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. In the normal case, existing regions will be expanded
244 * to accommodate the specified range. Sufficient regions should
245 * exist for expansion due to the previous call to region_chg
246 * with the same range. However, it is possible that region_del
247 * could have been called after region_chg and modifed the map
248 * in such a way that no region exists to be expanded. In this
249 * case, pull a region descriptor from the cache associated with
250 * the map and use that for the new range.
252 * Return the number of new huge pages added to the map. This
253 * number is greater than or equal to zero.
255 static long region_add(struct resv_map *resv, long f, long t)
257 struct list_head *head = &resv->regions;
258 struct file_region *rg, *nrg, *trg;
261 spin_lock(&resv->lock);
262 /* Locate the region we are either in or before. */
263 list_for_each_entry(rg, head, link)
268 * If no region exists which can be expanded to include the
269 * specified range, the list must have been modified by an
270 * interleving call to region_del(). Pull a region descriptor
271 * from the cache and use it for this range.
273 if (&rg->link == head || t < rg->from) {
274 VM_BUG_ON(resv->region_cache_count <= 0);
276 resv->region_cache_count--;
277 nrg = list_first_entry(&resv->region_cache, struct file_region,
279 list_del(&nrg->link);
283 list_add(&nrg->link, rg->link.prev);
289 /* Round our left edge to the current segment if it encloses us. */
293 /* Check for and consume any regions we now overlap with. */
295 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
296 if (&rg->link == head)
301 /* If this area reaches higher then extend our area to
302 * include it completely. If this is not the first area
303 * which we intend to reuse, free it. */
307 /* Decrement return value by the deleted range.
308 * Another range will span this area so that by
309 * end of routine add will be >= zero
311 add -= (rg->to - rg->from);
317 add += (nrg->from - f); /* Added to beginning of region */
319 add += t - nrg->to; /* Added to end of region */
323 resv->adds_in_progress--;
324 spin_unlock(&resv->lock);
330 * Examine the existing reserve map and determine how many
331 * huge pages in the specified range [f, t) are NOT currently
332 * represented. This routine is called before a subsequent
333 * call to region_add that will actually modify the reserve
334 * map to add the specified range [f, t). region_chg does
335 * not change the number of huge pages represented by the
336 * map. However, if the existing regions in the map can not
337 * be expanded to represent the new range, a new file_region
338 * structure is added to the map as a placeholder. This is
339 * so that the subsequent region_add call will have all the
340 * regions it needs and will not fail.
342 * Upon entry, region_chg will also examine the cache of region descriptors
343 * associated with the map. If there are not enough descriptors cached, one
344 * will be allocated for the in progress add operation.
346 * Returns the number of huge pages that need to be added to the existing
347 * reservation map for the range [f, t). This number is greater or equal to
348 * zero. -ENOMEM is returned if a new file_region structure or cache entry
349 * is needed and can not be allocated.
351 static long region_chg(struct resv_map *resv, long f, long t)
353 struct list_head *head = &resv->regions;
354 struct file_region *rg, *nrg = NULL;
358 spin_lock(&resv->lock);
360 resv->adds_in_progress++;
363 * Check for sufficient descriptors in the cache to accommodate
364 * the number of in progress add operations.
366 if (resv->adds_in_progress > resv->region_cache_count) {
367 struct file_region *trg;
369 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
370 /* Must drop lock to allocate a new descriptor. */
371 resv->adds_in_progress--;
372 spin_unlock(&resv->lock);
374 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
378 spin_lock(&resv->lock);
379 list_add(&trg->link, &resv->region_cache);
380 resv->region_cache_count++;
384 /* Locate the region we are before or in. */
385 list_for_each_entry(rg, head, link)
389 /* If we are below the current region then a new region is required.
390 * Subtle, allocate a new region at the position but make it zero
391 * size such that we can guarantee to record the reservation. */
392 if (&rg->link == head || t < rg->from) {
394 resv->adds_in_progress--;
395 spin_unlock(&resv->lock);
396 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
402 INIT_LIST_HEAD(&nrg->link);
406 list_add(&nrg->link, rg->link.prev);
411 /* Round our left edge to the current segment if it encloses us. */
416 /* Check for and consume any regions we now overlap with. */
417 list_for_each_entry(rg, rg->link.prev, link) {
418 if (&rg->link == head)
423 /* We overlap with this area, if it extends further than
424 * us then we must extend ourselves. Account for its
425 * existing reservation. */
430 chg -= rg->to - rg->from;
434 spin_unlock(&resv->lock);
435 /* We already know we raced and no longer need the new region */
439 spin_unlock(&resv->lock);
444 * Abort the in progress add operation. The adds_in_progress field
445 * of the resv_map keeps track of the operations in progress between
446 * calls to region_chg and region_add. Operations are sometimes
447 * aborted after the call to region_chg. In such cases, region_abort
448 * is called to decrement the adds_in_progress counter.
450 * NOTE: The range arguments [f, t) are not needed or used in this
451 * routine. They are kept to make reading the calling code easier as
452 * arguments will match the associated region_chg call.
454 static void region_abort(struct resv_map *resv, long f, long t)
456 spin_lock(&resv->lock);
457 VM_BUG_ON(!resv->region_cache_count);
458 resv->adds_in_progress--;
459 spin_unlock(&resv->lock);
463 * Truncate the reserve map at index 'end'. Modify/truncate any
464 * region which contains end. Delete any regions past end.
465 * Return the number of huge pages removed from the map.
467 static long region_truncate(struct resv_map *resv, long end)
469 struct list_head *head = &resv->regions;
470 struct file_region *rg, *trg;
473 spin_lock(&resv->lock);
474 /* Locate the region we are either in or before. */
475 list_for_each_entry(rg, head, link)
478 if (&rg->link == head)
481 /* If we are in the middle of a region then adjust it. */
482 if (end > rg->from) {
485 rg = list_entry(rg->link.next, typeof(*rg), link);
488 /* Drop any remaining regions. */
489 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
490 if (&rg->link == head)
492 chg += rg->to - rg->from;
498 spin_unlock(&resv->lock);
503 * Count and return the number of huge pages in the reserve map
504 * that intersect with the range [f, t).
506 static long region_count(struct resv_map *resv, long f, long t)
508 struct list_head *head = &resv->regions;
509 struct file_region *rg;
512 spin_lock(&resv->lock);
513 /* Locate each segment we overlap with, and count that overlap. */
514 list_for_each_entry(rg, head, link) {
523 seg_from = max(rg->from, f);
524 seg_to = min(rg->to, t);
526 chg += seg_to - seg_from;
528 spin_unlock(&resv->lock);
534 * Convert the address within this vma to the page offset within
535 * the mapping, in pagecache page units; huge pages here.
537 static pgoff_t vma_hugecache_offset(struct hstate *h,
538 struct vm_area_struct *vma, unsigned long address)
540 return ((address - vma->vm_start) >> huge_page_shift(h)) +
541 (vma->vm_pgoff >> huge_page_order(h));
544 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
545 unsigned long address)
547 return vma_hugecache_offset(hstate_vma(vma), vma, address);
551 * Return the size of the pages allocated when backing a VMA. In the majority
552 * cases this will be same size as used by the page table entries.
554 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
556 struct hstate *hstate;
558 if (!is_vm_hugetlb_page(vma))
561 hstate = hstate_vma(vma);
563 return 1UL << huge_page_shift(hstate);
565 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
568 * Return the page size being used by the MMU to back a VMA. In the majority
569 * of cases, the page size used by the kernel matches the MMU size. On
570 * architectures where it differs, an architecture-specific version of this
571 * function is required.
573 #ifndef vma_mmu_pagesize
574 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
576 return vma_kernel_pagesize(vma);
581 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
582 * bits of the reservation map pointer, which are always clear due to
585 #define HPAGE_RESV_OWNER (1UL << 0)
586 #define HPAGE_RESV_UNMAPPED (1UL << 1)
587 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
590 * These helpers are used to track how many pages are reserved for
591 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
592 * is guaranteed to have their future faults succeed.
594 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
595 * the reserve counters are updated with the hugetlb_lock held. It is safe
596 * to reset the VMA at fork() time as it is not in use yet and there is no
597 * chance of the global counters getting corrupted as a result of the values.
599 * The private mapping reservation is represented in a subtly different
600 * manner to a shared mapping. A shared mapping has a region map associated
601 * with the underlying file, this region map represents the backing file
602 * pages which have ever had a reservation assigned which this persists even
603 * after the page is instantiated. A private mapping has a region map
604 * associated with the original mmap which is attached to all VMAs which
605 * reference it, this region map represents those offsets which have consumed
606 * reservation ie. where pages have been instantiated.
608 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
610 return (unsigned long)vma->vm_private_data;
613 static void set_vma_private_data(struct vm_area_struct *vma,
616 vma->vm_private_data = (void *)value;
619 struct resv_map *resv_map_alloc(void)
621 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
622 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
624 if (!resv_map || !rg) {
630 kref_init(&resv_map->refs);
631 spin_lock_init(&resv_map->lock);
632 INIT_LIST_HEAD(&resv_map->regions);
634 resv_map->adds_in_progress = 0;
636 INIT_LIST_HEAD(&resv_map->region_cache);
637 list_add(&rg->link, &resv_map->region_cache);
638 resv_map->region_cache_count = 1;
643 void resv_map_release(struct kref *ref)
645 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
646 struct list_head *head = &resv_map->region_cache;
647 struct file_region *rg, *trg;
649 /* Clear out any active regions before we release the map. */
650 region_truncate(resv_map, 0);
652 /* ... and any entries left in the cache */
653 list_for_each_entry_safe(rg, trg, head, link) {
658 VM_BUG_ON(resv_map->adds_in_progress);
663 static inline struct resv_map *inode_resv_map(struct inode *inode)
665 return inode->i_mapping->private_data;
668 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
670 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
671 if (vma->vm_flags & VM_MAYSHARE) {
672 struct address_space *mapping = vma->vm_file->f_mapping;
673 struct inode *inode = mapping->host;
675 return inode_resv_map(inode);
678 return (struct resv_map *)(get_vma_private_data(vma) &
683 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
685 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
686 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
688 set_vma_private_data(vma, (get_vma_private_data(vma) &
689 HPAGE_RESV_MASK) | (unsigned long)map);
692 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
694 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
695 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
697 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
700 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
702 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
704 return (get_vma_private_data(vma) & flag) != 0;
707 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
708 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
710 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
711 if (!(vma->vm_flags & VM_MAYSHARE))
712 vma->vm_private_data = (void *)0;
715 /* Returns true if the VMA has associated reserve pages */
716 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
718 if (vma->vm_flags & VM_NORESERVE) {
720 * This address is already reserved by other process(chg == 0),
721 * so, we should decrement reserved count. Without decrementing,
722 * reserve count remains after releasing inode, because this
723 * allocated page will go into page cache and is regarded as
724 * coming from reserved pool in releasing step. Currently, we
725 * don't have any other solution to deal with this situation
726 * properly, so add work-around here.
728 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
734 /* Shared mappings always use reserves */
735 if (vma->vm_flags & VM_MAYSHARE)
739 * Only the process that called mmap() has reserves for
742 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
748 static void enqueue_huge_page(struct hstate *h, struct page *page)
750 int nid = page_to_nid(page);
751 list_move(&page->lru, &h->hugepage_freelists[nid]);
752 h->free_huge_pages++;
753 h->free_huge_pages_node[nid]++;
756 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
760 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
761 if (!is_migrate_isolate_page(page))
764 * if 'non-isolated free hugepage' not found on the list,
765 * the allocation fails.
767 if (&h->hugepage_freelists[nid] == &page->lru)
769 list_move(&page->lru, &h->hugepage_activelist);
770 set_page_refcounted(page);
771 h->free_huge_pages--;
772 h->free_huge_pages_node[nid]--;
776 /* Movability of hugepages depends on migration support. */
777 static inline gfp_t htlb_alloc_mask(struct hstate *h)
779 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
780 return GFP_HIGHUSER_MOVABLE;
785 static struct page *dequeue_huge_page_vma(struct hstate *h,
786 struct vm_area_struct *vma,
787 unsigned long address, int avoid_reserve,
790 struct page *page = NULL;
791 struct mempolicy *mpol;
792 nodemask_t *nodemask;
793 struct zonelist *zonelist;
796 unsigned int cpuset_mems_cookie;
799 * A child process with MAP_PRIVATE mappings created by their parent
800 * have no page reserves. This check ensures that reservations are
801 * not "stolen". The child may still get SIGKILLed
803 if (!vma_has_reserves(vma, chg) &&
804 h->free_huge_pages - h->resv_huge_pages == 0)
807 /* If reserves cannot be used, ensure enough pages are in the pool */
808 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
812 cpuset_mems_cookie = read_mems_allowed_begin();
813 zonelist = huge_zonelist(vma, address,
814 htlb_alloc_mask(h), &mpol, &nodemask);
816 for_each_zone_zonelist_nodemask(zone, z, zonelist,
817 MAX_NR_ZONES - 1, nodemask) {
818 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
819 page = dequeue_huge_page_node(h, zone_to_nid(zone));
823 if (!vma_has_reserves(vma, chg))
826 SetPagePrivate(page);
827 h->resv_huge_pages--;
834 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
843 * common helper functions for hstate_next_node_to_{alloc|free}.
844 * We may have allocated or freed a huge page based on a different
845 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
846 * be outside of *nodes_allowed. Ensure that we use an allowed
847 * node for alloc or free.
849 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
851 nid = next_node(nid, *nodes_allowed);
852 if (nid == MAX_NUMNODES)
853 nid = first_node(*nodes_allowed);
854 VM_BUG_ON(nid >= MAX_NUMNODES);
859 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
861 if (!node_isset(nid, *nodes_allowed))
862 nid = next_node_allowed(nid, nodes_allowed);
867 * returns the previously saved node ["this node"] from which to
868 * allocate a persistent huge page for the pool and advance the
869 * next node from which to allocate, handling wrap at end of node
872 static int hstate_next_node_to_alloc(struct hstate *h,
873 nodemask_t *nodes_allowed)
877 VM_BUG_ON(!nodes_allowed);
879 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
880 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
886 * helper for free_pool_huge_page() - return the previously saved
887 * node ["this node"] from which to free a huge page. Advance the
888 * next node id whether or not we find a free huge page to free so
889 * that the next attempt to free addresses the next node.
891 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
895 VM_BUG_ON(!nodes_allowed);
897 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
898 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
903 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
904 for (nr_nodes = nodes_weight(*mask); \
906 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
909 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
910 for (nr_nodes = nodes_weight(*mask); \
912 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
915 #if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
916 static void destroy_compound_gigantic_page(struct page *page,
920 int nr_pages = 1 << order;
921 struct page *p = page + 1;
923 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
925 set_page_refcounted(p);
926 p->first_page = NULL;
929 set_compound_order(page, 0);
930 __ClearPageHead(page);
933 static void free_gigantic_page(struct page *page, unsigned order)
935 free_contig_range(page_to_pfn(page), 1 << order);
938 static int __alloc_gigantic_page(unsigned long start_pfn,
939 unsigned long nr_pages)
941 unsigned long end_pfn = start_pfn + nr_pages;
942 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
945 static bool pfn_range_valid_gigantic(unsigned long start_pfn,
946 unsigned long nr_pages)
948 unsigned long i, end_pfn = start_pfn + nr_pages;
951 for (i = start_pfn; i < end_pfn; i++) {
955 page = pfn_to_page(i);
957 if (PageReserved(page))
960 if (page_count(page) > 0)
970 static bool zone_spans_last_pfn(const struct zone *zone,
971 unsigned long start_pfn, unsigned long nr_pages)
973 unsigned long last_pfn = start_pfn + nr_pages - 1;
974 return zone_spans_pfn(zone, last_pfn);
977 static struct page *alloc_gigantic_page(int nid, unsigned order)
979 unsigned long nr_pages = 1 << order;
980 unsigned long ret, pfn, flags;
983 z = NODE_DATA(nid)->node_zones;
984 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
985 spin_lock_irqsave(&z->lock, flags);
987 pfn = ALIGN(z->zone_start_pfn, nr_pages);
988 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
989 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
991 * We release the zone lock here because
992 * alloc_contig_range() will also lock the zone
993 * at some point. If there's an allocation
994 * spinning on this lock, it may win the race
995 * and cause alloc_contig_range() to fail...
997 spin_unlock_irqrestore(&z->lock, flags);
998 ret = __alloc_gigantic_page(pfn, nr_pages);
1000 return pfn_to_page(pfn);
1001 spin_lock_irqsave(&z->lock, flags);
1006 spin_unlock_irqrestore(&z->lock, flags);
1012 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1013 static void prep_compound_gigantic_page(struct page *page, unsigned long order);
1015 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1019 page = alloc_gigantic_page(nid, huge_page_order(h));
1021 prep_compound_gigantic_page(page, huge_page_order(h));
1022 prep_new_huge_page(h, page, nid);
1028 static int alloc_fresh_gigantic_page(struct hstate *h,
1029 nodemask_t *nodes_allowed)
1031 struct page *page = NULL;
1034 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1035 page = alloc_fresh_gigantic_page_node(h, node);
1043 static inline bool gigantic_page_supported(void) { return true; }
1045 static inline bool gigantic_page_supported(void) { return false; }
1046 static inline void free_gigantic_page(struct page *page, unsigned order) { }
1047 static inline void destroy_compound_gigantic_page(struct page *page,
1048 unsigned long order) { }
1049 static inline int alloc_fresh_gigantic_page(struct hstate *h,
1050 nodemask_t *nodes_allowed) { return 0; }
1053 static void update_and_free_page(struct hstate *h, struct page *page)
1057 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1061 h->nr_huge_pages_node[page_to_nid(page)]--;
1062 for (i = 0; i < pages_per_huge_page(h); i++) {
1063 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1064 1 << PG_referenced | 1 << PG_dirty |
1065 1 << PG_active | 1 << PG_private |
1068 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1069 set_compound_page_dtor(page, NULL);
1070 set_page_refcounted(page);
1071 if (hstate_is_gigantic(h)) {
1072 destroy_compound_gigantic_page(page, huge_page_order(h));
1073 free_gigantic_page(page, huge_page_order(h));
1075 __free_pages(page, huge_page_order(h));
1079 struct hstate *size_to_hstate(unsigned long size)
1083 for_each_hstate(h) {
1084 if (huge_page_size(h) == size)
1091 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1092 * to hstate->hugepage_activelist.)
1094 * This function can be called for tail pages, but never returns true for them.
1096 bool page_huge_active(struct page *page)
1098 VM_BUG_ON_PAGE(!PageHuge(page), page);
1099 return PageHead(page) && PagePrivate(&page[1]);
1102 /* never called for tail page */
1103 static void set_page_huge_active(struct page *page)
1105 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1106 SetPagePrivate(&page[1]);
1109 static void clear_page_huge_active(struct page *page)
1111 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1112 ClearPagePrivate(&page[1]);
1115 void free_huge_page(struct page *page)
1118 * Can't pass hstate in here because it is called from the
1119 * compound page destructor.
1121 struct hstate *h = page_hstate(page);
1122 int nid = page_to_nid(page);
1123 struct hugepage_subpool *spool =
1124 (struct hugepage_subpool *)page_private(page);
1125 bool restore_reserve;
1127 set_page_private(page, 0);
1128 page->mapping = NULL;
1129 BUG_ON(page_count(page));
1130 BUG_ON(page_mapcount(page));
1131 restore_reserve = PagePrivate(page);
1132 ClearPagePrivate(page);
1135 * A return code of zero implies that the subpool will be under its
1136 * minimum size if the reservation is not restored after page is free.
1137 * Therefore, force restore_reserve operation.
1139 if (hugepage_subpool_put_pages(spool, 1) == 0)
1140 restore_reserve = true;
1142 spin_lock(&hugetlb_lock);
1143 clear_page_huge_active(page);
1144 hugetlb_cgroup_uncharge_page(hstate_index(h),
1145 pages_per_huge_page(h), page);
1146 if (restore_reserve)
1147 h->resv_huge_pages++;
1149 if (h->surplus_huge_pages_node[nid]) {
1150 /* remove the page from active list */
1151 list_del(&page->lru);
1152 update_and_free_page(h, page);
1153 h->surplus_huge_pages--;
1154 h->surplus_huge_pages_node[nid]--;
1156 arch_clear_hugepage_flags(page);
1157 enqueue_huge_page(h, page);
1159 spin_unlock(&hugetlb_lock);
1162 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1164 INIT_LIST_HEAD(&page->lru);
1165 set_compound_page_dtor(page, free_huge_page);
1166 spin_lock(&hugetlb_lock);
1167 set_hugetlb_cgroup(page, NULL);
1169 h->nr_huge_pages_node[nid]++;
1170 spin_unlock(&hugetlb_lock);
1171 put_page(page); /* free it into the hugepage allocator */
1174 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
1177 int nr_pages = 1 << order;
1178 struct page *p = page + 1;
1180 /* we rely on prep_new_huge_page to set the destructor */
1181 set_compound_order(page, order);
1182 __SetPageHead(page);
1183 __ClearPageReserved(page);
1184 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1186 * For gigantic hugepages allocated through bootmem at
1187 * boot, it's safer to be consistent with the not-gigantic
1188 * hugepages and clear the PG_reserved bit from all tail pages
1189 * too. Otherwse drivers using get_user_pages() to access tail
1190 * pages may get the reference counting wrong if they see
1191 * PG_reserved set on a tail page (despite the head page not
1192 * having PG_reserved set). Enforcing this consistency between
1193 * head and tail pages allows drivers to optimize away a check
1194 * on the head page when they need know if put_page() is needed
1195 * after get_user_pages().
1197 __ClearPageReserved(p);
1198 set_page_count(p, 0);
1199 p->first_page = page;
1200 /* Make sure p->first_page is always valid for PageTail() */
1207 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1208 * transparent huge pages. See the PageTransHuge() documentation for more
1211 int PageHuge(struct page *page)
1213 if (!PageCompound(page))
1216 page = compound_head(page);
1217 return get_compound_page_dtor(page) == free_huge_page;
1219 EXPORT_SYMBOL_GPL(PageHuge);
1222 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1223 * normal or transparent huge pages.
1225 int PageHeadHuge(struct page *page_head)
1227 if (!PageHead(page_head))
1230 return get_compound_page_dtor(page_head) == free_huge_page;
1233 pgoff_t __basepage_index(struct page *page)
1235 struct page *page_head = compound_head(page);
1236 pgoff_t index = page_index(page_head);
1237 unsigned long compound_idx;
1239 if (!PageHuge(page_head))
1240 return page_index(page);
1242 if (compound_order(page_head) >= MAX_ORDER)
1243 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1245 compound_idx = page - page_head;
1247 return (index << compound_order(page_head)) + compound_idx;
1250 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1254 page = alloc_pages_exact_node(nid,
1255 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1256 __GFP_REPEAT|__GFP_NOWARN,
1257 huge_page_order(h));
1259 prep_new_huge_page(h, page, nid);
1265 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1271 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1272 page = alloc_fresh_huge_page_node(h, node);
1280 count_vm_event(HTLB_BUDDY_PGALLOC);
1282 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1288 * Free huge page from pool from next node to free.
1289 * Attempt to keep persistent huge pages more or less
1290 * balanced over allowed nodes.
1291 * Called with hugetlb_lock locked.
1293 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1299 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1301 * If we're returning unused surplus pages, only examine
1302 * nodes with surplus pages.
1304 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1305 !list_empty(&h->hugepage_freelists[node])) {
1307 list_entry(h->hugepage_freelists[node].next,
1309 list_del(&page->lru);
1310 h->free_huge_pages--;
1311 h->free_huge_pages_node[node]--;
1313 h->surplus_huge_pages--;
1314 h->surplus_huge_pages_node[node]--;
1316 update_and_free_page(h, page);
1326 * Dissolve a given free hugepage into free buddy pages. This function does
1327 * nothing for in-use (including surplus) hugepages.
1329 static void dissolve_free_huge_page(struct page *page)
1331 spin_lock(&hugetlb_lock);
1332 if (PageHuge(page) && !page_count(page)) {
1333 struct hstate *h = page_hstate(page);
1334 int nid = page_to_nid(page);
1335 list_del(&page->lru);
1336 h->free_huge_pages--;
1337 h->free_huge_pages_node[nid]--;
1338 update_and_free_page(h, page);
1340 spin_unlock(&hugetlb_lock);
1344 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1345 * make specified memory blocks removable from the system.
1346 * Note that start_pfn should aligned with (minimum) hugepage size.
1348 void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1352 if (!hugepages_supported())
1355 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order));
1356 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
1357 dissolve_free_huge_page(pfn_to_page(pfn));
1360 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
1365 if (hstate_is_gigantic(h))
1369 * Assume we will successfully allocate the surplus page to
1370 * prevent racing processes from causing the surplus to exceed
1373 * This however introduces a different race, where a process B
1374 * tries to grow the static hugepage pool while alloc_pages() is
1375 * called by process A. B will only examine the per-node
1376 * counters in determining if surplus huge pages can be
1377 * converted to normal huge pages in adjust_pool_surplus(). A
1378 * won't be able to increment the per-node counter, until the
1379 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1380 * no more huge pages can be converted from surplus to normal
1381 * state (and doesn't try to convert again). Thus, we have a
1382 * case where a surplus huge page exists, the pool is grown, and
1383 * the surplus huge page still exists after, even though it
1384 * should just have been converted to a normal huge page. This
1385 * does not leak memory, though, as the hugepage will be freed
1386 * once it is out of use. It also does not allow the counters to
1387 * go out of whack in adjust_pool_surplus() as we don't modify
1388 * the node values until we've gotten the hugepage and only the
1389 * per-node value is checked there.
1391 spin_lock(&hugetlb_lock);
1392 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1393 spin_unlock(&hugetlb_lock);
1397 h->surplus_huge_pages++;
1399 spin_unlock(&hugetlb_lock);
1401 if (nid == NUMA_NO_NODE)
1402 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
1403 __GFP_REPEAT|__GFP_NOWARN,
1404 huge_page_order(h));
1406 page = alloc_pages_exact_node(nid,
1407 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1408 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
1410 spin_lock(&hugetlb_lock);
1412 INIT_LIST_HEAD(&page->lru);
1413 r_nid = page_to_nid(page);
1414 set_compound_page_dtor(page, free_huge_page);
1415 set_hugetlb_cgroup(page, NULL);
1417 * We incremented the global counters already
1419 h->nr_huge_pages_node[r_nid]++;
1420 h->surplus_huge_pages_node[r_nid]++;
1421 __count_vm_event(HTLB_BUDDY_PGALLOC);
1424 h->surplus_huge_pages--;
1425 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1427 spin_unlock(&hugetlb_lock);
1433 * This allocation function is useful in the context where vma is irrelevant.
1434 * E.g. soft-offlining uses this function because it only cares physical
1435 * address of error page.
1437 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1439 struct page *page = NULL;
1441 spin_lock(&hugetlb_lock);
1442 if (h->free_huge_pages - h->resv_huge_pages > 0)
1443 page = dequeue_huge_page_node(h, nid);
1444 spin_unlock(&hugetlb_lock);
1447 page = alloc_buddy_huge_page(h, nid);
1453 * Increase the hugetlb pool such that it can accommodate a reservation
1456 static int gather_surplus_pages(struct hstate *h, int delta)
1458 struct list_head surplus_list;
1459 struct page *page, *tmp;
1461 int needed, allocated;
1462 bool alloc_ok = true;
1464 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1466 h->resv_huge_pages += delta;
1471 INIT_LIST_HEAD(&surplus_list);
1475 spin_unlock(&hugetlb_lock);
1476 for (i = 0; i < needed; i++) {
1477 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1482 list_add(&page->lru, &surplus_list);
1487 * After retaking hugetlb_lock, we need to recalculate 'needed'
1488 * because either resv_huge_pages or free_huge_pages may have changed.
1490 spin_lock(&hugetlb_lock);
1491 needed = (h->resv_huge_pages + delta) -
1492 (h->free_huge_pages + allocated);
1497 * We were not able to allocate enough pages to
1498 * satisfy the entire reservation so we free what
1499 * we've allocated so far.
1504 * The surplus_list now contains _at_least_ the number of extra pages
1505 * needed to accommodate the reservation. Add the appropriate number
1506 * of pages to the hugetlb pool and free the extras back to the buddy
1507 * allocator. Commit the entire reservation here to prevent another
1508 * process from stealing the pages as they are added to the pool but
1509 * before they are reserved.
1511 needed += allocated;
1512 h->resv_huge_pages += delta;
1515 /* Free the needed pages to the hugetlb pool */
1516 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1520 * This page is now managed by the hugetlb allocator and has
1521 * no users -- drop the buddy allocator's reference.
1523 put_page_testzero(page);
1524 VM_BUG_ON_PAGE(page_count(page), page);
1525 enqueue_huge_page(h, page);
1528 spin_unlock(&hugetlb_lock);
1530 /* Free unnecessary surplus pages to the buddy allocator */
1531 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1533 spin_lock(&hugetlb_lock);
1539 * When releasing a hugetlb pool reservation, any surplus pages that were
1540 * allocated to satisfy the reservation must be explicitly freed if they were
1542 * Called with hugetlb_lock held.
1544 static void return_unused_surplus_pages(struct hstate *h,
1545 unsigned long unused_resv_pages)
1547 unsigned long nr_pages;
1549 /* Uncommit the reservation */
1550 h->resv_huge_pages -= unused_resv_pages;
1552 /* Cannot return gigantic pages currently */
1553 if (hstate_is_gigantic(h))
1556 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1559 * We want to release as many surplus pages as possible, spread
1560 * evenly across all nodes with memory. Iterate across these nodes
1561 * until we can no longer free unreserved surplus pages. This occurs
1562 * when the nodes with surplus pages have no free pages.
1563 * free_pool_huge_page() will balance the the freed pages across the
1564 * on-line nodes with memory and will handle the hstate accounting.
1566 while (nr_pages--) {
1567 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1569 cond_resched_lock(&hugetlb_lock);
1575 * vma_needs_reservation, vma_commit_reservation and vma_abort_reservation
1576 * are used by the huge page allocation routines to manage reservations.
1578 * vma_needs_reservation is called to determine if the huge page at addr
1579 * within the vma has an associated reservation. If a reservation is
1580 * needed, the value 1 is returned. The caller is then responsible for
1581 * managing the global reservation and subpool usage counts. After
1582 * the huge page has been allocated, vma_commit_reservation is called
1583 * to add the page to the reservation map. If the reservation must be
1584 * aborted instead of committed, vma_abort_reservation is called.
1586 * In the normal case, vma_commit_reservation returns the same value
1587 * as the preceding vma_needs_reservation call. The only time this
1588 * is not the case is if a reserve map was changed between calls. It
1589 * is the responsibility of the caller to notice the difference and
1590 * take appropriate action.
1592 enum vma_resv_mode {
1597 static long __vma_reservation_common(struct hstate *h,
1598 struct vm_area_struct *vma, unsigned long addr,
1599 enum vma_resv_mode mode)
1601 struct resv_map *resv;
1605 resv = vma_resv_map(vma);
1609 idx = vma_hugecache_offset(h, vma, addr);
1611 case VMA_NEEDS_RESV:
1612 ret = region_chg(resv, idx, idx + 1);
1614 case VMA_COMMIT_RESV:
1615 ret = region_add(resv, idx, idx + 1);
1617 case VMA_ABORT_RESV:
1618 region_abort(resv, idx, idx + 1);
1625 if (vma->vm_flags & VM_MAYSHARE)
1628 return ret < 0 ? ret : 0;
1631 static long vma_needs_reservation(struct hstate *h,
1632 struct vm_area_struct *vma, unsigned long addr)
1634 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1637 static long vma_commit_reservation(struct hstate *h,
1638 struct vm_area_struct *vma, unsigned long addr)
1640 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1643 static void vma_abort_reservation(struct hstate *h,
1644 struct vm_area_struct *vma, unsigned long addr)
1646 (void)__vma_reservation_common(h, vma, addr, VMA_ABORT_RESV);
1649 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1650 unsigned long addr, int avoid_reserve)
1652 struct hugepage_subpool *spool = subpool_vma(vma);
1653 struct hstate *h = hstate_vma(vma);
1657 struct hugetlb_cgroup *h_cg;
1659 idx = hstate_index(h);
1661 * Processes that did not create the mapping will have no
1662 * reserves and will not have accounted against subpool
1663 * limit. Check that the subpool limit can be made before
1664 * satisfying the allocation MAP_NORESERVE mappings may also
1665 * need pages and subpool limit allocated allocated if no reserve
1668 chg = vma_needs_reservation(h, vma, addr);
1670 return ERR_PTR(-ENOMEM);
1671 if (chg || avoid_reserve)
1672 if (hugepage_subpool_get_pages(spool, 1) < 0) {
1673 vma_abort_reservation(h, vma, addr);
1674 return ERR_PTR(-ENOSPC);
1677 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1679 goto out_subpool_put;
1681 spin_lock(&hugetlb_lock);
1682 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
1684 spin_unlock(&hugetlb_lock);
1685 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1687 goto out_uncharge_cgroup;
1689 spin_lock(&hugetlb_lock);
1690 list_move(&page->lru, &h->hugepage_activelist);
1693 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1694 spin_unlock(&hugetlb_lock);
1696 set_page_private(page, (unsigned long)spool);
1698 commit = vma_commit_reservation(h, vma, addr);
1699 if (unlikely(chg > commit)) {
1701 * The page was added to the reservation map between
1702 * vma_needs_reservation and vma_commit_reservation.
1703 * This indicates a race with hugetlb_reserve_pages.
1704 * Adjust for the subpool count incremented above AND
1705 * in hugetlb_reserve_pages for the same page. Also,
1706 * the reservation count added in hugetlb_reserve_pages
1707 * no longer applies.
1711 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
1712 hugetlb_acct_memory(h, -rsv_adjust);
1716 out_uncharge_cgroup:
1717 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1719 if (chg || avoid_reserve)
1720 hugepage_subpool_put_pages(spool, 1);
1721 vma_abort_reservation(h, vma, addr);
1722 return ERR_PTR(-ENOSPC);
1726 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1727 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1728 * where no ERR_VALUE is expected to be returned.
1730 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1731 unsigned long addr, int avoid_reserve)
1733 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1739 int __weak alloc_bootmem_huge_page(struct hstate *h)
1741 struct huge_bootmem_page *m;
1744 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1747 addr = memblock_virt_alloc_try_nid_nopanic(
1748 huge_page_size(h), huge_page_size(h),
1749 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1752 * Use the beginning of the huge page to store the
1753 * huge_bootmem_page struct (until gather_bootmem
1754 * puts them into the mem_map).
1763 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
1764 /* Put them into a private list first because mem_map is not up yet */
1765 list_add(&m->list, &huge_boot_pages);
1770 static void __init prep_compound_huge_page(struct page *page, int order)
1772 if (unlikely(order > (MAX_ORDER - 1)))
1773 prep_compound_gigantic_page(page, order);
1775 prep_compound_page(page, order);
1778 /* Put bootmem huge pages into the standard lists after mem_map is up */
1779 static void __init gather_bootmem_prealloc(void)
1781 struct huge_bootmem_page *m;
1783 list_for_each_entry(m, &huge_boot_pages, list) {
1784 struct hstate *h = m->hstate;
1787 #ifdef CONFIG_HIGHMEM
1788 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1789 memblock_free_late(__pa(m),
1790 sizeof(struct huge_bootmem_page));
1792 page = virt_to_page(m);
1794 WARN_ON(page_count(page) != 1);
1795 prep_compound_huge_page(page, h->order);
1796 WARN_ON(PageReserved(page));
1797 prep_new_huge_page(h, page, page_to_nid(page));
1799 * If we had gigantic hugepages allocated at boot time, we need
1800 * to restore the 'stolen' pages to totalram_pages in order to
1801 * fix confusing memory reports from free(1) and another
1802 * side-effects, like CommitLimit going negative.
1804 if (hstate_is_gigantic(h))
1805 adjust_managed_page_count(page, 1 << h->order);
1809 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1813 for (i = 0; i < h->max_huge_pages; ++i) {
1814 if (hstate_is_gigantic(h)) {
1815 if (!alloc_bootmem_huge_page(h))
1817 } else if (!alloc_fresh_huge_page(h,
1818 &node_states[N_MEMORY]))
1821 h->max_huge_pages = i;
1824 static void __init hugetlb_init_hstates(void)
1828 for_each_hstate(h) {
1829 if (minimum_order > huge_page_order(h))
1830 minimum_order = huge_page_order(h);
1832 /* oversize hugepages were init'ed in early boot */
1833 if (!hstate_is_gigantic(h))
1834 hugetlb_hstate_alloc_pages(h);
1836 VM_BUG_ON(minimum_order == UINT_MAX);
1839 static char * __init memfmt(char *buf, unsigned long n)
1841 if (n >= (1UL << 30))
1842 sprintf(buf, "%lu GB", n >> 30);
1843 else if (n >= (1UL << 20))
1844 sprintf(buf, "%lu MB", n >> 20);
1846 sprintf(buf, "%lu KB", n >> 10);
1850 static void __init report_hugepages(void)
1854 for_each_hstate(h) {
1856 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1857 memfmt(buf, huge_page_size(h)),
1858 h->free_huge_pages);
1862 #ifdef CONFIG_HIGHMEM
1863 static void try_to_free_low(struct hstate *h, unsigned long count,
1864 nodemask_t *nodes_allowed)
1868 if (hstate_is_gigantic(h))
1871 for_each_node_mask(i, *nodes_allowed) {
1872 struct page *page, *next;
1873 struct list_head *freel = &h->hugepage_freelists[i];
1874 list_for_each_entry_safe(page, next, freel, lru) {
1875 if (count >= h->nr_huge_pages)
1877 if (PageHighMem(page))
1879 list_del(&page->lru);
1880 update_and_free_page(h, page);
1881 h->free_huge_pages--;
1882 h->free_huge_pages_node[page_to_nid(page)]--;
1887 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1888 nodemask_t *nodes_allowed)
1894 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1895 * balanced by operating on them in a round-robin fashion.
1896 * Returns 1 if an adjustment was made.
1898 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1903 VM_BUG_ON(delta != -1 && delta != 1);
1906 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1907 if (h->surplus_huge_pages_node[node])
1911 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1912 if (h->surplus_huge_pages_node[node] <
1913 h->nr_huge_pages_node[node])
1920 h->surplus_huge_pages += delta;
1921 h->surplus_huge_pages_node[node] += delta;
1925 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1926 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1927 nodemask_t *nodes_allowed)
1929 unsigned long min_count, ret;
1931 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1932 return h->max_huge_pages;
1935 * Increase the pool size
1936 * First take pages out of surplus state. Then make up the
1937 * remaining difference by allocating fresh huge pages.
1939 * We might race with alloc_buddy_huge_page() here and be unable
1940 * to convert a surplus huge page to a normal huge page. That is
1941 * not critical, though, it just means the overall size of the
1942 * pool might be one hugepage larger than it needs to be, but
1943 * within all the constraints specified by the sysctls.
1945 spin_lock(&hugetlb_lock);
1946 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1947 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1951 while (count > persistent_huge_pages(h)) {
1953 * If this allocation races such that we no longer need the
1954 * page, free_huge_page will handle it by freeing the page
1955 * and reducing the surplus.
1957 spin_unlock(&hugetlb_lock);
1958 if (hstate_is_gigantic(h))
1959 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
1961 ret = alloc_fresh_huge_page(h, nodes_allowed);
1962 spin_lock(&hugetlb_lock);
1966 /* Bail for signals. Probably ctrl-c from user */
1967 if (signal_pending(current))
1972 * Decrease the pool size
1973 * First return free pages to the buddy allocator (being careful
1974 * to keep enough around to satisfy reservations). Then place
1975 * pages into surplus state as needed so the pool will shrink
1976 * to the desired size as pages become free.
1978 * By placing pages into the surplus state independent of the
1979 * overcommit value, we are allowing the surplus pool size to
1980 * exceed overcommit. There are few sane options here. Since
1981 * alloc_buddy_huge_page() is checking the global counter,
1982 * though, we'll note that we're not allowed to exceed surplus
1983 * and won't grow the pool anywhere else. Not until one of the
1984 * sysctls are changed, or the surplus pages go out of use.
1986 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1987 min_count = max(count, min_count);
1988 try_to_free_low(h, min_count, nodes_allowed);
1989 while (min_count < persistent_huge_pages(h)) {
1990 if (!free_pool_huge_page(h, nodes_allowed, 0))
1992 cond_resched_lock(&hugetlb_lock);
1994 while (count < persistent_huge_pages(h)) {
1995 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1999 ret = persistent_huge_pages(h);
2000 spin_unlock(&hugetlb_lock);
2004 #define HSTATE_ATTR_RO(_name) \
2005 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2007 #define HSTATE_ATTR(_name) \
2008 static struct kobj_attribute _name##_attr = \
2009 __ATTR(_name, 0644, _name##_show, _name##_store)
2011 static struct kobject *hugepages_kobj;
2012 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2014 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2016 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2020 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2021 if (hstate_kobjs[i] == kobj) {
2023 *nidp = NUMA_NO_NODE;
2027 return kobj_to_node_hstate(kobj, nidp);
2030 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2031 struct kobj_attribute *attr, char *buf)
2034 unsigned long nr_huge_pages;
2037 h = kobj_to_hstate(kobj, &nid);
2038 if (nid == NUMA_NO_NODE)
2039 nr_huge_pages = h->nr_huge_pages;
2041 nr_huge_pages = h->nr_huge_pages_node[nid];
2043 return sprintf(buf, "%lu\n", nr_huge_pages);
2046 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2047 struct hstate *h, int nid,
2048 unsigned long count, size_t len)
2051 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2053 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2058 if (nid == NUMA_NO_NODE) {
2060 * global hstate attribute
2062 if (!(obey_mempolicy &&
2063 init_nodemask_of_mempolicy(nodes_allowed))) {
2064 NODEMASK_FREE(nodes_allowed);
2065 nodes_allowed = &node_states[N_MEMORY];
2067 } else if (nodes_allowed) {
2069 * per node hstate attribute: adjust count to global,
2070 * but restrict alloc/free to the specified node.
2072 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2073 init_nodemask_of_node(nodes_allowed, nid);
2075 nodes_allowed = &node_states[N_MEMORY];
2077 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2079 if (nodes_allowed != &node_states[N_MEMORY])
2080 NODEMASK_FREE(nodes_allowed);
2084 NODEMASK_FREE(nodes_allowed);
2088 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2089 struct kobject *kobj, const char *buf,
2093 unsigned long count;
2097 err = kstrtoul(buf, 10, &count);
2101 h = kobj_to_hstate(kobj, &nid);
2102 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2105 static ssize_t nr_hugepages_show(struct kobject *kobj,
2106 struct kobj_attribute *attr, char *buf)
2108 return nr_hugepages_show_common(kobj, attr, buf);
2111 static ssize_t nr_hugepages_store(struct kobject *kobj,
2112 struct kobj_attribute *attr, const char *buf, size_t len)
2114 return nr_hugepages_store_common(false, kobj, buf, len);
2116 HSTATE_ATTR(nr_hugepages);
2121 * hstate attribute for optionally mempolicy-based constraint on persistent
2122 * huge page alloc/free.
2124 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2125 struct kobj_attribute *attr, char *buf)
2127 return nr_hugepages_show_common(kobj, attr, buf);
2130 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2131 struct kobj_attribute *attr, const char *buf, size_t len)
2133 return nr_hugepages_store_common(true, kobj, buf, len);
2135 HSTATE_ATTR(nr_hugepages_mempolicy);
2139 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2140 struct kobj_attribute *attr, char *buf)
2142 struct hstate *h = kobj_to_hstate(kobj, NULL);
2143 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2146 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2147 struct kobj_attribute *attr, const char *buf, size_t count)
2150 unsigned long input;
2151 struct hstate *h = kobj_to_hstate(kobj, NULL);
2153 if (hstate_is_gigantic(h))
2156 err = kstrtoul(buf, 10, &input);
2160 spin_lock(&hugetlb_lock);
2161 h->nr_overcommit_huge_pages = input;
2162 spin_unlock(&hugetlb_lock);
2166 HSTATE_ATTR(nr_overcommit_hugepages);
2168 static ssize_t free_hugepages_show(struct kobject *kobj,
2169 struct kobj_attribute *attr, char *buf)
2172 unsigned long free_huge_pages;
2175 h = kobj_to_hstate(kobj, &nid);
2176 if (nid == NUMA_NO_NODE)
2177 free_huge_pages = h->free_huge_pages;
2179 free_huge_pages = h->free_huge_pages_node[nid];
2181 return sprintf(buf, "%lu\n", free_huge_pages);
2183 HSTATE_ATTR_RO(free_hugepages);
2185 static ssize_t resv_hugepages_show(struct kobject *kobj,
2186 struct kobj_attribute *attr, char *buf)
2188 struct hstate *h = kobj_to_hstate(kobj, NULL);
2189 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2191 HSTATE_ATTR_RO(resv_hugepages);
2193 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2194 struct kobj_attribute *attr, char *buf)
2197 unsigned long surplus_huge_pages;
2200 h = kobj_to_hstate(kobj, &nid);
2201 if (nid == NUMA_NO_NODE)
2202 surplus_huge_pages = h->surplus_huge_pages;
2204 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2206 return sprintf(buf, "%lu\n", surplus_huge_pages);
2208 HSTATE_ATTR_RO(surplus_hugepages);
2210 static struct attribute *hstate_attrs[] = {
2211 &nr_hugepages_attr.attr,
2212 &nr_overcommit_hugepages_attr.attr,
2213 &free_hugepages_attr.attr,
2214 &resv_hugepages_attr.attr,
2215 &surplus_hugepages_attr.attr,
2217 &nr_hugepages_mempolicy_attr.attr,
2222 static struct attribute_group hstate_attr_group = {
2223 .attrs = hstate_attrs,
2226 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2227 struct kobject **hstate_kobjs,
2228 struct attribute_group *hstate_attr_group)
2231 int hi = hstate_index(h);
2233 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2234 if (!hstate_kobjs[hi])
2237 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2239 kobject_put(hstate_kobjs[hi]);
2244 static void __init hugetlb_sysfs_init(void)
2249 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2250 if (!hugepages_kobj)
2253 for_each_hstate(h) {
2254 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2255 hstate_kobjs, &hstate_attr_group);
2257 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2264 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2265 * with node devices in node_devices[] using a parallel array. The array
2266 * index of a node device or _hstate == node id.
2267 * This is here to avoid any static dependency of the node device driver, in
2268 * the base kernel, on the hugetlb module.
2270 struct node_hstate {
2271 struct kobject *hugepages_kobj;
2272 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2274 struct node_hstate node_hstates[MAX_NUMNODES];
2277 * A subset of global hstate attributes for node devices
2279 static struct attribute *per_node_hstate_attrs[] = {
2280 &nr_hugepages_attr.attr,
2281 &free_hugepages_attr.attr,
2282 &surplus_hugepages_attr.attr,
2286 static struct attribute_group per_node_hstate_attr_group = {
2287 .attrs = per_node_hstate_attrs,
2291 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2292 * Returns node id via non-NULL nidp.
2294 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2298 for (nid = 0; nid < nr_node_ids; nid++) {
2299 struct node_hstate *nhs = &node_hstates[nid];
2301 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2302 if (nhs->hstate_kobjs[i] == kobj) {
2314 * Unregister hstate attributes from a single node device.
2315 * No-op if no hstate attributes attached.
2317 static void hugetlb_unregister_node(struct node *node)
2320 struct node_hstate *nhs = &node_hstates[node->dev.id];
2322 if (!nhs->hugepages_kobj)
2323 return; /* no hstate attributes */
2325 for_each_hstate(h) {
2326 int idx = hstate_index(h);
2327 if (nhs->hstate_kobjs[idx]) {
2328 kobject_put(nhs->hstate_kobjs[idx]);
2329 nhs->hstate_kobjs[idx] = NULL;
2333 kobject_put(nhs->hugepages_kobj);
2334 nhs->hugepages_kobj = NULL;
2338 * hugetlb module exit: unregister hstate attributes from node devices
2341 static void hugetlb_unregister_all_nodes(void)
2346 * disable node device registrations.
2348 register_hugetlbfs_with_node(NULL, NULL);
2351 * remove hstate attributes from any nodes that have them.
2353 for (nid = 0; nid < nr_node_ids; nid++)
2354 hugetlb_unregister_node(node_devices[nid]);
2358 * Register hstate attributes for a single node device.
2359 * No-op if attributes already registered.
2361 static void hugetlb_register_node(struct node *node)
2364 struct node_hstate *nhs = &node_hstates[node->dev.id];
2367 if (nhs->hugepages_kobj)
2368 return; /* already allocated */
2370 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2372 if (!nhs->hugepages_kobj)
2375 for_each_hstate(h) {
2376 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2378 &per_node_hstate_attr_group);
2380 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2381 h->name, node->dev.id);
2382 hugetlb_unregister_node(node);
2389 * hugetlb init time: register hstate attributes for all registered node
2390 * devices of nodes that have memory. All on-line nodes should have
2391 * registered their associated device by this time.
2393 static void __init hugetlb_register_all_nodes(void)
2397 for_each_node_state(nid, N_MEMORY) {
2398 struct node *node = node_devices[nid];
2399 if (node->dev.id == nid)
2400 hugetlb_register_node(node);
2404 * Let the node device driver know we're here so it can
2405 * [un]register hstate attributes on node hotplug.
2407 register_hugetlbfs_with_node(hugetlb_register_node,
2408 hugetlb_unregister_node);
2410 #else /* !CONFIG_NUMA */
2412 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2420 static void hugetlb_unregister_all_nodes(void) { }
2422 static void hugetlb_register_all_nodes(void) { }
2426 static void __exit hugetlb_exit(void)
2430 hugetlb_unregister_all_nodes();
2432 for_each_hstate(h) {
2433 kobject_put(hstate_kobjs[hstate_index(h)]);
2436 kobject_put(hugepages_kobj);
2437 kfree(htlb_fault_mutex_table);
2439 module_exit(hugetlb_exit);
2441 static int __init hugetlb_init(void)
2445 if (!hugepages_supported())
2448 if (!size_to_hstate(default_hstate_size)) {
2449 default_hstate_size = HPAGE_SIZE;
2450 if (!size_to_hstate(default_hstate_size))
2451 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2453 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2454 if (default_hstate_max_huge_pages)
2455 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2457 hugetlb_init_hstates();
2458 gather_bootmem_prealloc();
2461 hugetlb_sysfs_init();
2462 hugetlb_register_all_nodes();
2463 hugetlb_cgroup_file_init();
2466 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2468 num_fault_mutexes = 1;
2470 htlb_fault_mutex_table =
2471 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2472 BUG_ON(!htlb_fault_mutex_table);
2474 for (i = 0; i < num_fault_mutexes; i++)
2475 mutex_init(&htlb_fault_mutex_table[i]);
2478 module_init(hugetlb_init);
2480 /* Should be called on processing a hugepagesz=... option */
2481 void __init hugetlb_add_hstate(unsigned order)
2486 if (size_to_hstate(PAGE_SIZE << order)) {
2487 pr_warning("hugepagesz= specified twice, ignoring\n");
2490 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2492 h = &hstates[hugetlb_max_hstate++];
2494 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2495 h->nr_huge_pages = 0;
2496 h->free_huge_pages = 0;
2497 for (i = 0; i < MAX_NUMNODES; ++i)
2498 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2499 INIT_LIST_HEAD(&h->hugepage_activelist);
2500 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2501 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2502 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2503 huge_page_size(h)/1024);
2508 static int __init hugetlb_nrpages_setup(char *s)
2511 static unsigned long *last_mhp;
2514 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2515 * so this hugepages= parameter goes to the "default hstate".
2517 if (!hugetlb_max_hstate)
2518 mhp = &default_hstate_max_huge_pages;
2520 mhp = &parsed_hstate->max_huge_pages;
2522 if (mhp == last_mhp) {
2523 pr_warning("hugepages= specified twice without "
2524 "interleaving hugepagesz=, ignoring\n");
2528 if (sscanf(s, "%lu", mhp) <= 0)
2532 * Global state is always initialized later in hugetlb_init.
2533 * But we need to allocate >= MAX_ORDER hstates here early to still
2534 * use the bootmem allocator.
2536 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2537 hugetlb_hstate_alloc_pages(parsed_hstate);
2543 __setup("hugepages=", hugetlb_nrpages_setup);
2545 static int __init hugetlb_default_setup(char *s)
2547 default_hstate_size = memparse(s, &s);
2550 __setup("default_hugepagesz=", hugetlb_default_setup);
2552 static unsigned int cpuset_mems_nr(unsigned int *array)
2555 unsigned int nr = 0;
2557 for_each_node_mask(node, cpuset_current_mems_allowed)
2563 #ifdef CONFIG_SYSCTL
2564 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2565 struct ctl_table *table, int write,
2566 void __user *buffer, size_t *length, loff_t *ppos)
2568 struct hstate *h = &default_hstate;
2569 unsigned long tmp = h->max_huge_pages;
2572 if (!hugepages_supported())
2576 table->maxlen = sizeof(unsigned long);
2577 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2582 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2583 NUMA_NO_NODE, tmp, *length);
2588 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2589 void __user *buffer, size_t *length, loff_t *ppos)
2592 return hugetlb_sysctl_handler_common(false, table, write,
2593 buffer, length, ppos);
2597 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2598 void __user *buffer, size_t *length, loff_t *ppos)
2600 return hugetlb_sysctl_handler_common(true, table, write,
2601 buffer, length, ppos);
2603 #endif /* CONFIG_NUMA */
2605 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2606 void __user *buffer,
2607 size_t *length, loff_t *ppos)
2609 struct hstate *h = &default_hstate;
2613 if (!hugepages_supported())
2616 tmp = h->nr_overcommit_huge_pages;
2618 if (write && hstate_is_gigantic(h))
2622 table->maxlen = sizeof(unsigned long);
2623 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2628 spin_lock(&hugetlb_lock);
2629 h->nr_overcommit_huge_pages = tmp;
2630 spin_unlock(&hugetlb_lock);
2636 #endif /* CONFIG_SYSCTL */
2638 void hugetlb_report_meminfo(struct seq_file *m)
2640 struct hstate *h = &default_hstate;
2641 if (!hugepages_supported())
2644 "HugePages_Total: %5lu\n"
2645 "HugePages_Free: %5lu\n"
2646 "HugePages_Rsvd: %5lu\n"
2647 "HugePages_Surp: %5lu\n"
2648 "Hugepagesize: %8lu kB\n",
2652 h->surplus_huge_pages,
2653 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2656 int hugetlb_report_node_meminfo(int nid, char *buf)
2658 struct hstate *h = &default_hstate;
2659 if (!hugepages_supported())
2662 "Node %d HugePages_Total: %5u\n"
2663 "Node %d HugePages_Free: %5u\n"
2664 "Node %d HugePages_Surp: %5u\n",
2665 nid, h->nr_huge_pages_node[nid],
2666 nid, h->free_huge_pages_node[nid],
2667 nid, h->surplus_huge_pages_node[nid]);
2670 void hugetlb_show_meminfo(void)
2675 if (!hugepages_supported())
2678 for_each_node_state(nid, N_MEMORY)
2680 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2682 h->nr_huge_pages_node[nid],
2683 h->free_huge_pages_node[nid],
2684 h->surplus_huge_pages_node[nid],
2685 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2688 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2689 unsigned long hugetlb_total_pages(void)
2692 unsigned long nr_total_pages = 0;
2695 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2696 return nr_total_pages;
2699 static int hugetlb_acct_memory(struct hstate *h, long delta)
2703 spin_lock(&hugetlb_lock);
2705 * When cpuset is configured, it breaks the strict hugetlb page
2706 * reservation as the accounting is done on a global variable. Such
2707 * reservation is completely rubbish in the presence of cpuset because
2708 * the reservation is not checked against page availability for the
2709 * current cpuset. Application can still potentially OOM'ed by kernel
2710 * with lack of free htlb page in cpuset that the task is in.
2711 * Attempt to enforce strict accounting with cpuset is almost
2712 * impossible (or too ugly) because cpuset is too fluid that
2713 * task or memory node can be dynamically moved between cpusets.
2715 * The change of semantics for shared hugetlb mapping with cpuset is
2716 * undesirable. However, in order to preserve some of the semantics,
2717 * we fall back to check against current free page availability as
2718 * a best attempt and hopefully to minimize the impact of changing
2719 * semantics that cpuset has.
2722 if (gather_surplus_pages(h, delta) < 0)
2725 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2726 return_unused_surplus_pages(h, delta);
2733 return_unused_surplus_pages(h, (unsigned long) -delta);
2736 spin_unlock(&hugetlb_lock);
2740 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2742 struct resv_map *resv = vma_resv_map(vma);
2745 * This new VMA should share its siblings reservation map if present.
2746 * The VMA will only ever have a valid reservation map pointer where
2747 * it is being copied for another still existing VMA. As that VMA
2748 * has a reference to the reservation map it cannot disappear until
2749 * after this open call completes. It is therefore safe to take a
2750 * new reference here without additional locking.
2752 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2753 kref_get(&resv->refs);
2756 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2758 struct hstate *h = hstate_vma(vma);
2759 struct resv_map *resv = vma_resv_map(vma);
2760 struct hugepage_subpool *spool = subpool_vma(vma);
2761 unsigned long reserve, start, end;
2764 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2767 start = vma_hugecache_offset(h, vma, vma->vm_start);
2768 end = vma_hugecache_offset(h, vma, vma->vm_end);
2770 reserve = (end - start) - region_count(resv, start, end);
2772 kref_put(&resv->refs, resv_map_release);
2776 * Decrement reserve counts. The global reserve count may be
2777 * adjusted if the subpool has a minimum size.
2779 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
2780 hugetlb_acct_memory(h, -gbl_reserve);
2785 * We cannot handle pagefaults against hugetlb pages at all. They cause
2786 * handle_mm_fault() to try to instantiate regular-sized pages in the
2787 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2790 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2796 const struct vm_operations_struct hugetlb_vm_ops = {
2797 .fault = hugetlb_vm_op_fault,
2798 .open = hugetlb_vm_op_open,
2799 .close = hugetlb_vm_op_close,
2802 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2808 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2809 vma->vm_page_prot)));
2811 entry = huge_pte_wrprotect(mk_huge_pte(page,
2812 vma->vm_page_prot));
2814 entry = pte_mkyoung(entry);
2815 entry = pte_mkhuge(entry);
2816 entry = arch_make_huge_pte(entry, vma, page, writable);
2821 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2822 unsigned long address, pte_t *ptep)
2826 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2827 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2828 update_mmu_cache(vma, address, ptep);
2831 static int is_hugetlb_entry_migration(pte_t pte)
2835 if (huge_pte_none(pte) || pte_present(pte))
2837 swp = pte_to_swp_entry(pte);
2838 if (non_swap_entry(swp) && is_migration_entry(swp))
2844 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2848 if (huge_pte_none(pte) || pte_present(pte))
2850 swp = pte_to_swp_entry(pte);
2851 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2857 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2858 struct vm_area_struct *vma)
2860 pte_t *src_pte, *dst_pte, entry;
2861 struct page *ptepage;
2864 struct hstate *h = hstate_vma(vma);
2865 unsigned long sz = huge_page_size(h);
2866 unsigned long mmun_start; /* For mmu_notifiers */
2867 unsigned long mmun_end; /* For mmu_notifiers */
2870 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2872 mmun_start = vma->vm_start;
2873 mmun_end = vma->vm_end;
2875 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2877 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2878 spinlock_t *src_ptl, *dst_ptl;
2879 src_pte = huge_pte_offset(src, addr);
2882 dst_pte = huge_pte_alloc(dst, addr, sz);
2888 /* If the pagetables are shared don't copy or take references */
2889 if (dst_pte == src_pte)
2892 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2893 src_ptl = huge_pte_lockptr(h, src, src_pte);
2894 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
2895 entry = huge_ptep_get(src_pte);
2896 if (huge_pte_none(entry)) { /* skip none entry */
2898 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2899 is_hugetlb_entry_hwpoisoned(entry))) {
2900 swp_entry_t swp_entry = pte_to_swp_entry(entry);
2902 if (is_write_migration_entry(swp_entry) && cow) {
2904 * COW mappings require pages in both
2905 * parent and child to be set to read.
2907 make_migration_entry_read(&swp_entry);
2908 entry = swp_entry_to_pte(swp_entry);
2909 set_huge_pte_at(src, addr, src_pte, entry);
2911 set_huge_pte_at(dst, addr, dst_pte, entry);
2914 huge_ptep_set_wrprotect(src, addr, src_pte);
2915 mmu_notifier_invalidate_range(src, mmun_start,
2918 entry = huge_ptep_get(src_pte);
2919 ptepage = pte_page(entry);
2921 page_dup_rmap(ptepage);
2922 set_huge_pte_at(dst, addr, dst_pte, entry);
2924 spin_unlock(src_ptl);
2925 spin_unlock(dst_ptl);
2929 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
2934 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2935 unsigned long start, unsigned long end,
2936 struct page *ref_page)
2938 int force_flush = 0;
2939 struct mm_struct *mm = vma->vm_mm;
2940 unsigned long address;
2945 struct hstate *h = hstate_vma(vma);
2946 unsigned long sz = huge_page_size(h);
2947 const unsigned long mmun_start = start; /* For mmu_notifiers */
2948 const unsigned long mmun_end = end; /* For mmu_notifiers */
2950 WARN_ON(!is_vm_hugetlb_page(vma));
2951 BUG_ON(start & ~huge_page_mask(h));
2952 BUG_ON(end & ~huge_page_mask(h));
2954 tlb_start_vma(tlb, vma);
2955 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2958 for (; address < end; address += sz) {
2959 ptep = huge_pte_offset(mm, address);
2963 ptl = huge_pte_lock(h, mm, ptep);
2964 if (huge_pmd_unshare(mm, &address, ptep))
2967 pte = huge_ptep_get(ptep);
2968 if (huge_pte_none(pte))
2972 * Migrating hugepage or HWPoisoned hugepage is already
2973 * unmapped and its refcount is dropped, so just clear pte here.
2975 if (unlikely(!pte_present(pte))) {
2976 huge_pte_clear(mm, address, ptep);
2980 page = pte_page(pte);
2982 * If a reference page is supplied, it is because a specific
2983 * page is being unmapped, not a range. Ensure the page we
2984 * are about to unmap is the actual page of interest.
2987 if (page != ref_page)
2991 * Mark the VMA as having unmapped its page so that
2992 * future faults in this VMA will fail rather than
2993 * looking like data was lost
2995 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2998 pte = huge_ptep_get_and_clear(mm, address, ptep);
2999 tlb_remove_tlb_entry(tlb, ptep, address);
3000 if (huge_pte_dirty(pte))
3001 set_page_dirty(page);
3003 page_remove_rmap(page);
3004 force_flush = !__tlb_remove_page(tlb, page);
3010 /* Bail out after unmapping reference page if supplied */
3019 * mmu_gather ran out of room to batch pages, we break out of
3020 * the PTE lock to avoid doing the potential expensive TLB invalidate
3021 * and page-free while holding it.
3026 if (address < end && !ref_page)
3029 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3030 tlb_end_vma(tlb, vma);
3033 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3034 struct vm_area_struct *vma, unsigned long start,
3035 unsigned long end, struct page *ref_page)
3037 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3040 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3041 * test will fail on a vma being torn down, and not grab a page table
3042 * on its way out. We're lucky that the flag has such an appropriate
3043 * name, and can in fact be safely cleared here. We could clear it
3044 * before the __unmap_hugepage_range above, but all that's necessary
3045 * is to clear it before releasing the i_mmap_rwsem. This works
3046 * because in the context this is called, the VMA is about to be
3047 * destroyed and the i_mmap_rwsem is held.
3049 vma->vm_flags &= ~VM_MAYSHARE;
3052 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3053 unsigned long end, struct page *ref_page)
3055 struct mm_struct *mm;
3056 struct mmu_gather tlb;
3060 tlb_gather_mmu(&tlb, mm, start, end);
3061 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3062 tlb_finish_mmu(&tlb, start, end);
3066 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3067 * mappping it owns the reserve page for. The intention is to unmap the page
3068 * from other VMAs and let the children be SIGKILLed if they are faulting the
3071 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3072 struct page *page, unsigned long address)
3074 struct hstate *h = hstate_vma(vma);
3075 struct vm_area_struct *iter_vma;
3076 struct address_space *mapping;
3080 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3081 * from page cache lookup which is in HPAGE_SIZE units.
3083 address = address & huge_page_mask(h);
3084 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3086 mapping = file_inode(vma->vm_file)->i_mapping;
3089 * Take the mapping lock for the duration of the table walk. As
3090 * this mapping should be shared between all the VMAs,
3091 * __unmap_hugepage_range() is called as the lock is already held
3093 i_mmap_lock_write(mapping);
3094 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3095 /* Do not unmap the current VMA */
3096 if (iter_vma == vma)
3100 * Unmap the page from other VMAs without their own reserves.
3101 * They get marked to be SIGKILLed if they fault in these
3102 * areas. This is because a future no-page fault on this VMA
3103 * could insert a zeroed page instead of the data existing
3104 * from the time of fork. This would look like data corruption
3106 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3107 unmap_hugepage_range(iter_vma, address,
3108 address + huge_page_size(h), page);
3110 i_mmap_unlock_write(mapping);
3114 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3115 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3116 * cannot race with other handlers or page migration.
3117 * Keep the pte_same checks anyway to make transition from the mutex easier.
3119 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3120 unsigned long address, pte_t *ptep, pte_t pte,
3121 struct page *pagecache_page, spinlock_t *ptl)
3123 struct hstate *h = hstate_vma(vma);
3124 struct page *old_page, *new_page;
3125 int ret = 0, outside_reserve = 0;
3126 unsigned long mmun_start; /* For mmu_notifiers */
3127 unsigned long mmun_end; /* For mmu_notifiers */
3129 old_page = pte_page(pte);
3132 /* If no-one else is actually using this page, avoid the copy
3133 * and just make the page writable */
3134 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3135 page_move_anon_rmap(old_page, vma, address);
3136 set_huge_ptep_writable(vma, address, ptep);
3141 * If the process that created a MAP_PRIVATE mapping is about to
3142 * perform a COW due to a shared page count, attempt to satisfy
3143 * the allocation without using the existing reserves. The pagecache
3144 * page is used to determine if the reserve at this address was
3145 * consumed or not. If reserves were used, a partial faulted mapping
3146 * at the time of fork() could consume its reserves on COW instead
3147 * of the full address range.
3149 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3150 old_page != pagecache_page)
3151 outside_reserve = 1;
3153 page_cache_get(old_page);
3156 * Drop page table lock as buddy allocator may be called. It will
3157 * be acquired again before returning to the caller, as expected.
3160 new_page = alloc_huge_page(vma, address, outside_reserve);
3162 if (IS_ERR(new_page)) {
3164 * If a process owning a MAP_PRIVATE mapping fails to COW,
3165 * it is due to references held by a child and an insufficient
3166 * huge page pool. To guarantee the original mappers
3167 * reliability, unmap the page from child processes. The child
3168 * may get SIGKILLed if it later faults.
3170 if (outside_reserve) {
3171 page_cache_release(old_page);
3172 BUG_ON(huge_pte_none(pte));
3173 unmap_ref_private(mm, vma, old_page, address);
3174 BUG_ON(huge_pte_none(pte));
3176 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3178 pte_same(huge_ptep_get(ptep), pte)))
3179 goto retry_avoidcopy;
3181 * race occurs while re-acquiring page table
3182 * lock, and our job is done.
3187 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3188 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3189 goto out_release_old;
3193 * When the original hugepage is shared one, it does not have
3194 * anon_vma prepared.
3196 if (unlikely(anon_vma_prepare(vma))) {
3198 goto out_release_all;
3201 copy_user_huge_page(new_page, old_page, address, vma,
3202 pages_per_huge_page(h));
3203 __SetPageUptodate(new_page);
3204 set_page_huge_active(new_page);
3206 mmun_start = address & huge_page_mask(h);
3207 mmun_end = mmun_start + huge_page_size(h);
3208 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3211 * Retake the page table lock to check for racing updates
3212 * before the page tables are altered
3215 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3216 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3217 ClearPagePrivate(new_page);
3220 huge_ptep_clear_flush(vma, address, ptep);
3221 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3222 set_huge_pte_at(mm, address, ptep,
3223 make_huge_pte(vma, new_page, 1));
3224 page_remove_rmap(old_page);
3225 hugepage_add_new_anon_rmap(new_page, vma, address);
3226 /* Make the old page be freed below */
3227 new_page = old_page;
3230 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3232 page_cache_release(new_page);
3234 page_cache_release(old_page);
3236 spin_lock(ptl); /* Caller expects lock to be held */
3240 /* Return the pagecache page at a given address within a VMA */
3241 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3242 struct vm_area_struct *vma, unsigned long address)
3244 struct address_space *mapping;
3247 mapping = vma->vm_file->f_mapping;
3248 idx = vma_hugecache_offset(h, vma, address);
3250 return find_lock_page(mapping, idx);
3254 * Return whether there is a pagecache page to back given address within VMA.
3255 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3257 static bool hugetlbfs_pagecache_present(struct hstate *h,
3258 struct vm_area_struct *vma, unsigned long address)
3260 struct address_space *mapping;
3264 mapping = vma->vm_file->f_mapping;
3265 idx = vma_hugecache_offset(h, vma, address);
3267 page = find_get_page(mapping, idx);
3270 return page != NULL;
3273 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3274 struct address_space *mapping, pgoff_t idx,
3275 unsigned long address, pte_t *ptep, unsigned int flags)
3277 struct hstate *h = hstate_vma(vma);
3278 int ret = VM_FAULT_SIGBUS;
3286 * Currently, we are forced to kill the process in the event the
3287 * original mapper has unmapped pages from the child due to a failed
3288 * COW. Warn that such a situation has occurred as it may not be obvious
3290 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3291 pr_warning("PID %d killed due to inadequate hugepage pool\n",
3297 * Use page lock to guard against racing truncation
3298 * before we get page_table_lock.
3301 page = find_lock_page(mapping, idx);
3303 size = i_size_read(mapping->host) >> huge_page_shift(h);
3306 page = alloc_huge_page(vma, address, 0);
3308 ret = PTR_ERR(page);
3312 ret = VM_FAULT_SIGBUS;
3315 clear_huge_page(page, address, pages_per_huge_page(h));
3316 __SetPageUptodate(page);
3317 set_page_huge_active(page);
3319 if (vma->vm_flags & VM_MAYSHARE) {
3321 struct inode *inode = mapping->host;
3323 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3330 ClearPagePrivate(page);
3332 spin_lock(&inode->i_lock);
3333 inode->i_blocks += blocks_per_huge_page(h);
3334 spin_unlock(&inode->i_lock);
3337 if (unlikely(anon_vma_prepare(vma))) {
3339 goto backout_unlocked;
3345 * If memory error occurs between mmap() and fault, some process
3346 * don't have hwpoisoned swap entry for errored virtual address.
3347 * So we need to block hugepage fault by PG_hwpoison bit check.
3349 if (unlikely(PageHWPoison(page))) {
3350 ret = VM_FAULT_HWPOISON |
3351 VM_FAULT_SET_HINDEX(hstate_index(h));
3352 goto backout_unlocked;
3357 * If we are going to COW a private mapping later, we examine the
3358 * pending reservations for this page now. This will ensure that
3359 * any allocations necessary to record that reservation occur outside
3362 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3363 if (vma_needs_reservation(h, vma, address) < 0) {
3365 goto backout_unlocked;
3367 /* Just decrements count, does not deallocate */
3368 vma_abort_reservation(h, vma, address);
3371 ptl = huge_pte_lockptr(h, mm, ptep);
3373 size = i_size_read(mapping->host) >> huge_page_shift(h);
3378 if (!huge_pte_none(huge_ptep_get(ptep)))
3382 ClearPagePrivate(page);
3383 hugepage_add_new_anon_rmap(page, vma, address);
3385 page_dup_rmap(page);
3386 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3387 && (vma->vm_flags & VM_SHARED)));
3388 set_huge_pte_at(mm, address, ptep, new_pte);
3390 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3391 /* Optimization, do the COW without a second fault */
3392 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
3409 static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3410 struct vm_area_struct *vma,
3411 struct address_space *mapping,
3412 pgoff_t idx, unsigned long address)
3414 unsigned long key[2];
3417 if (vma->vm_flags & VM_SHARED) {
3418 key[0] = (unsigned long) mapping;
3421 key[0] = (unsigned long) mm;
3422 key[1] = address >> huge_page_shift(h);
3425 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3427 return hash & (num_fault_mutexes - 1);
3431 * For uniprocesor systems we always use a single mutex, so just
3432 * return 0 and avoid the hashing overhead.
3434 static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3435 struct vm_area_struct *vma,
3436 struct address_space *mapping,
3437 pgoff_t idx, unsigned long address)
3443 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3444 unsigned long address, unsigned int flags)
3451 struct page *page = NULL;
3452 struct page *pagecache_page = NULL;
3453 struct hstate *h = hstate_vma(vma);
3454 struct address_space *mapping;
3455 int need_wait_lock = 0;
3457 address &= huge_page_mask(h);
3459 ptep = huge_pte_offset(mm, address);
3461 entry = huge_ptep_get(ptep);
3462 if (unlikely(is_hugetlb_entry_migration(entry))) {
3463 migration_entry_wait_huge(vma, mm, ptep);
3465 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3466 return VM_FAULT_HWPOISON_LARGE |
3467 VM_FAULT_SET_HINDEX(hstate_index(h));
3470 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3472 return VM_FAULT_OOM;
3474 mapping = vma->vm_file->f_mapping;
3475 idx = vma_hugecache_offset(h, vma, address);
3478 * Serialize hugepage allocation and instantiation, so that we don't
3479 * get spurious allocation failures if two CPUs race to instantiate
3480 * the same page in the page cache.
3482 hash = fault_mutex_hash(h, mm, vma, mapping, idx, address);
3483 mutex_lock(&htlb_fault_mutex_table[hash]);
3485 entry = huge_ptep_get(ptep);
3486 if (huge_pte_none(entry)) {
3487 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3494 * entry could be a migration/hwpoison entry at this point, so this
3495 * check prevents the kernel from going below assuming that we have
3496 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3497 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3500 if (!pte_present(entry))
3504 * If we are going to COW the mapping later, we examine the pending
3505 * reservations for this page now. This will ensure that any
3506 * allocations necessary to record that reservation occur outside the
3507 * spinlock. For private mappings, we also lookup the pagecache
3508 * page now as it is used to determine if a reservation has been
3511 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3512 if (vma_needs_reservation(h, vma, address) < 0) {
3516 /* Just decrements count, does not deallocate */
3517 vma_abort_reservation(h, vma, address);
3519 if (!(vma->vm_flags & VM_MAYSHARE))
3520 pagecache_page = hugetlbfs_pagecache_page(h,
3524 ptl = huge_pte_lock(h, mm, ptep);
3526 /* Check for a racing update before calling hugetlb_cow */
3527 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3531 * hugetlb_cow() requires page locks of pte_page(entry) and
3532 * pagecache_page, so here we need take the former one
3533 * when page != pagecache_page or !pagecache_page.
3535 page = pte_page(entry);
3536 if (page != pagecache_page)
3537 if (!trylock_page(page)) {
3544 if (flags & FAULT_FLAG_WRITE) {
3545 if (!huge_pte_write(entry)) {
3546 ret = hugetlb_cow(mm, vma, address, ptep, entry,
3547 pagecache_page, ptl);
3550 entry = huge_pte_mkdirty(entry);
3552 entry = pte_mkyoung(entry);
3553 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3554 flags & FAULT_FLAG_WRITE))
3555 update_mmu_cache(vma, address, ptep);
3557 if (page != pagecache_page)
3563 if (pagecache_page) {
3564 unlock_page(pagecache_page);
3565 put_page(pagecache_page);
3568 mutex_unlock(&htlb_fault_mutex_table[hash]);
3570 * Generally it's safe to hold refcount during waiting page lock. But
3571 * here we just wait to defer the next page fault to avoid busy loop and
3572 * the page is not used after unlocked before returning from the current
3573 * page fault. So we are safe from accessing freed page, even if we wait
3574 * here without taking refcount.
3577 wait_on_page_locked(page);
3581 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3582 struct page **pages, struct vm_area_struct **vmas,
3583 unsigned long *position, unsigned long *nr_pages,
3584 long i, unsigned int flags)
3586 unsigned long pfn_offset;
3587 unsigned long vaddr = *position;
3588 unsigned long remainder = *nr_pages;
3589 struct hstate *h = hstate_vma(vma);
3591 while (vaddr < vma->vm_end && remainder) {
3593 spinlock_t *ptl = NULL;
3598 * If we have a pending SIGKILL, don't keep faulting pages and
3599 * potentially allocating memory.
3601 if (unlikely(fatal_signal_pending(current))) {
3607 * Some archs (sparc64, sh*) have multiple pte_ts to
3608 * each hugepage. We have to make sure we get the
3609 * first, for the page indexing below to work.
3611 * Note that page table lock is not held when pte is null.
3613 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3615 ptl = huge_pte_lock(h, mm, pte);
3616 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3619 * When coredumping, it suits get_dump_page if we just return
3620 * an error where there's an empty slot with no huge pagecache
3621 * to back it. This way, we avoid allocating a hugepage, and
3622 * the sparse dumpfile avoids allocating disk blocks, but its
3623 * huge holes still show up with zeroes where they need to be.
3625 if (absent && (flags & FOLL_DUMP) &&
3626 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3634 * We need call hugetlb_fault for both hugepages under migration
3635 * (in which case hugetlb_fault waits for the migration,) and
3636 * hwpoisoned hugepages (in which case we need to prevent the
3637 * caller from accessing to them.) In order to do this, we use
3638 * here is_swap_pte instead of is_hugetlb_entry_migration and
3639 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3640 * both cases, and because we can't follow correct pages
3641 * directly from any kind of swap entries.
3643 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3644 ((flags & FOLL_WRITE) &&
3645 !huge_pte_write(huge_ptep_get(pte)))) {
3650 ret = hugetlb_fault(mm, vma, vaddr,
3651 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3652 if (!(ret & VM_FAULT_ERROR))
3659 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3660 page = pte_page(huge_ptep_get(pte));
3663 pages[i] = mem_map_offset(page, pfn_offset);
3664 get_page_foll(pages[i]);
3674 if (vaddr < vma->vm_end && remainder &&
3675 pfn_offset < pages_per_huge_page(h)) {
3677 * We use pfn_offset to avoid touching the pageframes
3678 * of this compound page.
3684 *nr_pages = remainder;
3687 return i ? i : -EFAULT;
3690 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3691 unsigned long address, unsigned long end, pgprot_t newprot)
3693 struct mm_struct *mm = vma->vm_mm;
3694 unsigned long start = address;
3697 struct hstate *h = hstate_vma(vma);
3698 unsigned long pages = 0;
3700 BUG_ON(address >= end);
3701 flush_cache_range(vma, address, end);
3703 mmu_notifier_invalidate_range_start(mm, start, end);
3704 i_mmap_lock_write(vma->vm_file->f_mapping);
3705 for (; address < end; address += huge_page_size(h)) {
3707 ptep = huge_pte_offset(mm, address);
3710 ptl = huge_pte_lock(h, mm, ptep);
3711 if (huge_pmd_unshare(mm, &address, ptep)) {
3716 pte = huge_ptep_get(ptep);
3717 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3721 if (unlikely(is_hugetlb_entry_migration(pte))) {
3722 swp_entry_t entry = pte_to_swp_entry(pte);
3724 if (is_write_migration_entry(entry)) {
3727 make_migration_entry_read(&entry);
3728 newpte = swp_entry_to_pte(entry);
3729 set_huge_pte_at(mm, address, ptep, newpte);
3735 if (!huge_pte_none(pte)) {
3736 pte = huge_ptep_get_and_clear(mm, address, ptep);
3737 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3738 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3739 set_huge_pte_at(mm, address, ptep, pte);
3745 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
3746 * may have cleared our pud entry and done put_page on the page table:
3747 * once we release i_mmap_rwsem, another task can do the final put_page
3748 * and that page table be reused and filled with junk.
3750 flush_tlb_range(vma, start, end);
3751 mmu_notifier_invalidate_range(mm, start, end);
3752 i_mmap_unlock_write(vma->vm_file->f_mapping);
3753 mmu_notifier_invalidate_range_end(mm, start, end);
3755 return pages << h->order;
3758 int hugetlb_reserve_pages(struct inode *inode,
3760 struct vm_area_struct *vma,
3761 vm_flags_t vm_flags)
3764 struct hstate *h = hstate_inode(inode);
3765 struct hugepage_subpool *spool = subpool_inode(inode);
3766 struct resv_map *resv_map;
3770 * Only apply hugepage reservation if asked. At fault time, an
3771 * attempt will be made for VM_NORESERVE to allocate a page
3772 * without using reserves
3774 if (vm_flags & VM_NORESERVE)
3778 * Shared mappings base their reservation on the number of pages that
3779 * are already allocated on behalf of the file. Private mappings need
3780 * to reserve the full area even if read-only as mprotect() may be
3781 * called to make the mapping read-write. Assume !vma is a shm mapping
3783 if (!vma || vma->vm_flags & VM_MAYSHARE) {
3784 resv_map = inode_resv_map(inode);
3786 chg = region_chg(resv_map, from, to);
3789 resv_map = resv_map_alloc();
3795 set_vma_resv_map(vma, resv_map);
3796 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3805 * There must be enough pages in the subpool for the mapping. If
3806 * the subpool has a minimum size, there may be some global
3807 * reservations already in place (gbl_reserve).
3809 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
3810 if (gbl_reserve < 0) {
3816 * Check enough hugepages are available for the reservation.
3817 * Hand the pages back to the subpool if there are not
3819 ret = hugetlb_acct_memory(h, gbl_reserve);
3821 /* put back original number of pages, chg */
3822 (void)hugepage_subpool_put_pages(spool, chg);
3827 * Account for the reservations made. Shared mappings record regions
3828 * that have reservations as they are shared by multiple VMAs.
3829 * When the last VMA disappears, the region map says how much
3830 * the reservation was and the page cache tells how much of
3831 * the reservation was consumed. Private mappings are per-VMA and
3832 * only the consumed reservations are tracked. When the VMA
3833 * disappears, the original reservation is the VMA size and the
3834 * consumed reservations are stored in the map. Hence, nothing
3835 * else has to be done for private mappings here
3837 if (!vma || vma->vm_flags & VM_MAYSHARE) {
3838 long add = region_add(resv_map, from, to);
3840 if (unlikely(chg > add)) {
3842 * pages in this range were added to the reserve
3843 * map between region_chg and region_add. This
3844 * indicates a race with alloc_huge_page. Adjust
3845 * the subpool and reserve counts modified above
3846 * based on the difference.
3850 rsv_adjust = hugepage_subpool_put_pages(spool,
3852 hugetlb_acct_memory(h, -rsv_adjust);
3857 if (!vma || vma->vm_flags & VM_MAYSHARE)
3858 region_abort(resv_map, from, to);
3859 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3860 kref_put(&resv_map->refs, resv_map_release);
3864 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3866 struct hstate *h = hstate_inode(inode);
3867 struct resv_map *resv_map = inode_resv_map(inode);
3869 struct hugepage_subpool *spool = subpool_inode(inode);
3873 chg = region_truncate(resv_map, offset);
3874 spin_lock(&inode->i_lock);
3875 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3876 spin_unlock(&inode->i_lock);
3879 * If the subpool has a minimum size, the number of global
3880 * reservations to be released may be adjusted.
3882 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
3883 hugetlb_acct_memory(h, -gbl_reserve);
3886 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3887 static unsigned long page_table_shareable(struct vm_area_struct *svma,
3888 struct vm_area_struct *vma,
3889 unsigned long addr, pgoff_t idx)
3891 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3893 unsigned long sbase = saddr & PUD_MASK;
3894 unsigned long s_end = sbase + PUD_SIZE;
3896 /* Allow segments to share if only one is marked locked */
3897 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3898 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3901 * match the virtual addresses, permission and the alignment of the
3904 if (pmd_index(addr) != pmd_index(saddr) ||
3905 vm_flags != svm_flags ||
3906 sbase < svma->vm_start || svma->vm_end < s_end)
3912 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3914 unsigned long base = addr & PUD_MASK;
3915 unsigned long end = base + PUD_SIZE;
3918 * check on proper vm_flags and page table alignment
3920 if (vma->vm_flags & VM_MAYSHARE &&
3921 vma->vm_start <= base && end <= vma->vm_end)
3927 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
3928 * and returns the corresponding pte. While this is not necessary for the
3929 * !shared pmd case because we can allocate the pmd later as well, it makes the
3930 * code much cleaner. pmd allocation is essential for the shared case because
3931 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
3932 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
3933 * bad pmd for sharing.
3935 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3937 struct vm_area_struct *vma = find_vma(mm, addr);
3938 struct address_space *mapping = vma->vm_file->f_mapping;
3939 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3941 struct vm_area_struct *svma;
3942 unsigned long saddr;
3947 if (!vma_shareable(vma, addr))
3948 return (pte_t *)pmd_alloc(mm, pud, addr);
3950 i_mmap_lock_write(mapping);
3951 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3955 saddr = page_table_shareable(svma, vma, addr, idx);
3957 spte = huge_pte_offset(svma->vm_mm, saddr);
3960 get_page(virt_to_page(spte));
3969 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
3971 if (pud_none(*pud)) {
3972 pud_populate(mm, pud,
3973 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3975 put_page(virt_to_page(spte));
3980 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3981 i_mmap_unlock_write(mapping);
3986 * unmap huge page backed by shared pte.
3988 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
3989 * indicated by page_count > 1, unmap is achieved by clearing pud and
3990 * decrementing the ref count. If count == 1, the pte page is not shared.
3992 * called with page table lock held.
3994 * returns: 1 successfully unmapped a shared pte page
3995 * 0 the underlying pte page is not shared, or it is the last user
3997 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3999 pgd_t *pgd = pgd_offset(mm, *addr);
4000 pud_t *pud = pud_offset(pgd, *addr);
4002 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4003 if (page_count(virt_to_page(ptep)) == 1)
4007 put_page(virt_to_page(ptep));
4009 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4012 #define want_pmd_share() (1)
4013 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4014 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4019 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4023 #define want_pmd_share() (0)
4024 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4026 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4027 pte_t *huge_pte_alloc(struct mm_struct *mm,
4028 unsigned long addr, unsigned long sz)
4034 pgd = pgd_offset(mm, addr);
4035 pud = pud_alloc(mm, pgd, addr);
4037 if (sz == PUD_SIZE) {
4040 BUG_ON(sz != PMD_SIZE);
4041 if (want_pmd_share() && pud_none(*pud))
4042 pte = huge_pmd_share(mm, addr, pud);
4044 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4047 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
4052 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
4058 pgd = pgd_offset(mm, addr);
4059 if (pgd_present(*pgd)) {
4060 pud = pud_offset(pgd, addr);
4061 if (pud_present(*pud)) {
4063 return (pte_t *)pud;
4064 pmd = pmd_offset(pud, addr);
4067 return (pte_t *) pmd;
4070 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4073 * These functions are overwritable if your architecture needs its own
4076 struct page * __weak
4077 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4080 return ERR_PTR(-EINVAL);
4083 struct page * __weak
4084 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4085 pmd_t *pmd, int flags)
4087 struct page *page = NULL;
4090 ptl = pmd_lockptr(mm, pmd);
4093 * make sure that the address range covered by this pmd is not
4094 * unmapped from other threads.
4096 if (!pmd_huge(*pmd))
4098 if (pmd_present(*pmd)) {
4099 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4100 if (flags & FOLL_GET)
4103 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
4105 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4109 * hwpoisoned entry is treated as no_page_table in
4110 * follow_page_mask().
4118 struct page * __weak
4119 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4120 pud_t *pud, int flags)
4122 if (flags & FOLL_GET)
4125 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4128 #ifdef CONFIG_MEMORY_FAILURE
4131 * This function is called from memory failure code.
4132 * Assume the caller holds page lock of the head page.
4134 int dequeue_hwpoisoned_huge_page(struct page *hpage)
4136 struct hstate *h = page_hstate(hpage);
4137 int nid = page_to_nid(hpage);
4140 spin_lock(&hugetlb_lock);
4142 * Just checking !page_huge_active is not enough, because that could be
4143 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4145 if (!page_huge_active(hpage) && !page_count(hpage)) {
4147 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4148 * but dangling hpage->lru can trigger list-debug warnings
4149 * (this happens when we call unpoison_memory() on it),
4150 * so let it point to itself with list_del_init().
4152 list_del_init(&hpage->lru);
4153 set_page_refcounted(hpage);
4154 h->free_huge_pages--;
4155 h->free_huge_pages_node[nid]--;
4158 spin_unlock(&hugetlb_lock);
4163 bool isolate_huge_page(struct page *page, struct list_head *list)
4167 VM_BUG_ON_PAGE(!PageHead(page), page);
4168 spin_lock(&hugetlb_lock);
4169 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4173 clear_page_huge_active(page);
4174 list_move_tail(&page->lru, list);
4176 spin_unlock(&hugetlb_lock);
4180 void putback_active_hugepage(struct page *page)
4182 VM_BUG_ON_PAGE(!PageHead(page), page);
4183 spin_lock(&hugetlb_lock);
4184 set_page_huge_active(page);
4185 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4186 spin_unlock(&hugetlb_lock);