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 struct mutex *hugetlb_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 * Delete the specified range [f, t) from the reserve map. If the
464 * t parameter is LONG_MAX, this indicates that ALL regions after f
465 * should be deleted. Locate the regions which intersect [f, t)
466 * and either trim, delete or split the existing regions.
468 * Returns the number of huge pages deleted from the reserve map.
469 * In the normal case, the return value is zero or more. In the
470 * case where a region must be split, a new region descriptor must
471 * be allocated. If the allocation fails, -ENOMEM will be returned.
472 * NOTE: If the parameter t == LONG_MAX, then we will never split
473 * a region and possibly return -ENOMEM. Callers specifying
474 * t == LONG_MAX do not need to check for -ENOMEM error.
476 static long region_del(struct resv_map *resv, long f, long t)
478 struct list_head *head = &resv->regions;
479 struct file_region *rg, *trg;
480 struct file_region *nrg = NULL;
484 spin_lock(&resv->lock);
485 list_for_each_entry_safe(rg, trg, head, link) {
491 if (f > rg->from && t < rg->to) { /* Must split region */
493 * Check for an entry in the cache before dropping
494 * lock and attempting allocation.
497 resv->region_cache_count > resv->adds_in_progress) {
498 nrg = list_first_entry(&resv->region_cache,
501 list_del(&nrg->link);
502 resv->region_cache_count--;
506 spin_unlock(&resv->lock);
507 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
515 /* New entry for end of split region */
518 INIT_LIST_HEAD(&nrg->link);
520 /* Original entry is trimmed */
523 list_add(&nrg->link, &rg->link);
528 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
529 del += rg->to - rg->from;
535 if (f <= rg->from) { /* Trim beginning of region */
538 } else { /* Trim end of region */
544 spin_unlock(&resv->lock);
550 * A rare out of memory error was encountered which prevented removal of
551 * the reserve map region for a page. The huge page itself was free'ed
552 * and removed from the page cache. This routine will adjust the subpool
553 * usage count, and the global reserve count if needed. By incrementing
554 * these counts, the reserve map entry which could not be deleted will
555 * appear as a "reserved" entry instead of simply dangling with incorrect
558 void hugetlb_fix_reserve_counts(struct inode *inode, bool restore_reserve)
560 struct hugepage_subpool *spool = subpool_inode(inode);
563 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
564 if (restore_reserve && rsv_adjust) {
565 struct hstate *h = hstate_inode(inode);
567 hugetlb_acct_memory(h, 1);
572 * Count and return the number of huge pages in the reserve map
573 * that intersect with the range [f, t).
575 static long region_count(struct resv_map *resv, long f, long t)
577 struct list_head *head = &resv->regions;
578 struct file_region *rg;
581 spin_lock(&resv->lock);
582 /* Locate each segment we overlap with, and count that overlap. */
583 list_for_each_entry(rg, head, link) {
592 seg_from = max(rg->from, f);
593 seg_to = min(rg->to, t);
595 chg += seg_to - seg_from;
597 spin_unlock(&resv->lock);
603 * Convert the address within this vma to the page offset within
604 * the mapping, in pagecache page units; huge pages here.
606 static pgoff_t vma_hugecache_offset(struct hstate *h,
607 struct vm_area_struct *vma, unsigned long address)
609 return ((address - vma->vm_start) >> huge_page_shift(h)) +
610 (vma->vm_pgoff >> huge_page_order(h));
613 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
614 unsigned long address)
616 return vma_hugecache_offset(hstate_vma(vma), vma, address);
620 * Return the size of the pages allocated when backing a VMA. In the majority
621 * cases this will be same size as used by the page table entries.
623 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
625 struct hstate *hstate;
627 if (!is_vm_hugetlb_page(vma))
630 hstate = hstate_vma(vma);
632 return 1UL << huge_page_shift(hstate);
634 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
637 * Return the page size being used by the MMU to back a VMA. In the majority
638 * of cases, the page size used by the kernel matches the MMU size. On
639 * architectures where it differs, an architecture-specific version of this
640 * function is required.
642 #ifndef vma_mmu_pagesize
643 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
645 return vma_kernel_pagesize(vma);
650 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
651 * bits of the reservation map pointer, which are always clear due to
654 #define HPAGE_RESV_OWNER (1UL << 0)
655 #define HPAGE_RESV_UNMAPPED (1UL << 1)
656 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
659 * These helpers are used to track how many pages are reserved for
660 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
661 * is guaranteed to have their future faults succeed.
663 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
664 * the reserve counters are updated with the hugetlb_lock held. It is safe
665 * to reset the VMA at fork() time as it is not in use yet and there is no
666 * chance of the global counters getting corrupted as a result of the values.
668 * The private mapping reservation is represented in a subtly different
669 * manner to a shared mapping. A shared mapping has a region map associated
670 * with the underlying file, this region map represents the backing file
671 * pages which have ever had a reservation assigned which this persists even
672 * after the page is instantiated. A private mapping has a region map
673 * associated with the original mmap which is attached to all VMAs which
674 * reference it, this region map represents those offsets which have consumed
675 * reservation ie. where pages have been instantiated.
677 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
679 return (unsigned long)vma->vm_private_data;
682 static void set_vma_private_data(struct vm_area_struct *vma,
685 vma->vm_private_data = (void *)value;
688 struct resv_map *resv_map_alloc(void)
690 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
691 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
693 if (!resv_map || !rg) {
699 kref_init(&resv_map->refs);
700 spin_lock_init(&resv_map->lock);
701 INIT_LIST_HEAD(&resv_map->regions);
703 resv_map->adds_in_progress = 0;
705 INIT_LIST_HEAD(&resv_map->region_cache);
706 list_add(&rg->link, &resv_map->region_cache);
707 resv_map->region_cache_count = 1;
712 void resv_map_release(struct kref *ref)
714 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
715 struct list_head *head = &resv_map->region_cache;
716 struct file_region *rg, *trg;
718 /* Clear out any active regions before we release the map. */
719 region_del(resv_map, 0, LONG_MAX);
721 /* ... and any entries left in the cache */
722 list_for_each_entry_safe(rg, trg, head, link) {
727 VM_BUG_ON(resv_map->adds_in_progress);
732 static inline struct resv_map *inode_resv_map(struct inode *inode)
734 return inode->i_mapping->private_data;
737 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
739 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
740 if (vma->vm_flags & VM_MAYSHARE) {
741 struct address_space *mapping = vma->vm_file->f_mapping;
742 struct inode *inode = mapping->host;
744 return inode_resv_map(inode);
747 return (struct resv_map *)(get_vma_private_data(vma) &
752 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
754 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
755 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
757 set_vma_private_data(vma, (get_vma_private_data(vma) &
758 HPAGE_RESV_MASK) | (unsigned long)map);
761 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
763 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
764 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
766 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
769 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
771 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
773 return (get_vma_private_data(vma) & flag) != 0;
776 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
777 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
779 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
780 if (!(vma->vm_flags & VM_MAYSHARE))
781 vma->vm_private_data = (void *)0;
784 /* Returns true if the VMA has associated reserve pages */
785 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
787 if (vma->vm_flags & VM_NORESERVE) {
789 * This address is already reserved by other process(chg == 0),
790 * so, we should decrement reserved count. Without decrementing,
791 * reserve count remains after releasing inode, because this
792 * allocated page will go into page cache and is regarded as
793 * coming from reserved pool in releasing step. Currently, we
794 * don't have any other solution to deal with this situation
795 * properly, so add work-around here.
797 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
803 /* Shared mappings always use reserves */
804 if (vma->vm_flags & VM_MAYSHARE)
808 * Only the process that called mmap() has reserves for
811 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
817 static void enqueue_huge_page(struct hstate *h, struct page *page)
819 int nid = page_to_nid(page);
820 list_move(&page->lru, &h->hugepage_freelists[nid]);
821 h->free_huge_pages++;
822 h->free_huge_pages_node[nid]++;
825 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
829 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
830 if (!is_migrate_isolate_page(page))
833 * if 'non-isolated free hugepage' not found on the list,
834 * the allocation fails.
836 if (&h->hugepage_freelists[nid] == &page->lru)
838 list_move(&page->lru, &h->hugepage_activelist);
839 set_page_refcounted(page);
840 h->free_huge_pages--;
841 h->free_huge_pages_node[nid]--;
845 /* Movability of hugepages depends on migration support. */
846 static inline gfp_t htlb_alloc_mask(struct hstate *h)
848 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
849 return GFP_HIGHUSER_MOVABLE;
854 static struct page *dequeue_huge_page_vma(struct hstate *h,
855 struct vm_area_struct *vma,
856 unsigned long address, int avoid_reserve,
859 struct page *page = NULL;
860 struct mempolicy *mpol;
861 nodemask_t *nodemask;
862 struct zonelist *zonelist;
865 unsigned int cpuset_mems_cookie;
868 * A child process with MAP_PRIVATE mappings created by their parent
869 * have no page reserves. This check ensures that reservations are
870 * not "stolen". The child may still get SIGKILLed
872 if (!vma_has_reserves(vma, chg) &&
873 h->free_huge_pages - h->resv_huge_pages == 0)
876 /* If reserves cannot be used, ensure enough pages are in the pool */
877 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
881 cpuset_mems_cookie = read_mems_allowed_begin();
882 zonelist = huge_zonelist(vma, address,
883 htlb_alloc_mask(h), &mpol, &nodemask);
885 for_each_zone_zonelist_nodemask(zone, z, zonelist,
886 MAX_NR_ZONES - 1, nodemask) {
887 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
888 page = dequeue_huge_page_node(h, zone_to_nid(zone));
892 if (!vma_has_reserves(vma, chg))
895 SetPagePrivate(page);
896 h->resv_huge_pages--;
903 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
912 * common helper functions for hstate_next_node_to_{alloc|free}.
913 * We may have allocated or freed a huge page based on a different
914 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
915 * be outside of *nodes_allowed. Ensure that we use an allowed
916 * node for alloc or free.
918 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
920 nid = next_node(nid, *nodes_allowed);
921 if (nid == MAX_NUMNODES)
922 nid = first_node(*nodes_allowed);
923 VM_BUG_ON(nid >= MAX_NUMNODES);
928 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
930 if (!node_isset(nid, *nodes_allowed))
931 nid = next_node_allowed(nid, nodes_allowed);
936 * returns the previously saved node ["this node"] from which to
937 * allocate a persistent huge page for the pool and advance the
938 * next node from which to allocate, handling wrap at end of node
941 static int hstate_next_node_to_alloc(struct hstate *h,
942 nodemask_t *nodes_allowed)
946 VM_BUG_ON(!nodes_allowed);
948 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
949 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
955 * helper for free_pool_huge_page() - return the previously saved
956 * node ["this node"] from which to free a huge page. Advance the
957 * next node id whether or not we find a free huge page to free so
958 * that the next attempt to free addresses the next node.
960 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
964 VM_BUG_ON(!nodes_allowed);
966 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
967 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
972 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
973 for (nr_nodes = nodes_weight(*mask); \
975 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
978 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
979 for (nr_nodes = nodes_weight(*mask); \
981 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
984 #if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
985 static void destroy_compound_gigantic_page(struct page *page,
989 int nr_pages = 1 << order;
990 struct page *p = page + 1;
992 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
994 set_page_refcounted(p);
995 p->first_page = NULL;
998 set_compound_order(page, 0);
999 __ClearPageHead(page);
1002 static void free_gigantic_page(struct page *page, unsigned order)
1004 free_contig_range(page_to_pfn(page), 1 << order);
1007 static int __alloc_gigantic_page(unsigned long start_pfn,
1008 unsigned long nr_pages)
1010 unsigned long end_pfn = start_pfn + nr_pages;
1011 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1014 static bool pfn_range_valid_gigantic(unsigned long start_pfn,
1015 unsigned long nr_pages)
1017 unsigned long i, end_pfn = start_pfn + nr_pages;
1020 for (i = start_pfn; i < end_pfn; i++) {
1024 page = pfn_to_page(i);
1026 if (PageReserved(page))
1029 if (page_count(page) > 0)
1039 static bool zone_spans_last_pfn(const struct zone *zone,
1040 unsigned long start_pfn, unsigned long nr_pages)
1042 unsigned long last_pfn = start_pfn + nr_pages - 1;
1043 return zone_spans_pfn(zone, last_pfn);
1046 static struct page *alloc_gigantic_page(int nid, unsigned order)
1048 unsigned long nr_pages = 1 << order;
1049 unsigned long ret, pfn, flags;
1052 z = NODE_DATA(nid)->node_zones;
1053 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
1054 spin_lock_irqsave(&z->lock, flags);
1056 pfn = ALIGN(z->zone_start_pfn, nr_pages);
1057 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
1058 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
1060 * We release the zone lock here because
1061 * alloc_contig_range() will also lock the zone
1062 * at some point. If there's an allocation
1063 * spinning on this lock, it may win the race
1064 * and cause alloc_contig_range() to fail...
1066 spin_unlock_irqrestore(&z->lock, flags);
1067 ret = __alloc_gigantic_page(pfn, nr_pages);
1069 return pfn_to_page(pfn);
1070 spin_lock_irqsave(&z->lock, flags);
1075 spin_unlock_irqrestore(&z->lock, flags);
1081 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1082 static void prep_compound_gigantic_page(struct page *page, unsigned long order);
1084 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1088 page = alloc_gigantic_page(nid, huge_page_order(h));
1090 prep_compound_gigantic_page(page, huge_page_order(h));
1091 prep_new_huge_page(h, page, nid);
1097 static int alloc_fresh_gigantic_page(struct hstate *h,
1098 nodemask_t *nodes_allowed)
1100 struct page *page = NULL;
1103 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1104 page = alloc_fresh_gigantic_page_node(h, node);
1112 static inline bool gigantic_page_supported(void) { return true; }
1114 static inline bool gigantic_page_supported(void) { return false; }
1115 static inline void free_gigantic_page(struct page *page, unsigned order) { }
1116 static inline void destroy_compound_gigantic_page(struct page *page,
1117 unsigned long order) { }
1118 static inline int alloc_fresh_gigantic_page(struct hstate *h,
1119 nodemask_t *nodes_allowed) { return 0; }
1122 static void update_and_free_page(struct hstate *h, struct page *page)
1126 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1130 h->nr_huge_pages_node[page_to_nid(page)]--;
1131 for (i = 0; i < pages_per_huge_page(h); i++) {
1132 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1133 1 << PG_referenced | 1 << PG_dirty |
1134 1 << PG_active | 1 << PG_private |
1137 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1138 set_compound_page_dtor(page, NULL);
1139 set_page_refcounted(page);
1140 if (hstate_is_gigantic(h)) {
1141 destroy_compound_gigantic_page(page, huge_page_order(h));
1142 free_gigantic_page(page, huge_page_order(h));
1144 __free_pages(page, huge_page_order(h));
1148 struct hstate *size_to_hstate(unsigned long size)
1152 for_each_hstate(h) {
1153 if (huge_page_size(h) == size)
1160 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1161 * to hstate->hugepage_activelist.)
1163 * This function can be called for tail pages, but never returns true for them.
1165 bool page_huge_active(struct page *page)
1167 VM_BUG_ON_PAGE(!PageHuge(page), page);
1168 return PageHead(page) && PagePrivate(&page[1]);
1171 /* never called for tail page */
1172 static void set_page_huge_active(struct page *page)
1174 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1175 SetPagePrivate(&page[1]);
1178 static void clear_page_huge_active(struct page *page)
1180 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1181 ClearPagePrivate(&page[1]);
1184 void free_huge_page(struct page *page)
1187 * Can't pass hstate in here because it is called from the
1188 * compound page destructor.
1190 struct hstate *h = page_hstate(page);
1191 int nid = page_to_nid(page);
1192 struct hugepage_subpool *spool =
1193 (struct hugepage_subpool *)page_private(page);
1194 bool restore_reserve;
1196 set_page_private(page, 0);
1197 page->mapping = NULL;
1198 BUG_ON(page_count(page));
1199 BUG_ON(page_mapcount(page));
1200 restore_reserve = PagePrivate(page);
1201 ClearPagePrivate(page);
1204 * A return code of zero implies that the subpool will be under its
1205 * minimum size if the reservation is not restored after page is free.
1206 * Therefore, force restore_reserve operation.
1208 if (hugepage_subpool_put_pages(spool, 1) == 0)
1209 restore_reserve = true;
1211 spin_lock(&hugetlb_lock);
1212 clear_page_huge_active(page);
1213 hugetlb_cgroup_uncharge_page(hstate_index(h),
1214 pages_per_huge_page(h), page);
1215 if (restore_reserve)
1216 h->resv_huge_pages++;
1218 if (h->surplus_huge_pages_node[nid]) {
1219 /* remove the page from active list */
1220 list_del(&page->lru);
1221 update_and_free_page(h, page);
1222 h->surplus_huge_pages--;
1223 h->surplus_huge_pages_node[nid]--;
1225 arch_clear_hugepage_flags(page);
1226 enqueue_huge_page(h, page);
1228 spin_unlock(&hugetlb_lock);
1231 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1233 INIT_LIST_HEAD(&page->lru);
1234 set_compound_page_dtor(page, free_huge_page);
1235 spin_lock(&hugetlb_lock);
1236 set_hugetlb_cgroup(page, NULL);
1238 h->nr_huge_pages_node[nid]++;
1239 spin_unlock(&hugetlb_lock);
1240 put_page(page); /* free it into the hugepage allocator */
1243 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
1246 int nr_pages = 1 << order;
1247 struct page *p = page + 1;
1249 /* we rely on prep_new_huge_page to set the destructor */
1250 set_compound_order(page, order);
1251 __SetPageHead(page);
1252 __ClearPageReserved(page);
1253 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1255 * For gigantic hugepages allocated through bootmem at
1256 * boot, it's safer to be consistent with the not-gigantic
1257 * hugepages and clear the PG_reserved bit from all tail pages
1258 * too. Otherwse drivers using get_user_pages() to access tail
1259 * pages may get the reference counting wrong if they see
1260 * PG_reserved set on a tail page (despite the head page not
1261 * having PG_reserved set). Enforcing this consistency between
1262 * head and tail pages allows drivers to optimize away a check
1263 * on the head page when they need know if put_page() is needed
1264 * after get_user_pages().
1266 __ClearPageReserved(p);
1267 set_page_count(p, 0);
1268 p->first_page = page;
1269 /* Make sure p->first_page is always valid for PageTail() */
1276 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1277 * transparent huge pages. See the PageTransHuge() documentation for more
1280 int PageHuge(struct page *page)
1282 if (!PageCompound(page))
1285 page = compound_head(page);
1286 return get_compound_page_dtor(page) == free_huge_page;
1288 EXPORT_SYMBOL_GPL(PageHuge);
1291 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1292 * normal or transparent huge pages.
1294 int PageHeadHuge(struct page *page_head)
1296 if (!PageHead(page_head))
1299 return get_compound_page_dtor(page_head) == free_huge_page;
1302 pgoff_t __basepage_index(struct page *page)
1304 struct page *page_head = compound_head(page);
1305 pgoff_t index = page_index(page_head);
1306 unsigned long compound_idx;
1308 if (!PageHuge(page_head))
1309 return page_index(page);
1311 if (compound_order(page_head) >= MAX_ORDER)
1312 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1314 compound_idx = page - page_head;
1316 return (index << compound_order(page_head)) + compound_idx;
1319 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1323 page = alloc_pages_exact_node(nid,
1324 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1325 __GFP_REPEAT|__GFP_NOWARN,
1326 huge_page_order(h));
1328 prep_new_huge_page(h, page, nid);
1334 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1340 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1341 page = alloc_fresh_huge_page_node(h, node);
1349 count_vm_event(HTLB_BUDDY_PGALLOC);
1351 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1357 * Free huge page from pool from next node to free.
1358 * Attempt to keep persistent huge pages more or less
1359 * balanced over allowed nodes.
1360 * Called with hugetlb_lock locked.
1362 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1368 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1370 * If we're returning unused surplus pages, only examine
1371 * nodes with surplus pages.
1373 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1374 !list_empty(&h->hugepage_freelists[node])) {
1376 list_entry(h->hugepage_freelists[node].next,
1378 list_del(&page->lru);
1379 h->free_huge_pages--;
1380 h->free_huge_pages_node[node]--;
1382 h->surplus_huge_pages--;
1383 h->surplus_huge_pages_node[node]--;
1385 update_and_free_page(h, page);
1395 * Dissolve a given free hugepage into free buddy pages. This function does
1396 * nothing for in-use (including surplus) hugepages.
1398 static void dissolve_free_huge_page(struct page *page)
1400 spin_lock(&hugetlb_lock);
1401 if (PageHuge(page) && !page_count(page)) {
1402 struct hstate *h = page_hstate(page);
1403 int nid = page_to_nid(page);
1404 list_del(&page->lru);
1405 h->free_huge_pages--;
1406 h->free_huge_pages_node[nid]--;
1407 update_and_free_page(h, page);
1409 spin_unlock(&hugetlb_lock);
1413 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1414 * make specified memory blocks removable from the system.
1415 * Note that start_pfn should aligned with (minimum) hugepage size.
1417 void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1421 if (!hugepages_supported())
1424 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order));
1425 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
1426 dissolve_free_huge_page(pfn_to_page(pfn));
1429 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
1434 if (hstate_is_gigantic(h))
1438 * Assume we will successfully allocate the surplus page to
1439 * prevent racing processes from causing the surplus to exceed
1442 * This however introduces a different race, where a process B
1443 * tries to grow the static hugepage pool while alloc_pages() is
1444 * called by process A. B will only examine the per-node
1445 * counters in determining if surplus huge pages can be
1446 * converted to normal huge pages in adjust_pool_surplus(). A
1447 * won't be able to increment the per-node counter, until the
1448 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1449 * no more huge pages can be converted from surplus to normal
1450 * state (and doesn't try to convert again). Thus, we have a
1451 * case where a surplus huge page exists, the pool is grown, and
1452 * the surplus huge page still exists after, even though it
1453 * should just have been converted to a normal huge page. This
1454 * does not leak memory, though, as the hugepage will be freed
1455 * once it is out of use. It also does not allow the counters to
1456 * go out of whack in adjust_pool_surplus() as we don't modify
1457 * the node values until we've gotten the hugepage and only the
1458 * per-node value is checked there.
1460 spin_lock(&hugetlb_lock);
1461 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1462 spin_unlock(&hugetlb_lock);
1466 h->surplus_huge_pages++;
1468 spin_unlock(&hugetlb_lock);
1470 if (nid == NUMA_NO_NODE)
1471 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
1472 __GFP_REPEAT|__GFP_NOWARN,
1473 huge_page_order(h));
1475 page = alloc_pages_exact_node(nid,
1476 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1477 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
1479 spin_lock(&hugetlb_lock);
1481 INIT_LIST_HEAD(&page->lru);
1482 r_nid = page_to_nid(page);
1483 set_compound_page_dtor(page, free_huge_page);
1484 set_hugetlb_cgroup(page, NULL);
1486 * We incremented the global counters already
1488 h->nr_huge_pages_node[r_nid]++;
1489 h->surplus_huge_pages_node[r_nid]++;
1490 __count_vm_event(HTLB_BUDDY_PGALLOC);
1493 h->surplus_huge_pages--;
1494 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1496 spin_unlock(&hugetlb_lock);
1502 * This allocation function is useful in the context where vma is irrelevant.
1503 * E.g. soft-offlining uses this function because it only cares physical
1504 * address of error page.
1506 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1508 struct page *page = NULL;
1510 spin_lock(&hugetlb_lock);
1511 if (h->free_huge_pages - h->resv_huge_pages > 0)
1512 page = dequeue_huge_page_node(h, nid);
1513 spin_unlock(&hugetlb_lock);
1516 page = alloc_buddy_huge_page(h, nid);
1522 * Increase the hugetlb pool such that it can accommodate a reservation
1525 static int gather_surplus_pages(struct hstate *h, int delta)
1527 struct list_head surplus_list;
1528 struct page *page, *tmp;
1530 int needed, allocated;
1531 bool alloc_ok = true;
1533 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1535 h->resv_huge_pages += delta;
1540 INIT_LIST_HEAD(&surplus_list);
1544 spin_unlock(&hugetlb_lock);
1545 for (i = 0; i < needed; i++) {
1546 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1551 list_add(&page->lru, &surplus_list);
1556 * After retaking hugetlb_lock, we need to recalculate 'needed'
1557 * because either resv_huge_pages or free_huge_pages may have changed.
1559 spin_lock(&hugetlb_lock);
1560 needed = (h->resv_huge_pages + delta) -
1561 (h->free_huge_pages + allocated);
1566 * We were not able to allocate enough pages to
1567 * satisfy the entire reservation so we free what
1568 * we've allocated so far.
1573 * The surplus_list now contains _at_least_ the number of extra pages
1574 * needed to accommodate the reservation. Add the appropriate number
1575 * of pages to the hugetlb pool and free the extras back to the buddy
1576 * allocator. Commit the entire reservation here to prevent another
1577 * process from stealing the pages as they are added to the pool but
1578 * before they are reserved.
1580 needed += allocated;
1581 h->resv_huge_pages += delta;
1584 /* Free the needed pages to the hugetlb pool */
1585 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1589 * This page is now managed by the hugetlb allocator and has
1590 * no users -- drop the buddy allocator's reference.
1592 put_page_testzero(page);
1593 VM_BUG_ON_PAGE(page_count(page), page);
1594 enqueue_huge_page(h, page);
1597 spin_unlock(&hugetlb_lock);
1599 /* Free unnecessary surplus pages to the buddy allocator */
1600 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1602 spin_lock(&hugetlb_lock);
1608 * When releasing a hugetlb pool reservation, any surplus pages that were
1609 * allocated to satisfy the reservation must be explicitly freed if they were
1611 * Called with hugetlb_lock held.
1613 static void return_unused_surplus_pages(struct hstate *h,
1614 unsigned long unused_resv_pages)
1616 unsigned long nr_pages;
1618 /* Uncommit the reservation */
1619 h->resv_huge_pages -= unused_resv_pages;
1621 /* Cannot return gigantic pages currently */
1622 if (hstate_is_gigantic(h))
1625 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1628 * We want to release as many surplus pages as possible, spread
1629 * evenly across all nodes with memory. Iterate across these nodes
1630 * until we can no longer free unreserved surplus pages. This occurs
1631 * when the nodes with surplus pages have no free pages.
1632 * free_pool_huge_page() will balance the the freed pages across the
1633 * on-line nodes with memory and will handle the hstate accounting.
1635 while (nr_pages--) {
1636 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1638 cond_resched_lock(&hugetlb_lock);
1644 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1645 * are used by the huge page allocation routines to manage reservations.
1647 * vma_needs_reservation is called to determine if the huge page at addr
1648 * within the vma has an associated reservation. If a reservation is
1649 * needed, the value 1 is returned. The caller is then responsible for
1650 * managing the global reservation and subpool usage counts. After
1651 * the huge page has been allocated, vma_commit_reservation is called
1652 * to add the page to the reservation map. If the page allocation fails,
1653 * the reservation must be ended instead of committed. vma_end_reservation
1654 * is called in such cases.
1656 * In the normal case, vma_commit_reservation returns the same value
1657 * as the preceding vma_needs_reservation call. The only time this
1658 * is not the case is if a reserve map was changed between calls. It
1659 * is the responsibility of the caller to notice the difference and
1660 * take appropriate action.
1662 enum vma_resv_mode {
1667 static long __vma_reservation_common(struct hstate *h,
1668 struct vm_area_struct *vma, unsigned long addr,
1669 enum vma_resv_mode mode)
1671 struct resv_map *resv;
1675 resv = vma_resv_map(vma);
1679 idx = vma_hugecache_offset(h, vma, addr);
1681 case VMA_NEEDS_RESV:
1682 ret = region_chg(resv, idx, idx + 1);
1684 case VMA_COMMIT_RESV:
1685 ret = region_add(resv, idx, idx + 1);
1688 region_abort(resv, idx, idx + 1);
1695 if (vma->vm_flags & VM_MAYSHARE)
1698 return ret < 0 ? ret : 0;
1701 static long vma_needs_reservation(struct hstate *h,
1702 struct vm_area_struct *vma, unsigned long addr)
1704 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1707 static long vma_commit_reservation(struct hstate *h,
1708 struct vm_area_struct *vma, unsigned long addr)
1710 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1713 static void vma_end_reservation(struct hstate *h,
1714 struct vm_area_struct *vma, unsigned long addr)
1716 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1719 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1720 unsigned long addr, int avoid_reserve)
1722 struct hugepage_subpool *spool = subpool_vma(vma);
1723 struct hstate *h = hstate_vma(vma);
1727 struct hugetlb_cgroup *h_cg;
1729 idx = hstate_index(h);
1731 * Processes that did not create the mapping will have no
1732 * reserves and will not have accounted against subpool
1733 * limit. Check that the subpool limit can be made before
1734 * satisfying the allocation MAP_NORESERVE mappings may also
1735 * need pages and subpool limit allocated allocated if no reserve
1738 chg = vma_needs_reservation(h, vma, addr);
1740 return ERR_PTR(-ENOMEM);
1741 if (chg || avoid_reserve)
1742 if (hugepage_subpool_get_pages(spool, 1) < 0) {
1743 vma_end_reservation(h, vma, addr);
1744 return ERR_PTR(-ENOSPC);
1747 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1749 goto out_subpool_put;
1751 spin_lock(&hugetlb_lock);
1752 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
1754 spin_unlock(&hugetlb_lock);
1755 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1757 goto out_uncharge_cgroup;
1759 spin_lock(&hugetlb_lock);
1760 list_move(&page->lru, &h->hugepage_activelist);
1763 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1764 spin_unlock(&hugetlb_lock);
1766 set_page_private(page, (unsigned long)spool);
1768 commit = vma_commit_reservation(h, vma, addr);
1769 if (unlikely(chg > commit)) {
1771 * The page was added to the reservation map between
1772 * vma_needs_reservation and vma_commit_reservation.
1773 * This indicates a race with hugetlb_reserve_pages.
1774 * Adjust for the subpool count incremented above AND
1775 * in hugetlb_reserve_pages for the same page. Also,
1776 * the reservation count added in hugetlb_reserve_pages
1777 * no longer applies.
1781 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
1782 hugetlb_acct_memory(h, -rsv_adjust);
1786 out_uncharge_cgroup:
1787 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1789 if (chg || avoid_reserve)
1790 hugepage_subpool_put_pages(spool, 1);
1791 vma_end_reservation(h, vma, addr);
1792 return ERR_PTR(-ENOSPC);
1796 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1797 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1798 * where no ERR_VALUE is expected to be returned.
1800 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1801 unsigned long addr, int avoid_reserve)
1803 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1809 int __weak alloc_bootmem_huge_page(struct hstate *h)
1811 struct huge_bootmem_page *m;
1814 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1817 addr = memblock_virt_alloc_try_nid_nopanic(
1818 huge_page_size(h), huge_page_size(h),
1819 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1822 * Use the beginning of the huge page to store the
1823 * huge_bootmem_page struct (until gather_bootmem
1824 * puts them into the mem_map).
1833 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
1834 /* Put them into a private list first because mem_map is not up yet */
1835 list_add(&m->list, &huge_boot_pages);
1840 static void __init prep_compound_huge_page(struct page *page, int order)
1842 if (unlikely(order > (MAX_ORDER - 1)))
1843 prep_compound_gigantic_page(page, order);
1845 prep_compound_page(page, order);
1848 /* Put bootmem huge pages into the standard lists after mem_map is up */
1849 static void __init gather_bootmem_prealloc(void)
1851 struct huge_bootmem_page *m;
1853 list_for_each_entry(m, &huge_boot_pages, list) {
1854 struct hstate *h = m->hstate;
1857 #ifdef CONFIG_HIGHMEM
1858 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1859 memblock_free_late(__pa(m),
1860 sizeof(struct huge_bootmem_page));
1862 page = virt_to_page(m);
1864 WARN_ON(page_count(page) != 1);
1865 prep_compound_huge_page(page, h->order);
1866 WARN_ON(PageReserved(page));
1867 prep_new_huge_page(h, page, page_to_nid(page));
1869 * If we had gigantic hugepages allocated at boot time, we need
1870 * to restore the 'stolen' pages to totalram_pages in order to
1871 * fix confusing memory reports from free(1) and another
1872 * side-effects, like CommitLimit going negative.
1874 if (hstate_is_gigantic(h))
1875 adjust_managed_page_count(page, 1 << h->order);
1879 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1883 for (i = 0; i < h->max_huge_pages; ++i) {
1884 if (hstate_is_gigantic(h)) {
1885 if (!alloc_bootmem_huge_page(h))
1887 } else if (!alloc_fresh_huge_page(h,
1888 &node_states[N_MEMORY]))
1891 h->max_huge_pages = i;
1894 static void __init hugetlb_init_hstates(void)
1898 for_each_hstate(h) {
1899 if (minimum_order > huge_page_order(h))
1900 minimum_order = huge_page_order(h);
1902 /* oversize hugepages were init'ed in early boot */
1903 if (!hstate_is_gigantic(h))
1904 hugetlb_hstate_alloc_pages(h);
1906 VM_BUG_ON(minimum_order == UINT_MAX);
1909 static char * __init memfmt(char *buf, unsigned long n)
1911 if (n >= (1UL << 30))
1912 sprintf(buf, "%lu GB", n >> 30);
1913 else if (n >= (1UL << 20))
1914 sprintf(buf, "%lu MB", n >> 20);
1916 sprintf(buf, "%lu KB", n >> 10);
1920 static void __init report_hugepages(void)
1924 for_each_hstate(h) {
1926 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1927 memfmt(buf, huge_page_size(h)),
1928 h->free_huge_pages);
1932 #ifdef CONFIG_HIGHMEM
1933 static void try_to_free_low(struct hstate *h, unsigned long count,
1934 nodemask_t *nodes_allowed)
1938 if (hstate_is_gigantic(h))
1941 for_each_node_mask(i, *nodes_allowed) {
1942 struct page *page, *next;
1943 struct list_head *freel = &h->hugepage_freelists[i];
1944 list_for_each_entry_safe(page, next, freel, lru) {
1945 if (count >= h->nr_huge_pages)
1947 if (PageHighMem(page))
1949 list_del(&page->lru);
1950 update_and_free_page(h, page);
1951 h->free_huge_pages--;
1952 h->free_huge_pages_node[page_to_nid(page)]--;
1957 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1958 nodemask_t *nodes_allowed)
1964 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1965 * balanced by operating on them in a round-robin fashion.
1966 * Returns 1 if an adjustment was made.
1968 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1973 VM_BUG_ON(delta != -1 && delta != 1);
1976 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1977 if (h->surplus_huge_pages_node[node])
1981 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1982 if (h->surplus_huge_pages_node[node] <
1983 h->nr_huge_pages_node[node])
1990 h->surplus_huge_pages += delta;
1991 h->surplus_huge_pages_node[node] += delta;
1995 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1996 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1997 nodemask_t *nodes_allowed)
1999 unsigned long min_count, ret;
2001 if (hstate_is_gigantic(h) && !gigantic_page_supported())
2002 return h->max_huge_pages;
2005 * Increase the pool size
2006 * First take pages out of surplus state. Then make up the
2007 * remaining difference by allocating fresh huge pages.
2009 * We might race with alloc_buddy_huge_page() here and be unable
2010 * to convert a surplus huge page to a normal huge page. That is
2011 * not critical, though, it just means the overall size of the
2012 * pool might be one hugepage larger than it needs to be, but
2013 * within all the constraints specified by the sysctls.
2015 spin_lock(&hugetlb_lock);
2016 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2017 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2021 while (count > persistent_huge_pages(h)) {
2023 * If this allocation races such that we no longer need the
2024 * page, free_huge_page will handle it by freeing the page
2025 * and reducing the surplus.
2027 spin_unlock(&hugetlb_lock);
2028 if (hstate_is_gigantic(h))
2029 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2031 ret = alloc_fresh_huge_page(h, nodes_allowed);
2032 spin_lock(&hugetlb_lock);
2036 /* Bail for signals. Probably ctrl-c from user */
2037 if (signal_pending(current))
2042 * Decrease the pool size
2043 * First return free pages to the buddy allocator (being careful
2044 * to keep enough around to satisfy reservations). Then place
2045 * pages into surplus state as needed so the pool will shrink
2046 * to the desired size as pages become free.
2048 * By placing pages into the surplus state independent of the
2049 * overcommit value, we are allowing the surplus pool size to
2050 * exceed overcommit. There are few sane options here. Since
2051 * alloc_buddy_huge_page() is checking the global counter,
2052 * though, we'll note that we're not allowed to exceed surplus
2053 * and won't grow the pool anywhere else. Not until one of the
2054 * sysctls are changed, or the surplus pages go out of use.
2056 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2057 min_count = max(count, min_count);
2058 try_to_free_low(h, min_count, nodes_allowed);
2059 while (min_count < persistent_huge_pages(h)) {
2060 if (!free_pool_huge_page(h, nodes_allowed, 0))
2062 cond_resched_lock(&hugetlb_lock);
2064 while (count < persistent_huge_pages(h)) {
2065 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2069 ret = persistent_huge_pages(h);
2070 spin_unlock(&hugetlb_lock);
2074 #define HSTATE_ATTR_RO(_name) \
2075 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2077 #define HSTATE_ATTR(_name) \
2078 static struct kobj_attribute _name##_attr = \
2079 __ATTR(_name, 0644, _name##_show, _name##_store)
2081 static struct kobject *hugepages_kobj;
2082 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2084 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2086 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2090 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2091 if (hstate_kobjs[i] == kobj) {
2093 *nidp = NUMA_NO_NODE;
2097 return kobj_to_node_hstate(kobj, nidp);
2100 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2101 struct kobj_attribute *attr, char *buf)
2104 unsigned long nr_huge_pages;
2107 h = kobj_to_hstate(kobj, &nid);
2108 if (nid == NUMA_NO_NODE)
2109 nr_huge_pages = h->nr_huge_pages;
2111 nr_huge_pages = h->nr_huge_pages_node[nid];
2113 return sprintf(buf, "%lu\n", nr_huge_pages);
2116 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2117 struct hstate *h, int nid,
2118 unsigned long count, size_t len)
2121 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2123 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2128 if (nid == NUMA_NO_NODE) {
2130 * global hstate attribute
2132 if (!(obey_mempolicy &&
2133 init_nodemask_of_mempolicy(nodes_allowed))) {
2134 NODEMASK_FREE(nodes_allowed);
2135 nodes_allowed = &node_states[N_MEMORY];
2137 } else if (nodes_allowed) {
2139 * per node hstate attribute: adjust count to global,
2140 * but restrict alloc/free to the specified node.
2142 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2143 init_nodemask_of_node(nodes_allowed, nid);
2145 nodes_allowed = &node_states[N_MEMORY];
2147 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2149 if (nodes_allowed != &node_states[N_MEMORY])
2150 NODEMASK_FREE(nodes_allowed);
2154 NODEMASK_FREE(nodes_allowed);
2158 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2159 struct kobject *kobj, const char *buf,
2163 unsigned long count;
2167 err = kstrtoul(buf, 10, &count);
2171 h = kobj_to_hstate(kobj, &nid);
2172 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2175 static ssize_t nr_hugepages_show(struct kobject *kobj,
2176 struct kobj_attribute *attr, char *buf)
2178 return nr_hugepages_show_common(kobj, attr, buf);
2181 static ssize_t nr_hugepages_store(struct kobject *kobj,
2182 struct kobj_attribute *attr, const char *buf, size_t len)
2184 return nr_hugepages_store_common(false, kobj, buf, len);
2186 HSTATE_ATTR(nr_hugepages);
2191 * hstate attribute for optionally mempolicy-based constraint on persistent
2192 * huge page alloc/free.
2194 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2195 struct kobj_attribute *attr, char *buf)
2197 return nr_hugepages_show_common(kobj, attr, buf);
2200 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2201 struct kobj_attribute *attr, const char *buf, size_t len)
2203 return nr_hugepages_store_common(true, kobj, buf, len);
2205 HSTATE_ATTR(nr_hugepages_mempolicy);
2209 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2210 struct kobj_attribute *attr, char *buf)
2212 struct hstate *h = kobj_to_hstate(kobj, NULL);
2213 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2216 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2217 struct kobj_attribute *attr, const char *buf, size_t count)
2220 unsigned long input;
2221 struct hstate *h = kobj_to_hstate(kobj, NULL);
2223 if (hstate_is_gigantic(h))
2226 err = kstrtoul(buf, 10, &input);
2230 spin_lock(&hugetlb_lock);
2231 h->nr_overcommit_huge_pages = input;
2232 spin_unlock(&hugetlb_lock);
2236 HSTATE_ATTR(nr_overcommit_hugepages);
2238 static ssize_t free_hugepages_show(struct kobject *kobj,
2239 struct kobj_attribute *attr, char *buf)
2242 unsigned long free_huge_pages;
2245 h = kobj_to_hstate(kobj, &nid);
2246 if (nid == NUMA_NO_NODE)
2247 free_huge_pages = h->free_huge_pages;
2249 free_huge_pages = h->free_huge_pages_node[nid];
2251 return sprintf(buf, "%lu\n", free_huge_pages);
2253 HSTATE_ATTR_RO(free_hugepages);
2255 static ssize_t resv_hugepages_show(struct kobject *kobj,
2256 struct kobj_attribute *attr, char *buf)
2258 struct hstate *h = kobj_to_hstate(kobj, NULL);
2259 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2261 HSTATE_ATTR_RO(resv_hugepages);
2263 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2264 struct kobj_attribute *attr, char *buf)
2267 unsigned long surplus_huge_pages;
2270 h = kobj_to_hstate(kobj, &nid);
2271 if (nid == NUMA_NO_NODE)
2272 surplus_huge_pages = h->surplus_huge_pages;
2274 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2276 return sprintf(buf, "%lu\n", surplus_huge_pages);
2278 HSTATE_ATTR_RO(surplus_hugepages);
2280 static struct attribute *hstate_attrs[] = {
2281 &nr_hugepages_attr.attr,
2282 &nr_overcommit_hugepages_attr.attr,
2283 &free_hugepages_attr.attr,
2284 &resv_hugepages_attr.attr,
2285 &surplus_hugepages_attr.attr,
2287 &nr_hugepages_mempolicy_attr.attr,
2292 static struct attribute_group hstate_attr_group = {
2293 .attrs = hstate_attrs,
2296 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2297 struct kobject **hstate_kobjs,
2298 struct attribute_group *hstate_attr_group)
2301 int hi = hstate_index(h);
2303 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2304 if (!hstate_kobjs[hi])
2307 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2309 kobject_put(hstate_kobjs[hi]);
2314 static void __init hugetlb_sysfs_init(void)
2319 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2320 if (!hugepages_kobj)
2323 for_each_hstate(h) {
2324 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2325 hstate_kobjs, &hstate_attr_group);
2327 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2334 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2335 * with node devices in node_devices[] using a parallel array. The array
2336 * index of a node device or _hstate == node id.
2337 * This is here to avoid any static dependency of the node device driver, in
2338 * the base kernel, on the hugetlb module.
2340 struct node_hstate {
2341 struct kobject *hugepages_kobj;
2342 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2344 struct node_hstate node_hstates[MAX_NUMNODES];
2347 * A subset of global hstate attributes for node devices
2349 static struct attribute *per_node_hstate_attrs[] = {
2350 &nr_hugepages_attr.attr,
2351 &free_hugepages_attr.attr,
2352 &surplus_hugepages_attr.attr,
2356 static struct attribute_group per_node_hstate_attr_group = {
2357 .attrs = per_node_hstate_attrs,
2361 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2362 * Returns node id via non-NULL nidp.
2364 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2368 for (nid = 0; nid < nr_node_ids; nid++) {
2369 struct node_hstate *nhs = &node_hstates[nid];
2371 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2372 if (nhs->hstate_kobjs[i] == kobj) {
2384 * Unregister hstate attributes from a single node device.
2385 * No-op if no hstate attributes attached.
2387 static void hugetlb_unregister_node(struct node *node)
2390 struct node_hstate *nhs = &node_hstates[node->dev.id];
2392 if (!nhs->hugepages_kobj)
2393 return; /* no hstate attributes */
2395 for_each_hstate(h) {
2396 int idx = hstate_index(h);
2397 if (nhs->hstate_kobjs[idx]) {
2398 kobject_put(nhs->hstate_kobjs[idx]);
2399 nhs->hstate_kobjs[idx] = NULL;
2403 kobject_put(nhs->hugepages_kobj);
2404 nhs->hugepages_kobj = NULL;
2408 * hugetlb module exit: unregister hstate attributes from node devices
2411 static void hugetlb_unregister_all_nodes(void)
2416 * disable node device registrations.
2418 register_hugetlbfs_with_node(NULL, NULL);
2421 * remove hstate attributes from any nodes that have them.
2423 for (nid = 0; nid < nr_node_ids; nid++)
2424 hugetlb_unregister_node(node_devices[nid]);
2428 * Register hstate attributes for a single node device.
2429 * No-op if attributes already registered.
2431 static void hugetlb_register_node(struct node *node)
2434 struct node_hstate *nhs = &node_hstates[node->dev.id];
2437 if (nhs->hugepages_kobj)
2438 return; /* already allocated */
2440 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2442 if (!nhs->hugepages_kobj)
2445 for_each_hstate(h) {
2446 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2448 &per_node_hstate_attr_group);
2450 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2451 h->name, node->dev.id);
2452 hugetlb_unregister_node(node);
2459 * hugetlb init time: register hstate attributes for all registered node
2460 * devices of nodes that have memory. All on-line nodes should have
2461 * registered their associated device by this time.
2463 static void __init hugetlb_register_all_nodes(void)
2467 for_each_node_state(nid, N_MEMORY) {
2468 struct node *node = node_devices[nid];
2469 if (node->dev.id == nid)
2470 hugetlb_register_node(node);
2474 * Let the node device driver know we're here so it can
2475 * [un]register hstate attributes on node hotplug.
2477 register_hugetlbfs_with_node(hugetlb_register_node,
2478 hugetlb_unregister_node);
2480 #else /* !CONFIG_NUMA */
2482 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2490 static void hugetlb_unregister_all_nodes(void) { }
2492 static void hugetlb_register_all_nodes(void) { }
2496 static void __exit hugetlb_exit(void)
2500 hugetlb_unregister_all_nodes();
2502 for_each_hstate(h) {
2503 kobject_put(hstate_kobjs[hstate_index(h)]);
2506 kobject_put(hugepages_kobj);
2507 kfree(hugetlb_fault_mutex_table);
2509 module_exit(hugetlb_exit);
2511 static int __init hugetlb_init(void)
2515 if (!hugepages_supported())
2518 if (!size_to_hstate(default_hstate_size)) {
2519 default_hstate_size = HPAGE_SIZE;
2520 if (!size_to_hstate(default_hstate_size))
2521 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2523 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2524 if (default_hstate_max_huge_pages)
2525 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2527 hugetlb_init_hstates();
2528 gather_bootmem_prealloc();
2531 hugetlb_sysfs_init();
2532 hugetlb_register_all_nodes();
2533 hugetlb_cgroup_file_init();
2536 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2538 num_fault_mutexes = 1;
2540 hugetlb_fault_mutex_table =
2541 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2542 BUG_ON(!hugetlb_fault_mutex_table);
2544 for (i = 0; i < num_fault_mutexes; i++)
2545 mutex_init(&hugetlb_fault_mutex_table[i]);
2548 module_init(hugetlb_init);
2550 /* Should be called on processing a hugepagesz=... option */
2551 void __init hugetlb_add_hstate(unsigned order)
2556 if (size_to_hstate(PAGE_SIZE << order)) {
2557 pr_warning("hugepagesz= specified twice, ignoring\n");
2560 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2562 h = &hstates[hugetlb_max_hstate++];
2564 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2565 h->nr_huge_pages = 0;
2566 h->free_huge_pages = 0;
2567 for (i = 0; i < MAX_NUMNODES; ++i)
2568 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2569 INIT_LIST_HEAD(&h->hugepage_activelist);
2570 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2571 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2572 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2573 huge_page_size(h)/1024);
2578 static int __init hugetlb_nrpages_setup(char *s)
2581 static unsigned long *last_mhp;
2584 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2585 * so this hugepages= parameter goes to the "default hstate".
2587 if (!hugetlb_max_hstate)
2588 mhp = &default_hstate_max_huge_pages;
2590 mhp = &parsed_hstate->max_huge_pages;
2592 if (mhp == last_mhp) {
2593 pr_warning("hugepages= specified twice without "
2594 "interleaving hugepagesz=, ignoring\n");
2598 if (sscanf(s, "%lu", mhp) <= 0)
2602 * Global state is always initialized later in hugetlb_init.
2603 * But we need to allocate >= MAX_ORDER hstates here early to still
2604 * use the bootmem allocator.
2606 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2607 hugetlb_hstate_alloc_pages(parsed_hstate);
2613 __setup("hugepages=", hugetlb_nrpages_setup);
2615 static int __init hugetlb_default_setup(char *s)
2617 default_hstate_size = memparse(s, &s);
2620 __setup("default_hugepagesz=", hugetlb_default_setup);
2622 static unsigned int cpuset_mems_nr(unsigned int *array)
2625 unsigned int nr = 0;
2627 for_each_node_mask(node, cpuset_current_mems_allowed)
2633 #ifdef CONFIG_SYSCTL
2634 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2635 struct ctl_table *table, int write,
2636 void __user *buffer, size_t *length, loff_t *ppos)
2638 struct hstate *h = &default_hstate;
2639 unsigned long tmp = h->max_huge_pages;
2642 if (!hugepages_supported())
2646 table->maxlen = sizeof(unsigned long);
2647 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2652 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2653 NUMA_NO_NODE, tmp, *length);
2658 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2659 void __user *buffer, size_t *length, loff_t *ppos)
2662 return hugetlb_sysctl_handler_common(false, table, write,
2663 buffer, length, ppos);
2667 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2668 void __user *buffer, size_t *length, loff_t *ppos)
2670 return hugetlb_sysctl_handler_common(true, table, write,
2671 buffer, length, ppos);
2673 #endif /* CONFIG_NUMA */
2675 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2676 void __user *buffer,
2677 size_t *length, loff_t *ppos)
2679 struct hstate *h = &default_hstate;
2683 if (!hugepages_supported())
2686 tmp = h->nr_overcommit_huge_pages;
2688 if (write && hstate_is_gigantic(h))
2692 table->maxlen = sizeof(unsigned long);
2693 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2698 spin_lock(&hugetlb_lock);
2699 h->nr_overcommit_huge_pages = tmp;
2700 spin_unlock(&hugetlb_lock);
2706 #endif /* CONFIG_SYSCTL */
2708 void hugetlb_report_meminfo(struct seq_file *m)
2710 struct hstate *h = &default_hstate;
2711 if (!hugepages_supported())
2714 "HugePages_Total: %5lu\n"
2715 "HugePages_Free: %5lu\n"
2716 "HugePages_Rsvd: %5lu\n"
2717 "HugePages_Surp: %5lu\n"
2718 "Hugepagesize: %8lu kB\n",
2722 h->surplus_huge_pages,
2723 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2726 int hugetlb_report_node_meminfo(int nid, char *buf)
2728 struct hstate *h = &default_hstate;
2729 if (!hugepages_supported())
2732 "Node %d HugePages_Total: %5u\n"
2733 "Node %d HugePages_Free: %5u\n"
2734 "Node %d HugePages_Surp: %5u\n",
2735 nid, h->nr_huge_pages_node[nid],
2736 nid, h->free_huge_pages_node[nid],
2737 nid, h->surplus_huge_pages_node[nid]);
2740 void hugetlb_show_meminfo(void)
2745 if (!hugepages_supported())
2748 for_each_node_state(nid, N_MEMORY)
2750 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2752 h->nr_huge_pages_node[nid],
2753 h->free_huge_pages_node[nid],
2754 h->surplus_huge_pages_node[nid],
2755 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2758 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2759 unsigned long hugetlb_total_pages(void)
2762 unsigned long nr_total_pages = 0;
2765 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2766 return nr_total_pages;
2769 static int hugetlb_acct_memory(struct hstate *h, long delta)
2773 spin_lock(&hugetlb_lock);
2775 * When cpuset is configured, it breaks the strict hugetlb page
2776 * reservation as the accounting is done on a global variable. Such
2777 * reservation is completely rubbish in the presence of cpuset because
2778 * the reservation is not checked against page availability for the
2779 * current cpuset. Application can still potentially OOM'ed by kernel
2780 * with lack of free htlb page in cpuset that the task is in.
2781 * Attempt to enforce strict accounting with cpuset is almost
2782 * impossible (or too ugly) because cpuset is too fluid that
2783 * task or memory node can be dynamically moved between cpusets.
2785 * The change of semantics for shared hugetlb mapping with cpuset is
2786 * undesirable. However, in order to preserve some of the semantics,
2787 * we fall back to check against current free page availability as
2788 * a best attempt and hopefully to minimize the impact of changing
2789 * semantics that cpuset has.
2792 if (gather_surplus_pages(h, delta) < 0)
2795 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2796 return_unused_surplus_pages(h, delta);
2803 return_unused_surplus_pages(h, (unsigned long) -delta);
2806 spin_unlock(&hugetlb_lock);
2810 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2812 struct resv_map *resv = vma_resv_map(vma);
2815 * This new VMA should share its siblings reservation map if present.
2816 * The VMA will only ever have a valid reservation map pointer where
2817 * it is being copied for another still existing VMA. As that VMA
2818 * has a reference to the reservation map it cannot disappear until
2819 * after this open call completes. It is therefore safe to take a
2820 * new reference here without additional locking.
2822 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2823 kref_get(&resv->refs);
2826 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2828 struct hstate *h = hstate_vma(vma);
2829 struct resv_map *resv = vma_resv_map(vma);
2830 struct hugepage_subpool *spool = subpool_vma(vma);
2831 unsigned long reserve, start, end;
2834 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2837 start = vma_hugecache_offset(h, vma, vma->vm_start);
2838 end = vma_hugecache_offset(h, vma, vma->vm_end);
2840 reserve = (end - start) - region_count(resv, start, end);
2842 kref_put(&resv->refs, resv_map_release);
2846 * Decrement reserve counts. The global reserve count may be
2847 * adjusted if the subpool has a minimum size.
2849 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
2850 hugetlb_acct_memory(h, -gbl_reserve);
2855 * We cannot handle pagefaults against hugetlb pages at all. They cause
2856 * handle_mm_fault() to try to instantiate regular-sized pages in the
2857 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2860 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2866 const struct vm_operations_struct hugetlb_vm_ops = {
2867 .fault = hugetlb_vm_op_fault,
2868 .open = hugetlb_vm_op_open,
2869 .close = hugetlb_vm_op_close,
2872 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2878 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2879 vma->vm_page_prot)));
2881 entry = huge_pte_wrprotect(mk_huge_pte(page,
2882 vma->vm_page_prot));
2884 entry = pte_mkyoung(entry);
2885 entry = pte_mkhuge(entry);
2886 entry = arch_make_huge_pte(entry, vma, page, writable);
2891 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2892 unsigned long address, pte_t *ptep)
2896 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2897 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2898 update_mmu_cache(vma, address, ptep);
2901 static int is_hugetlb_entry_migration(pte_t pte)
2905 if (huge_pte_none(pte) || pte_present(pte))
2907 swp = pte_to_swp_entry(pte);
2908 if (non_swap_entry(swp) && is_migration_entry(swp))
2914 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2918 if (huge_pte_none(pte) || pte_present(pte))
2920 swp = pte_to_swp_entry(pte);
2921 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2927 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2928 struct vm_area_struct *vma)
2930 pte_t *src_pte, *dst_pte, entry;
2931 struct page *ptepage;
2934 struct hstate *h = hstate_vma(vma);
2935 unsigned long sz = huge_page_size(h);
2936 unsigned long mmun_start; /* For mmu_notifiers */
2937 unsigned long mmun_end; /* For mmu_notifiers */
2940 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2942 mmun_start = vma->vm_start;
2943 mmun_end = vma->vm_end;
2945 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2947 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2948 spinlock_t *src_ptl, *dst_ptl;
2949 src_pte = huge_pte_offset(src, addr);
2952 dst_pte = huge_pte_alloc(dst, addr, sz);
2958 /* If the pagetables are shared don't copy or take references */
2959 if (dst_pte == src_pte)
2962 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2963 src_ptl = huge_pte_lockptr(h, src, src_pte);
2964 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
2965 entry = huge_ptep_get(src_pte);
2966 if (huge_pte_none(entry)) { /* skip none entry */
2968 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2969 is_hugetlb_entry_hwpoisoned(entry))) {
2970 swp_entry_t swp_entry = pte_to_swp_entry(entry);
2972 if (is_write_migration_entry(swp_entry) && cow) {
2974 * COW mappings require pages in both
2975 * parent and child to be set to read.
2977 make_migration_entry_read(&swp_entry);
2978 entry = swp_entry_to_pte(swp_entry);
2979 set_huge_pte_at(src, addr, src_pte, entry);
2981 set_huge_pte_at(dst, addr, dst_pte, entry);
2984 huge_ptep_set_wrprotect(src, addr, src_pte);
2985 mmu_notifier_invalidate_range(src, mmun_start,
2988 entry = huge_ptep_get(src_pte);
2989 ptepage = pte_page(entry);
2991 page_dup_rmap(ptepage);
2992 set_huge_pte_at(dst, addr, dst_pte, entry);
2994 spin_unlock(src_ptl);
2995 spin_unlock(dst_ptl);
2999 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3004 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3005 unsigned long start, unsigned long end,
3006 struct page *ref_page)
3008 int force_flush = 0;
3009 struct mm_struct *mm = vma->vm_mm;
3010 unsigned long address;
3015 struct hstate *h = hstate_vma(vma);
3016 unsigned long sz = huge_page_size(h);
3017 const unsigned long mmun_start = start; /* For mmu_notifiers */
3018 const unsigned long mmun_end = end; /* For mmu_notifiers */
3020 WARN_ON(!is_vm_hugetlb_page(vma));
3021 BUG_ON(start & ~huge_page_mask(h));
3022 BUG_ON(end & ~huge_page_mask(h));
3024 tlb_start_vma(tlb, vma);
3025 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3028 for (; address < end; address += sz) {
3029 ptep = huge_pte_offset(mm, address);
3033 ptl = huge_pte_lock(h, mm, ptep);
3034 if (huge_pmd_unshare(mm, &address, ptep))
3037 pte = huge_ptep_get(ptep);
3038 if (huge_pte_none(pte))
3042 * Migrating hugepage or HWPoisoned hugepage is already
3043 * unmapped and its refcount is dropped, so just clear pte here.
3045 if (unlikely(!pte_present(pte))) {
3046 huge_pte_clear(mm, address, ptep);
3050 page = pte_page(pte);
3052 * If a reference page is supplied, it is because a specific
3053 * page is being unmapped, not a range. Ensure the page we
3054 * are about to unmap is the actual page of interest.
3057 if (page != ref_page)
3061 * Mark the VMA as having unmapped its page so that
3062 * future faults in this VMA will fail rather than
3063 * looking like data was lost
3065 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3068 pte = huge_ptep_get_and_clear(mm, address, ptep);
3069 tlb_remove_tlb_entry(tlb, ptep, address);
3070 if (huge_pte_dirty(pte))
3071 set_page_dirty(page);
3073 page_remove_rmap(page);
3074 force_flush = !__tlb_remove_page(tlb, page);
3080 /* Bail out after unmapping reference page if supplied */
3089 * mmu_gather ran out of room to batch pages, we break out of
3090 * the PTE lock to avoid doing the potential expensive TLB invalidate
3091 * and page-free while holding it.
3096 if (address < end && !ref_page)
3099 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3100 tlb_end_vma(tlb, vma);
3103 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3104 struct vm_area_struct *vma, unsigned long start,
3105 unsigned long end, struct page *ref_page)
3107 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3110 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3111 * test will fail on a vma being torn down, and not grab a page table
3112 * on its way out. We're lucky that the flag has such an appropriate
3113 * name, and can in fact be safely cleared here. We could clear it
3114 * before the __unmap_hugepage_range above, but all that's necessary
3115 * is to clear it before releasing the i_mmap_rwsem. This works
3116 * because in the context this is called, the VMA is about to be
3117 * destroyed and the i_mmap_rwsem is held.
3119 vma->vm_flags &= ~VM_MAYSHARE;
3122 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3123 unsigned long end, struct page *ref_page)
3125 struct mm_struct *mm;
3126 struct mmu_gather tlb;
3130 tlb_gather_mmu(&tlb, mm, start, end);
3131 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3132 tlb_finish_mmu(&tlb, start, end);
3136 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3137 * mappping it owns the reserve page for. The intention is to unmap the page
3138 * from other VMAs and let the children be SIGKILLed if they are faulting the
3141 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3142 struct page *page, unsigned long address)
3144 struct hstate *h = hstate_vma(vma);
3145 struct vm_area_struct *iter_vma;
3146 struct address_space *mapping;
3150 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3151 * from page cache lookup which is in HPAGE_SIZE units.
3153 address = address & huge_page_mask(h);
3154 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3156 mapping = file_inode(vma->vm_file)->i_mapping;
3159 * Take the mapping lock for the duration of the table walk. As
3160 * this mapping should be shared between all the VMAs,
3161 * __unmap_hugepage_range() is called as the lock is already held
3163 i_mmap_lock_write(mapping);
3164 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3165 /* Do not unmap the current VMA */
3166 if (iter_vma == vma)
3170 * Unmap the page from other VMAs without their own reserves.
3171 * They get marked to be SIGKILLed if they fault in these
3172 * areas. This is because a future no-page fault on this VMA
3173 * could insert a zeroed page instead of the data existing
3174 * from the time of fork. This would look like data corruption
3176 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3177 unmap_hugepage_range(iter_vma, address,
3178 address + huge_page_size(h), page);
3180 i_mmap_unlock_write(mapping);
3184 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3185 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3186 * cannot race with other handlers or page migration.
3187 * Keep the pte_same checks anyway to make transition from the mutex easier.
3189 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3190 unsigned long address, pte_t *ptep, pte_t pte,
3191 struct page *pagecache_page, spinlock_t *ptl)
3193 struct hstate *h = hstate_vma(vma);
3194 struct page *old_page, *new_page;
3195 int ret = 0, outside_reserve = 0;
3196 unsigned long mmun_start; /* For mmu_notifiers */
3197 unsigned long mmun_end; /* For mmu_notifiers */
3199 old_page = pte_page(pte);
3202 /* If no-one else is actually using this page, avoid the copy
3203 * and just make the page writable */
3204 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3205 page_move_anon_rmap(old_page, vma, address);
3206 set_huge_ptep_writable(vma, address, ptep);
3211 * If the process that created a MAP_PRIVATE mapping is about to
3212 * perform a COW due to a shared page count, attempt to satisfy
3213 * the allocation without using the existing reserves. The pagecache
3214 * page is used to determine if the reserve at this address was
3215 * consumed or not. If reserves were used, a partial faulted mapping
3216 * at the time of fork() could consume its reserves on COW instead
3217 * of the full address range.
3219 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3220 old_page != pagecache_page)
3221 outside_reserve = 1;
3223 page_cache_get(old_page);
3226 * Drop page table lock as buddy allocator may be called. It will
3227 * be acquired again before returning to the caller, as expected.
3230 new_page = alloc_huge_page(vma, address, outside_reserve);
3232 if (IS_ERR(new_page)) {
3234 * If a process owning a MAP_PRIVATE mapping fails to COW,
3235 * it is due to references held by a child and an insufficient
3236 * huge page pool. To guarantee the original mappers
3237 * reliability, unmap the page from child processes. The child
3238 * may get SIGKILLed if it later faults.
3240 if (outside_reserve) {
3241 page_cache_release(old_page);
3242 BUG_ON(huge_pte_none(pte));
3243 unmap_ref_private(mm, vma, old_page, address);
3244 BUG_ON(huge_pte_none(pte));
3246 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3248 pte_same(huge_ptep_get(ptep), pte)))
3249 goto retry_avoidcopy;
3251 * race occurs while re-acquiring page table
3252 * lock, and our job is done.
3257 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3258 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3259 goto out_release_old;
3263 * When the original hugepage is shared one, it does not have
3264 * anon_vma prepared.
3266 if (unlikely(anon_vma_prepare(vma))) {
3268 goto out_release_all;
3271 copy_user_huge_page(new_page, old_page, address, vma,
3272 pages_per_huge_page(h));
3273 __SetPageUptodate(new_page);
3274 set_page_huge_active(new_page);
3276 mmun_start = address & huge_page_mask(h);
3277 mmun_end = mmun_start + huge_page_size(h);
3278 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3281 * Retake the page table lock to check for racing updates
3282 * before the page tables are altered
3285 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3286 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3287 ClearPagePrivate(new_page);
3290 huge_ptep_clear_flush(vma, address, ptep);
3291 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3292 set_huge_pte_at(mm, address, ptep,
3293 make_huge_pte(vma, new_page, 1));
3294 page_remove_rmap(old_page);
3295 hugepage_add_new_anon_rmap(new_page, vma, address);
3296 /* Make the old page be freed below */
3297 new_page = old_page;
3300 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3302 page_cache_release(new_page);
3304 page_cache_release(old_page);
3306 spin_lock(ptl); /* Caller expects lock to be held */
3310 /* Return the pagecache page at a given address within a VMA */
3311 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3312 struct vm_area_struct *vma, unsigned long address)
3314 struct address_space *mapping;
3317 mapping = vma->vm_file->f_mapping;
3318 idx = vma_hugecache_offset(h, vma, address);
3320 return find_lock_page(mapping, idx);
3324 * Return whether there is a pagecache page to back given address within VMA.
3325 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3327 static bool hugetlbfs_pagecache_present(struct hstate *h,
3328 struct vm_area_struct *vma, unsigned long address)
3330 struct address_space *mapping;
3334 mapping = vma->vm_file->f_mapping;
3335 idx = vma_hugecache_offset(h, vma, address);
3337 page = find_get_page(mapping, idx);
3340 return page != NULL;
3343 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3344 struct address_space *mapping, pgoff_t idx,
3345 unsigned long address, pte_t *ptep, unsigned int flags)
3347 struct hstate *h = hstate_vma(vma);
3348 int ret = VM_FAULT_SIGBUS;
3356 * Currently, we are forced to kill the process in the event the
3357 * original mapper has unmapped pages from the child due to a failed
3358 * COW. Warn that such a situation has occurred as it may not be obvious
3360 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3361 pr_warning("PID %d killed due to inadequate hugepage pool\n",
3367 * Use page lock to guard against racing truncation
3368 * before we get page_table_lock.
3371 page = find_lock_page(mapping, idx);
3373 size = i_size_read(mapping->host) >> huge_page_shift(h);
3376 page = alloc_huge_page(vma, address, 0);
3378 ret = PTR_ERR(page);
3382 ret = VM_FAULT_SIGBUS;
3385 clear_huge_page(page, address, pages_per_huge_page(h));
3386 __SetPageUptodate(page);
3387 set_page_huge_active(page);
3389 if (vma->vm_flags & VM_MAYSHARE) {
3391 struct inode *inode = mapping->host;
3393 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3400 ClearPagePrivate(page);
3402 spin_lock(&inode->i_lock);
3403 inode->i_blocks += blocks_per_huge_page(h);
3404 spin_unlock(&inode->i_lock);
3407 if (unlikely(anon_vma_prepare(vma))) {
3409 goto backout_unlocked;
3415 * If memory error occurs between mmap() and fault, some process
3416 * don't have hwpoisoned swap entry for errored virtual address.
3417 * So we need to block hugepage fault by PG_hwpoison bit check.
3419 if (unlikely(PageHWPoison(page))) {
3420 ret = VM_FAULT_HWPOISON |
3421 VM_FAULT_SET_HINDEX(hstate_index(h));
3422 goto backout_unlocked;
3427 * If we are going to COW a private mapping later, we examine the
3428 * pending reservations for this page now. This will ensure that
3429 * any allocations necessary to record that reservation occur outside
3432 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3433 if (vma_needs_reservation(h, vma, address) < 0) {
3435 goto backout_unlocked;
3437 /* Just decrements count, does not deallocate */
3438 vma_end_reservation(h, vma, address);
3441 ptl = huge_pte_lockptr(h, mm, ptep);
3443 size = i_size_read(mapping->host) >> huge_page_shift(h);
3448 if (!huge_pte_none(huge_ptep_get(ptep)))
3452 ClearPagePrivate(page);
3453 hugepage_add_new_anon_rmap(page, vma, address);
3455 page_dup_rmap(page);
3456 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3457 && (vma->vm_flags & VM_SHARED)));
3458 set_huge_pte_at(mm, address, ptep, new_pte);
3460 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3461 /* Optimization, do the COW without a second fault */
3462 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
3479 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3480 struct vm_area_struct *vma,
3481 struct address_space *mapping,
3482 pgoff_t idx, unsigned long address)
3484 unsigned long key[2];
3487 if (vma->vm_flags & VM_SHARED) {
3488 key[0] = (unsigned long) mapping;
3491 key[0] = (unsigned long) mm;
3492 key[1] = address >> huge_page_shift(h);
3495 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3497 return hash & (num_fault_mutexes - 1);
3501 * For uniprocesor systems we always use a single mutex, so just
3502 * return 0 and avoid the hashing overhead.
3504 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3505 struct vm_area_struct *vma,
3506 struct address_space *mapping,
3507 pgoff_t idx, unsigned long address)
3513 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3514 unsigned long address, unsigned int flags)
3521 struct page *page = NULL;
3522 struct page *pagecache_page = NULL;
3523 struct hstate *h = hstate_vma(vma);
3524 struct address_space *mapping;
3525 int need_wait_lock = 0;
3527 address &= huge_page_mask(h);
3529 ptep = huge_pte_offset(mm, address);
3531 entry = huge_ptep_get(ptep);
3532 if (unlikely(is_hugetlb_entry_migration(entry))) {
3533 migration_entry_wait_huge(vma, mm, ptep);
3535 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3536 return VM_FAULT_HWPOISON_LARGE |
3537 VM_FAULT_SET_HINDEX(hstate_index(h));
3540 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3542 return VM_FAULT_OOM;
3544 mapping = vma->vm_file->f_mapping;
3545 idx = vma_hugecache_offset(h, vma, address);
3548 * Serialize hugepage allocation and instantiation, so that we don't
3549 * get spurious allocation failures if two CPUs race to instantiate
3550 * the same page in the page cache.
3552 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3553 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3555 entry = huge_ptep_get(ptep);
3556 if (huge_pte_none(entry)) {
3557 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3564 * entry could be a migration/hwpoison entry at this point, so this
3565 * check prevents the kernel from going below assuming that we have
3566 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3567 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3570 if (!pte_present(entry))
3574 * If we are going to COW the mapping later, we examine the pending
3575 * reservations for this page now. This will ensure that any
3576 * allocations necessary to record that reservation occur outside the
3577 * spinlock. For private mappings, we also lookup the pagecache
3578 * page now as it is used to determine if a reservation has been
3581 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3582 if (vma_needs_reservation(h, vma, address) < 0) {
3586 /* Just decrements count, does not deallocate */
3587 vma_end_reservation(h, vma, address);
3589 if (!(vma->vm_flags & VM_MAYSHARE))
3590 pagecache_page = hugetlbfs_pagecache_page(h,
3594 ptl = huge_pte_lock(h, mm, ptep);
3596 /* Check for a racing update before calling hugetlb_cow */
3597 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3601 * hugetlb_cow() requires page locks of pte_page(entry) and
3602 * pagecache_page, so here we need take the former one
3603 * when page != pagecache_page or !pagecache_page.
3605 page = pte_page(entry);
3606 if (page != pagecache_page)
3607 if (!trylock_page(page)) {
3614 if (flags & FAULT_FLAG_WRITE) {
3615 if (!huge_pte_write(entry)) {
3616 ret = hugetlb_cow(mm, vma, address, ptep, entry,
3617 pagecache_page, ptl);
3620 entry = huge_pte_mkdirty(entry);
3622 entry = pte_mkyoung(entry);
3623 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3624 flags & FAULT_FLAG_WRITE))
3625 update_mmu_cache(vma, address, ptep);
3627 if (page != pagecache_page)
3633 if (pagecache_page) {
3634 unlock_page(pagecache_page);
3635 put_page(pagecache_page);
3638 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3640 * Generally it's safe to hold refcount during waiting page lock. But
3641 * here we just wait to defer the next page fault to avoid busy loop and
3642 * the page is not used after unlocked before returning from the current
3643 * page fault. So we are safe from accessing freed page, even if we wait
3644 * here without taking refcount.
3647 wait_on_page_locked(page);
3651 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3652 struct page **pages, struct vm_area_struct **vmas,
3653 unsigned long *position, unsigned long *nr_pages,
3654 long i, unsigned int flags)
3656 unsigned long pfn_offset;
3657 unsigned long vaddr = *position;
3658 unsigned long remainder = *nr_pages;
3659 struct hstate *h = hstate_vma(vma);
3661 while (vaddr < vma->vm_end && remainder) {
3663 spinlock_t *ptl = NULL;
3668 * If we have a pending SIGKILL, don't keep faulting pages and
3669 * potentially allocating memory.
3671 if (unlikely(fatal_signal_pending(current))) {
3677 * Some archs (sparc64, sh*) have multiple pte_ts to
3678 * each hugepage. We have to make sure we get the
3679 * first, for the page indexing below to work.
3681 * Note that page table lock is not held when pte is null.
3683 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3685 ptl = huge_pte_lock(h, mm, pte);
3686 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3689 * When coredumping, it suits get_dump_page if we just return
3690 * an error where there's an empty slot with no huge pagecache
3691 * to back it. This way, we avoid allocating a hugepage, and
3692 * the sparse dumpfile avoids allocating disk blocks, but its
3693 * huge holes still show up with zeroes where they need to be.
3695 if (absent && (flags & FOLL_DUMP) &&
3696 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3704 * We need call hugetlb_fault for both hugepages under migration
3705 * (in which case hugetlb_fault waits for the migration,) and
3706 * hwpoisoned hugepages (in which case we need to prevent the
3707 * caller from accessing to them.) In order to do this, we use
3708 * here is_swap_pte instead of is_hugetlb_entry_migration and
3709 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3710 * both cases, and because we can't follow correct pages
3711 * directly from any kind of swap entries.
3713 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3714 ((flags & FOLL_WRITE) &&
3715 !huge_pte_write(huge_ptep_get(pte)))) {
3720 ret = hugetlb_fault(mm, vma, vaddr,
3721 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3722 if (!(ret & VM_FAULT_ERROR))
3729 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3730 page = pte_page(huge_ptep_get(pte));
3733 pages[i] = mem_map_offset(page, pfn_offset);
3734 get_page_foll(pages[i]);
3744 if (vaddr < vma->vm_end && remainder &&
3745 pfn_offset < pages_per_huge_page(h)) {
3747 * We use pfn_offset to avoid touching the pageframes
3748 * of this compound page.
3754 *nr_pages = remainder;
3757 return i ? i : -EFAULT;
3760 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3761 unsigned long address, unsigned long end, pgprot_t newprot)
3763 struct mm_struct *mm = vma->vm_mm;
3764 unsigned long start = address;
3767 struct hstate *h = hstate_vma(vma);
3768 unsigned long pages = 0;
3770 BUG_ON(address >= end);
3771 flush_cache_range(vma, address, end);
3773 mmu_notifier_invalidate_range_start(mm, start, end);
3774 i_mmap_lock_write(vma->vm_file->f_mapping);
3775 for (; address < end; address += huge_page_size(h)) {
3777 ptep = huge_pte_offset(mm, address);
3780 ptl = huge_pte_lock(h, mm, ptep);
3781 if (huge_pmd_unshare(mm, &address, ptep)) {
3786 pte = huge_ptep_get(ptep);
3787 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3791 if (unlikely(is_hugetlb_entry_migration(pte))) {
3792 swp_entry_t entry = pte_to_swp_entry(pte);
3794 if (is_write_migration_entry(entry)) {
3797 make_migration_entry_read(&entry);
3798 newpte = swp_entry_to_pte(entry);
3799 set_huge_pte_at(mm, address, ptep, newpte);
3805 if (!huge_pte_none(pte)) {
3806 pte = huge_ptep_get_and_clear(mm, address, ptep);
3807 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3808 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3809 set_huge_pte_at(mm, address, ptep, pte);
3815 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
3816 * may have cleared our pud entry and done put_page on the page table:
3817 * once we release i_mmap_rwsem, another task can do the final put_page
3818 * and that page table be reused and filled with junk.
3820 flush_tlb_range(vma, start, end);
3821 mmu_notifier_invalidate_range(mm, start, end);
3822 i_mmap_unlock_write(vma->vm_file->f_mapping);
3823 mmu_notifier_invalidate_range_end(mm, start, end);
3825 return pages << h->order;
3828 int hugetlb_reserve_pages(struct inode *inode,
3830 struct vm_area_struct *vma,
3831 vm_flags_t vm_flags)
3834 struct hstate *h = hstate_inode(inode);
3835 struct hugepage_subpool *spool = subpool_inode(inode);
3836 struct resv_map *resv_map;
3840 * Only apply hugepage reservation if asked. At fault time, an
3841 * attempt will be made for VM_NORESERVE to allocate a page
3842 * without using reserves
3844 if (vm_flags & VM_NORESERVE)
3848 * Shared mappings base their reservation on the number of pages that
3849 * are already allocated on behalf of the file. Private mappings need
3850 * to reserve the full area even if read-only as mprotect() may be
3851 * called to make the mapping read-write. Assume !vma is a shm mapping
3853 if (!vma || vma->vm_flags & VM_MAYSHARE) {
3854 resv_map = inode_resv_map(inode);
3856 chg = region_chg(resv_map, from, to);
3859 resv_map = resv_map_alloc();
3865 set_vma_resv_map(vma, resv_map);
3866 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3875 * There must be enough pages in the subpool for the mapping. If
3876 * the subpool has a minimum size, there may be some global
3877 * reservations already in place (gbl_reserve).
3879 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
3880 if (gbl_reserve < 0) {
3886 * Check enough hugepages are available for the reservation.
3887 * Hand the pages back to the subpool if there are not
3889 ret = hugetlb_acct_memory(h, gbl_reserve);
3891 /* put back original number of pages, chg */
3892 (void)hugepage_subpool_put_pages(spool, chg);
3897 * Account for the reservations made. Shared mappings record regions
3898 * that have reservations as they are shared by multiple VMAs.
3899 * When the last VMA disappears, the region map says how much
3900 * the reservation was and the page cache tells how much of
3901 * the reservation was consumed. Private mappings are per-VMA and
3902 * only the consumed reservations are tracked. When the VMA
3903 * disappears, the original reservation is the VMA size and the
3904 * consumed reservations are stored in the map. Hence, nothing
3905 * else has to be done for private mappings here
3907 if (!vma || vma->vm_flags & VM_MAYSHARE) {
3908 long add = region_add(resv_map, from, to);
3910 if (unlikely(chg > add)) {
3912 * pages in this range were added to the reserve
3913 * map between region_chg and region_add. This
3914 * indicates a race with alloc_huge_page. Adjust
3915 * the subpool and reserve counts modified above
3916 * based on the difference.
3920 rsv_adjust = hugepage_subpool_put_pages(spool,
3922 hugetlb_acct_memory(h, -rsv_adjust);
3927 if (!vma || vma->vm_flags & VM_MAYSHARE)
3928 region_abort(resv_map, from, to);
3929 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3930 kref_put(&resv_map->refs, resv_map_release);
3934 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
3937 struct hstate *h = hstate_inode(inode);
3938 struct resv_map *resv_map = inode_resv_map(inode);
3940 struct hugepage_subpool *spool = subpool_inode(inode);
3944 chg = region_del(resv_map, start, end);
3946 * region_del() can fail in the rare case where a region
3947 * must be split and another region descriptor can not be
3948 * allocated. If end == LONG_MAX, it will not fail.
3954 spin_lock(&inode->i_lock);
3955 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3956 spin_unlock(&inode->i_lock);
3959 * If the subpool has a minimum size, the number of global
3960 * reservations to be released may be adjusted.
3962 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
3963 hugetlb_acct_memory(h, -gbl_reserve);
3968 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3969 static unsigned long page_table_shareable(struct vm_area_struct *svma,
3970 struct vm_area_struct *vma,
3971 unsigned long addr, pgoff_t idx)
3973 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3975 unsigned long sbase = saddr & PUD_MASK;
3976 unsigned long s_end = sbase + PUD_SIZE;
3978 /* Allow segments to share if only one is marked locked */
3979 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3980 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3983 * match the virtual addresses, permission and the alignment of the
3986 if (pmd_index(addr) != pmd_index(saddr) ||
3987 vm_flags != svm_flags ||
3988 sbase < svma->vm_start || svma->vm_end < s_end)
3994 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3996 unsigned long base = addr & PUD_MASK;
3997 unsigned long end = base + PUD_SIZE;
4000 * check on proper vm_flags and page table alignment
4002 if (vma->vm_flags & VM_MAYSHARE &&
4003 vma->vm_start <= base && end <= vma->vm_end)
4009 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4010 * and returns the corresponding pte. While this is not necessary for the
4011 * !shared pmd case because we can allocate the pmd later as well, it makes the
4012 * code much cleaner. pmd allocation is essential for the shared case because
4013 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4014 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4015 * bad pmd for sharing.
4017 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4019 struct vm_area_struct *vma = find_vma(mm, addr);
4020 struct address_space *mapping = vma->vm_file->f_mapping;
4021 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4023 struct vm_area_struct *svma;
4024 unsigned long saddr;
4029 if (!vma_shareable(vma, addr))
4030 return (pte_t *)pmd_alloc(mm, pud, addr);
4032 i_mmap_lock_write(mapping);
4033 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4037 saddr = page_table_shareable(svma, vma, addr, idx);
4039 spte = huge_pte_offset(svma->vm_mm, saddr);
4042 get_page(virt_to_page(spte));
4051 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
4053 if (pud_none(*pud)) {
4054 pud_populate(mm, pud,
4055 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4057 put_page(virt_to_page(spte));
4062 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4063 i_mmap_unlock_write(mapping);
4068 * unmap huge page backed by shared pte.
4070 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4071 * indicated by page_count > 1, unmap is achieved by clearing pud and
4072 * decrementing the ref count. If count == 1, the pte page is not shared.
4074 * called with page table lock held.
4076 * returns: 1 successfully unmapped a shared pte page
4077 * 0 the underlying pte page is not shared, or it is the last user
4079 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4081 pgd_t *pgd = pgd_offset(mm, *addr);
4082 pud_t *pud = pud_offset(pgd, *addr);
4084 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4085 if (page_count(virt_to_page(ptep)) == 1)
4089 put_page(virt_to_page(ptep));
4091 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4094 #define want_pmd_share() (1)
4095 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4096 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4101 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4105 #define want_pmd_share() (0)
4106 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4108 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4109 pte_t *huge_pte_alloc(struct mm_struct *mm,
4110 unsigned long addr, unsigned long sz)
4116 pgd = pgd_offset(mm, addr);
4117 pud = pud_alloc(mm, pgd, addr);
4119 if (sz == PUD_SIZE) {
4122 BUG_ON(sz != PMD_SIZE);
4123 if (want_pmd_share() && pud_none(*pud))
4124 pte = huge_pmd_share(mm, addr, pud);
4126 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4129 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
4134 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
4140 pgd = pgd_offset(mm, addr);
4141 if (pgd_present(*pgd)) {
4142 pud = pud_offset(pgd, addr);
4143 if (pud_present(*pud)) {
4145 return (pte_t *)pud;
4146 pmd = pmd_offset(pud, addr);
4149 return (pte_t *) pmd;
4152 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4155 * These functions are overwritable if your architecture needs its own
4158 struct page * __weak
4159 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4162 return ERR_PTR(-EINVAL);
4165 struct page * __weak
4166 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4167 pmd_t *pmd, int flags)
4169 struct page *page = NULL;
4172 ptl = pmd_lockptr(mm, pmd);
4175 * make sure that the address range covered by this pmd is not
4176 * unmapped from other threads.
4178 if (!pmd_huge(*pmd))
4180 if (pmd_present(*pmd)) {
4181 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4182 if (flags & FOLL_GET)
4185 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
4187 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4191 * hwpoisoned entry is treated as no_page_table in
4192 * follow_page_mask().
4200 struct page * __weak
4201 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4202 pud_t *pud, int flags)
4204 if (flags & FOLL_GET)
4207 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4210 #ifdef CONFIG_MEMORY_FAILURE
4213 * This function is called from memory failure code.
4214 * Assume the caller holds page lock of the head page.
4216 int dequeue_hwpoisoned_huge_page(struct page *hpage)
4218 struct hstate *h = page_hstate(hpage);
4219 int nid = page_to_nid(hpage);
4222 spin_lock(&hugetlb_lock);
4224 * Just checking !page_huge_active is not enough, because that could be
4225 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4227 if (!page_huge_active(hpage) && !page_count(hpage)) {
4229 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4230 * but dangling hpage->lru can trigger list-debug warnings
4231 * (this happens when we call unpoison_memory() on it),
4232 * so let it point to itself with list_del_init().
4234 list_del_init(&hpage->lru);
4235 set_page_refcounted(hpage);
4236 h->free_huge_pages--;
4237 h->free_huge_pages_node[nid]--;
4240 spin_unlock(&hugetlb_lock);
4245 bool isolate_huge_page(struct page *page, struct list_head *list)
4249 VM_BUG_ON_PAGE(!PageHead(page), page);
4250 spin_lock(&hugetlb_lock);
4251 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4255 clear_page_huge_active(page);
4256 list_move_tail(&page->lru, list);
4258 spin_unlock(&hugetlb_lock);
4262 void putback_active_hugepage(struct page *page)
4264 VM_BUG_ON_PAGE(!PageHead(page), page);
4265 spin_lock(&hugetlb_lock);
4266 set_page_huge_active(page);
4267 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4268 spin_unlock(&hugetlb_lock);