlb_lock);
1047 clear_page_huge_active(page);
1048 hugetlb_cgroup_uncharge_page(hstate_index(h),
1049 pages_per_huge_page(h), page);
1050 if (restore_reserve)
1051 h->resv_huge_pages++;
1053 if (h->surplus_huge_pages_node[nid]) {
1054 /* remove the page from active list */
1055 list_del(&page->lru);
1056 update_and_free_page(h, page);
1057 h->surplus_huge_pages--;
1058 h->surplus_huge_pages_node[nid]--;
1060 arch_clear_hugepage_flags(page);
1061 enqueue_huge_page(h, page);
1063 spin_unlock(&hugetlb_lock);
1066 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1068 INIT_LIST_HEAD(&page->lru);
1069 set_compound_page_dtor(page, free_huge_page);
1070 spin_lock(&hugetlb_lock);
1071 set_hugetlb_cgroup(page, NULL);
1073 h->nr_huge_pages_node[nid]++;
1074 spin_unlock(&hugetlb_lock);
1075 put_page(page); /* free it into the hugepage allocator */
1078 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
1081 int nr_pages = 1 << order;
1082 struct page *p = page + 1;
1084 /* we rely on prep_new_huge_page to set the destructor */
1085 set_compound_order(page, order);
1086 __SetPageHead(page);
1087 __ClearPageReserved(page);
1088 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1090 * For gigantic hugepages allocated through bootmem at
1091 * boot, it's safer to be consistent with the not-gigantic
1092 * hugepages and clear the PG_reserved bit from all tail pages
1093 * too. Otherwse drivers using get_user_pages() to access tail
1094 * pages may get the reference counting wrong if they see
1095 * PG_reserved set on a tail page (despite the head page not
1096 * having PG_reserved set). Enforcing this consistency between
1097 * head and tail pages allows drivers to optimize away a check
1098 * on the head page when they need know if put_page() is needed
1099 * after get_user_pages().
1101 __ClearPageReserved(p);
1102 set_page_count(p, 0);
1103 p->first_page = page;
1104 /* Make sure p->first_page is always valid for PageTail() */
1111 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1112 * transparent huge pages. See the PageTransHuge() documentation for more
1115 int PageHuge(struct page *page)
1117 if (!PageCompound(page))
1120 page = compound_head(page);
1121 return get_compound_page_dtor(page) == free_huge_page;
1123 EXPORT_SYMBOL_GPL(PageHuge);
1126 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1127 * normal or transparent huge pages.
1129 int PageHeadHuge(struct page *page_head)
1131 if (!PageHead(page_head))
1134 return get_compound_page_dtor(page_head) == free_huge_page;
1137 pgoff_t __basepage_index(struct page *page)
1139 struct page *page_head = compound_head(page);
1140 pgoff_t index = page_index(page_head);
1141 unsigned long compound_idx;
1143 if (!PageHuge(page_head))
1144 return page_index(page);
1146 if (compound_order(page_head) >= MAX_ORDER)
1147 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1149 compound_idx = page - page_head;
1151 return (index << compound_order(page_head)) + compound_idx;
1154 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1158 page = alloc_pages_exact_node(nid,
1159 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1160 __GFP_REPEAT|__GFP_NOWARN,
1161 huge_page_order(h));
1163 if (arch_prepare_hugepage(page)) {
1164 __free_pages(page, huge_page_order(h));
1167 prep_new_huge_page(h, page, nid);
1173 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1179 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1180 page = alloc_fresh_huge_page_node(h, node);
1188 count_vm_event(HTLB_BUDDY_PGALLOC);
1190 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1196 * Free huge page from pool from next node to free.
1197 * Attempt to keep persistent huge pages more or less
1198 * balanced over allowed nodes.
1199 * Called with hugetlb_lock locked.
1201 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1207 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1209 * If we're returning unused surplus pages, only examine
1210 * nodes with surplus pages.
1212 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1213 !list_empty(&h->hugepage_freelists[node])) {
1215 list_entry(h->hugepage_freelists[node].next,
1217 list_del(&page->lru);
1218 h->free_huge_pages--;
1219 h->free_huge_pages_node[node]--;
1221 h->surplus_huge_pages--;
1222 h->surplus_huge_pages_node[node]--;
1224 update_and_free_page(h, page);
1234 * Dissolve a given free hugepage into free buddy pages. This function does
1235 * nothing for in-use (including surplus) hugepages.
1237 static void dissolve_free_huge_page(struct page *page)
1239 spin_lock(&hugetlb_lock);
1240 if (PageHuge(page) && !page_count(page)) {
1241 struct hstate *h = page_hstate(page);
1242 int nid = page_to_nid(page);
1243 list_del(&page->lru);
1244 h->free_huge_pages--;
1245 h->free_huge_pages_node[nid]--;
1246 update_and_free_page(h, page);
1248 spin_unlock(&hugetlb_lock);
1252 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1253 * make specified memory blocks removable from the system.
1254 * Note that start_pfn should aligned with (minimum) hugepage size.
1256 void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1260 if (!hugepages_supported())
1263 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order));
1264 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
1265 dissolve_free_huge_page(pfn_to_page(pfn));
1268 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
1273 if (hstate_is_gigantic(h))
1277 * Assume we will successfully allocate the surplus page to
1278 * prevent racing processes from causing the surplus to exceed
1281 * This however introduces a different race, where a process B
1282 * tries to grow the static hugepage pool while alloc_pages() is
1283 * called by process A. B will only examine the per-node
1284 * counters in determining if surplus huge pages can be
1285 * converted to normal huge pages in adjust_pool_surplus(). A
1286 * won't be able to increment the per-node counter, until the
1287 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1288 * no more huge pages can be converted from surplus to normal
1289 * state (and doesn't try to convert again). Thus, we have a
1290 * case where a surplus huge page exists, the pool is grown, and
1291 * the surplus huge page still exists after, even though it
1292 * should just have been converted to a normal huge page. This
1293 * does not leak memory, though, as the hugepage will be freed
1294 * once it is out of use. It also does not allow the counters to
1295 * go out of whack in adjust_pool_surplus() as we don't modify
1296 * the node values until we've gotten the hugepage and only the
1297 * per-node value is checked there.
1299 spin_lock(&hugetlb_lock);
1300 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1301 spin_unlock(&hugetlb_lock);
1305 h->surplus_huge_pages++;
1307 spin_unlock(&hugetlb_lock);
1309 if (nid == NUMA_NO_NODE)
1310 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
1311 __GFP_REPEAT|__GFP_NOWARN,
1312 huge_page_order(h));
1314 page = alloc_pages_exact_node(nid,
1315 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1316 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
1318 if (page && arch_prepare_hugepage(page)) {
1319 __free_pages(page, huge_page_order(h));
1323 spin_lock(&hugetlb_lock);
1325 INIT_LIST_HEAD(&page->lru);
1326 r_nid = page_to_nid(page);
1327 set_compound_page_dtor(page, free_huge_page);
1328 set_hugetlb_cgroup(page, NULL);
1330 * We incremented the global counters already
1332 h->nr_huge_pages_node[r_nid]++;
1333 h->surplus_huge_pages_node[r_nid]++;
1334 __count_vm_event(HTLB_BUDDY_PGALLOC);
1337 h->surplus_huge_pages--;
1338 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1340 spin_unlock(&hugetlb_lock);
1346 * This allocation function is useful in the context where vma is irrelevant.
1347 * E.g. soft-offlining uses this function because it only cares physical
1348 * address of error page.
1350 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1352 struct page *page = NULL;
1354 spin_lock(&hugetlb_lock);
1355 if (h->free_huge_pages - h->resv_huge_pages > 0)
1356 page = dequeue_huge_page_node(h, nid);
1357 spin_unlock(&hugetlb_lock);
1360 page = alloc_buddy_huge_page(h, nid);
1366 * Increase the hugetlb pool such that it can accommodate a reservation
1369 static int gather_surplus_pages(struct hstate *h, int delta)
1371 struct list_head surplus_list;
1372 struct page *page, *tmp;
1374 int needed, allocated;
1375 bool alloc_ok = true;
1377 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1379 h->resv_huge_pages += delta;
1384 INIT_LIST_HEAD(&surplus_list);
1388 spin_unlock(&hugetlb_lock);
1389 for (i = 0; i < needed; i++) {
1390 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1395 list_add(&page->lru, &surplus_list);
1400 * After retaking hugetlb_lock, we need to recalculate 'needed'
1401 * because either resv_huge_pages or free_huge_pages may have changed.
1403 spin_lock(&hugetlb_lock);
1404 needed = (h->resv_huge_pages + delta) -
1405 (h->free_huge_pages + allocated);
1410 * We were not able to allocate enough pages to
1411 * satisfy the entire reservation so we free what
1412 * we've allocated so far.
1417 * The surplus_list now contains _at_least_ the number of extra pages
1418 * needed to accommodate the reservation. Add the appropriate number
1419 * of pages to the hugetlb pool and free the extras back to the buddy
1420 * allocator. Commit the entire reservation here to prevent another
1421 * process from stealing the pages as they are added to the pool but
1422 * before they are reserved.
1424 needed += allocated;
1425 h->resv_huge_pages += delta;
1428 /* Free the needed pages to the hugetlb pool */
1429 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1433 * This page is now managed by the hugetlb allocator and has
1434 * no users -- drop the buddy allocator's reference.
1436 put_page_testzero(page);
1437 VM_BUG_ON_PAGE(page_count(page), page);
1438 enqueue_huge_page(h, page);
1441 spin_unlock(&hugetlb_lock);
1443 /* Free unnecessary surplus pages to the buddy allocator */
1444 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1446 spin_lock(&hugetlb_lock);
1452 * When releasing a hugetlb pool reservation, any surplus pages that were
1453 * allocated to satisfy the reservation must be explicitly freed if they were
1455 * Called with hugetlb_lock held.
1457 static void return_unused_surplus_pages(struct hstate *h,
1458 unsigned long unused_resv_pages)
1460 unsigned long nr_pages;
1462 /* Uncommit the reservation */
1463 h->resv_huge_pages -= unused_resv_pages;
1465 /* Cannot return gigantic pages currently */
1466 if (hstate_is_gigantic(h))
1469 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1472 * We want to release as many surplus pages as possible, spread
1473 * evenly across all nodes with memory. Iterate across these nodes
1474 * until we can no longer free unreserved surplus pages. This occurs
1475 * when the nodes with surplus pages have no free pages.
1476 * free_pool_huge_page() will balance the the freed pages across the
1477 * on-line nodes with memory and will handle the hstate accounting.
1479 while (nr_pages--) {
1480 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1482 cond_resched_lock(&hugetlb_lock);
1487 * vma_needs_reservation and vma_commit_reservation are used by the huge
1488 * page allocation routines to manage reservations.
1490 * vma_needs_reservation is called to determine if the huge page at addr
1491 * within the vma has an associated reservation. If a reservation is
1492 * needed, the value 1 is returned. The caller is then responsible for
1493 * managing the global reservation and subpool usage counts. After
1494 * the huge page has been allocated, vma_commit_reservation is called
1495 * to add the page to the reservation map.
1497 * In the normal case, vma_commit_reservation returns the same value
1498 * as the preceding vma_needs_reservation call. The only time this
1499 * is not the case is if a reserve map was changed between calls. It
1500 * is the responsibility of the caller to notice the difference and
1501 * take appropriate action.
1503 static long __vma_reservation_common(struct hstate *h,
1504 struct vm_area_struct *vma, unsigned long addr,
1507 struct resv_map *resv;
1511 resv = vma_resv_map(vma);
1515 idx = vma_hugecache_offset(h, vma, addr);
1517 ret = region_add(resv, idx, idx + 1);
1519 ret = region_chg(resv, idx, idx + 1);
1521 if (vma->vm_flags & VM_MAYSHARE)
1524 return ret < 0 ? ret : 0;
1527 static long vma_needs_reservation(struct hstate *h,
1528 struct vm_area_struct *vma, unsigned long addr)
1530 return __vma_reservation_common(h, vma, addr, false);
1533 static long vma_commit_reservation(struct hstate *h,
1534 struct vm_area_struct *vma, unsigned long addr)
1536 return __vma_reservation_common(h, vma, addr, true);
1539 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1540 unsigned long addr, int avoid_reserve)
1542 struct hugepage_subpool *spool = subpool_vma(vma);
1543 struct hstate *h = hstate_vma(vma);
1547 struct hugetlb_cgroup *h_cg;
1549 idx = hstate_index(h);
1551 * Processes that did not create the mapping will have no
1552 * reserves and will not have accounted against subpool
1553 * limit. Check that the subpool limit can be made before
1554 * satisfying the allocation MAP_NORESERVE mappings may also
1555 * need pages and subpool limit allocated allocated if no reserve
1558 chg = vma_needs_reservation(h, vma, addr);
1560 return ERR_PTR(-ENOMEM);
1561 if (chg || avoid_reserve)
1562 if (hugepage_subpool_get_pages(spool, 1) < 0)
1563 return ERR_PTR(-ENOSPC);
1565 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1567 goto out_subpool_put;
1569 spin_lock(&hugetlb_lock);
1570 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
1572 spin_unlock(&hugetlb_lock);
1573 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1575 goto out_uncharge_cgroup;
1577 spin_lock(&hugetlb_lock);
1578 list_move(&page->lru, &h->hugepage_activelist);
1581 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1582 spin_unlock(&hugetlb_lock);
1584 set_page_private(page, (unsigned long)spool);
1586 vma_commit_reservation(h, vma, addr);
1589 out_uncharge_cgroup:
1590 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1592 if (chg || avoid_reserve)
1593 hugepage_subpool_put_pages(spool, 1);
1594 return ERR_PTR(-ENOSPC);
1598 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1599 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1600 * where no ERR_VALUE is expected to be returned.
1602 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1603 unsigned long addr, int avoid_reserve)
1605 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1611 int __weak alloc_bootmem_huge_page(struct hstate *h)
1613 struct huge_bootmem_page *m;
1616 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1619 addr = memblock_virt_alloc_try_nid_nopanic(
1620 huge_page_size(h), huge_page_size(h),
1621 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1624 * Use the beginning of the huge page to store the
1625 * huge_bootmem_page struct (until gather_bootmem
1626 * puts them into the mem_map).
1635 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
1636 /* Put them into a private list first because mem_map is not up yet */
1637 list_add(&m->list, &huge_boot_pages);
1642 static void __init prep_compound_huge_page(struct page *page, int order)
1644 if (unlikely(order > (MAX_ORDER - 1)))
1645 prep_compound_gigantic_page(page, order);
1647 prep_compound_page(page, order);
1650 /* Put bootmem huge pages into the standard lists after mem_map is up */
1651 static void __init gather_bootmem_prealloc(void)
1653 struct huge_bootmem_page *m;
1655 list_for_each_entry(m, &huge_boot_pages, list) {
1656 struct hstate *h = m->hstate;
1659 #ifdef CONFIG_HIGHMEM
1660 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1661 memblock_free_late(__pa(m),
1662 sizeof(struct huge_bootmem_page));
1664 page = virt_to_page(m);
1666 WARN_ON(page_count(page) != 1);
1667 prep_compound_huge_page(page, h->order);
1668 WARN_ON(PageReserved(page));
1669 prep_new_huge_page(h, page, page_to_nid(page));
1671 * If we had gigantic hugepages allocated at boot time, we need
1672 * to restore the 'stolen' pages to totalram_pages in order to
1673 * fix confusing memory reports from free(1) and another
1674 * side-effects, like CommitLimit going negative.
1676 if (hstate_is_gigantic(h))
1677 adjust_managed_page_count(page, 1 << h->order);
1681 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1685 for (i = 0; i < h->max_huge_pages; ++i) {
1686 if (hstate_is_gigantic(h)) {
1687 if (!alloc_bootmem_huge_page(h))
1689 } else if (!alloc_fresh_huge_page(h,
1690 &node_states[N_MEMORY]))
1693 h->max_huge_pages = i;
1696 static void __init hugetlb_init_hstates(void)
1700 for_each_hstate(h) {
1701 if (minimum_order > huge_page_order(h))
1702 minimum_order = huge_page_order(h);
1704 /* oversize hugepages were init'ed in early boot */
1705 if (!hstate_is_gigantic(h))
1706 hugetlb_hstate_alloc_pages(h);
1708 VM_BUG_ON(minimum_order == UINT_MAX);
1711 static char * __init memfmt(char *buf, unsigned long n)
1713 if (n >= (1UL << 30))
1714 sprintf(buf, "%lu GB", n >> 30);
1715 else if (n >= (1UL << 20))
1716 sprintf(buf, "%lu MB", n >> 20);
1718 sprintf(buf, "%lu KB", n >> 10);
1722 static void __init report_hugepages(void)
1726 for_each_hstate(h) {
1728 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1729 memfmt(buf, huge_page_size(h)),
1730 h->free_huge_pages);
1734 #ifdef CONFIG_HIGHMEM
1735 static void try_to_free_low(struct hstate *h, unsigned long count,
1736 nodemask_t *nodes_allowed)
1740 if (hstate_is_gigantic(h))
1743 for_each_node_mask(i, *nodes_allowed) {
1744 struct page *page, *next;
1745 struct list_head *freel = &h->hugepage_freelists[i];
1746 list_for_each_entry_safe(page, next, freel, lru) {
1747 if (count >= h->nr_huge_pages)
1749 if (PageHighMem(page))
1751 list_del(&page->lru);
1752 update_and_free_page(h, page);
1753 h->free_huge_pages--;
1754 h->free_huge_pages_node[page_to_nid(page)]--;
1759 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1760 nodemask_t *nodes_allowed)
1766 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1767 * balanced by operating on them in a round-robin fashion.
1768 * Returns 1 if an adjustment was made.
1770 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1775 VM_BUG_ON(delta != -1 && delta != 1);
1778 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1779 if (h->surplus_huge_pages_node[node])
1783 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1784 if (h->surplus_huge_pages_node[node] <
1785 h->nr_huge_pages_node[node])
1792 h->surplus_huge_pages += delta;
1793 h->surplus_huge_pages_node[node] += delta;
1797 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1798 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1799 nodemask_t *nodes_allowed)
1801 unsigned long min_count, ret;
1803 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1804 return h->max_huge_pages;
1807 * Increase the pool size
1808 * First take pages out of surplus state. Then make up the
1809 * remaining difference by allocating fresh huge pages.
1811 * We might race with alloc_buddy_huge_page() here and be unable
1812 * to convert a surplus huge page to a normal huge page. That is
1813 * not critical, though, it just means the overall size of the
1814 * pool might be one hugepage larger than it needs to be, but
1815 * within all the constraints specified by the sysctls.
1817 spin_lock(&hugetlb_lock);
1818 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1819 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1823 while (count > persistent_huge_pages(h)) {
1825 * If this allocation races such that we no longer need the
1826 * page, free_huge_page will handle it by freeing the page
1827 * and reducing the surplus.
1829 spin_unlock(&hugetlb_lock);
1830 if (hstate_is_gigantic(h))
1831 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
1833 ret = alloc_fresh_huge_page(h, nodes_allowed);
1834 spin_lock(&hugetlb_lock);
1838 /* Bail for signals. Probably ctrl-c from user */
1839 if (signal_pending(current))
1844 * Decrease the pool size
1845 * First return free pages to the buddy allocator (being careful
1846 * to keep enough around to satisfy reservations). Then place
1847 * pages into surplus state as needed so the pool will shrink
1848 * to the desired size as pages become free.
1850 * By placing pages into the surplus state independent of the
1851 * overcommit value, we are allowing the surplus pool size to
1852 * exceed overcommit. There are few sane options here. Since
1853 * alloc_buddy_huge_page() is checking the global counter,
1854 * though, we'll note that we're not allowed to exceed surplus
1855 * and won't grow the pool anywhere else. Not until one of the
1856 * sysctls are changed, or the surplus pages go out of use.
1858 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1859 min_count = max(count, min_count);
1860 try_to_free_low(h, min_count, nodes_allowed);
1861 while (min_count < persistent_huge_pages(h)) {
1862 if (!free_pool_huge_page(h, nodes_allowed, 0))
1864 cond_resched_lock(&hugetlb_lock);
1866 while (count < persistent_huge_pages(h)) {
1867 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1871 ret = persistent_huge_pages(h);
1872 spin_unlock(&hugetlb_lock);
1876 #define HSTATE_ATTR_RO(_name) \
1877 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1879 #define HSTATE_ATTR(_name) \
1880 static struct kobj_attribute _name##_attr = \
1881 __ATTR(_name, 0644, _name##_show, _name##_store)
1883 static struct kobject *hugepages_kobj;
1884 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1886 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1888 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1892 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1893 if (hstate_kobjs[i] == kobj) {
1895 *nidp = NUMA_NO_NODE;
1899 return kobj_to_node_hstate(kobj, nidp);
1902 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1903 struct kobj_attribute *attr, char *buf)
1906 unsigned long nr_huge_pages;
1909 h = kobj_to_hstate(kobj, &nid);
1910 if (nid == NUMA_NO_NODE)
1911 nr_huge_pages = h->nr_huge_pages;
1913 nr_huge_pages = h->nr_huge_pages_node[nid];
1915 return sprintf(buf, "%lu\n", nr_huge_pages);
1918 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
1919 struct hstate *h, int nid,
1920 unsigned long count, size_t len)
1923 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1925 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
1930 if (nid == NUMA_NO_NODE) {
1932 * global hstate attribute
1934 if (!(obey_mempolicy &&
1935 init_nodemask_of_mempolicy(nodes_allowed))) {
1936 NODEMASK_FREE(nodes_allowed);
1937 nodes_allowed = &node_states[N_MEMORY];
1939 } else if (nodes_allowed) {
1941 * per node hstate attribute: adjust count to global,
1942 * but restrict alloc/free to the specified node.
1944 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1945 init_nodemask_of_node(nodes_allowed, nid);
1947 nodes_allowed = &node_states[N_MEMORY];
1949 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1951 if (nodes_allowed != &node_states[N_MEMORY])
1952 NODEMASK_FREE(nodes_allowed);
1956 NODEMASK_FREE(nodes_allowed);
1960 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1961 struct kobject *kobj, const char *buf,
1965 unsigned long count;
1969 err = kstrtoul(buf, 10, &count);
1973 h = kobj_to_hstate(kobj, &nid);
1974 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
1977 static ssize_t nr_hugepages_show(struct kobject *kobj,
1978 struct kobj_attribute *attr, char *buf)
1980 return nr_hugepages_show_common(kobj, attr, buf);
1983 static ssize_t nr_hugepages_store(struct kobject *kobj,
1984 struct kobj_attribute *attr, const char *buf, size_t len)
1986 return nr_hugepages_store_common(false, kobj, buf, len);
1988 HSTATE_ATTR(nr_hugepages);
1993 * hstate attribute for optionally mempolicy-based constraint on persistent
1994 * huge page alloc/free.
1996 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1997 struct kobj_attribute *attr, char *buf)
1999 return nr_hugepages_show_common(kobj, attr, buf);
2002 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2003 struct kobj_attribute *attr, const char *buf, size_t len)
2005 return nr_hugepages_store_common(true, kobj, buf, len);
2007 HSTATE_ATTR(nr_hugepages_mempolicy);
2011 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2012 struct kobj_attribute *attr, char *buf)
2014 struct hstate *h = kobj_to_hstate(kobj, NULL);
2015 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2018 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2019 struct kobj_attribute *attr, const char *buf, size_t count)
2022 unsigned long input;
2023 struct hstate *h = kobj_to_hstate(kobj, NULL);
2025 if (hstate_is_gigantic(h))
2028 err = kstrtoul(buf, 10, &input);
2032 spin_lock(&hugetlb_lock);
2033 h->nr_overcommit_huge_pages = input;
2034 spin_unlock(&hugetlb_lock);
2038 HSTATE_ATTR(nr_overcommit_hugepages);
2040 static ssize_t free_hugepages_show(struct kobject *kobj,
2041 struct kobj_attribute *attr, char *buf)
2044 unsigned long free_huge_pages;
2047 h = kobj_to_hstate(kobj, &nid);
2048 if (nid == NUMA_NO_NODE)
2049 free_huge_pages = h->free_huge_pages;
2051 free_huge_pages = h->free_huge_pages_node[nid];
2053 return sprintf(buf, "%lu\n", free_huge_pages);
2055 HSTATE_ATTR_RO(free_hugepages);
2057 static ssize_t resv_hugepages_show(struct kobject *kobj,
2058 struct kobj_attribute *attr, char *buf)
2060 struct hstate *h = kobj_to_hstate(kobj, NULL);
2061 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2063 HSTATE_ATTR_RO(resv_hugepages);
2065 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2066 struct kobj_attribute *attr, char *buf)
2069 unsigned long surplus_huge_pages;
2072 h = kobj_to_hstate(kobj, &nid);
2073 if (nid == NUMA_NO_NODE)
2074 surplus_huge_pages = h->surplus_huge_pages;
2076 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2078 return sprintf(buf, "%lu\n", surplus_huge_pages);
2080 HSTATE_ATTR_RO(surplus_hugepages);
2082 static struct attribute *hstate_attrs[] = {
2083 &nr_hugepages_attr.attr,
2084 &nr_overcommit_hugepages_attr.attr,
2085 &free_hugepages_attr.attr,
2086 &resv_hugepages_attr.attr,
2087 &surplus_hugepages_attr.attr,
2089 &nr_hugepages_mempolicy_attr.attr,
2094 static struct attribute_group hstate_attr_group = {
2095 .attrs = hstate_attrs,
2098 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2099 struct kobject **hstate_kobjs,
2100 struct attribute_group *hstate_attr_group)
2103 int hi = hstate_index(h);
2105 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2106 if (!hstate_kobjs[hi])
2109 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2111 kobject_put(hstate_kobjs[hi]);
2116 static void __init hugetlb_sysfs_init(void)
2121 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2122 if (!hugepages_kobj)
2125 for_each_hstate(h) {
2126 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2127 hstate_kobjs, &hstate_attr_group);
2129 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2136 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2137 * with node devices in node_devices[] using a parallel array. The array
2138 * index of a node device or _hstate == node id.
2139 * This is here to avoid any static dependency of the node device driver, in
2140 * the base kernel, on the hugetlb module.
2142 struct node_hstate {
2143 struct kobject *hugepages_kobj;
2144 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2146 struct node_hstate node_hstates[MAX_NUMNODES];
2149 * A subset of global hstate attributes for node devices
2151 static struct attribute *per_node_hstate_attrs[] = {
2152 &nr_hugepages_attr.attr,
2153 &free_hugepages_attr.attr,
2154 &surplus_hugepages_attr.attr,
2158 static struct attribute_group per_node_hstate_attr_group = {
2159 .attrs = per_node_hstate_attrs,
2163 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2164 * Returns node id via non-NULL nidp.
2166 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2170 for (nid = 0; nid < nr_node_ids; nid++) {
2171 struct node_hstate *nhs = &node_hstates[nid];
2173 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2174 if (nhs->hstate_kobjs[i] == kobj) {
2186 * Unregister hstate attributes from a single node device.
2187 * No-op if no hstate attributes attached.
2189 static void hugetlb_unregister_node(struct node *node)
2192 struct node_hstate *nhs = &node_hstates[node->dev.id];
2194 if (!nhs->hugepages_kobj)
2195 return; /* no hstate attributes */
2197 for_each_hstate(h) {
2198 int idx = hstate_index(h);
2199 if (nhs->hstate_kobjs[idx]) {
2200 kobject_put(nhs->hstate_kobjs[idx]);
2201 nhs->hstate_kobjs[idx] = NULL;
2205 kobject_put(nhs->hugepages_kobj);
2206 nhs->hugepages_kobj = NULL;
2210 * hugetlb module exit: unregister hstate attributes from node devices
2213 static void hugetlb_unregister_all_nodes(void)
2218 * disable node device registrations.
2220 register_hugetlbfs_with_node(NULL, NULL);
2223 * remove hstate attributes from any nodes that have them.
2225 for (nid = 0; nid < nr_node_ids; nid++)
2226 hugetlb_unregister_node(node_devices[nid]);
2230 * Register hstate attributes for a single node device.
2231 * No-op if attributes already registered.
2233 static void hugetlb_register_node(struct node *node)
2236 struct node_hstate *nhs = &node_hstates[node->dev.id];
2239 if (nhs->hugepages_kobj)
2240 return; /* already allocated */
2242 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2244 if (!nhs->hugepages_kobj)
2247 for_each_hstate(h) {
2248 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2250 &per_node_hstate_attr_group);
2252 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2253 h->name, node->dev.id);
2254 hugetlb_unregister_node(node);
2261 * hugetlb init time: register hstate attributes for all registered node
2262 * devices of nodes that have memory. All on-line nodes should have
2263 * registered their associated device by this time.
2265 static void __init hugetlb_register_all_nodes(void)
2269 for_each_node_state(nid, N_MEMORY) {
2270 struct node *node = node_devices[nid];
2271 if (node->dev.id == nid)
2272 hugetlb_register_node(node);
2276 * Let the node device driver know we're here so it can
2277 * [un]register hstate attributes on node hotplug.
2279 register_hugetlbfs_with_node(hugetlb_register_node,
2280 hugetlb_unregister_node);
2282 #else /* !CONFIG_NUMA */
2284 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2292 static void hugetlb_unregister_all_nodes(void) { }
2294 static void hugetlb_register_all_nodes(void) { }
2298 static void __exit hugetlb_exit(void)
2302 hugetlb_unregister_all_nodes();
2304 for_each_hstate(h) {
2305 kobject_put(hstate_kobjs[hstate_index(h)]);
2308 kobject_put(hugepages_kobj);
2309 kfree(htlb_fault_mutex_table);
2311 module_exit(hugetlb_exit);
2313 static int __init hugetlb_init(void)
2317 if (!hugepages_supported())
2320 if (!size_to_hstate(default_hstate_size)) {
2321 default_hstate_size = HPAGE_SIZE;
2322 if (!size_to_hstate(default_hstate_size))
2323 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2325 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2326 if (default_hstate_max_huge_pages)
2327 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2329 hugetlb_init_hstates();
2330 gather_bootmem_prealloc();
2333 hugetlb_sysfs_init();
2334 hugetlb_register_all_nodes();
2335 hugetlb_cgroup_file_init();
2338 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2340 num_fault_mutexes = 1;
2342 htlb_fault_mutex_table =
2343 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2344 BUG_ON(!htlb_fault_mutex_table);
2346 for (i = 0; i < num_fault_mutexes; i++)
2347 mutex_init(&htlb_fault_mutex_table[i]);
2350 module_init(hugetlb_init);
2352 /* Should be called on processing a hugepagesz=... option */
2353 void __init hugetlb_add_hstate(unsigned order)
2358 if (size_to_hstate(PAGE_SIZE << order)) {
2359 pr_warning("hugepagesz= specified twice, ignoring\n");
2362 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2364 h = &hstates[hugetlb_max_hstate++];
2366 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2367 h->nr_huge_pages = 0;
2368 h->free_huge_pages = 0;
2369 for (i = 0; i < MAX_NUMNODES; ++i)
2370 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2371 INIT_LIST_HEAD(&h->hugepage_activelist);
2372 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2373 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2374 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2375 huge_page_size(h)/1024);
2380 static int __init hugetlb_nrpages_setup(char *s)
2383 static unsigned long *last_mhp;
2386 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2387 * so this hugepages= parameter goes to the "default hstate".
2389 if (!hugetlb_max_hstate)
2390 mhp = &default_hstate_max_huge_pages;
2392 mhp = &parsed_hstate->max_huge_pages;
2394 if (mhp == last_mhp) {
2395 pr_warning("hugepages= specified twice without "
2396 "interleaving hugepagesz=, ignoring\n");
2400 if (sscanf(s, "%lu", mhp) <= 0)
2404 * Global state is always initialized later in hugetlb_init.
2405 * But we need to allocate >= MAX_ORDER hstates here early to still
2406 * use the bootmem allocator.
2408 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2409 hugetlb_hstate_alloc_pages(parsed_hstate);
2415 __setup("hugepages=", hugetlb_nrpages_setup);
2417 static int __init hugetlb_default_setup(char *s)
2419 default_hstate_size = memparse(s, &s);
2422 __setup("default_hugepagesz=", hugetlb_default_setup);
2424 static unsigned int cpuset_mems_nr(unsigned int *array)
2427 unsigned int nr = 0;
2429 for_each_node_mask(node, cpuset_current_mems_allowed)
2435 #ifdef CONFIG_SYSCTL
2436 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2437 struct ctl_table *table, int write,
2438 void __user *buffer, size_t *length, loff_t *ppos)
2440 struct hstate *h = &default_hstate;
2441 unsigned long tmp = h->max_huge_pages;
2444 if (!hugepages_supported())
2448 table->maxlen = sizeof(unsigned long);
2449 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2454 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2455 NUMA_NO_NODE, tmp, *length);
2460 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2461 void __user *buffer, size_t *length, loff_t *ppos)
2464 return hugetlb_sysctl_handler_common(false, table, write,
2465 buffer, length, ppos);
2469 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2470 void __user *buffer, size_t *length, loff_t *ppos)
2472 return hugetlb_sysctl_handler_common(true, table, write,
2473 buffer, length, ppos);
2475 #endif /* CONFIG_NUMA */
2477 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2478 void __user *buffer,
2479 size_t *length, loff_t *ppos)
2481 struct hstate *h = &default_hstate;
2485 if (!hugepages_supported())
2488 tmp = h->nr_overcommit_huge_pages;
2490 if (write && hstate_is_gigantic(h))
2494 table->maxlen = sizeof(unsigned long);
2495 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2500 spin_lock(&hugetlb_lock);
2501 h->nr_overcommit_huge_pages = tmp;
2502 spin_unlock(&hugetlb_lock);
2508 #endif /* CONFIG_SYSCTL */
2510 void hugetlb_report_meminfo(struct seq_file *m)
2512 struct hstate *h = &default_hstate;
2513 if (!hugepages_supported())
2516 "HugePages_Total: %5lu\n"
2517 "HugePages_Free: %5lu\n"
2518 "HugePages_Rsvd: %5lu\n"
2519 "HugePages_Surp: %5lu\n"
2520 "Hugepagesize: %8lu kB\n",
2524 h->surplus_huge_pages,
2525 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2528 int hugetlb_report_node_meminfo(int nid, char *buf)
2530 struct hstate *h = &default_hstate;
2531 if (!hugepages_supported())
2534 "Node %d HugePages_Total: %5u\n"
2535 "Node %d HugePages_Free: %5u\n"
2536 "Node %d HugePages_Surp: %5u\n",
2537 nid, h->nr_huge_pages_node[nid],
2538 nid, h->free_huge_pages_node[nid],
2539 nid, h->surplus_huge_pages_node[nid]);
2542 void hugetlb_show_meminfo(void)
2547 if (!hugepages_supported())
2550 for_each_node_state(nid, N_MEMORY)
2552 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2554 h->nr_huge_pages_node[nid],
2555 h->free_huge_pages_node[nid],
2556 h->surplus_huge_pages_node[nid],
2557 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2560 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2561 unsigned long hugetlb_total_pages(void)
2564 unsigned long nr_total_pages = 0;
2567 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2568 return nr_total_pages;
2571 static int hugetlb_acct_memory(struct hstate *h, long delta)
2575 spin_lock(&hugetlb_lock);
2577 * When cpuset is configured, it breaks the strict hugetlb page
2578 * reservation as the accounting is done on a global variable. Such
2579 * reservation is completely rubbish in the presence of cpuset because
2580 * the reservation is not checked against page availability for the
2581 * current cpuset. Application can still potentially OOM'ed by kernel
2582 * with lack of free htlb page in cpuset that the task is in.
2583 * Attempt to enforce strict accounting with cpuset is almost
2584 * impossible (or too ugly) because cpuset is too fluid that
2585 * task or memory node can be dynamically moved between cpusets.
2587 * The change of semantics for shared hugetlb mapping with cpuset is
2588 * undesirable. However, in order to preserve some of the semantics,
2589 * we fall back to check against current free page availability as
2590 * a best attempt and hopefully to minimize the impact of changing
2591 * semantics that cpuset has.
2594 if (gather_surplus_pages(h, delta) < 0)
2597 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2598 return_unused_surplus_pages(h, delta);
2605 return_unused_surplus_pages(h, (unsigned long) -delta);
2608 spin_unlock(&hugetlb_lock);
2612 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2614 struct resv_map *resv = vma_resv_map(vma);
2617 * This new VMA should share its siblings reservation map if present.
2618 * The VMA will only ever have a valid reservation map pointer where
2619 * it is being copied for another still existing VMA. As that VMA
2620 * has a reference to the reservation map it cannot disappear until
2621 * after this open call completes. It is therefore safe to take a
2622 * new reference here without additional locking.
2624 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2625 kref_get(&resv->refs);
2628 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2630 struct hstate *h = hstate_vma(vma);
2631 struct resv_map *resv = vma_resv_map(vma);
2632 struct hugepage_subpool *spool = subpool_vma(vma);
2633 unsigned long reserve, start, end;
2636 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2639 start = vma_hugecache_offset(h, vma, vma->vm_start);
2640 end = vma_hugecache_offset(h, vma, vma->vm_end);
2642 reserve = (end - start) - region_count(resv, start, end);
2644 kref_put(&resv->refs, resv_map_release);
2648 * Decrement reserve counts. The global reserve count may be
2649 * adjusted if the subpool has a minimum size.
2651 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
2652 hugetlb_acct_memory(h, -gbl_reserve);
2657 * We cannot handle pagefaults against hugetlb pages at all. They cause
2658 * handle_mm_fault() to try to instantiate regular-sized pages in the
2659 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2662 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2668 const struct vm_operations_struct hugetlb_vm_ops = {
2669 .fault = hugetlb_vm_op_fault,
2670 .open = hugetlb_vm_op_open,
2671 .close = hugetlb_vm_op_close,
2674 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2680 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2681 vma->vm_page_prot)));
2683 entry = huge_pte_wrprotect(mk_huge_pte(page,
2684 vma->vm_page_prot));
2686 entry = pte_mkyoung(entry);
2687 entry = pte_mkhuge(entry);
2688 entry = arch_make_huge_pte(entry, vma, page, writable);
2693 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2694 unsigned long address, pte_t *ptep)
2698 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2699 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2700 update_mmu_cache(vma, address, ptep);
2703 static int is_hugetlb_entry_migration(pte_t pte)
2707 if (huge_pte_none(pte) || pte_present(pte))
2709 swp = pte_to_swp_entry(pte);
2710 if (non_swap_entry(swp) && is_migration_entry(swp))
2716 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2720 if (huge_pte_none(pte) || pte_present(pte))
2722 swp = pte_to_swp_entry(pte);
2723 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2729 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2730 struct vm_area_struct *vma)
2732 pte_t *src_pte, *dst_pte, entry;
2733 struct page *ptepage;
2736 struct hstate *h = hstate_vma(vma);
2737 unsigned long sz = huge_page_size(h);
2738 unsigned long mmun_start; /* For mmu_notifiers */
2739 unsigned long mmun_end; /* For mmu_notifiers */
2742 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2744 mmun_start = vma->vm_start;
2745 mmun_end = vma->vm_end;
2747 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2749 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2750 spinlock_t *src_ptl, *dst_ptl;
2751 src_pte = huge_pte_offset(src, addr);
2754 dst_pte = huge_pte_alloc(dst, addr, sz);
2760 /* If the pagetables are shared don't copy or take references */
2761 if (dst_pte == src_pte)
2764 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2765 src_ptl = huge_pte_lockptr(h, src, src_pte);
2766 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
2767 entry = huge_ptep_get(src_pte);
2768 if (huge_pte_none(entry)) { /* skip none entry */
2770 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2771 is_hugetlb_entry_hwpoisoned(entry))) {
2772 swp_entry_t swp_entry = pte_to_swp_entry(entry);
2774 if (is_write_migration_entry(swp_entry) && cow) {
2776 * COW mappings require pages in both
2777 * parent and child to be set to read.
2779 make_migration_entry_read(&swp_entry);
2780 entry = swp_entry_to_pte(swp_entry);
2781 set_huge_pte_at(src, addr, src_pte, entry);
2783 set_huge_pte_at(dst, addr, dst_pte, entry);
2786 huge_ptep_set_wrprotect(src, addr, src_pte);
2787 mmu_notifier_invalidate_range(src, mmun_start,
2790 entry = huge_ptep_get(src_pte);
2791 ptepage = pte_page(entry);
2793 page_dup_rmap(ptepage);
2794 set_huge_pte_at(dst, addr, dst_pte, entry);
2796 spin_unlock(src_ptl);
2797 spin_unlock(dst_ptl);
2801 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
2806 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2807 unsigned long start, unsigned long end,
2808 struct page *ref_page)
2810 int force_flush = 0;
2811 struct mm_struct *mm = vma->vm_mm;
2812 unsigned long address;
2817 struct hstate *h = hstate_vma(vma);
2818 unsigned long sz = huge_page_size(h);
2819 const unsigned long mmun_start = start; /* For mmu_notifiers */
2820 const unsigned long mmun_end = end; /* For mmu_notifiers */
2822 WARN_ON(!is_vm_hugetlb_page(vma));
2823 BUG_ON(start & ~huge_page_mask(h));
2824 BUG_ON(end & ~huge_page_mask(h));
2826 tlb_start_vma(tlb, vma);
2827 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2830 for (; address < end; address += sz) {
2831 ptep = huge_pte_offset(mm, address);
2835 ptl = huge_pte_lock(h, mm, ptep);
2836 if (huge_pmd_unshare(mm, &address, ptep))
2839 pte = huge_ptep_get(ptep);
2840 if (huge_pte_none(pte))
2844 * Migrating hugepage or HWPoisoned hugepage is already
2845 * unmapped and its refcount is dropped, so just clear pte here.
2847 if (unlikely(!pte_present(pte))) {
2848 huge_pte_clear(mm, address, ptep);
2852 page = pte_page(pte);
2854 * If a reference page is supplied, it is because a specific
2855 * page is being unmapped, not a range. Ensure the page we
2856 * are about to unmap is the actual page of interest.
2859 if (page != ref_page)
2863 * Mark the VMA as having unmapped its page so that
2864 * future faults in this VMA will fail rather than
2865 * looking like data was lost
2867 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2870 pte = huge_ptep_get_and_clear(mm, address, ptep);
2871 tlb_remove_tlb_entry(tlb, ptep, address);
2872 if (huge_pte_dirty(pte))
2873 set_page_dirty(page);
2875 page_remove_rmap(page);
2876 force_flush = !__tlb_remove_page(tlb, page);
2882 /* Bail out after unmapping reference page if supplied */
2891 * mmu_gather ran out of room to batch pages, we break out of
2892 * the PTE lock to avoid doing the potential expensive TLB invalidate
2893 * and page-free while holding it.
2898 if (address < end && !ref_page)
2901 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2902 tlb_end_vma(tlb, vma);
2905 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2906 struct vm_area_struct *vma, unsigned long start,
2907 unsigned long end, struct page *ref_page)
2909 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2912 * Clear this flag so that x86's huge_pmd_share page_table_shareable
2913 * test will fail on a vma being torn down, and not grab a page table
2914 * on its way out. We're lucky that the flag has such an appropriate
2915 * name, and can in fact be safely cleared here. We could clear it
2916 * before the __unmap_hugepage_range above, but all that's necessary
2917 * is to clear it before releasing the i_mmap_rwsem. This works
2918 * because in the context this is called, the VMA is about to be
2919 * destroyed and the i_mmap_rwsem is held.
2921 vma->vm_flags &= ~VM_MAYSHARE;
2924 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2925 unsigned long end, struct page *ref_page)
2927 struct mm_struct *mm;
2928 struct mmu_gather tlb;
2932 tlb_gather_mmu(&tlb, mm, start, end);
2933 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2934 tlb_finish_mmu(&tlb, start, end);
2938 * This is called when the original mapper is failing to COW a MAP_PRIVATE
2939 * mappping it owns the reserve page for. The intention is to unmap the page
2940 * from other VMAs and let the children be SIGKILLed if they are faulting the
2943 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2944 struct page *page, unsigned long address)
2946 struct hstate *h = hstate_vma(vma);
2947 struct vm_area_struct *iter_vma;
2948 struct address_space *mapping;
2952 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2953 * from page cache lookup which is in HPAGE_SIZE units.
2955 address = address & huge_page_mask(h);
2956 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2958 mapping = file_inode(vma->vm_file)->i_mapping;
2961 * Take the mapping lock for the duration of the table walk. As
2962 * this mapping should be shared between all the VMAs,
2963 * __unmap_hugepage_range() is called as the lock is already held
2965 i_mmap_lock_write(mapping);
2966 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2967 /* Do not unmap the current VMA */
2968 if (iter_vma == vma)
2972 * Unmap the page from other VMAs without their own reserves.
2973 * They get marked to be SIGKILLed if they fault in these
2974 * areas. This is because a future no-page fault on this VMA
2975 * could insert a zeroed page instead of the data existing
2976 * from the time of fork. This would look like data corruption
2978 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2979 unmap_hugepage_range(iter_vma, address,
2980 address + huge_page_size(h), page);
2982 i_mmap_unlock_write(mapping);
2986 * Hugetlb_cow() should be called with page lock of the original hugepage held.
2987 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2988 * cannot race with other handlers or page migration.
2989 * Keep the pte_same checks anyway to make transition from the mutex easier.
2991 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2992 unsigned long address, pte_t *ptep, pte_t pte,
2993 struct page *pagecache_page, spinlock_t *ptl)
2995 struct hstate *h = hstate_vma(vma);
2996 struct page *old_page, *new_page;
2997 int ret = 0, outside_reserve = 0;
2998 unsigned long mmun_start; /* For mmu_notifiers */
2999 unsigned long mmun_end; /* For mmu_notifiers */
3001 old_page = pte_page(pte);
3004 /* If no-one else is actually using this page, avoid the copy
3005 * and just make the page writable */
3006 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3007 page_move_anon_rmap(old_page, vma, address);
3008 set_huge_ptep_writable(vma, address, ptep);
3013 * If the process that created a MAP_PRIVATE mapping is about to
3014 * perform a COW due to a shared page count, attempt to satisfy
3015 * the allocation without using the existing reserves. The pagecache
3016 * page is used to determine if the reserve at this address was
3017 * consumed or not. If reserves were used, a partial faulted mapping
3018 * at the time of fork() could consume its reserves on COW instead
3019 * of the full address range.
3021 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3022 old_page != pagecache_page)
3023 outside_reserve = 1;
3025 page_cache_get(old_page);
3028 * Drop page table lock as buddy allocator may be called. It will
3029 * be acquired again before returning to the caller, as expected.
3032 new_page = alloc_huge_page(vma, address, outside_reserve);
3034 if (IS_ERR(new_page)) {
3036 * If a process owning a MAP_PRIVATE mapping fails to COW,
3037 * it is due to references held by a child and an insufficient
3038 * huge page pool. To guarantee the original mappers
3039 * reliability, unmap the page from child processes. The child
3040 * may get SIGKILLed if it later faults.
3042 if (outside_reserve) {
3043 page_cache_release(old_page);
3044 BUG_ON(huge_pte_none(pte));
3045 unmap_ref_private(mm, vma, old_page, address);
3046 BUG_ON(huge_pte_none(pte));
3048 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3050 pte_same(huge_ptep_get(ptep), pte)))
3051 goto retry_avoidcopy;
3053 * race occurs while re-acquiring page table
3054 * lock, and our job is done.
3059 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3060 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3061 goto out_release_old;
3065 * When the original hugepage is shared one, it does not have
3066 * anon_vma prepared.
3068 if (unlikely(anon_vma_prepare(vma))) {
3070 goto out_release_all;
3073 copy_user_huge_page(new_page, old_page, address, vma,
3074 pages_per_huge_page(h));
3075 __SetPageUptodate(new_page);
3076 set_page_huge_active(new_page);
3078 mmun_start = address & huge_page_mask(h);
3079 mmun_end = mmun_start + huge_page_size(h);
3080 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3083 * Retake the page table lock to check for racing updates
3084 * before the page tables are altered
3087 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3088 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3089 ClearPagePrivate(new_page);
3092 huge_ptep_clear_flush(vma, address, ptep);
3093 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3094 set_huge_pte_at(mm, address, ptep,
3095 make_huge_pte(vma, new_page, 1));
3096 page_remove_rmap(old_page);
3097 hugepage_add_new_anon_rmap(new_page, vma, address);
3098 /* Make the old page be freed below */
3099 new_page = old_page;
3102 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3104 page_cache_release(new_page);
3106 page_cache_release(old_page);
3108 spin_lock(ptl); /* Caller expects lock to be held */
3112 /* Return the pagecache page at a given address within a VMA */
3113 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3114 struct vm_area_struct *vma, unsigned long address)
3116 struct address_space *mapping;
3119 mapping = vma->vm_file->f_mapping;
3120 idx = vma_hugecache_offset(h, vma, address);
3122 return find_lock_page(mapping, idx);
3126 * Return whether there is a pagecache page to back given address within VMA.
3127 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3129 static bool hugetlbfs_pagecache_present(struct hstate *h,
3130 struct vm_area_struct *vma, unsigned long address)
3132 struct address_space *mapping;
3136 mapping = vma->vm_file->f_mapping;
3137 idx = vma_hugecache_offset(h, vma, address);
3139 page = find_get_page(mapping, idx);
3142 return page != NULL;
3145 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3146 struct address_space *mapping, pgoff_t idx,
3147 unsigned long address, pte_t *ptep, unsigned int flags)
3149 struct hstate *h = hstate_vma(vma);
3150 int ret = VM_FAULT_SIGBUS;
3158 * Currently, we are forced to kill the process in the event the
3159 * original mapper has unmapped pages from the child due to a failed
3160 * COW. Warn that such a situation has occurred as it may not be obvious
3162 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3163 pr_warning("PID %d killed due to inadequate hugepage pool\n",
3169 * Use page lock to guard against racing truncation
3170 * before we get page_table_lock.
3173 page = find_lock_page(mapping, idx);
3175 size = i_size_read(mapping->host) >> huge_page_shift(h);
3178 page = alloc_huge_page(vma, address, 0);
3180 ret = PTR_ERR(page);
3184 ret = VM_FAULT_SIGBUS;
3187 clear_huge_page(page, address, pages_per_huge_page(h));
3188 __SetPageUptodate(page);
3189 set_page_huge_active(page);
3191 if (vma->vm_flags & VM_MAYSHARE) {
3193 struct inode *inode = mapping->host;
3195 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3202 ClearPagePrivate(page);
3204 spin_lock(&inode->i_lock);
3205 inode->i_blocks += blocks_per_huge_page(h);
3206 spin_unlock(&inode->i_lock);
3209 if (unlikely(anon_vma_prepare(vma))) {
3211 goto backout_unlocked;
3217 * If memory error occurs between mmap() and fault, some process
3218 * don't have hwpoisoned swap entry for errored virtual address.
3219 * So we need to block hugepage fault by PG_hwpoison bit check.
3221 if (unlikely(PageHWPoison(page))) {
3222 ret = VM_FAULT_HWPOISON |
3223 VM_FAULT_SET_HINDEX(hstate_index(h));
3224 goto backout_unlocked;
3229 * If we are going to COW a private mapping later, we examine the
3230 * pending reservations for this page now. This will ensure that
3231 * any allocations necessary to record that reservation occur outside
3234 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
3235 if (vma_needs_reservation(h, vma, address) < 0) {
3237 goto backout_unlocked;
3240 ptl = huge_pte_lockptr(h, mm, ptep);
3242 size = i_size_read(mapping->host) >> huge_page_shift(h);
3247 if (!huge_pte_none(huge_ptep_get(ptep)))
3251 ClearPagePrivate(page);
3252 hugepage_add_new_anon_rmap(page, vma, address);
3254 page_dup_rmap(page);
3255 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3256 && (vma->vm_flags & VM_SHARED)));
3257 set_huge_pte_at(mm, address, ptep, new_pte);
3259 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3260 /* Optimization, do the COW without a second fault */
3261 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
3278 static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3279 struct vm_area_struct *vma,
3280 struct address_space *mapping,
3281 pgoff_t idx, unsigned long address)
3283 unsigned long key[2];
3286 if (vma->vm_flags & VM_SHARED) {
3287 key[0] = (unsigned long) mapping;
3290 key[0] = (unsigned long) mm;
3291 key[1] = address >> huge_page_shift(h);
3294 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3296 return hash & (num_fault_mutexes - 1);
3300 * For uniprocesor systems we always use a single mutex, so just
3301 * return 0 and avoid the hashing overhead.
3303 static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3304 struct vm_area_struct *vma,
3305 struct address_space *mapping,
3306 pgoff_t idx, unsigned long address)
3312 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3313 unsigned long address, unsigned int flags)
3320 struct page *page = NULL;
3321 struct page *pagecache_page = NULL;
3322 struct hstate *h = hstate_vma(vma);
3323 struct address_space *mapping;
3324 int need_wait_lock = 0;
3326 address &= huge_page_mask(h);
3328 ptep = huge_pte_offset(mm, address);
3330 entry = huge_ptep_get(ptep);
3331 if (unlikely(is_hugetlb_entry_migration(entry))) {
3332 migration_entry_wait_huge(vma, mm, ptep);
3334 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3335 return VM_FAULT_HWPOISON_LARGE |
3336 VM_FAULT_SET_HINDEX(hstate_index(h));
3339 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3341 return VM_FAULT_OOM;
3343 mapping = vma->vm_file->f_mapping;
3344 idx = vma_hugecache_offset(h, vma, address);
3347 * Serialize hugepage allocation and instantiation, so that we don't
3348 * get spurious allocation failures if two CPUs race to instantiate
3349 * the same page in the page cache.
3351 hash = fault_mutex_hash(h, mm, vma, mapping, idx, address);
3352 mutex_lock(&htlb_fault_mutex_table[hash]);
3354 entry = huge_ptep_get(ptep);
3355 if (huge_pte_none(entry)) {
3356 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3363 * entry could be a migration/hwpoison entry at this point, so this
3364 * check prevents the kernel from going below assuming that we have
3365 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3366 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3369 if (!pte_present(entry))
3373 * If we are going to COW the mapping later, we examine the pending
3374 * reservations for this page now. This will ensure that any
3375 * allocations necessary to record that reservation occur outside the
3376 * spinlock. For private mappings, we also lookup the pagecache
3377 * page now as it is used to determine if a reservation has been
3380 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3381 if (vma_needs_reservation(h, vma, address) < 0) {
3386 if (!(vma->vm_flags & VM_MAYSHARE))
3387 pagecache_page = hugetlbfs_pagecache_page(h,
3391 ptl = huge_pte_lock(h, mm, ptep);
3393 /* Check for a racing update before calling hugetlb_cow */
3394 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3398 * hugetlb_cow() requires page locks of pte_page(entry) and
3399 * pagecache_page, so here we need take the former one
3400 * when page != pagecache_page or !pagecache_page.
3402 page = pte_page(entry);
3403 if (page != pagecache_page)
3404 if (!trylock_page(page)) {
3411 if (flags & FAULT_FLAG_WRITE) {
3412 if (!huge_pte_write(entry)) {
3413 ret = hugetlb_cow(mm, vma, address, ptep, entry,
3414 pagecache_page, ptl);
3417 entry = huge_pte_mkdirty(entry);
3419 entry = pte_mkyoung(entry);
3420 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3421 flags & FAULT_FLAG_WRITE))
3422 update_mmu_cache(vma, address, ptep);
3424 if (page != pagecache_page)
3430 if (pagecache_page) {
3431 unlock_page(pagecache_page);
3432 put_page(pagecache_page);
3435 mutex_unlock(&htlb_fault_mutex_table[hash]);
3437 * Generally it's safe to hold refcount during waiting page lock. But
3438 * here we just wait to defer the next page fault to avoid busy loop and
3439 * the page is not used after unlocked before returning from the current
3440 * page fault. So we are safe from accessing freed page, even if we wait
3441 * here without taking refcount.
3444 wait_on_page_locked(page);
3448 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3449 struct page **pages, struct vm_area_struct **vmas,
3450 unsigned long *position, unsigned long *nr_pages,
3451 long i, unsigned int flags)
3453 unsigned long pfn_offset;
3454 unsigned long vaddr = *position;
3455 unsigned long remainder = *nr_pages;
3456 struct hstate *h = hstate_vma(vma);
3458 while (vaddr < vma->vm_end && remainder) {
3460 spinlock_t *ptl = NULL;
3465 * If we have a pending SIGKILL, don't keep faulting pages and
3466 * potentially allocating memory.
3468 if (unlikely(fatal_signal_pending(current))) {
3474 * Some archs (sparc64, sh*) have multiple pte_ts to
3475 * each hugepage. We have to make sure we get the
3476 * first, for the page indexing below to work.
3478 * Note that page table lock is not held when pte is null.
3480 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3482 ptl = huge_pte_lock(h, mm, pte);
3483 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3486 * When coredumping, it suits get_dump_page if we just return
3487 * an error where there's an empty slot with no huge pagecache
3488 * to back it. This way, we avoid allocating a hugepage, and
3489 * the sparse dumpfile avoids allocating disk blocks, but its
3490 * huge holes still show up with zeroes where they need to be.
3492 if (absent && (flags & FOLL_DUMP) &&
3493 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3501 * We need call hugetlb_fault for both hugepages under migration
3502 * (in which case hugetlb_fault waits for the migration,) and
3503 * hwpoisoned hugepages (in which case we need to prevent the
3504 * caller from accessing to them.) In order to do this, we use
3505 * here is_swap_pte instead of is_hugetlb_entry_migration and
3506 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3507 * both cases, and because we can't follow correct pages
3508 * directly from any kind of swap entries.
3510 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3511 ((flags & FOLL_WRITE) &&
3512 !huge_pte_write(huge_ptep_get(pte)))) {
3517 ret = hugetlb_fault(mm, vma, vaddr,
3518 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3519 if (!(ret & VM_FAULT_ERROR))
3526 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3527 page = pte_page(huge_ptep_get(pte));
3530 pages[i] = mem_map_offset(page, pfn_offset);
3531 get_page_foll(pages[i]);
3541 if (vaddr < vma->vm_end && remainder &&
3542 pfn_offset < pages_per_huge_page(h)) {
3544 * We use pfn_offset to avoid touching the pageframes
3545 * of this compound page.
3551 *nr_pages = remainder;
3554 return i ? i : -EFAULT;
3557 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3558 unsigned long address, unsigned long end, pgprot_t newprot)
3560 struct mm_struct *mm = vma->vm_mm;
3561 unsigned long start = address;
3564 struct hstate *h = hstate_vma(vma);
3565 unsigned long pages = 0;
3567 BUG_ON(address >= end);
3568 flush_cache_range(vma, address, end);
3570 mmu_notifier_invalidate_range_start(mm, start, end);
3571 i_mmap_lock_write(vma->vm_file->f_mapping);
3572 for (; address < end; address += huge_page_size(h)) {
3574 ptep = huge_pte_offset(mm, address);
3577 ptl = huge_pte_lock(h, mm, ptep);
3578 if (huge_pmd_unshare(mm, &address, ptep)) {
3583 pte = huge_ptep_get(ptep);
3584 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3588 if (unlikely(is_hugetlb_entry_migration(pte))) {
3589 swp_entry_t entry = pte_to_swp_entry(pte);
3591 if (is_write_migration_entry(entry)) {
3594 make_migration_entry_read(&entry);
3595 newpte = swp_entry_to_pte(entry);
3596 set_huge_pte_at(mm, address, ptep, newpte);
3602 if (!huge_pte_none(pte)) {
3603 pte = huge_ptep_get_and_clear(mm, address, ptep);
3604 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3605 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3606 set_huge_pte_at(mm, address, ptep, pte);
3612 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
3613 * may have cleared our pud entry and done put_page on the page table:
3614 * once we release i_mmap_rwsem, another task can do the final put_page
3615 * and that page table be reused and filled with junk.
3617 flush_tlb_range(vma, start, end);
3618 mmu_notifier_invalidate_range(mm, start, end);
3619 i_mmap_unlock_write(vma->vm_file->f_mapping);
3620 mmu_notifier_invalidate_range_end(mm, start, end);
3622 return pages << h->order;
3625 int hugetlb_reserve_pages(struct inode *inode,
3627 struct vm_area_struct *vma,
3628 vm_flags_t vm_flags)
3631 struct hstate *h = hstate_inode(inode);
3632 struct hugepage_subpool *spool = subpool_inode(inode);
3633 struct resv_map *resv_map;
3637 * Only apply hugepage reservation if asked. At fault time, an
3638 * attempt will be made for VM_NORESERVE to allocate a page
3639 * without using reserves
3641 if (vm_flags & VM_NORESERVE)
3645 * Shared mappings base their reservation on the number of pages that
3646 * are already allocated on behalf of the file. Private mappings need
3647 * to reserve the full area even if read-only as mprotect() may be
3648 * called to make the mapping read-write. Assume !vma is a shm mapping
3650 if (!vma || vma->vm_flags & VM_MAYSHARE) {
3651 resv_map = inode_resv_map(inode);
3653 chg = region_chg(resv_map, from, to);
3656 resv_map = resv_map_alloc();
3662 set_vma_resv_map(vma, resv_map);
3663 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3672 * There must be enough pages in the subpool for the mapping. If
3673 * the subpool has a minimum size, there may be some global
3674 * reservations already in place (gbl_reserve).
3676 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
3677 if (gbl_reserve < 0) {
3683 * Check enough hugepages are available for the reservation.
3684 * Hand the pages back to the subpool if there are not
3686 ret = hugetlb_acct_memory(h, gbl_reserve);
3688 /* put back original number of pages, chg */
3689 (void)hugepage_subpool_put_pages(spool, chg);
3694 * Account for the reservations made. Shared mappings record regions
3695 * that have reservations as they are shared by multiple VMAs.
3696 * When the last VMA disappears, the region map says how much
3697 * the reservation was and the page cache tells how much of
3698 * the reservation was consumed. Private mappings are per-VMA and
3699 * only the consumed reservations are tracked. When the VMA
3700 * disappears, the original reservation is the VMA size and the
3701 * consumed reservations are stored in the map. Hence, nothing
3702 * else has to be done for private mappings here
3704 if (!vma || vma->vm_flags & VM_MAYSHARE)
3705 region_add(resv_map, from, to);
3708 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3709 kref_put(&resv_map->refs, resv_map_release);
3713 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3715 struct hstate *h = hstate_inode(inode);
3716 struct resv_map *resv_map = inode_resv_map(inode);
3718 struct hugepage_subpool *spool = subpool_inode(inode);
3722 chg = region_truncate(resv_map, offset);
3723 spin_lock(&inode->i_lock);
3724 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3725 spin_unlock(&inode->i_lock);
3728 * If the subpool has a minimum size, the number of global
3729 * reservations to be released may be adjusted.
3731 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
3732 hugetlb_acct_memory(h, -gbl_reserve);
3735 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3736 static unsigned long page_table_shareable(struct vm_area_struct *svma,
3737 struct vm_area_struct *vma,
3738 unsigned long addr, pgoff_t idx)
3740 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3742 unsigned long sbase = saddr & PUD_MASK;
3743 unsigned long s_end = sbase + PUD_SIZE;
3745 /* Allow segments to share if only one is marked locked */
3746 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3747 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3750 * match the virtual addresses, permission and the alignment of the
3753 if (pmd_index(addr) != pmd_index(saddr) ||
3754 vm_flags != svm_flags ||
3755 sbase < svma->vm_start || svma->vm_end < s_end)
3761 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3763 unsigned long base = addr & PUD_MASK;
3764 unsigned long end = base + PUD_SIZE;
3767 * check on proper vm_flags and page table alignment
3769 if (vma->vm_flags & VM_MAYSHARE &&
3770 vma->vm_start <= base && end <= vma->vm_end)
3776 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3777 * and returns the corresponding pte. While this is not necessary for the
3778 * !shared pmd case because we can allocate the pmd later as well, it makes the
3779 * code much cleaner. pmd allocation is essential for the shared case because
3780 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
3781 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3782 * bad pmd for sharing.
3784 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3786 struct vm_area_struct *vma = find_vma(mm, addr);
3787 struct address_space *mapping = vma->vm_file->f_mapping;
3788 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3790 struct vm_area_struct *svma;
3791 unsigned long saddr;
3796 if (!vma_shareable(vma, addr))
3797 return (pte_t *)pmd_alloc(mm, pud, addr);
3799 i_mmap_lock_write(mapping);
3800 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3804 saddr = page_table_shareable(svma, vma, addr, idx);
3806 spte = huge_pte_offset(svma->vm_mm, saddr);
3809 get_page(virt_to_page(spte));
3818 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
3820 if (pud_none(*pud)) {
3821 pud_populate(mm, pud,
3822 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3824 put_page(virt_to_page(spte));
3829 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3830 i_mmap_unlock_write(mapping);
3835 * unmap huge page backed by shared pte.
3837 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3838 * indicated by page_count > 1, unmap is achieved by clearing pud and
3839 * decrementing the ref count. If count == 1, the pte page is not shared.
3841 * called with page table lock held.
3843 * returns: 1 successfully unmapped a shared pte page
3844 * 0 the underlying pte page is not shared, or it is the last user
3846 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3848 pgd_t *pgd = pgd_offset(mm, *addr);
3849 pud_t *pud = pud_offset(pgd, *addr);
3851 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3852 if (page_count(virt_to_page(ptep)) == 1)
3856 put_page(virt_to_page(ptep));
3858 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3861 #define want_pmd_share() (1)
3862 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3863 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3868 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3872 #define want_pmd_share() (0)
3873 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
3875 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3876 pte_t *huge_pte_alloc(struct mm_struct *mm,
3877 unsigned long addr, unsigned long sz)
3883 pgd = pgd_offset(mm, addr);
3884 pud = pud_alloc(mm, pgd, addr);
3886 if (sz == PUD_SIZE) {
3889 BUG_ON(sz != PMD_SIZE);
3890 if (want_pmd_share() && pud_none(*pud))
3891 pte = huge_pmd_share(mm, addr, pud);
3893 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3896 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3901 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3907 pgd = pgd_offset(mm, addr);
3908 if (pgd_present(*pgd)) {
3909 pud = pud_offset(pgd, addr);
3910 if (pud_present(*pud)) {
3912 return (pte_t *)pud;
3913 pmd = pmd_offset(pud, addr);
3916 return (pte_t *) pmd;
3919 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
3922 * These functions are overwritable if your architecture needs its own
3925 struct page * __weak
3926 follow_huge_addr(struct mm_struct *mm, unsigned long address,
3929 return ERR_PTR(-EINVAL);
3932 struct page * __weak
3933 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3934 pmd_t *pmd, int flags)
3936 struct page *page = NULL;
3939 ptl = pmd_lockptr(mm, pmd);
3942 * make sure that the address range covered by this pmd is not
3943 * unmapped from other threads.
3945 if (!pmd_huge(*pmd))
3947 if (pmd_present(*pmd)) {
3948 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
3949 if (flags & FOLL_GET)
3952 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
3954 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
3958 * hwpoisoned entry is treated as no_page_table in
3959 * follow_page_mask().
3967 struct page * __weak
3968 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3969 pud_t *pud, int flags)
3971 if (flags & FOLL_GET)
3974 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
3977 #ifdef CONFIG_MEMORY_FAILURE
3980 * This function is called from memory failure code.
3981 * Assume the caller holds page lock of the head page.
3983 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3985 struct hstate *h = page_hstate(hpage);
3986 int nid = page_to_nid(hpage);
3989 spin_lock(&hugetlb_lock);
3991 * Just checking !page_huge_active is not enough, because that could be
3992 * an isolated/hwpoisoned hugepage (which have >0 refcount).
3994 if (!page_huge_active(hpage) && !page_count(hpage)) {
3996 * Hwpoisoned hugepage isn't linked to activelist or freelist,
3997 * but dangling hpage->lru can trigger list-debug warnings
3998 * (this happens when we call unpoison_memory() on it),
3999 * so let it point to itself with list_del_init().
4001 list_del_init(&hpage->lru);
4002 set_page_refcounted(hpage);
4003 h->free_huge_pages--;
4004 h->free_huge_pages_node[nid]--;
4007 spin_unlock(&hugetlb_lock);
4012 bool isolate_huge_page(struct page *page, struct list_head *list)
4016 VM_BUG_ON_PAGE(!PageHead(page), page);
4017 spin_lock(&hugetlb_lock);
4018 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4022 clear_page_huge_active(page);
4023 list_move_tail(&page->lru, list);
4025 spin_unlock(&hugetlb_lock);
4029 void putback_active_hugepage(struct page *page)
4031 VM_BUG_ON_PAGE(!PageHead(page), page);
4032 spin_lock(&hugetlb_lock);
4033 set_page_huge_active(page);
4034 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4035 spin_unlock(&hugetlb_lock);