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) {
806 * We know VM_NORESERVE is not set. Therefore, there SHOULD
807 * be a region map for all pages. The only situation where
808 * there is no region map is if a hole was punched via
809 * fallocate. In this case, there really are no reverves to
810 * use. This situation is indicated if chg != 0.
819 * Only the process that called mmap() has reserves for
822 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
828 static void enqueue_huge_page(struct hstate *h, struct page *page)
830 int nid = page_to_nid(page);
831 list_move(&page->lru, &h->hugepage_freelists[nid]);
832 h->free_huge_pages++;
833 h->free_huge_pages_node[nid]++;
836 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
840 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
841 if (!is_migrate_isolate_page(page))
844 * if 'non-isolated free hugepage' not found on the list,
845 * the allocation fails.
847 if (&h->hugepage_freelists[nid] == &page->lru)
849 list_move(&page->lru, &h->hugepage_activelist);
850 set_page_refcounted(page);
851 h->free_huge_pages--;
852 h->free_huge_pages_node[nid]--;
856 /* Movability of hugepages depends on migration support. */
857 static inline gfp_t htlb_alloc_mask(struct hstate *h)
859 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
860 return GFP_HIGHUSER_MOVABLE;
865 static struct page *dequeue_huge_page_vma(struct hstate *h,
866 struct vm_area_struct *vma,
867 unsigned long address, int avoid_reserve,
870 struct page *page = NULL;
871 struct mempolicy *mpol;
872 nodemask_t *nodemask;
873 struct zonelist *zonelist;
876 unsigned int cpuset_mems_cookie;
879 * A child process with MAP_PRIVATE mappings created by their parent
880 * have no page reserves. This check ensures that reservations are
881 * not "stolen". The child may still get SIGKILLed
883 if (!vma_has_reserves(vma, chg) &&
884 h->free_huge_pages - h->resv_huge_pages == 0)
887 /* If reserves cannot be used, ensure enough pages are in the pool */
888 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
892 cpuset_mems_cookie = read_mems_allowed_begin();
893 zonelist = huge_zonelist(vma, address,
894 htlb_alloc_mask(h), &mpol, &nodemask);
896 for_each_zone_zonelist_nodemask(zone, z, zonelist,
897 MAX_NR_ZONES - 1, nodemask) {
898 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
899 page = dequeue_huge_page_node(h, zone_to_nid(zone));
903 if (!vma_has_reserves(vma, chg))
906 SetPagePrivate(page);
907 h->resv_huge_pages--;
914 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
923 * common helper functions for hstate_next_node_to_{alloc|free}.
924 * We may have allocated or freed a huge page based on a different
925 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
926 * be outside of *nodes_allowed. Ensure that we use an allowed
927 * node for alloc or free.
929 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
931 nid = next_node(nid, *nodes_allowed);
932 if (nid == MAX_NUMNODES)
933 nid = first_node(*nodes_allowed);
934 VM_BUG_ON(nid >= MAX_NUMNODES);
939 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
941 if (!node_isset(nid, *nodes_allowed))
942 nid = next_node_allowed(nid, nodes_allowed);
947 * returns the previously saved node ["this node"] from which to
948 * allocate a persistent huge page for the pool and advance the
949 * next node from which to allocate, handling wrap at end of node
952 static int hstate_next_node_to_alloc(struct hstate *h,
953 nodemask_t *nodes_allowed)
957 VM_BUG_ON(!nodes_allowed);
959 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
960 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
966 * helper for free_pool_huge_page() - return the previously saved
967 * node ["this node"] from which to free a huge page. Advance the
968 * next node id whether or not we find a free huge page to free so
969 * that the next attempt to free addresses the next node.
971 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
975 VM_BUG_ON(!nodes_allowed);
977 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
978 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
983 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
984 for (nr_nodes = nodes_weight(*mask); \
986 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
989 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
990 for (nr_nodes = nodes_weight(*mask); \
992 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
995 #if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
996 static void destroy_compound_gigantic_page(struct page *page,
1000 int nr_pages = 1 << order;
1001 struct page *p = page + 1;
1003 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1005 set_page_refcounted(p);
1006 p->first_page = NULL;
1009 set_compound_order(page, 0);
1010 __ClearPageHead(page);
1013 static void free_gigantic_page(struct page *page, unsigned order)
1015 free_contig_range(page_to_pfn(page), 1 << order);
1018 static int __alloc_gigantic_page(unsigned long start_pfn,
1019 unsigned long nr_pages)
1021 unsigned long end_pfn = start_pfn + nr_pages;
1022 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1025 static bool pfn_range_valid_gigantic(unsigned long start_pfn,
1026 unsigned long nr_pages)
1028 unsigned long i, end_pfn = start_pfn + nr_pages;
1031 for (i = start_pfn; i < end_pfn; i++) {
1035 page = pfn_to_page(i);
1037 if (PageReserved(page))
1040 if (page_count(page) > 0)
1050 static bool zone_spans_last_pfn(const struct zone *zone,
1051 unsigned long start_pfn, unsigned long nr_pages)
1053 unsigned long last_pfn = start_pfn + nr_pages - 1;
1054 return zone_spans_pfn(zone, last_pfn);
1057 static struct page *alloc_gigantic_page(int nid, unsigned order)
1059 unsigned long nr_pages = 1 << order;
1060 unsigned long ret, pfn, flags;
1063 z = NODE_DATA(nid)->node_zones;
1064 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
1065 spin_lock_irqsave(&z->lock, flags);
1067 pfn = ALIGN(z->zone_start_pfn, nr_pages);
1068 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
1069 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
1071 * We release the zone lock here because
1072 * alloc_contig_range() will also lock the zone
1073 * at some point. If there's an allocation
1074 * spinning on this lock, it may win the race
1075 * and cause alloc_contig_range() to fail...
1077 spin_unlock_irqrestore(&z->lock, flags);
1078 ret = __alloc_gigantic_page(pfn, nr_pages);
1080 return pfn_to_page(pfn);
1081 spin_lock_irqsave(&z->lock, flags);
1086 spin_unlock_irqrestore(&z->lock, flags);
1092 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1093 static void prep_compound_gigantic_page(struct page *page, unsigned long order);
1095 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1099 page = alloc_gigantic_page(nid, huge_page_order(h));
1101 prep_compound_gigantic_page(page, huge_page_order(h));
1102 prep_new_huge_page(h, page, nid);
1108 static int alloc_fresh_gigantic_page(struct hstate *h,
1109 nodemask_t *nodes_allowed)
1111 struct page *page = NULL;
1114 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1115 page = alloc_fresh_gigantic_page_node(h, node);
1123 static inline bool gigantic_page_supported(void) { return true; }
1125 static inline bool gigantic_page_supported(void) { return false; }
1126 static inline void free_gigantic_page(struct page *page, unsigned order) { }
1127 static inline void destroy_compound_gigantic_page(struct page *page,
1128 unsigned long order) { }
1129 static inline int alloc_fresh_gigantic_page(struct hstate *h,
1130 nodemask_t *nodes_allowed) { return 0; }
1133 static void update_and_free_page(struct hstate *h, struct page *page)
1137 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1141 h->nr_huge_pages_node[page_to_nid(page)]--;
1142 for (i = 0; i < pages_per_huge_page(h); i++) {
1143 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1144 1 << PG_referenced | 1 << PG_dirty |
1145 1 << PG_active | 1 << PG_private |
1148 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1149 set_compound_page_dtor(page, NULL);
1150 set_page_refcounted(page);
1151 if (hstate_is_gigantic(h)) {
1152 destroy_compound_gigantic_page(page, huge_page_order(h));
1153 free_gigantic_page(page, huge_page_order(h));
1155 __free_pages(page, huge_page_order(h));
1159 struct hstate *size_to_hstate(unsigned long size)
1163 for_each_hstate(h) {
1164 if (huge_page_size(h) == size)
1171 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1172 * to hstate->hugepage_activelist.)
1174 * This function can be called for tail pages, but never returns true for them.
1176 bool page_huge_active(struct page *page)
1178 VM_BUG_ON_PAGE(!PageHuge(page), page);
1179 return PageHead(page) && PagePrivate(&page[1]);
1182 /* never called for tail page */
1183 static void set_page_huge_active(struct page *page)
1185 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1186 SetPagePrivate(&page[1]);
1189 static void clear_page_huge_active(struct page *page)
1191 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1192 ClearPagePrivate(&page[1]);
1195 void free_huge_page(struct page *page)
1198 * Can't pass hstate in here because it is called from the
1199 * compound page destructor.
1201 struct hstate *h = page_hstate(page);
1202 int nid = page_to_nid(page);
1203 struct hugepage_subpool *spool =
1204 (struct hugepage_subpool *)page_private(page);
1205 bool restore_reserve;
1207 set_page_private(page, 0);
1208 page->mapping = NULL;
1209 BUG_ON(page_count(page));
1210 BUG_ON(page_mapcount(page));
1211 restore_reserve = PagePrivate(page);
1212 ClearPagePrivate(page);
1215 * A return code of zero implies that the subpool will be under its
1216 * minimum size if the reservation is not restored after page is free.
1217 * Therefore, force restore_reserve operation.
1219 if (hugepage_subpool_put_pages(spool, 1) == 0)
1220 restore_reserve = true;
1222 spin_lock(&hugetlb_lock);
1223 clear_page_huge_active(page);
1224 hugetlb_cgroup_uncharge_page(hstate_index(h),
1225 pages_per_huge_page(h), page);
1226 if (restore_reserve)
1227 h->resv_huge_pages++;
1229 if (h->surplus_huge_pages_node[nid]) {
1230 /* remove the page from active list */
1231 list_del(&page->lru);
1232 update_and_free_page(h, page);
1233 h->surplus_huge_pages--;
1234 h->surplus_huge_pages_node[nid]--;
1236 arch_clear_hugepage_flags(page);
1237 enqueue_huge_page(h, page);
1239 spin_unlock(&hugetlb_lock);
1242 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1244 INIT_LIST_HEAD(&page->lru);
1245 set_compound_page_dtor(page, free_huge_page);
1246 spin_lock(&hugetlb_lock);
1247 set_hugetlb_cgroup(page, NULL);
1249 h->nr_huge_pages_node[nid]++;
1250 spin_unlock(&hugetlb_lock);
1251 put_page(page); /* free it into the hugepage allocator */
1254 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
1257 int nr_pages = 1 << order;
1258 struct page *p = page + 1;
1260 /* we rely on prep_new_huge_page to set the destructor */
1261 set_compound_order(page, order);
1262 __SetPageHead(page);
1263 __ClearPageReserved(page);
1264 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1266 * For gigantic hugepages allocated through bootmem at
1267 * boot, it's safer to be consistent with the not-gigantic
1268 * hugepages and clear the PG_reserved bit from all tail pages
1269 * too. Otherwse drivers using get_user_pages() to access tail
1270 * pages may get the reference counting wrong if they see
1271 * PG_reserved set on a tail page (despite the head page not
1272 * having PG_reserved set). Enforcing this consistency between
1273 * head and tail pages allows drivers to optimize away a check
1274 * on the head page when they need know if put_page() is needed
1275 * after get_user_pages().
1277 __ClearPageReserved(p);
1278 set_page_count(p, 0);
1279 p->first_page = page;
1280 /* Make sure p->first_page is always valid for PageTail() */
1287 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1288 * transparent huge pages. See the PageTransHuge() documentation for more
1291 int PageHuge(struct page *page)
1293 if (!PageCompound(page))
1296 page = compound_head(page);
1297 return get_compound_page_dtor(page) == free_huge_page;
1299 EXPORT_SYMBOL_GPL(PageHuge);
1302 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1303 * normal or transparent huge pages.
1305 int PageHeadHuge(struct page *page_head)
1307 if (!PageHead(page_head))
1310 return get_compound_page_dtor(page_head) == free_huge_page;
1313 pgoff_t __basepage_index(struct page *page)
1315 struct page *page_head = compound_head(page);
1316 pgoff_t index = page_index(page_head);
1317 unsigned long compound_idx;
1319 if (!PageHuge(page_head))
1320 return page_index(page);
1322 if (compound_order(page_head) >= MAX_ORDER)
1323 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1325 compound_idx = page - page_head;
1327 return (index << compound_order(page_head)) + compound_idx;
1330 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1334 page = alloc_pages_exact_node(nid,
1335 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1336 __GFP_REPEAT|__GFP_NOWARN,
1337 huge_page_order(h));
1339 prep_new_huge_page(h, page, nid);
1345 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1351 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1352 page = alloc_fresh_huge_page_node(h, node);
1360 count_vm_event(HTLB_BUDDY_PGALLOC);
1362 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1368 * Free huge page from pool from next node to free.
1369 * Attempt to keep persistent huge pages more or less
1370 * balanced over allowed nodes.
1371 * Called with hugetlb_lock locked.
1373 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1379 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1381 * If we're returning unused surplus pages, only examine
1382 * nodes with surplus pages.
1384 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1385 !list_empty(&h->hugepage_freelists[node])) {
1387 list_entry(h->hugepage_freelists[node].next,
1389 list_del(&page->lru);
1390 h->free_huge_pages--;
1391 h->free_huge_pages_node[node]--;
1393 h->surplus_huge_pages--;
1394 h->surplus_huge_pages_node[node]--;
1396 update_and_free_page(h, page);
1406 * Dissolve a given free hugepage into free buddy pages. This function does
1407 * nothing for in-use (including surplus) hugepages.
1409 static void dissolve_free_huge_page(struct page *page)
1411 spin_lock(&hugetlb_lock);
1412 if (PageHuge(page) && !page_count(page)) {
1413 struct hstate *h = page_hstate(page);
1414 int nid = page_to_nid(page);
1415 list_del(&page->lru);
1416 h->free_huge_pages--;
1417 h->free_huge_pages_node[nid]--;
1418 update_and_free_page(h, page);
1420 spin_unlock(&hugetlb_lock);
1424 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1425 * make specified memory blocks removable from the system.
1426 * Note that start_pfn should aligned with (minimum) hugepage size.
1428 void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1432 if (!hugepages_supported())
1435 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order));
1436 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
1437 dissolve_free_huge_page(pfn_to_page(pfn));
1440 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
1445 if (hstate_is_gigantic(h))
1449 * Assume we will successfully allocate the surplus page to
1450 * prevent racing processes from causing the surplus to exceed
1453 * This however introduces a different race, where a process B
1454 * tries to grow the static hugepage pool while alloc_pages() is
1455 * called by process A. B will only examine the per-node
1456 * counters in determining if surplus huge pages can be
1457 * converted to normal huge pages in adjust_pool_surplus(). A
1458 * won't be able to increment the per-node counter, until the
1459 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1460 * no more huge pages can be converted from surplus to normal
1461 * state (and doesn't try to convert again). Thus, we have a
1462 * case where a surplus huge page exists, the pool is grown, and
1463 * the surplus huge page still exists after, even though it
1464 * should just have been converted to a normal huge page. This
1465 * does not leak memory, though, as the hugepage will be freed
1466 * once it is out of use. It also does not allow the counters to
1467 * go out of whack in adjust_pool_surplus() as we don't modify
1468 * the node values until we've gotten the hugepage and only the
1469 * per-node value is checked there.
1471 spin_lock(&hugetlb_lock);
1472 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1473 spin_unlock(&hugetlb_lock);
1477 h->surplus_huge_pages++;
1479 spin_unlock(&hugetlb_lock);
1481 if (nid == NUMA_NO_NODE)
1482 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
1483 __GFP_REPEAT|__GFP_NOWARN,
1484 huge_page_order(h));
1486 page = alloc_pages_exact_node(nid,
1487 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1488 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
1490 spin_lock(&hugetlb_lock);
1492 INIT_LIST_HEAD(&page->lru);
1493 r_nid = page_to_nid(page);
1494 set_compound_page_dtor(page, free_huge_page);
1495 set_hugetlb_cgroup(page, NULL);
1497 * We incremented the global counters already
1499 h->nr_huge_pages_node[r_nid]++;
1500 h->surplus_huge_pages_node[r_nid]++;
1501 __count_vm_event(HTLB_BUDDY_PGALLOC);
1504 h->surplus_huge_pages--;
1505 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1507 spin_unlock(&hugetlb_lock);
1513 * This allocation function is useful in the context where vma is irrelevant.
1514 * E.g. soft-offlining uses this function because it only cares physical
1515 * address of error page.
1517 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1519 struct page *page = NULL;
1521 spin_lock(&hugetlb_lock);
1522 if (h->free_huge_pages - h->resv_huge_pages > 0)
1523 page = dequeue_huge_page_node(h, nid);
1524 spin_unlock(&hugetlb_lock);
1527 page = alloc_buddy_huge_page(h, nid);
1533 * Increase the hugetlb pool such that it can accommodate a reservation
1536 static int gather_surplus_pages(struct hstate *h, int delta)
1538 struct list_head surplus_list;
1539 struct page *page, *tmp;
1541 int needed, allocated;
1542 bool alloc_ok = true;
1544 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1546 h->resv_huge_pages += delta;
1551 INIT_LIST_HEAD(&surplus_list);
1555 spin_unlock(&hugetlb_lock);
1556 for (i = 0; i < needed; i++) {
1557 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1562 list_add(&page->lru, &surplus_list);
1567 * After retaking hugetlb_lock, we need to recalculate 'needed'
1568 * because either resv_huge_pages or free_huge_pages may have changed.
1570 spin_lock(&hugetlb_lock);
1571 needed = (h->resv_huge_pages + delta) -
1572 (h->free_huge_pages + allocated);
1577 * We were not able to allocate enough pages to
1578 * satisfy the entire reservation so we free what
1579 * we've allocated so far.
1584 * The surplus_list now contains _at_least_ the number of extra pages
1585 * needed to accommodate the reservation. Add the appropriate number
1586 * of pages to the hugetlb pool and free the extras back to the buddy
1587 * allocator. Commit the entire reservation here to prevent another
1588 * process from stealing the pages as they are added to the pool but
1589 * before they are reserved.
1591 needed += allocated;
1592 h->resv_huge_pages += delta;
1595 /* Free the needed pages to the hugetlb pool */
1596 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1600 * This page is now managed by the hugetlb allocator and has
1601 * no users -- drop the buddy allocator's reference.
1603 put_page_testzero(page);
1604 VM_BUG_ON_PAGE(page_count(page), page);
1605 enqueue_huge_page(h, page);
1608 spin_unlock(&hugetlb_lock);
1610 /* Free unnecessary surplus pages to the buddy allocator */
1611 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1613 spin_lock(&hugetlb_lock);
1619 * When releasing a hugetlb pool reservation, any surplus pages that were
1620 * allocated to satisfy the reservation must be explicitly freed if they were
1622 * Called with hugetlb_lock held.
1624 static void return_unused_surplus_pages(struct hstate *h,
1625 unsigned long unused_resv_pages)
1627 unsigned long nr_pages;
1629 /* Uncommit the reservation */
1630 h->resv_huge_pages -= unused_resv_pages;
1632 /* Cannot return gigantic pages currently */
1633 if (hstate_is_gigantic(h))
1636 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1639 * We want to release as many surplus pages as possible, spread
1640 * evenly across all nodes with memory. Iterate across these nodes
1641 * until we can no longer free unreserved surplus pages. This occurs
1642 * when the nodes with surplus pages have no free pages.
1643 * free_pool_huge_page() will balance the the freed pages across the
1644 * on-line nodes with memory and will handle the hstate accounting.
1646 while (nr_pages--) {
1647 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1649 cond_resched_lock(&hugetlb_lock);
1655 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1656 * are used by the huge page allocation routines to manage reservations.
1658 * vma_needs_reservation is called to determine if the huge page at addr
1659 * within the vma has an associated reservation. If a reservation is
1660 * needed, the value 1 is returned. The caller is then responsible for
1661 * managing the global reservation and subpool usage counts. After
1662 * the huge page has been allocated, vma_commit_reservation is called
1663 * to add the page to the reservation map. If the page allocation fails,
1664 * the reservation must be ended instead of committed. vma_end_reservation
1665 * is called in such cases.
1667 * In the normal case, vma_commit_reservation returns the same value
1668 * as the preceding vma_needs_reservation call. The only time this
1669 * is not the case is if a reserve map was changed between calls. It
1670 * is the responsibility of the caller to notice the difference and
1671 * take appropriate action.
1673 enum vma_resv_mode {
1678 static long __vma_reservation_common(struct hstate *h,
1679 struct vm_area_struct *vma, unsigned long addr,
1680 enum vma_resv_mode mode)
1682 struct resv_map *resv;
1686 resv = vma_resv_map(vma);
1690 idx = vma_hugecache_offset(h, vma, addr);
1692 case VMA_NEEDS_RESV:
1693 ret = region_chg(resv, idx, idx + 1);
1695 case VMA_COMMIT_RESV:
1696 ret = region_add(resv, idx, idx + 1);
1699 region_abort(resv, idx, idx + 1);
1706 if (vma->vm_flags & VM_MAYSHARE)
1709 return ret < 0 ? ret : 0;
1712 static long vma_needs_reservation(struct hstate *h,
1713 struct vm_area_struct *vma, unsigned long addr)
1715 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1718 static long vma_commit_reservation(struct hstate *h,
1719 struct vm_area_struct *vma, unsigned long addr)
1721 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1724 static void vma_end_reservation(struct hstate *h,
1725 struct vm_area_struct *vma, unsigned long addr)
1727 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1730 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1731 unsigned long addr, int avoid_reserve)
1733 struct hugepage_subpool *spool = subpool_vma(vma);
1734 struct hstate *h = hstate_vma(vma);
1738 struct hugetlb_cgroup *h_cg;
1740 idx = hstate_index(h);
1742 * Processes that did not create the mapping will have no
1743 * reserves and will not have accounted against subpool
1744 * limit. Check that the subpool limit can be made before
1745 * satisfying the allocation MAP_NORESERVE mappings may also
1746 * need pages and subpool limit allocated allocated if no reserve
1749 chg = vma_needs_reservation(h, vma, addr);
1751 return ERR_PTR(-ENOMEM);
1752 if (chg || avoid_reserve)
1753 if (hugepage_subpool_get_pages(spool, 1) < 0) {
1754 vma_end_reservation(h, vma, addr);
1755 return ERR_PTR(-ENOSPC);
1758 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1760 goto out_subpool_put;
1762 spin_lock(&hugetlb_lock);
1763 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
1765 spin_unlock(&hugetlb_lock);
1766 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1768 goto out_uncharge_cgroup;
1770 spin_lock(&hugetlb_lock);
1771 list_move(&page->lru, &h->hugepage_activelist);
1774 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1775 spin_unlock(&hugetlb_lock);
1777 set_page_private(page, (unsigned long)spool);
1779 commit = vma_commit_reservation(h, vma, addr);
1780 if (unlikely(chg > commit)) {
1782 * The page was added to the reservation map between
1783 * vma_needs_reservation and vma_commit_reservation.
1784 * This indicates a race with hugetlb_reserve_pages.
1785 * Adjust for the subpool count incremented above AND
1786 * in hugetlb_reserve_pages for the same page. Also,
1787 * the reservation count added in hugetlb_reserve_pages
1788 * no longer applies.
1792 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
1793 hugetlb_acct_memory(h, -rsv_adjust);
1797 out_uncharge_cgroup:
1798 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1800 if (chg || avoid_reserve)
1801 hugepage_subpool_put_pages(spool, 1);
1802 vma_end_reservation(h, vma, addr);
1803 return ERR_PTR(-ENOSPC);
1807 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1808 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1809 * where no ERR_VALUE is expected to be returned.
1811 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1812 unsigned long addr, int avoid_reserve)
1814 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1820 int __weak alloc_bootmem_huge_page(struct hstate *h)
1822 struct huge_bootmem_page *m;
1825 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1828 addr = memblock_virt_alloc_try_nid_nopanic(
1829 huge_page_size(h), huge_page_size(h),
1830 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1833 * Use the beginning of the huge page to store the
1834 * huge_bootmem_page struct (until gather_bootmem
1835 * puts them into the mem_map).
1844 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
1845 /* Put them into a private list first because mem_map is not up yet */
1846 list_add(&m->list, &huge_boot_pages);
1851 static void __init prep_compound_huge_page(struct page *page, int order)
1853 if (unlikely(order > (MAX_ORDER - 1)))
1854 prep_compound_gigantic_page(page, order);
1856 prep_compound_page(page, order);
1859 /* Put bootmem huge pages into the standard lists after mem_map is up */
1860 static void __init gather_bootmem_prealloc(void)
1862 struct huge_bootmem_page *m;
1864 list_for_each_entry(m, &huge_boot_pages, list) {
1865 struct hstate *h = m->hstate;
1868 #ifdef CONFIG_HIGHMEM
1869 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1870 memblock_free_late(__pa(m),
1871 sizeof(struct huge_bootmem_page));
1873 page = virt_to_page(m);
1875 WARN_ON(page_count(page) != 1);
1876 prep_compound_huge_page(page, h->order);
1877 WARN_ON(PageReserved(page));
1878 prep_new_huge_page(h, page, page_to_nid(page));
1880 * If we had gigantic hugepages allocated at boot time, we need
1881 * to restore the 'stolen' pages to totalram_pages in order to
1882 * fix confusing memory reports from free(1) and another
1883 * side-effects, like CommitLimit going negative.
1885 if (hstate_is_gigantic(h))
1886 adjust_managed_page_count(page, 1 << h->order);
1890 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1894 for (i = 0; i < h->max_huge_pages; ++i) {
1895 if (hstate_is_gigantic(h)) {
1896 if (!alloc_bootmem_huge_page(h))
1898 } else if (!alloc_fresh_huge_page(h,
1899 &node_states[N_MEMORY]))
1902 h->max_huge_pages = i;
1905 static void __init hugetlb_init_hstates(void)
1909 for_each_hstate(h) {
1910 if (minimum_order > huge_page_order(h))
1911 minimum_order = huge_page_order(h);
1913 /* oversize hugepages were init'ed in early boot */
1914 if (!hstate_is_gigantic(h))
1915 hugetlb_hstate_alloc_pages(h);
1917 VM_BUG_ON(minimum_order == UINT_MAX);
1920 static char * __init memfmt(char *buf, unsigned long n)
1922 if (n >= (1UL << 30))
1923 sprintf(buf, "%lu GB", n >> 30);
1924 else if (n >= (1UL << 20))
1925 sprintf(buf, "%lu MB", n >> 20);
1927 sprintf(buf, "%lu KB", n >> 10);
1931 static void __init report_hugepages(void)
1935 for_each_hstate(h) {
1937 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1938 memfmt(buf, huge_page_size(h)),
1939 h->free_huge_pages);
1943 #ifdef CONFIG_HIGHMEM
1944 static void try_to_free_low(struct hstate *h, unsigned long count,
1945 nodemask_t *nodes_allowed)
1949 if (hstate_is_gigantic(h))
1952 for_each_node_mask(i, *nodes_allowed) {
1953 struct page *page, *next;
1954 struct list_head *freel = &h->hugepage_freelists[i];
1955 list_for_each_entry_safe(page, next, freel, lru) {
1956 if (count >= h->nr_huge_pages)
1958 if (PageHighMem(page))
1960 list_del(&page->lru);
1961 update_and_free_page(h, page);
1962 h->free_huge_pages--;
1963 h->free_huge_pages_node[page_to_nid(page)]--;
1968 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1969 nodemask_t *nodes_allowed)
1975 * Increment or decrement surplus_huge_pages. Keep node-specific counters
1976 * balanced by operating on them in a round-robin fashion.
1977 * Returns 1 if an adjustment was made.
1979 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1984 VM_BUG_ON(delta != -1 && delta != 1);
1987 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1988 if (h->surplus_huge_pages_node[node])
1992 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1993 if (h->surplus_huge_pages_node[node] <
1994 h->nr_huge_pages_node[node])
2001 h->surplus_huge_pages += delta;
2002 h->surplus_huge_pages_node[node] += delta;
2006 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2007 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2008 nodemask_t *nodes_allowed)
2010 unsigned long min_count, ret;
2012 if (hstate_is_gigantic(h) && !gigantic_page_supported())
2013 return h->max_huge_pages;
2016 * Increase the pool size
2017 * First take pages out of surplus state. Then make up the
2018 * remaining difference by allocating fresh huge pages.
2020 * We might race with alloc_buddy_huge_page() here and be unable
2021 * to convert a surplus huge page to a normal huge page. That is
2022 * not critical, though, it just means the overall size of the
2023 * pool might be one hugepage larger than it needs to be, but
2024 * within all the constraints specified by the sysctls.
2026 spin_lock(&hugetlb_lock);
2027 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2028 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2032 while (count > persistent_huge_pages(h)) {
2034 * If this allocation races such that we no longer need the
2035 * page, free_huge_page will handle it by freeing the page
2036 * and reducing the surplus.
2038 spin_unlock(&hugetlb_lock);
2039 if (hstate_is_gigantic(h))
2040 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2042 ret = alloc_fresh_huge_page(h, nodes_allowed);
2043 spin_lock(&hugetlb_lock);
2047 /* Bail for signals. Probably ctrl-c from user */
2048 if (signal_pending(current))
2053 * Decrease the pool size
2054 * First return free pages to the buddy allocator (being careful
2055 * to keep enough around to satisfy reservations). Then place
2056 * pages into surplus state as needed so the pool will shrink
2057 * to the desired size as pages become free.
2059 * By placing pages into the surplus state independent of the
2060 * overcommit value, we are allowing the surplus pool size to
2061 * exceed overcommit. There are few sane options here. Since
2062 * alloc_buddy_huge_page() is checking the global counter,
2063 * though, we'll note that we're not allowed to exceed surplus
2064 * and won't grow the pool anywhere else. Not until one of the
2065 * sysctls are changed, or the surplus pages go out of use.
2067 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2068 min_count = max(count, min_count);
2069 try_to_free_low(h, min_count, nodes_allowed);
2070 while (min_count < persistent_huge_pages(h)) {
2071 if (!free_pool_huge_page(h, nodes_allowed, 0))
2073 cond_resched_lock(&hugetlb_lock);
2075 while (count < persistent_huge_pages(h)) {
2076 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2080 ret = persistent_huge_pages(h);
2081 spin_unlock(&hugetlb_lock);
2085 #define HSTATE_ATTR_RO(_name) \
2086 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2088 #define HSTATE_ATTR(_name) \
2089 static struct kobj_attribute _name##_attr = \
2090 __ATTR(_name, 0644, _name##_show, _name##_store)
2092 static struct kobject *hugepages_kobj;
2093 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2095 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2097 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2101 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2102 if (hstate_kobjs[i] == kobj) {
2104 *nidp = NUMA_NO_NODE;
2108 return kobj_to_node_hstate(kobj, nidp);
2111 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2112 struct kobj_attribute *attr, char *buf)
2115 unsigned long nr_huge_pages;
2118 h = kobj_to_hstate(kobj, &nid);
2119 if (nid == NUMA_NO_NODE)
2120 nr_huge_pages = h->nr_huge_pages;
2122 nr_huge_pages = h->nr_huge_pages_node[nid];
2124 return sprintf(buf, "%lu\n", nr_huge_pages);
2127 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2128 struct hstate *h, int nid,
2129 unsigned long count, size_t len)
2132 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2134 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2139 if (nid == NUMA_NO_NODE) {
2141 * global hstate attribute
2143 if (!(obey_mempolicy &&
2144 init_nodemask_of_mempolicy(nodes_allowed))) {
2145 NODEMASK_FREE(nodes_allowed);
2146 nodes_allowed = &node_states[N_MEMORY];
2148 } else if (nodes_allowed) {
2150 * per node hstate attribute: adjust count to global,
2151 * but restrict alloc/free to the specified node.
2153 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2154 init_nodemask_of_node(nodes_allowed, nid);
2156 nodes_allowed = &node_states[N_MEMORY];
2158 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2160 if (nodes_allowed != &node_states[N_MEMORY])
2161 NODEMASK_FREE(nodes_allowed);
2165 NODEMASK_FREE(nodes_allowed);
2169 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2170 struct kobject *kobj, const char *buf,
2174 unsigned long count;
2178 err = kstrtoul(buf, 10, &count);
2182 h = kobj_to_hstate(kobj, &nid);
2183 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2186 static ssize_t nr_hugepages_show(struct kobject *kobj,
2187 struct kobj_attribute *attr, char *buf)
2189 return nr_hugepages_show_common(kobj, attr, buf);
2192 static ssize_t nr_hugepages_store(struct kobject *kobj,
2193 struct kobj_attribute *attr, const char *buf, size_t len)
2195 return nr_hugepages_store_common(false, kobj, buf, len);
2197 HSTATE_ATTR(nr_hugepages);
2202 * hstate attribute for optionally mempolicy-based constraint on persistent
2203 * huge page alloc/free.
2205 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2206 struct kobj_attribute *attr, char *buf)
2208 return nr_hugepages_show_common(kobj, attr, buf);
2211 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2212 struct kobj_attribute *attr, const char *buf, size_t len)
2214 return nr_hugepages_store_common(true, kobj, buf, len);
2216 HSTATE_ATTR(nr_hugepages_mempolicy);
2220 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2221 struct kobj_attribute *attr, char *buf)
2223 struct hstate *h = kobj_to_hstate(kobj, NULL);
2224 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2227 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2228 struct kobj_attribute *attr, const char *buf, size_t count)
2231 unsigned long input;
2232 struct hstate *h = kobj_to_hstate(kobj, NULL);
2234 if (hstate_is_gigantic(h))
2237 err = kstrtoul(buf, 10, &input);
2241 spin_lock(&hugetlb_lock);
2242 h->nr_overcommit_huge_pages = input;
2243 spin_unlock(&hugetlb_lock);
2247 HSTATE_ATTR(nr_overcommit_hugepages);
2249 static ssize_t free_hugepages_show(struct kobject *kobj,
2250 struct kobj_attribute *attr, char *buf)
2253 unsigned long free_huge_pages;
2256 h = kobj_to_hstate(kobj, &nid);
2257 if (nid == NUMA_NO_NODE)
2258 free_huge_pages = h->free_huge_pages;
2260 free_huge_pages = h->free_huge_pages_node[nid];
2262 return sprintf(buf, "%lu\n", free_huge_pages);
2264 HSTATE_ATTR_RO(free_hugepages);
2266 static ssize_t resv_hugepages_show(struct kobject *kobj,
2267 struct kobj_attribute *attr, char *buf)
2269 struct hstate *h = kobj_to_hstate(kobj, NULL);
2270 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2272 HSTATE_ATTR_RO(resv_hugepages);
2274 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2275 struct kobj_attribute *attr, char *buf)
2278 unsigned long surplus_huge_pages;
2281 h = kobj_to_hstate(kobj, &nid);
2282 if (nid == NUMA_NO_NODE)
2283 surplus_huge_pages = h->surplus_huge_pages;
2285 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2287 return sprintf(buf, "%lu\n", surplus_huge_pages);
2289 HSTATE_ATTR_RO(surplus_hugepages);
2291 static struct attribute *hstate_attrs[] = {
2292 &nr_hugepages_attr.attr,
2293 &nr_overcommit_hugepages_attr.attr,
2294 &free_hugepages_attr.attr,
2295 &resv_hugepages_attr.attr,
2296 &surplus_hugepages_attr.attr,
2298 &nr_hugepages_mempolicy_attr.attr,
2303 static struct attribute_group hstate_attr_group = {
2304 .attrs = hstate_attrs,
2307 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2308 struct kobject **hstate_kobjs,
2309 struct attribute_group *hstate_attr_group)
2312 int hi = hstate_index(h);
2314 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2315 if (!hstate_kobjs[hi])
2318 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2320 kobject_put(hstate_kobjs[hi]);
2325 static void __init hugetlb_sysfs_init(void)
2330 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2331 if (!hugepages_kobj)
2334 for_each_hstate(h) {
2335 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2336 hstate_kobjs, &hstate_attr_group);
2338 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2345 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2346 * with node devices in node_devices[] using a parallel array. The array
2347 * index of a node device or _hstate == node id.
2348 * This is here to avoid any static dependency of the node device driver, in
2349 * the base kernel, on the hugetlb module.
2351 struct node_hstate {
2352 struct kobject *hugepages_kobj;
2353 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2355 struct node_hstate node_hstates[MAX_NUMNODES];
2358 * A subset of global hstate attributes for node devices
2360 static struct attribute *per_node_hstate_attrs[] = {
2361 &nr_hugepages_attr.attr,
2362 &free_hugepages_attr.attr,
2363 &surplus_hugepages_attr.attr,
2367 static struct attribute_group per_node_hstate_attr_group = {
2368 .attrs = per_node_hstate_attrs,
2372 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2373 * Returns node id via non-NULL nidp.
2375 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2379 for (nid = 0; nid < nr_node_ids; nid++) {
2380 struct node_hstate *nhs = &node_hstates[nid];
2382 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2383 if (nhs->hstate_kobjs[i] == kobj) {
2395 * Unregister hstate attributes from a single node device.
2396 * No-op if no hstate attributes attached.
2398 static void hugetlb_unregister_node(struct node *node)
2401 struct node_hstate *nhs = &node_hstates[node->dev.id];
2403 if (!nhs->hugepages_kobj)
2404 return; /* no hstate attributes */
2406 for_each_hstate(h) {
2407 int idx = hstate_index(h);
2408 if (nhs->hstate_kobjs[idx]) {
2409 kobject_put(nhs->hstate_kobjs[idx]);
2410 nhs->hstate_kobjs[idx] = NULL;
2414 kobject_put(nhs->hugepages_kobj);
2415 nhs->hugepages_kobj = NULL;
2419 * hugetlb module exit: unregister hstate attributes from node devices
2422 static void hugetlb_unregister_all_nodes(void)
2427 * disable node device registrations.
2429 register_hugetlbfs_with_node(NULL, NULL);
2432 * remove hstate attributes from any nodes that have them.
2434 for (nid = 0; nid < nr_node_ids; nid++)
2435 hugetlb_unregister_node(node_devices[nid]);
2439 * Register hstate attributes for a single node device.
2440 * No-op if attributes already registered.
2442 static void hugetlb_register_node(struct node *node)
2445 struct node_hstate *nhs = &node_hstates[node->dev.id];
2448 if (nhs->hugepages_kobj)
2449 return; /* already allocated */
2451 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2453 if (!nhs->hugepages_kobj)
2456 for_each_hstate(h) {
2457 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2459 &per_node_hstate_attr_group);
2461 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2462 h->name, node->dev.id);
2463 hugetlb_unregister_node(node);
2470 * hugetlb init time: register hstate attributes for all registered node
2471 * devices of nodes that have memory. All on-line nodes should have
2472 * registered their associated device by this time.
2474 static void __init hugetlb_register_all_nodes(void)
2478 for_each_node_state(nid, N_MEMORY) {
2479 struct node *node = node_devices[nid];
2480 if (node->dev.id == nid)
2481 hugetlb_register_node(node);
2485 * Let the node device driver know we're here so it can
2486 * [un]register hstate attributes on node hotplug.
2488 register_hugetlbfs_with_node(hugetlb_register_node,
2489 hugetlb_unregister_node);
2491 #else /* !CONFIG_NUMA */
2493 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2501 static void hugetlb_unregister_all_nodes(void) { }
2503 static void hugetlb_register_all_nodes(void) { }
2507 static void __exit hugetlb_exit(void)
2511 hugetlb_unregister_all_nodes();
2513 for_each_hstate(h) {
2514 kobject_put(hstate_kobjs[hstate_index(h)]);
2517 kobject_put(hugepages_kobj);
2518 kfree(hugetlb_fault_mutex_table);
2520 module_exit(hugetlb_exit);
2522 static int __init hugetlb_init(void)
2526 if (!hugepages_supported())
2529 if (!size_to_hstate(default_hstate_size)) {
2530 default_hstate_size = HPAGE_SIZE;
2531 if (!size_to_hstate(default_hstate_size))
2532 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2534 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2535 if (default_hstate_max_huge_pages)
2536 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2538 hugetlb_init_hstates();
2539 gather_bootmem_prealloc();
2542 hugetlb_sysfs_init();
2543 hugetlb_register_all_nodes();
2544 hugetlb_cgroup_file_init();
2547 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2549 num_fault_mutexes = 1;
2551 hugetlb_fault_mutex_table =
2552 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2553 BUG_ON(!hugetlb_fault_mutex_table);
2555 for (i = 0; i < num_fault_mutexes; i++)
2556 mutex_init(&hugetlb_fault_mutex_table[i]);
2559 module_init(hugetlb_init);
2561 /* Should be called on processing a hugepagesz=... option */
2562 void __init hugetlb_add_hstate(unsigned order)
2567 if (size_to_hstate(PAGE_SIZE << order)) {
2568 pr_warning("hugepagesz= specified twice, ignoring\n");
2571 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2573 h = &hstates[hugetlb_max_hstate++];
2575 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2576 h->nr_huge_pages = 0;
2577 h->free_huge_pages = 0;
2578 for (i = 0; i < MAX_NUMNODES; ++i)
2579 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2580 INIT_LIST_HEAD(&h->hugepage_activelist);
2581 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2582 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2583 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2584 huge_page_size(h)/1024);
2589 static int __init hugetlb_nrpages_setup(char *s)
2592 static unsigned long *last_mhp;
2595 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2596 * so this hugepages= parameter goes to the "default hstate".
2598 if (!hugetlb_max_hstate)
2599 mhp = &default_hstate_max_huge_pages;
2601 mhp = &parsed_hstate->max_huge_pages;
2603 if (mhp == last_mhp) {
2604 pr_warning("hugepages= specified twice without "
2605 "interleaving hugepagesz=, ignoring\n");
2609 if (sscanf(s, "%lu", mhp) <= 0)
2613 * Global state is always initialized later in hugetlb_init.
2614 * But we need to allocate >= MAX_ORDER hstates here early to still
2615 * use the bootmem allocator.
2617 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2618 hugetlb_hstate_alloc_pages(parsed_hstate);
2624 __setup("hugepages=", hugetlb_nrpages_setup);
2626 static int __init hugetlb_default_setup(char *s)
2628 default_hstate_size = memparse(s, &s);
2631 __setup("default_hugepagesz=", hugetlb_default_setup);
2633 static unsigned int cpuset_mems_nr(unsigned int *array)
2636 unsigned int nr = 0;
2638 for_each_node_mask(node, cpuset_current_mems_allowed)
2644 #ifdef CONFIG_SYSCTL
2645 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2646 struct ctl_table *table, int write,
2647 void __user *buffer, size_t *length, loff_t *ppos)
2649 struct hstate *h = &default_hstate;
2650 unsigned long tmp = h->max_huge_pages;
2653 if (!hugepages_supported())
2657 table->maxlen = sizeof(unsigned long);
2658 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2663 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2664 NUMA_NO_NODE, tmp, *length);
2669 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2670 void __user *buffer, size_t *length, loff_t *ppos)
2673 return hugetlb_sysctl_handler_common(false, table, write,
2674 buffer, length, ppos);
2678 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2679 void __user *buffer, size_t *length, loff_t *ppos)
2681 return hugetlb_sysctl_handler_common(true, table, write,
2682 buffer, length, ppos);
2684 #endif /* CONFIG_NUMA */
2686 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2687 void __user *buffer,
2688 size_t *length, loff_t *ppos)
2690 struct hstate *h = &default_hstate;
2694 if (!hugepages_supported())
2697 tmp = h->nr_overcommit_huge_pages;
2699 if (write && hstate_is_gigantic(h))
2703 table->maxlen = sizeof(unsigned long);
2704 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2709 spin_lock(&hugetlb_lock);
2710 h->nr_overcommit_huge_pages = tmp;
2711 spin_unlock(&hugetlb_lock);
2717 #endif /* CONFIG_SYSCTL */
2719 void hugetlb_report_meminfo(struct seq_file *m)
2721 struct hstate *h = &default_hstate;
2722 if (!hugepages_supported())
2725 "HugePages_Total: %5lu\n"
2726 "HugePages_Free: %5lu\n"
2727 "HugePages_Rsvd: %5lu\n"
2728 "HugePages_Surp: %5lu\n"
2729 "Hugepagesize: %8lu kB\n",
2733 h->surplus_huge_pages,
2734 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2737 int hugetlb_report_node_meminfo(int nid, char *buf)
2739 struct hstate *h = &default_hstate;
2740 if (!hugepages_supported())
2743 "Node %d HugePages_Total: %5u\n"
2744 "Node %d HugePages_Free: %5u\n"
2745 "Node %d HugePages_Surp: %5u\n",
2746 nid, h->nr_huge_pages_node[nid],
2747 nid, h->free_huge_pages_node[nid],
2748 nid, h->surplus_huge_pages_node[nid]);
2751 void hugetlb_show_meminfo(void)
2756 if (!hugepages_supported())
2759 for_each_node_state(nid, N_MEMORY)
2761 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2763 h->nr_huge_pages_node[nid],
2764 h->free_huge_pages_node[nid],
2765 h->surplus_huge_pages_node[nid],
2766 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2769 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2770 unsigned long hugetlb_total_pages(void)
2773 unsigned long nr_total_pages = 0;
2776 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2777 return nr_total_pages;
2780 static int hugetlb_acct_memory(struct hstate *h, long delta)
2784 spin_lock(&hugetlb_lock);
2786 * When cpuset is configured, it breaks the strict hugetlb page
2787 * reservation as the accounting is done on a global variable. Such
2788 * reservation is completely rubbish in the presence of cpuset because
2789 * the reservation is not checked against page availability for the
2790 * current cpuset. Application can still potentially OOM'ed by kernel
2791 * with lack of free htlb page in cpuset that the task is in.
2792 * Attempt to enforce strict accounting with cpuset is almost
2793 * impossible (or too ugly) because cpuset is too fluid that
2794 * task or memory node can be dynamically moved between cpusets.
2796 * The change of semantics for shared hugetlb mapping with cpuset is
2797 * undesirable. However, in order to preserve some of the semantics,
2798 * we fall back to check against current free page availability as
2799 * a best attempt and hopefully to minimize the impact of changing
2800 * semantics that cpuset has.
2803 if (gather_surplus_pages(h, delta) < 0)
2806 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2807 return_unused_surplus_pages(h, delta);
2814 return_unused_surplus_pages(h, (unsigned long) -delta);
2817 spin_unlock(&hugetlb_lock);
2821 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2823 struct resv_map *resv = vma_resv_map(vma);
2826 * This new VMA should share its siblings reservation map if present.
2827 * The VMA will only ever have a valid reservation map pointer where
2828 * it is being copied for another still existing VMA. As that VMA
2829 * has a reference to the reservation map it cannot disappear until
2830 * after this open call completes. It is therefore safe to take a
2831 * new reference here without additional locking.
2833 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2834 kref_get(&resv->refs);
2837 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2839 struct hstate *h = hstate_vma(vma);
2840 struct resv_map *resv = vma_resv_map(vma);
2841 struct hugepage_subpool *spool = subpool_vma(vma);
2842 unsigned long reserve, start, end;
2845 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2848 start = vma_hugecache_offset(h, vma, vma->vm_start);
2849 end = vma_hugecache_offset(h, vma, vma->vm_end);
2851 reserve = (end - start) - region_count(resv, start, end);
2853 kref_put(&resv->refs, resv_map_release);
2857 * Decrement reserve counts. The global reserve count may be
2858 * adjusted if the subpool has a minimum size.
2860 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
2861 hugetlb_acct_memory(h, -gbl_reserve);
2866 * We cannot handle pagefaults against hugetlb pages at all. They cause
2867 * handle_mm_fault() to try to instantiate regular-sized pages in the
2868 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2871 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2877 const struct vm_operations_struct hugetlb_vm_ops = {
2878 .fault = hugetlb_vm_op_fault,
2879 .open = hugetlb_vm_op_open,
2880 .close = hugetlb_vm_op_close,
2883 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2889 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2890 vma->vm_page_prot)));
2892 entry = huge_pte_wrprotect(mk_huge_pte(page,
2893 vma->vm_page_prot));
2895 entry = pte_mkyoung(entry);
2896 entry = pte_mkhuge(entry);
2897 entry = arch_make_huge_pte(entry, vma, page, writable);
2902 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2903 unsigned long address, pte_t *ptep)
2907 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2908 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2909 update_mmu_cache(vma, address, ptep);
2912 static int is_hugetlb_entry_migration(pte_t pte)
2916 if (huge_pte_none(pte) || pte_present(pte))
2918 swp = pte_to_swp_entry(pte);
2919 if (non_swap_entry(swp) && is_migration_entry(swp))
2925 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2929 if (huge_pte_none(pte) || pte_present(pte))
2931 swp = pte_to_swp_entry(pte);
2932 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2938 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2939 struct vm_area_struct *vma)
2941 pte_t *src_pte, *dst_pte, entry;
2942 struct page *ptepage;
2945 struct hstate *h = hstate_vma(vma);
2946 unsigned long sz = huge_page_size(h);
2947 unsigned long mmun_start; /* For mmu_notifiers */
2948 unsigned long mmun_end; /* For mmu_notifiers */
2951 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2953 mmun_start = vma->vm_start;
2954 mmun_end = vma->vm_end;
2956 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2958 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2959 spinlock_t *src_ptl, *dst_ptl;
2960 src_pte = huge_pte_offset(src, addr);
2963 dst_pte = huge_pte_alloc(dst, addr, sz);
2969 /* If the pagetables are shared don't copy or take references */
2970 if (dst_pte == src_pte)
2973 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2974 src_ptl = huge_pte_lockptr(h, src, src_pte);
2975 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
2976 entry = huge_ptep_get(src_pte);
2977 if (huge_pte_none(entry)) { /* skip none entry */
2979 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2980 is_hugetlb_entry_hwpoisoned(entry))) {
2981 swp_entry_t swp_entry = pte_to_swp_entry(entry);
2983 if (is_write_migration_entry(swp_entry) && cow) {
2985 * COW mappings require pages in both
2986 * parent and child to be set to read.
2988 make_migration_entry_read(&swp_entry);
2989 entry = swp_entry_to_pte(swp_entry);
2990 set_huge_pte_at(src, addr, src_pte, entry);
2992 set_huge_pte_at(dst, addr, dst_pte, entry);
2995 huge_ptep_set_wrprotect(src, addr, src_pte);
2996 mmu_notifier_invalidate_range(src, mmun_start,
2999 entry = huge_ptep_get(src_pte);
3000 ptepage = pte_page(entry);
3002 page_dup_rmap(ptepage);
3003 set_huge_pte_at(dst, addr, dst_pte, entry);
3005 spin_unlock(src_ptl);
3006 spin_unlock(dst_ptl);
3010 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3015 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3016 unsigned long start, unsigned long end,
3017 struct page *ref_page)
3019 int force_flush = 0;
3020 struct mm_struct *mm = vma->vm_mm;
3021 unsigned long address;
3026 struct hstate *h = hstate_vma(vma);
3027 unsigned long sz = huge_page_size(h);
3028 const unsigned long mmun_start = start; /* For mmu_notifiers */
3029 const unsigned long mmun_end = end; /* For mmu_notifiers */
3031 WARN_ON(!is_vm_hugetlb_page(vma));
3032 BUG_ON(start & ~huge_page_mask(h));
3033 BUG_ON(end & ~huge_page_mask(h));
3035 tlb_start_vma(tlb, vma);
3036 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3039 for (; address < end; address += sz) {
3040 ptep = huge_pte_offset(mm, address);
3044 ptl = huge_pte_lock(h, mm, ptep);
3045 if (huge_pmd_unshare(mm, &address, ptep))
3048 pte = huge_ptep_get(ptep);
3049 if (huge_pte_none(pte))
3053 * Migrating hugepage or HWPoisoned hugepage is already
3054 * unmapped and its refcount is dropped, so just clear pte here.
3056 if (unlikely(!pte_present(pte))) {
3057 huge_pte_clear(mm, address, ptep);
3061 page = pte_page(pte);
3063 * If a reference page is supplied, it is because a specific
3064 * page is being unmapped, not a range. Ensure the page we
3065 * are about to unmap is the actual page of interest.
3068 if (page != ref_page)
3072 * Mark the VMA as having unmapped its page so that
3073 * future faults in this VMA will fail rather than
3074 * looking like data was lost
3076 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3079 pte = huge_ptep_get_and_clear(mm, address, ptep);
3080 tlb_remove_tlb_entry(tlb, ptep, address);
3081 if (huge_pte_dirty(pte))
3082 set_page_dirty(page);
3084 page_remove_rmap(page);
3085 force_flush = !__tlb_remove_page(tlb, page);
3091 /* Bail out after unmapping reference page if supplied */
3100 * mmu_gather ran out of room to batch pages, we break out of
3101 * the PTE lock to avoid doing the potential expensive TLB invalidate
3102 * and page-free while holding it.
3107 if (address < end && !ref_page)
3110 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3111 tlb_end_vma(tlb, vma);
3114 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3115 struct vm_area_struct *vma, unsigned long start,
3116 unsigned long end, struct page *ref_page)
3118 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3121 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3122 * test will fail on a vma being torn down, and not grab a page table
3123 * on its way out. We're lucky that the flag has such an appropriate
3124 * name, and can in fact be safely cleared here. We could clear it
3125 * before the __unmap_hugepage_range above, but all that's necessary
3126 * is to clear it before releasing the i_mmap_rwsem. This works
3127 * because in the context this is called, the VMA is about to be
3128 * destroyed and the i_mmap_rwsem is held.
3130 vma->vm_flags &= ~VM_MAYSHARE;
3133 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3134 unsigned long end, struct page *ref_page)
3136 struct mm_struct *mm;
3137 struct mmu_gather tlb;
3141 tlb_gather_mmu(&tlb, mm, start, end);
3142 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3143 tlb_finish_mmu(&tlb, start, end);
3147 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3148 * mappping it owns the reserve page for. The intention is to unmap the page
3149 * from other VMAs and let the children be SIGKILLed if they are faulting the
3152 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3153 struct page *page, unsigned long address)
3155 struct hstate *h = hstate_vma(vma);
3156 struct vm_area_struct *iter_vma;
3157 struct address_space *mapping;
3161 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3162 * from page cache lookup which is in HPAGE_SIZE units.
3164 address = address & huge_page_mask(h);
3165 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3167 mapping = file_inode(vma->vm_file)->i_mapping;
3170 * Take the mapping lock for the duration of the table walk. As
3171 * this mapping should be shared between all the VMAs,
3172 * __unmap_hugepage_range() is called as the lock is already held
3174 i_mmap_lock_write(mapping);
3175 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3176 /* Do not unmap the current VMA */
3177 if (iter_vma == vma)
3181 * Unmap the page from other VMAs without their own reserves.
3182 * They get marked to be SIGKILLed if they fault in these
3183 * areas. This is because a future no-page fault on this VMA
3184 * could insert a zeroed page instead of the data existing
3185 * from the time of fork. This would look like data corruption
3187 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3188 unmap_hugepage_range(iter_vma, address,
3189 address + huge_page_size(h), page);
3191 i_mmap_unlock_write(mapping);
3195 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3196 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3197 * cannot race with other handlers or page migration.
3198 * Keep the pte_same checks anyway to make transition from the mutex easier.
3200 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3201 unsigned long address, pte_t *ptep, pte_t pte,
3202 struct page *pagecache_page, spinlock_t *ptl)
3204 struct hstate *h = hstate_vma(vma);
3205 struct page *old_page, *new_page;
3206 int ret = 0, outside_reserve = 0;
3207 unsigned long mmun_start; /* For mmu_notifiers */
3208 unsigned long mmun_end; /* For mmu_notifiers */
3210 old_page = pte_page(pte);
3213 /* If no-one else is actually using this page, avoid the copy
3214 * and just make the page writable */
3215 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3216 page_move_anon_rmap(old_page, vma, address);
3217 set_huge_ptep_writable(vma, address, ptep);
3222 * If the process that created a MAP_PRIVATE mapping is about to
3223 * perform a COW due to a shared page count, attempt to satisfy
3224 * the allocation without using the existing reserves. The pagecache
3225 * page is used to determine if the reserve at this address was
3226 * consumed or not. If reserves were used, a partial faulted mapping
3227 * at the time of fork() could consume its reserves on COW instead
3228 * of the full address range.
3230 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3231 old_page != pagecache_page)
3232 outside_reserve = 1;
3234 page_cache_get(old_page);
3237 * Drop page table lock as buddy allocator may be called. It will
3238 * be acquired again before returning to the caller, as expected.
3241 new_page = alloc_huge_page(vma, address, outside_reserve);
3243 if (IS_ERR(new_page)) {
3245 * If a process owning a MAP_PRIVATE mapping fails to COW,
3246 * it is due to references held by a child and an insufficient
3247 * huge page pool. To guarantee the original mappers
3248 * reliability, unmap the page from child processes. The child
3249 * may get SIGKILLed if it later faults.
3251 if (outside_reserve) {
3252 page_cache_release(old_page);
3253 BUG_ON(huge_pte_none(pte));
3254 unmap_ref_private(mm, vma, old_page, address);
3255 BUG_ON(huge_pte_none(pte));
3257 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3259 pte_same(huge_ptep_get(ptep), pte)))
3260 goto retry_avoidcopy;
3262 * race occurs while re-acquiring page table
3263 * lock, and our job is done.
3268 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3269 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3270 goto out_release_old;
3274 * When the original hugepage is shared one, it does not have
3275 * anon_vma prepared.
3277 if (unlikely(anon_vma_prepare(vma))) {
3279 goto out_release_all;
3282 copy_user_huge_page(new_page, old_page, address, vma,
3283 pages_per_huge_page(h));
3284 __SetPageUptodate(new_page);
3285 set_page_huge_active(new_page);
3287 mmun_start = address & huge_page_mask(h);
3288 mmun_end = mmun_start + huge_page_size(h);
3289 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3292 * Retake the page table lock to check for racing updates
3293 * before the page tables are altered
3296 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3297 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3298 ClearPagePrivate(new_page);
3301 huge_ptep_clear_flush(vma, address, ptep);
3302 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3303 set_huge_pte_at(mm, address, ptep,
3304 make_huge_pte(vma, new_page, 1));
3305 page_remove_rmap(old_page);
3306 hugepage_add_new_anon_rmap(new_page, vma, address);
3307 /* Make the old page be freed below */
3308 new_page = old_page;
3311 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3313 page_cache_release(new_page);
3315 page_cache_release(old_page);
3317 spin_lock(ptl); /* Caller expects lock to be held */
3321 /* Return the pagecache page at a given address within a VMA */
3322 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3323 struct vm_area_struct *vma, unsigned long address)
3325 struct address_space *mapping;
3328 mapping = vma->vm_file->f_mapping;
3329 idx = vma_hugecache_offset(h, vma, address);
3331 return find_lock_page(mapping, idx);
3335 * Return whether there is a pagecache page to back given address within VMA.
3336 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3338 static bool hugetlbfs_pagecache_present(struct hstate *h,
3339 struct vm_area_struct *vma, unsigned long address)
3341 struct address_space *mapping;
3345 mapping = vma->vm_file->f_mapping;
3346 idx = vma_hugecache_offset(h, vma, address);
3348 page = find_get_page(mapping, idx);
3351 return page != NULL;
3354 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3355 struct address_space *mapping, pgoff_t idx,
3356 unsigned long address, pte_t *ptep, unsigned int flags)
3358 struct hstate *h = hstate_vma(vma);
3359 int ret = VM_FAULT_SIGBUS;
3367 * Currently, we are forced to kill the process in the event the
3368 * original mapper has unmapped pages from the child due to a failed
3369 * COW. Warn that such a situation has occurred as it may not be obvious
3371 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3372 pr_warning("PID %d killed due to inadequate hugepage pool\n",
3378 * Use page lock to guard against racing truncation
3379 * before we get page_table_lock.
3382 page = find_lock_page(mapping, idx);
3384 size = i_size_read(mapping->host) >> huge_page_shift(h);
3387 page = alloc_huge_page(vma, address, 0);
3389 ret = PTR_ERR(page);
3393 ret = VM_FAULT_SIGBUS;
3396 clear_huge_page(page, address, pages_per_huge_page(h));
3397 __SetPageUptodate(page);
3398 set_page_huge_active(page);
3400 if (vma->vm_flags & VM_MAYSHARE) {
3402 struct inode *inode = mapping->host;
3404 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3411 ClearPagePrivate(page);
3413 spin_lock(&inode->i_lock);
3414 inode->i_blocks += blocks_per_huge_page(h);
3415 spin_unlock(&inode->i_lock);
3418 if (unlikely(anon_vma_prepare(vma))) {
3420 goto backout_unlocked;
3426 * If memory error occurs between mmap() and fault, some process
3427 * don't have hwpoisoned swap entry for errored virtual address.
3428 * So we need to block hugepage fault by PG_hwpoison bit check.
3430 if (unlikely(PageHWPoison(page))) {
3431 ret = VM_FAULT_HWPOISON |
3432 VM_FAULT_SET_HINDEX(hstate_index(h));
3433 goto backout_unlocked;
3438 * If we are going to COW a private mapping later, we examine the
3439 * pending reservations for this page now. This will ensure that
3440 * any allocations necessary to record that reservation occur outside
3443 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3444 if (vma_needs_reservation(h, vma, address) < 0) {
3446 goto backout_unlocked;
3448 /* Just decrements count, does not deallocate */
3449 vma_end_reservation(h, vma, address);
3452 ptl = huge_pte_lockptr(h, mm, ptep);
3454 size = i_size_read(mapping->host) >> huge_page_shift(h);
3459 if (!huge_pte_none(huge_ptep_get(ptep)))
3463 ClearPagePrivate(page);
3464 hugepage_add_new_anon_rmap(page, vma, address);
3466 page_dup_rmap(page);
3467 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3468 && (vma->vm_flags & VM_SHARED)));
3469 set_huge_pte_at(mm, address, ptep, new_pte);
3471 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3472 /* Optimization, do the COW without a second fault */
3473 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
3490 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3491 struct vm_area_struct *vma,
3492 struct address_space *mapping,
3493 pgoff_t idx, unsigned long address)
3495 unsigned long key[2];
3498 if (vma->vm_flags & VM_SHARED) {
3499 key[0] = (unsigned long) mapping;
3502 key[0] = (unsigned long) mm;
3503 key[1] = address >> huge_page_shift(h);
3506 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3508 return hash & (num_fault_mutexes - 1);
3512 * For uniprocesor systems we always use a single mutex, so just
3513 * return 0 and avoid the hashing overhead.
3515 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3516 struct vm_area_struct *vma,
3517 struct address_space *mapping,
3518 pgoff_t idx, unsigned long address)
3524 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3525 unsigned long address, unsigned int flags)
3532 struct page *page = NULL;
3533 struct page *pagecache_page = NULL;
3534 struct hstate *h = hstate_vma(vma);
3535 struct address_space *mapping;
3536 int need_wait_lock = 0;
3538 address &= huge_page_mask(h);
3540 ptep = huge_pte_offset(mm, address);
3542 entry = huge_ptep_get(ptep);
3543 if (unlikely(is_hugetlb_entry_migration(entry))) {
3544 migration_entry_wait_huge(vma, mm, ptep);
3546 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3547 return VM_FAULT_HWPOISON_LARGE |
3548 VM_FAULT_SET_HINDEX(hstate_index(h));
3551 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3553 return VM_FAULT_OOM;
3555 mapping = vma->vm_file->f_mapping;
3556 idx = vma_hugecache_offset(h, vma, address);
3559 * Serialize hugepage allocation and instantiation, so that we don't
3560 * get spurious allocation failures if two CPUs race to instantiate
3561 * the same page in the page cache.
3563 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3564 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3566 entry = huge_ptep_get(ptep);
3567 if (huge_pte_none(entry)) {
3568 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3575 * entry could be a migration/hwpoison entry at this point, so this
3576 * check prevents the kernel from going below assuming that we have
3577 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3578 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3581 if (!pte_present(entry))
3585 * If we are going to COW the mapping later, we examine the pending
3586 * reservations for this page now. This will ensure that any
3587 * allocations necessary to record that reservation occur outside the
3588 * spinlock. For private mappings, we also lookup the pagecache
3589 * page now as it is used to determine if a reservation has been
3592 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3593 if (vma_needs_reservation(h, vma, address) < 0) {
3597 /* Just decrements count, does not deallocate */
3598 vma_end_reservation(h, vma, address);
3600 if (!(vma->vm_flags & VM_MAYSHARE))
3601 pagecache_page = hugetlbfs_pagecache_page(h,
3605 ptl = huge_pte_lock(h, mm, ptep);
3607 /* Check for a racing update before calling hugetlb_cow */
3608 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3612 * hugetlb_cow() requires page locks of pte_page(entry) and
3613 * pagecache_page, so here we need take the former one
3614 * when page != pagecache_page or !pagecache_page.
3616 page = pte_page(entry);
3617 if (page != pagecache_page)
3618 if (!trylock_page(page)) {
3625 if (flags & FAULT_FLAG_WRITE) {
3626 if (!huge_pte_write(entry)) {
3627 ret = hugetlb_cow(mm, vma, address, ptep, entry,
3628 pagecache_page, ptl);
3631 entry = huge_pte_mkdirty(entry);
3633 entry = pte_mkyoung(entry);
3634 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3635 flags & FAULT_FLAG_WRITE))
3636 update_mmu_cache(vma, address, ptep);
3638 if (page != pagecache_page)
3644 if (pagecache_page) {
3645 unlock_page(pagecache_page);
3646 put_page(pagecache_page);
3649 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3651 * Generally it's safe to hold refcount during waiting page lock. But
3652 * here we just wait to defer the next page fault to avoid busy loop and
3653 * the page is not used after unlocked before returning from the current
3654 * page fault. So we are safe from accessing freed page, even if we wait
3655 * here without taking refcount.
3658 wait_on_page_locked(page);
3662 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3663 struct page **pages, struct vm_area_struct **vmas,
3664 unsigned long *position, unsigned long *nr_pages,
3665 long i, unsigned int flags)
3667 unsigned long pfn_offset;
3668 unsigned long vaddr = *position;
3669 unsigned long remainder = *nr_pages;
3670 struct hstate *h = hstate_vma(vma);
3672 while (vaddr < vma->vm_end && remainder) {
3674 spinlock_t *ptl = NULL;
3679 * If we have a pending SIGKILL, don't keep faulting pages and
3680 * potentially allocating memory.
3682 if (unlikely(fatal_signal_pending(current))) {
3688 * Some archs (sparc64, sh*) have multiple pte_ts to
3689 * each hugepage. We have to make sure we get the
3690 * first, for the page indexing below to work.
3692 * Note that page table lock is not held when pte is null.
3694 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3696 ptl = huge_pte_lock(h, mm, pte);
3697 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3700 * When coredumping, it suits get_dump_page if we just return
3701 * an error where there's an empty slot with no huge pagecache
3702 * to back it. This way, we avoid allocating a hugepage, and
3703 * the sparse dumpfile avoids allocating disk blocks, but its
3704 * huge holes still show up with zeroes where they need to be.
3706 if (absent && (flags & FOLL_DUMP) &&
3707 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3715 * We need call hugetlb_fault for both hugepages under migration
3716 * (in which case hugetlb_fault waits for the migration,) and
3717 * hwpoisoned hugepages (in which case we need to prevent the
3718 * caller from accessing to them.) In order to do this, we use
3719 * here is_swap_pte instead of is_hugetlb_entry_migration and
3720 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3721 * both cases, and because we can't follow correct pages
3722 * directly from any kind of swap entries.
3724 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3725 ((flags & FOLL_WRITE) &&
3726 !huge_pte_write(huge_ptep_get(pte)))) {
3731 ret = hugetlb_fault(mm, vma, vaddr,
3732 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3733 if (!(ret & VM_FAULT_ERROR))
3740 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3741 page = pte_page(huge_ptep_get(pte));
3744 pages[i] = mem_map_offset(page, pfn_offset);
3745 get_page_foll(pages[i]);
3755 if (vaddr < vma->vm_end && remainder &&
3756 pfn_offset < pages_per_huge_page(h)) {
3758 * We use pfn_offset to avoid touching the pageframes
3759 * of this compound page.
3765 *nr_pages = remainder;
3768 return i ? i : -EFAULT;
3771 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3772 unsigned long address, unsigned long end, pgprot_t newprot)
3774 struct mm_struct *mm = vma->vm_mm;
3775 unsigned long start = address;
3778 struct hstate *h = hstate_vma(vma);
3779 unsigned long pages = 0;
3781 BUG_ON(address >= end);
3782 flush_cache_range(vma, address, end);
3784 mmu_notifier_invalidate_range_start(mm, start, end);
3785 i_mmap_lock_write(vma->vm_file->f_mapping);
3786 for (; address < end; address += huge_page_size(h)) {
3788 ptep = huge_pte_offset(mm, address);
3791 ptl = huge_pte_lock(h, mm, ptep);
3792 if (huge_pmd_unshare(mm, &address, ptep)) {
3797 pte = huge_ptep_get(ptep);
3798 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3802 if (unlikely(is_hugetlb_entry_migration(pte))) {
3803 swp_entry_t entry = pte_to_swp_entry(pte);
3805 if (is_write_migration_entry(entry)) {
3808 make_migration_entry_read(&entry);
3809 newpte = swp_entry_to_pte(entry);
3810 set_huge_pte_at(mm, address, ptep, newpte);
3816 if (!huge_pte_none(pte)) {
3817 pte = huge_ptep_get_and_clear(mm, address, ptep);
3818 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3819 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3820 set_huge_pte_at(mm, address, ptep, pte);
3826 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
3827 * may have cleared our pud entry and done put_page on the page table:
3828 * once we release i_mmap_rwsem, another task can do the final put_page
3829 * and that page table be reused and filled with junk.
3831 flush_tlb_range(vma, start, end);
3832 mmu_notifier_invalidate_range(mm, start, end);
3833 i_mmap_unlock_write(vma->vm_file->f_mapping);
3834 mmu_notifier_invalidate_range_end(mm, start, end);
3836 return pages << h->order;
3839 int hugetlb_reserve_pages(struct inode *inode,
3841 struct vm_area_struct *vma,
3842 vm_flags_t vm_flags)
3845 struct hstate *h = hstate_inode(inode);
3846 struct hugepage_subpool *spool = subpool_inode(inode);
3847 struct resv_map *resv_map;
3851 * Only apply hugepage reservation if asked. At fault time, an
3852 * attempt will be made for VM_NORESERVE to allocate a page
3853 * without using reserves
3855 if (vm_flags & VM_NORESERVE)
3859 * Shared mappings base their reservation on the number of pages that
3860 * are already allocated on behalf of the file. Private mappings need
3861 * to reserve the full area even if read-only as mprotect() may be
3862 * called to make the mapping read-write. Assume !vma is a shm mapping
3864 if (!vma || vma->vm_flags & VM_MAYSHARE) {
3865 resv_map = inode_resv_map(inode);
3867 chg = region_chg(resv_map, from, to);
3870 resv_map = resv_map_alloc();
3876 set_vma_resv_map(vma, resv_map);
3877 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3886 * There must be enough pages in the subpool for the mapping. If
3887 * the subpool has a minimum size, there may be some global
3888 * reservations already in place (gbl_reserve).
3890 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
3891 if (gbl_reserve < 0) {
3897 * Check enough hugepages are available for the reservation.
3898 * Hand the pages back to the subpool if there are not
3900 ret = hugetlb_acct_memory(h, gbl_reserve);
3902 /* put back original number of pages, chg */
3903 (void)hugepage_subpool_put_pages(spool, chg);
3908 * Account for the reservations made. Shared mappings record regions
3909 * that have reservations as they are shared by multiple VMAs.
3910 * When the last VMA disappears, the region map says how much
3911 * the reservation was and the page cache tells how much of
3912 * the reservation was consumed. Private mappings are per-VMA and
3913 * only the consumed reservations are tracked. When the VMA
3914 * disappears, the original reservation is the VMA size and the
3915 * consumed reservations are stored in the map. Hence, nothing
3916 * else has to be done for private mappings here
3918 if (!vma || vma->vm_flags & VM_MAYSHARE) {
3919 long add = region_add(resv_map, from, to);
3921 if (unlikely(chg > add)) {
3923 * pages in this range were added to the reserve
3924 * map between region_chg and region_add. This
3925 * indicates a race with alloc_huge_page. Adjust
3926 * the subpool and reserve counts modified above
3927 * based on the difference.
3931 rsv_adjust = hugepage_subpool_put_pages(spool,
3933 hugetlb_acct_memory(h, -rsv_adjust);
3938 if (!vma || vma->vm_flags & VM_MAYSHARE)
3939 region_abort(resv_map, from, to);
3940 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3941 kref_put(&resv_map->refs, resv_map_release);
3945 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
3948 struct hstate *h = hstate_inode(inode);
3949 struct resv_map *resv_map = inode_resv_map(inode);
3951 struct hugepage_subpool *spool = subpool_inode(inode);
3955 chg = region_del(resv_map, start, end);
3957 * region_del() can fail in the rare case where a region
3958 * must be split and another region descriptor can not be
3959 * allocated. If end == LONG_MAX, it will not fail.
3965 spin_lock(&inode->i_lock);
3966 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3967 spin_unlock(&inode->i_lock);
3970 * If the subpool has a minimum size, the number of global
3971 * reservations to be released may be adjusted.
3973 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
3974 hugetlb_acct_memory(h, -gbl_reserve);
3979 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3980 static unsigned long page_table_shareable(struct vm_area_struct *svma,
3981 struct vm_area_struct *vma,
3982 unsigned long addr, pgoff_t idx)
3984 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3986 unsigned long sbase = saddr & PUD_MASK;
3987 unsigned long s_end = sbase + PUD_SIZE;
3989 /* Allow segments to share if only one is marked locked */
3990 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3991 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3994 * match the virtual addresses, permission and the alignment of the
3997 if (pmd_index(addr) != pmd_index(saddr) ||
3998 vm_flags != svm_flags ||
3999 sbase < svma->vm_start || svma->vm_end < s_end)
4005 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4007 unsigned long base = addr & PUD_MASK;
4008 unsigned long end = base + PUD_SIZE;
4011 * check on proper vm_flags and page table alignment
4013 if (vma->vm_flags & VM_MAYSHARE &&
4014 vma->vm_start <= base && end <= vma->vm_end)
4020 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4021 * and returns the corresponding pte. While this is not necessary for the
4022 * !shared pmd case because we can allocate the pmd later as well, it makes the
4023 * code much cleaner. pmd allocation is essential for the shared case because
4024 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4025 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4026 * bad pmd for sharing.
4028 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4030 struct vm_area_struct *vma = find_vma(mm, addr);
4031 struct address_space *mapping = vma->vm_file->f_mapping;
4032 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4034 struct vm_area_struct *svma;
4035 unsigned long saddr;
4040 if (!vma_shareable(vma, addr))
4041 return (pte_t *)pmd_alloc(mm, pud, addr);
4043 i_mmap_lock_write(mapping);
4044 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4048 saddr = page_table_shareable(svma, vma, addr, idx);
4050 spte = huge_pte_offset(svma->vm_mm, saddr);
4053 get_page(virt_to_page(spte));
4062 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
4064 if (pud_none(*pud)) {
4065 pud_populate(mm, pud,
4066 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4068 put_page(virt_to_page(spte));
4073 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4074 i_mmap_unlock_write(mapping);
4079 * unmap huge page backed by shared pte.
4081 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4082 * indicated by page_count > 1, unmap is achieved by clearing pud and
4083 * decrementing the ref count. If count == 1, the pte page is not shared.
4085 * called with page table lock held.
4087 * returns: 1 successfully unmapped a shared pte page
4088 * 0 the underlying pte page is not shared, or it is the last user
4090 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4092 pgd_t *pgd = pgd_offset(mm, *addr);
4093 pud_t *pud = pud_offset(pgd, *addr);
4095 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4096 if (page_count(virt_to_page(ptep)) == 1)
4100 put_page(virt_to_page(ptep));
4102 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4105 #define want_pmd_share() (1)
4106 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4107 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4112 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4116 #define want_pmd_share() (0)
4117 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4119 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4120 pte_t *huge_pte_alloc(struct mm_struct *mm,
4121 unsigned long addr, unsigned long sz)
4127 pgd = pgd_offset(mm, addr);
4128 pud = pud_alloc(mm, pgd, addr);
4130 if (sz == PUD_SIZE) {
4133 BUG_ON(sz != PMD_SIZE);
4134 if (want_pmd_share() && pud_none(*pud))
4135 pte = huge_pmd_share(mm, addr, pud);
4137 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4140 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
4145 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
4151 pgd = pgd_offset(mm, addr);
4152 if (pgd_present(*pgd)) {
4153 pud = pud_offset(pgd, addr);
4154 if (pud_present(*pud)) {
4156 return (pte_t *)pud;
4157 pmd = pmd_offset(pud, addr);
4160 return (pte_t *) pmd;
4163 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4166 * These functions are overwritable if your architecture needs its own
4169 struct page * __weak
4170 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4173 return ERR_PTR(-EINVAL);
4176 struct page * __weak
4177 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4178 pmd_t *pmd, int flags)
4180 struct page *page = NULL;
4183 ptl = pmd_lockptr(mm, pmd);
4186 * make sure that the address range covered by this pmd is not
4187 * unmapped from other threads.
4189 if (!pmd_huge(*pmd))
4191 if (pmd_present(*pmd)) {
4192 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4193 if (flags & FOLL_GET)
4196 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
4198 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4202 * hwpoisoned entry is treated as no_page_table in
4203 * follow_page_mask().
4211 struct page * __weak
4212 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4213 pud_t *pud, int flags)
4215 if (flags & FOLL_GET)
4218 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4221 #ifdef CONFIG_MEMORY_FAILURE
4224 * This function is called from memory failure code.
4225 * Assume the caller holds page lock of the head page.
4227 int dequeue_hwpoisoned_huge_page(struct page *hpage)
4229 struct hstate *h = page_hstate(hpage);
4230 int nid = page_to_nid(hpage);
4233 spin_lock(&hugetlb_lock);
4235 * Just checking !page_huge_active is not enough, because that could be
4236 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4238 if (!page_huge_active(hpage) && !page_count(hpage)) {
4240 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4241 * but dangling hpage->lru can trigger list-debug warnings
4242 * (this happens when we call unpoison_memory() on it),
4243 * so let it point to itself with list_del_init().
4245 list_del_init(&hpage->lru);
4246 set_page_refcounted(hpage);
4247 h->free_huge_pages--;
4248 h->free_huge_pages_node[nid]--;
4251 spin_unlock(&hugetlb_lock);
4256 bool isolate_huge_page(struct page *page, struct list_head *list)
4260 VM_BUG_ON_PAGE(!PageHead(page), page);
4261 spin_lock(&hugetlb_lock);
4262 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4266 clear_page_huge_active(page);
4267 list_move_tail(&page->lru, list);
4269 spin_unlock(&hugetlb_lock);
4273 void putback_active_hugepage(struct page *page)
4275 VM_BUG_ON_PAGE(!PageHead(page), page);
4276 spin_lock(&hugetlb_lock);
4277 set_page_huge_active(page);
4278 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4279 spin_unlock(&hugetlb_lock);