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);
380 spin_lock(&resv->lock);
381 list_add(&trg->link, &resv->region_cache);
382 resv->region_cache_count++;
386 /* Locate the region we are before or in. */
387 list_for_each_entry(rg, head, link)
391 /* If we are below the current region then a new region is required.
392 * Subtle, allocate a new region at the position but make it zero
393 * size such that we can guarantee to record the reservation. */
394 if (&rg->link == head || t < rg->from) {
396 resv->adds_in_progress--;
397 spin_unlock(&resv->lock);
398 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
404 INIT_LIST_HEAD(&nrg->link);
408 list_add(&nrg->link, rg->link.prev);
413 /* Round our left edge to the current segment if it encloses us. */
418 /* Check for and consume any regions we now overlap with. */
419 list_for_each_entry(rg, rg->link.prev, link) {
420 if (&rg->link == head)
425 /* We overlap with this area, if it extends further than
426 * us then we must extend ourselves. Account for its
427 * existing reservation. */
432 chg -= rg->to - rg->from;
436 spin_unlock(&resv->lock);
437 /* We already know we raced and no longer need the new region */
441 spin_unlock(&resv->lock);
446 * Abort the in progress add operation. The adds_in_progress field
447 * of the resv_map keeps track of the operations in progress between
448 * calls to region_chg and region_add. Operations are sometimes
449 * aborted after the call to region_chg. In such cases, region_abort
450 * is called to decrement the adds_in_progress counter.
452 * NOTE: The range arguments [f, t) are not needed or used in this
453 * routine. They are kept to make reading the calling code easier as
454 * arguments will match the associated region_chg call.
456 static void region_abort(struct resv_map *resv, long f, long t)
458 spin_lock(&resv->lock);
459 VM_BUG_ON(!resv->region_cache_count);
460 resv->adds_in_progress--;
461 spin_unlock(&resv->lock);
465 * Delete the specified range [f, t) from the reserve map. If the
466 * t parameter is LONG_MAX, this indicates that ALL regions after f
467 * should be deleted. Locate the regions which intersect [f, t)
468 * and either trim, delete or split the existing regions.
470 * Returns the number of huge pages deleted from the reserve map.
471 * In the normal case, the return value is zero or more. In the
472 * case where a region must be split, a new region descriptor must
473 * be allocated. If the allocation fails, -ENOMEM will be returned.
474 * NOTE: If the parameter t == LONG_MAX, then we will never split
475 * a region and possibly return -ENOMEM. Callers specifying
476 * t == LONG_MAX do not need to check for -ENOMEM error.
478 static long region_del(struct resv_map *resv, long f, long t)
480 struct list_head *head = &resv->regions;
481 struct file_region *rg, *trg;
482 struct file_region *nrg = NULL;
486 spin_lock(&resv->lock);
487 list_for_each_entry_safe(rg, trg, head, link) {
489 * Skip regions before the range to be deleted. file_region
490 * ranges are normally of the form [from, to). However, there
491 * may be a "placeholder" entry in the map which is of the form
492 * (from, to) with from == to. Check for placeholder entries
493 * at the beginning of the range to be deleted.
495 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
501 if (f > rg->from && t < rg->to) { /* Must split region */
503 * Check for an entry in the cache before dropping
504 * lock and attempting allocation.
507 resv->region_cache_count > resv->adds_in_progress) {
508 nrg = list_first_entry(&resv->region_cache,
511 list_del(&nrg->link);
512 resv->region_cache_count--;
516 spin_unlock(&resv->lock);
517 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
525 /* New entry for end of split region */
528 INIT_LIST_HEAD(&nrg->link);
530 /* Original entry is trimmed */
533 list_add(&nrg->link, &rg->link);
538 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
539 del += rg->to - rg->from;
545 if (f <= rg->from) { /* Trim beginning of region */
548 } else { /* Trim end of region */
554 spin_unlock(&resv->lock);
560 * A rare out of memory error was encountered which prevented removal of
561 * the reserve map region for a page. The huge page itself was free'ed
562 * and removed from the page cache. This routine will adjust the subpool
563 * usage count, and the global reserve count if needed. By incrementing
564 * these counts, the reserve map entry which could not be deleted will
565 * appear as a "reserved" entry instead of simply dangling with incorrect
568 void hugetlb_fix_reserve_counts(struct inode *inode, bool restore_reserve)
570 struct hugepage_subpool *spool = subpool_inode(inode);
573 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
574 if (restore_reserve && rsv_adjust) {
575 struct hstate *h = hstate_inode(inode);
577 hugetlb_acct_memory(h, 1);
582 * Count and return the number of huge pages in the reserve map
583 * that intersect with the range [f, t).
585 static long region_count(struct resv_map *resv, long f, long t)
587 struct list_head *head = &resv->regions;
588 struct file_region *rg;
591 spin_lock(&resv->lock);
592 /* Locate each segment we overlap with, and count that overlap. */
593 list_for_each_entry(rg, head, link) {
602 seg_from = max(rg->from, f);
603 seg_to = min(rg->to, t);
605 chg += seg_to - seg_from;
607 spin_unlock(&resv->lock);
613 * Convert the address within this vma to the page offset within
614 * the mapping, in pagecache page units; huge pages here.
616 static pgoff_t vma_hugecache_offset(struct hstate *h,
617 struct vm_area_struct *vma, unsigned long address)
619 return ((address - vma->vm_start) >> huge_page_shift(h)) +
620 (vma->vm_pgoff >> huge_page_order(h));
623 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
624 unsigned long address)
626 return vma_hugecache_offset(hstate_vma(vma), vma, address);
630 * Return the size of the pages allocated when backing a VMA. In the majority
631 * cases this will be same size as used by the page table entries.
633 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
635 struct hstate *hstate;
637 if (!is_vm_hugetlb_page(vma))
640 hstate = hstate_vma(vma);
642 return 1UL << huge_page_shift(hstate);
644 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
647 * Return the page size being used by the MMU to back a VMA. In the majority
648 * of cases, the page size used by the kernel matches the MMU size. On
649 * architectures where it differs, an architecture-specific version of this
650 * function is required.
652 #ifndef vma_mmu_pagesize
653 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
655 return vma_kernel_pagesize(vma);
660 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
661 * bits of the reservation map pointer, which are always clear due to
664 #define HPAGE_RESV_OWNER (1UL << 0)
665 #define HPAGE_RESV_UNMAPPED (1UL << 1)
666 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
669 * These helpers are used to track how many pages are reserved for
670 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
671 * is guaranteed to have their future faults succeed.
673 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
674 * the reserve counters are updated with the hugetlb_lock held. It is safe
675 * to reset the VMA at fork() time as it is not in use yet and there is no
676 * chance of the global counters getting corrupted as a result of the values.
678 * The private mapping reservation is represented in a subtly different
679 * manner to a shared mapping. A shared mapping has a region map associated
680 * with the underlying file, this region map represents the backing file
681 * pages which have ever had a reservation assigned which this persists even
682 * after the page is instantiated. A private mapping has a region map
683 * associated with the original mmap which is attached to all VMAs which
684 * reference it, this region map represents those offsets which have consumed
685 * reservation ie. where pages have been instantiated.
687 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
689 return (unsigned long)vma->vm_private_data;
692 static void set_vma_private_data(struct vm_area_struct *vma,
695 vma->vm_private_data = (void *)value;
698 struct resv_map *resv_map_alloc(void)
700 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
701 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
703 if (!resv_map || !rg) {
709 kref_init(&resv_map->refs);
710 spin_lock_init(&resv_map->lock);
711 INIT_LIST_HEAD(&resv_map->regions);
713 resv_map->adds_in_progress = 0;
715 INIT_LIST_HEAD(&resv_map->region_cache);
716 list_add(&rg->link, &resv_map->region_cache);
717 resv_map->region_cache_count = 1;
722 void resv_map_release(struct kref *ref)
724 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
725 struct list_head *head = &resv_map->region_cache;
726 struct file_region *rg, *trg;
728 /* Clear out any active regions before we release the map. */
729 region_del(resv_map, 0, LONG_MAX);
731 /* ... and any entries left in the cache */
732 list_for_each_entry_safe(rg, trg, head, link) {
737 VM_BUG_ON(resv_map->adds_in_progress);
742 static inline struct resv_map *inode_resv_map(struct inode *inode)
744 return inode->i_mapping->private_data;
747 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
749 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
750 if (vma->vm_flags & VM_MAYSHARE) {
751 struct address_space *mapping = vma->vm_file->f_mapping;
752 struct inode *inode = mapping->host;
754 return inode_resv_map(inode);
757 return (struct resv_map *)(get_vma_private_data(vma) &
762 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
764 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
765 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
767 set_vma_private_data(vma, (get_vma_private_data(vma) &
768 HPAGE_RESV_MASK) | (unsigned long)map);
771 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
773 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
774 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
776 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
779 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
781 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
783 return (get_vma_private_data(vma) & flag) != 0;
786 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
787 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
789 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
790 if (!(vma->vm_flags & VM_MAYSHARE))
791 vma->vm_private_data = (void *)0;
794 /* Returns true if the VMA has associated reserve pages */
795 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
797 if (vma->vm_flags & VM_NORESERVE) {
799 * This address is already reserved by other process(chg == 0),
800 * so, we should decrement reserved count. Without decrementing,
801 * reserve count remains after releasing inode, because this
802 * allocated page will go into page cache and is regarded as
803 * coming from reserved pool in releasing step. Currently, we
804 * don't have any other solution to deal with this situation
805 * properly, so add work-around here.
807 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
813 /* Shared mappings always use reserves */
814 if (vma->vm_flags & VM_MAYSHARE) {
816 * We know VM_NORESERVE is not set. Therefore, there SHOULD
817 * be a region map for all pages. The only situation where
818 * there is no region map is if a hole was punched via
819 * fallocate. In this case, there really are no reverves to
820 * use. This situation is indicated if chg != 0.
829 * Only the process that called mmap() has reserves for
832 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
838 static void enqueue_huge_page(struct hstate *h, struct page *page)
840 int nid = page_to_nid(page);
841 list_move(&page->lru, &h->hugepage_freelists[nid]);
842 h->free_huge_pages++;
843 h->free_huge_pages_node[nid]++;
846 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
850 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
851 if (!is_migrate_isolate_page(page))
854 * if 'non-isolated free hugepage' not found on the list,
855 * the allocation fails.
857 if (&h->hugepage_freelists[nid] == &page->lru)
859 list_move(&page->lru, &h->hugepage_activelist);
860 set_page_refcounted(page);
861 h->free_huge_pages--;
862 h->free_huge_pages_node[nid]--;
866 /* Movability of hugepages depends on migration support. */
867 static inline gfp_t htlb_alloc_mask(struct hstate *h)
869 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
870 return GFP_HIGHUSER_MOVABLE;
875 static struct page *dequeue_huge_page_vma(struct hstate *h,
876 struct vm_area_struct *vma,
877 unsigned long address, int avoid_reserve,
880 struct page *page = NULL;
881 struct mempolicy *mpol;
882 nodemask_t *nodemask;
883 struct zonelist *zonelist;
886 unsigned int cpuset_mems_cookie;
889 * A child process with MAP_PRIVATE mappings created by their parent
890 * have no page reserves. This check ensures that reservations are
891 * not "stolen". The child may still get SIGKILLed
893 if (!vma_has_reserves(vma, chg) &&
894 h->free_huge_pages - h->resv_huge_pages == 0)
897 /* If reserves cannot be used, ensure enough pages are in the pool */
898 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
902 cpuset_mems_cookie = read_mems_allowed_begin();
903 zonelist = huge_zonelist(vma, address,
904 htlb_alloc_mask(h), &mpol, &nodemask);
906 for_each_zone_zonelist_nodemask(zone, z, zonelist,
907 MAX_NR_ZONES - 1, nodemask) {
908 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
909 page = dequeue_huge_page_node(h, zone_to_nid(zone));
913 if (!vma_has_reserves(vma, chg))
916 SetPagePrivate(page);
917 h->resv_huge_pages--;
924 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
933 * common helper functions for hstate_next_node_to_{alloc|free}.
934 * We may have allocated or freed a huge page based on a different
935 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
936 * be outside of *nodes_allowed. Ensure that we use an allowed
937 * node for alloc or free.
939 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
941 nid = next_node(nid, *nodes_allowed);
942 if (nid == MAX_NUMNODES)
943 nid = first_node(*nodes_allowed);
944 VM_BUG_ON(nid >= MAX_NUMNODES);
949 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
951 if (!node_isset(nid, *nodes_allowed))
952 nid = next_node_allowed(nid, nodes_allowed);
957 * returns the previously saved node ["this node"] from which to
958 * allocate a persistent huge page for the pool and advance the
959 * next node from which to allocate, handling wrap at end of node
962 static int hstate_next_node_to_alloc(struct hstate *h,
963 nodemask_t *nodes_allowed)
967 VM_BUG_ON(!nodes_allowed);
969 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
970 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
976 * helper for free_pool_huge_page() - return the previously saved
977 * node ["this node"] from which to free a huge page. Advance the
978 * next node id whether or not we find a free huge page to free so
979 * that the next attempt to free addresses the next node.
981 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
985 VM_BUG_ON(!nodes_allowed);
987 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
988 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
993 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
994 for (nr_nodes = nodes_weight(*mask); \
996 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
999 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1000 for (nr_nodes = nodes_weight(*mask); \
1002 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1005 #if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
1006 static void destroy_compound_gigantic_page(struct page *page,
1010 int nr_pages = 1 << order;
1011 struct page *p = page + 1;
1013 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1014 clear_compound_head(p);
1015 set_page_refcounted(p);
1018 set_compound_order(page, 0);
1019 __ClearPageHead(page);
1022 static void free_gigantic_page(struct page *page, unsigned int order)
1024 free_contig_range(page_to_pfn(page), 1 << order);
1027 static int __alloc_gigantic_page(unsigned long start_pfn,
1028 unsigned long nr_pages)
1030 unsigned long end_pfn = start_pfn + nr_pages;
1031 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1034 static bool pfn_range_valid_gigantic(unsigned long start_pfn,
1035 unsigned long nr_pages)
1037 unsigned long i, end_pfn = start_pfn + nr_pages;
1040 for (i = start_pfn; i < end_pfn; i++) {
1044 page = pfn_to_page(i);
1046 if (PageReserved(page))
1049 if (page_count(page) > 0)
1059 static bool zone_spans_last_pfn(const struct zone *zone,
1060 unsigned long start_pfn, unsigned long nr_pages)
1062 unsigned long last_pfn = start_pfn + nr_pages - 1;
1063 return zone_spans_pfn(zone, last_pfn);
1066 static struct page *alloc_gigantic_page(int nid, unsigned int order)
1068 unsigned long nr_pages = 1 << order;
1069 unsigned long ret, pfn, flags;
1072 z = NODE_DATA(nid)->node_zones;
1073 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
1074 spin_lock_irqsave(&z->lock, flags);
1076 pfn = ALIGN(z->zone_start_pfn, nr_pages);
1077 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
1078 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
1080 * We release the zone lock here because
1081 * alloc_contig_range() will also lock the zone
1082 * at some point. If there's an allocation
1083 * spinning on this lock, it may win the race
1084 * and cause alloc_contig_range() to fail...
1086 spin_unlock_irqrestore(&z->lock, flags);
1087 ret = __alloc_gigantic_page(pfn, nr_pages);
1089 return pfn_to_page(pfn);
1090 spin_lock_irqsave(&z->lock, flags);
1095 spin_unlock_irqrestore(&z->lock, flags);
1101 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1102 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1104 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1108 page = alloc_gigantic_page(nid, huge_page_order(h));
1110 prep_compound_gigantic_page(page, huge_page_order(h));
1111 prep_new_huge_page(h, page, nid);
1117 static int alloc_fresh_gigantic_page(struct hstate *h,
1118 nodemask_t *nodes_allowed)
1120 struct page *page = NULL;
1123 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1124 page = alloc_fresh_gigantic_page_node(h, node);
1132 static inline bool gigantic_page_supported(void) { return true; }
1134 static inline bool gigantic_page_supported(void) { return false; }
1135 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1136 static inline void destroy_compound_gigantic_page(struct page *page,
1137 unsigned int order) { }
1138 static inline int alloc_fresh_gigantic_page(struct hstate *h,
1139 nodemask_t *nodes_allowed) { return 0; }
1142 static void update_and_free_page(struct hstate *h, struct page *page)
1146 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1150 h->nr_huge_pages_node[page_to_nid(page)]--;
1151 for (i = 0; i < pages_per_huge_page(h); i++) {
1152 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1153 1 << PG_referenced | 1 << PG_dirty |
1154 1 << PG_active | 1 << PG_private |
1157 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1158 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1159 set_page_refcounted(page);
1160 if (hstate_is_gigantic(h)) {
1161 destroy_compound_gigantic_page(page, huge_page_order(h));
1162 free_gigantic_page(page, huge_page_order(h));
1164 __free_pages(page, huge_page_order(h));
1168 struct hstate *size_to_hstate(unsigned long size)
1172 for_each_hstate(h) {
1173 if (huge_page_size(h) == size)
1180 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1181 * to hstate->hugepage_activelist.)
1183 * This function can be called for tail pages, but never returns true for them.
1185 bool page_huge_active(struct page *page)
1187 VM_BUG_ON_PAGE(!PageHuge(page), page);
1188 return PageHead(page) && PagePrivate(&page[1]);
1191 /* never called for tail page */
1192 static void set_page_huge_active(struct page *page)
1194 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1195 SetPagePrivate(&page[1]);
1198 static void clear_page_huge_active(struct page *page)
1200 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1201 ClearPagePrivate(&page[1]);
1204 void free_huge_page(struct page *page)
1207 * Can't pass hstate in here because it is called from the
1208 * compound page destructor.
1210 struct hstate *h = page_hstate(page);
1211 int nid = page_to_nid(page);
1212 struct hugepage_subpool *spool =
1213 (struct hugepage_subpool *)page_private(page);
1214 bool restore_reserve;
1216 set_page_private(page, 0);
1217 page->mapping = NULL;
1218 BUG_ON(page_count(page));
1219 BUG_ON(page_mapcount(page));
1220 restore_reserve = PagePrivate(page);
1221 ClearPagePrivate(page);
1224 * A return code of zero implies that the subpool will be under its
1225 * minimum size if the reservation is not restored after page is free.
1226 * Therefore, force restore_reserve operation.
1228 if (hugepage_subpool_put_pages(spool, 1) == 0)
1229 restore_reserve = true;
1231 spin_lock(&hugetlb_lock);
1232 clear_page_huge_active(page);
1233 hugetlb_cgroup_uncharge_page(hstate_index(h),
1234 pages_per_huge_page(h), page);
1235 if (restore_reserve)
1236 h->resv_huge_pages++;
1238 if (h->surplus_huge_pages_node[nid]) {
1239 /* remove the page from active list */
1240 list_del(&page->lru);
1241 update_and_free_page(h, page);
1242 h->surplus_huge_pages--;
1243 h->surplus_huge_pages_node[nid]--;
1245 arch_clear_hugepage_flags(page);
1246 enqueue_huge_page(h, page);
1248 spin_unlock(&hugetlb_lock);
1251 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1253 INIT_LIST_HEAD(&page->lru);
1254 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1255 spin_lock(&hugetlb_lock);
1256 set_hugetlb_cgroup(page, NULL);
1258 h->nr_huge_pages_node[nid]++;
1259 spin_unlock(&hugetlb_lock);
1260 put_page(page); /* free it into the hugepage allocator */
1263 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1266 int nr_pages = 1 << order;
1267 struct page *p = page + 1;
1269 /* we rely on prep_new_huge_page to set the destructor */
1270 set_compound_order(page, order);
1271 __SetPageHead(page);
1272 __ClearPageReserved(page);
1273 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1275 * For gigantic hugepages allocated through bootmem at
1276 * boot, it's safer to be consistent with the not-gigantic
1277 * hugepages and clear the PG_reserved bit from all tail pages
1278 * too. Otherwse drivers using get_user_pages() to access tail
1279 * pages may get the reference counting wrong if they see
1280 * PG_reserved set on a tail page (despite the head page not
1281 * having PG_reserved set). Enforcing this consistency between
1282 * head and tail pages allows drivers to optimize away a check
1283 * on the head page when they need know if put_page() is needed
1284 * after get_user_pages().
1286 __ClearPageReserved(p);
1287 set_page_count(p, 0);
1288 set_compound_head(p, page);
1293 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1294 * transparent huge pages. See the PageTransHuge() documentation for more
1297 int PageHuge(struct page *page)
1299 if (!PageCompound(page))
1302 page = compound_head(page);
1303 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1305 EXPORT_SYMBOL_GPL(PageHuge);
1308 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1309 * normal or transparent huge pages.
1311 int PageHeadHuge(struct page *page_head)
1313 if (!PageHead(page_head))
1316 return get_compound_page_dtor(page_head) == free_huge_page;
1319 pgoff_t __basepage_index(struct page *page)
1321 struct page *page_head = compound_head(page);
1322 pgoff_t index = page_index(page_head);
1323 unsigned long compound_idx;
1325 if (!PageHuge(page_head))
1326 return page_index(page);
1328 if (compound_order(page_head) >= MAX_ORDER)
1329 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1331 compound_idx = page - page_head;
1333 return (index << compound_order(page_head)) + compound_idx;
1336 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1340 page = __alloc_pages_node(nid,
1341 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1342 __GFP_REPEAT|__GFP_NOWARN,
1343 huge_page_order(h));
1345 prep_new_huge_page(h, page, nid);
1351 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1357 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1358 page = alloc_fresh_huge_page_node(h, node);
1366 count_vm_event(HTLB_BUDDY_PGALLOC);
1368 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1374 * Free huge page from pool from next node to free.
1375 * Attempt to keep persistent huge pages more or less
1376 * balanced over allowed nodes.
1377 * Called with hugetlb_lock locked.
1379 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1385 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1387 * If we're returning unused surplus pages, only examine
1388 * nodes with surplus pages.
1390 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1391 !list_empty(&h->hugepage_freelists[node])) {
1393 list_entry(h->hugepage_freelists[node].next,
1395 list_del(&page->lru);
1396 h->free_huge_pages--;
1397 h->free_huge_pages_node[node]--;
1399 h->surplus_huge_pages--;
1400 h->surplus_huge_pages_node[node]--;
1402 update_and_free_page(h, page);
1412 * Dissolve a given free hugepage into free buddy pages. This function does
1413 * nothing for in-use (including surplus) hugepages.
1415 static void dissolve_free_huge_page(struct page *page)
1417 spin_lock(&hugetlb_lock);
1418 if (PageHuge(page) && !page_count(page)) {
1419 struct page *head = compound_head(page);
1420 struct hstate *h = page_hstate(head);
1421 int nid = page_to_nid(head);
1422 list_del(&head->lru);
1423 h->free_huge_pages--;
1424 h->free_huge_pages_node[nid]--;
1425 update_and_free_page(h, head);
1427 spin_unlock(&hugetlb_lock);
1431 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1432 * make specified memory blocks removable from the system.
1433 * Note that this will dissolve a free gigantic hugepage completely, if any
1434 * part of it lies within the given range.
1436 void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1440 if (!hugepages_supported())
1443 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
1444 dissolve_free_huge_page(pfn_to_page(pfn));
1448 * There are 3 ways this can get called:
1449 * 1. With vma+addr: we use the VMA's memory policy
1450 * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge
1451 * page from any node, and let the buddy allocator itself figure
1453 * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page
1454 * strictly from 'nid'
1456 static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1457 struct vm_area_struct *vma, unsigned long addr, int nid)
1459 int order = huge_page_order(h);
1460 gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN;
1461 unsigned int cpuset_mems_cookie;
1464 * We need a VMA to get a memory policy. If we do not
1465 * have one, we use the 'nid' argument.
1467 * The mempolicy stuff below has some non-inlined bits
1468 * and calls ->vm_ops. That makes it hard to optimize at
1469 * compile-time, even when NUMA is off and it does
1470 * nothing. This helps the compiler optimize it out.
1472 if (!IS_ENABLED(CONFIG_NUMA) || !vma) {
1474 * If a specific node is requested, make sure to
1475 * get memory from there, but only when a node
1476 * is explicitly specified.
1478 if (nid != NUMA_NO_NODE)
1479 gfp |= __GFP_THISNODE;
1481 * Make sure to call something that can handle
1484 return alloc_pages_node(nid, gfp, order);
1488 * OK, so we have a VMA. Fetch the mempolicy and try to
1489 * allocate a huge page with it. We will only reach this
1490 * when CONFIG_NUMA=y.
1494 struct mempolicy *mpol;
1495 struct zonelist *zl;
1496 nodemask_t *nodemask;
1498 cpuset_mems_cookie = read_mems_allowed_begin();
1499 zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask);
1500 mpol_cond_put(mpol);
1501 page = __alloc_pages_nodemask(gfp, order, zl, nodemask);
1504 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1510 * There are two ways to allocate a huge page:
1511 * 1. When you have a VMA and an address (like a fault)
1512 * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1514 * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in
1515 * this case which signifies that the allocation should be done with
1516 * respect for the VMA's memory policy.
1518 * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1519 * implies that memory policies will not be taken in to account.
1521 static struct page *__alloc_buddy_huge_page(struct hstate *h,
1522 struct vm_area_struct *vma, unsigned long addr, int nid)
1527 if (hstate_is_gigantic(h))
1531 * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1532 * This makes sure the caller is picking _one_ of the modes with which
1533 * we can call this function, not both.
1535 if (vma || (addr != -1)) {
1536 VM_WARN_ON_ONCE(addr == -1);
1537 VM_WARN_ON_ONCE(nid != NUMA_NO_NODE);
1540 * Assume we will successfully allocate the surplus page to
1541 * prevent racing processes from causing the surplus to exceed
1544 * This however introduces a different race, where a process B
1545 * tries to grow the static hugepage pool while alloc_pages() is
1546 * called by process A. B will only examine the per-node
1547 * counters in determining if surplus huge pages can be
1548 * converted to normal huge pages in adjust_pool_surplus(). A
1549 * won't be able to increment the per-node counter, until the
1550 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1551 * no more huge pages can be converted from surplus to normal
1552 * state (and doesn't try to convert again). Thus, we have a
1553 * case where a surplus huge page exists, the pool is grown, and
1554 * the surplus huge page still exists after, even though it
1555 * should just have been converted to a normal huge page. This
1556 * does not leak memory, though, as the hugepage will be freed
1557 * once it is out of use. It also does not allow the counters to
1558 * go out of whack in adjust_pool_surplus() as we don't modify
1559 * the node values until we've gotten the hugepage and only the
1560 * per-node value is checked there.
1562 spin_lock(&hugetlb_lock);
1563 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1564 spin_unlock(&hugetlb_lock);
1568 h->surplus_huge_pages++;
1570 spin_unlock(&hugetlb_lock);
1572 page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid);
1574 spin_lock(&hugetlb_lock);
1576 INIT_LIST_HEAD(&page->lru);
1577 r_nid = page_to_nid(page);
1578 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1579 set_hugetlb_cgroup(page, NULL);
1581 * We incremented the global counters already
1583 h->nr_huge_pages_node[r_nid]++;
1584 h->surplus_huge_pages_node[r_nid]++;
1585 __count_vm_event(HTLB_BUDDY_PGALLOC);
1588 h->surplus_huge_pages--;
1589 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1591 spin_unlock(&hugetlb_lock);
1597 * Allocate a huge page from 'nid'. Note, 'nid' may be
1598 * NUMA_NO_NODE, which means that it may be allocated
1602 struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid)
1604 unsigned long addr = -1;
1606 return __alloc_buddy_huge_page(h, NULL, addr, nid);
1610 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1613 struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1614 struct vm_area_struct *vma, unsigned long addr)
1616 return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE);
1620 * This allocation function is useful in the context where vma is irrelevant.
1621 * E.g. soft-offlining uses this function because it only cares physical
1622 * address of error page.
1624 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1626 struct page *page = NULL;
1628 spin_lock(&hugetlb_lock);
1629 if (h->free_huge_pages - h->resv_huge_pages > 0)
1630 page = dequeue_huge_page_node(h, nid);
1631 spin_unlock(&hugetlb_lock);
1634 page = __alloc_buddy_huge_page_no_mpol(h, nid);
1640 * Increase the hugetlb pool such that it can accommodate a reservation
1643 static int gather_surplus_pages(struct hstate *h, int delta)
1645 struct list_head surplus_list;
1646 struct page *page, *tmp;
1648 int needed, allocated;
1649 bool alloc_ok = true;
1651 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1653 h->resv_huge_pages += delta;
1658 INIT_LIST_HEAD(&surplus_list);
1662 spin_unlock(&hugetlb_lock);
1663 for (i = 0; i < needed; i++) {
1664 page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE);
1669 list_add(&page->lru, &surplus_list);
1674 * After retaking hugetlb_lock, we need to recalculate 'needed'
1675 * because either resv_huge_pages or free_huge_pages may have changed.
1677 spin_lock(&hugetlb_lock);
1678 needed = (h->resv_huge_pages + delta) -
1679 (h->free_huge_pages + allocated);
1684 * We were not able to allocate enough pages to
1685 * satisfy the entire reservation so we free what
1686 * we've allocated so far.
1691 * The surplus_list now contains _at_least_ the number of extra pages
1692 * needed to accommodate the reservation. Add the appropriate number
1693 * of pages to the hugetlb pool and free the extras back to the buddy
1694 * allocator. Commit the entire reservation here to prevent another
1695 * process from stealing the pages as they are added to the pool but
1696 * before they are reserved.
1698 needed += allocated;
1699 h->resv_huge_pages += delta;
1702 /* Free the needed pages to the hugetlb pool */
1703 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1707 * This page is now managed by the hugetlb allocator and has
1708 * no users -- drop the buddy allocator's reference.
1710 put_page_testzero(page);
1711 VM_BUG_ON_PAGE(page_count(page), page);
1712 enqueue_huge_page(h, page);
1715 spin_unlock(&hugetlb_lock);
1717 /* Free unnecessary surplus pages to the buddy allocator */
1718 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1720 spin_lock(&hugetlb_lock);
1726 * This routine has two main purposes:
1727 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1728 * in unused_resv_pages. This corresponds to the prior adjustments made
1729 * to the associated reservation map.
1730 * 2) Free any unused surplus pages that may have been allocated to satisfy
1731 * the reservation. As many as unused_resv_pages may be freed.
1733 * Called with hugetlb_lock held. However, the lock could be dropped (and
1734 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1735 * we must make sure nobody else can claim pages we are in the process of
1736 * freeing. Do this by ensuring resv_huge_page always is greater than the
1737 * number of huge pages we plan to free when dropping the lock.
1739 static void return_unused_surplus_pages(struct hstate *h,
1740 unsigned long unused_resv_pages)
1742 unsigned long nr_pages;
1744 /* Cannot return gigantic pages currently */
1745 if (hstate_is_gigantic(h))
1749 * Part (or even all) of the reservation could have been backed
1750 * by pre-allocated pages. Only free surplus pages.
1752 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1755 * We want to release as many surplus pages as possible, spread
1756 * evenly across all nodes with memory. Iterate across these nodes
1757 * until we can no longer free unreserved surplus pages. This occurs
1758 * when the nodes with surplus pages have no free pages.
1759 * free_pool_huge_page() will balance the the freed pages across the
1760 * on-line nodes with memory and will handle the hstate accounting.
1762 * Note that we decrement resv_huge_pages as we free the pages. If
1763 * we drop the lock, resv_huge_pages will still be sufficiently large
1764 * to cover subsequent pages we may free.
1766 while (nr_pages--) {
1767 h->resv_huge_pages--;
1768 unused_resv_pages--;
1769 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1771 cond_resched_lock(&hugetlb_lock);
1775 /* Fully uncommit the reservation */
1776 h->resv_huge_pages -= unused_resv_pages;
1781 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1782 * are used by the huge page allocation routines to manage reservations.
1784 * vma_needs_reservation is called to determine if the huge page at addr
1785 * within the vma has an associated reservation. If a reservation is
1786 * needed, the value 1 is returned. The caller is then responsible for
1787 * managing the global reservation and subpool usage counts. After
1788 * the huge page has been allocated, vma_commit_reservation is called
1789 * to add the page to the reservation map. If the page allocation fails,
1790 * the reservation must be ended instead of committed. vma_end_reservation
1791 * is called in such cases.
1793 * In the normal case, vma_commit_reservation returns the same value
1794 * as the preceding vma_needs_reservation call. The only time this
1795 * is not the case is if a reserve map was changed between calls. It
1796 * is the responsibility of the caller to notice the difference and
1797 * take appropriate action.
1799 enum vma_resv_mode {
1804 static long __vma_reservation_common(struct hstate *h,
1805 struct vm_area_struct *vma, unsigned long addr,
1806 enum vma_resv_mode mode)
1808 struct resv_map *resv;
1812 resv = vma_resv_map(vma);
1816 idx = vma_hugecache_offset(h, vma, addr);
1818 case VMA_NEEDS_RESV:
1819 ret = region_chg(resv, idx, idx + 1);
1821 case VMA_COMMIT_RESV:
1822 ret = region_add(resv, idx, idx + 1);
1825 region_abort(resv, idx, idx + 1);
1832 if (vma->vm_flags & VM_MAYSHARE)
1835 return ret < 0 ? ret : 0;
1838 static long vma_needs_reservation(struct hstate *h,
1839 struct vm_area_struct *vma, unsigned long addr)
1841 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1844 static long vma_commit_reservation(struct hstate *h,
1845 struct vm_area_struct *vma, unsigned long addr)
1847 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1850 static void vma_end_reservation(struct hstate *h,
1851 struct vm_area_struct *vma, unsigned long addr)
1853 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1856 struct page *alloc_huge_page(struct vm_area_struct *vma,
1857 unsigned long addr, int avoid_reserve)
1859 struct hugepage_subpool *spool = subpool_vma(vma);
1860 struct hstate *h = hstate_vma(vma);
1862 long map_chg, map_commit;
1865 struct hugetlb_cgroup *h_cg;
1867 idx = hstate_index(h);
1869 * Examine the region/reserve map to determine if the process
1870 * has a reservation for the page to be allocated. A return
1871 * code of zero indicates a reservation exists (no change).
1873 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
1875 return ERR_PTR(-ENOMEM);
1878 * Processes that did not create the mapping will have no
1879 * reserves as indicated by the region/reserve map. Check
1880 * that the allocation will not exceed the subpool limit.
1881 * Allocations for MAP_NORESERVE mappings also need to be
1882 * checked against any subpool limit.
1884 if (map_chg || avoid_reserve) {
1885 gbl_chg = hugepage_subpool_get_pages(spool, 1);
1887 vma_end_reservation(h, vma, addr);
1888 return ERR_PTR(-ENOSPC);
1892 * Even though there was no reservation in the region/reserve
1893 * map, there could be reservations associated with the
1894 * subpool that can be used. This would be indicated if the
1895 * return value of hugepage_subpool_get_pages() is zero.
1896 * However, if avoid_reserve is specified we still avoid even
1897 * the subpool reservations.
1903 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1905 goto out_subpool_put;
1907 spin_lock(&hugetlb_lock);
1909 * glb_chg is passed to indicate whether or not a page must be taken
1910 * from the global free pool (global change). gbl_chg == 0 indicates
1911 * a reservation exists for the allocation.
1913 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
1915 spin_unlock(&hugetlb_lock);
1916 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
1918 goto out_uncharge_cgroup;
1919 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
1920 SetPagePrivate(page);
1921 h->resv_huge_pages--;
1923 spin_lock(&hugetlb_lock);
1924 list_move(&page->lru, &h->hugepage_activelist);
1927 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1928 spin_unlock(&hugetlb_lock);
1930 set_page_private(page, (unsigned long)spool);
1932 map_commit = vma_commit_reservation(h, vma, addr);
1933 if (unlikely(map_chg > map_commit)) {
1935 * The page was added to the reservation map between
1936 * vma_needs_reservation and vma_commit_reservation.
1937 * This indicates a race with hugetlb_reserve_pages.
1938 * Adjust for the subpool count incremented above AND
1939 * in hugetlb_reserve_pages for the same page. Also,
1940 * the reservation count added in hugetlb_reserve_pages
1941 * no longer applies.
1945 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
1946 hugetlb_acct_memory(h, -rsv_adjust);
1950 out_uncharge_cgroup:
1951 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1953 if (map_chg || avoid_reserve)
1954 hugepage_subpool_put_pages(spool, 1);
1955 vma_end_reservation(h, vma, addr);
1956 return ERR_PTR(-ENOSPC);
1960 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1961 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1962 * where no ERR_VALUE is expected to be returned.
1964 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1965 unsigned long addr, int avoid_reserve)
1967 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1973 int __weak alloc_bootmem_huge_page(struct hstate *h)
1975 struct huge_bootmem_page *m;
1978 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1981 addr = memblock_virt_alloc_try_nid_nopanic(
1982 huge_page_size(h), huge_page_size(h),
1983 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1986 * Use the beginning of the huge page to store the
1987 * huge_bootmem_page struct (until gather_bootmem
1988 * puts them into the mem_map).
1997 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
1998 /* Put them into a private list first because mem_map is not up yet */
1999 list_add(&m->list, &huge_boot_pages);
2004 static void __init prep_compound_huge_page(struct page *page,
2007 if (unlikely(order > (MAX_ORDER - 1)))
2008 prep_compound_gigantic_page(page, order);
2010 prep_compound_page(page, order);
2013 /* Put bootmem huge pages into the standard lists after mem_map is up */
2014 static void __init gather_bootmem_prealloc(void)
2016 struct huge_bootmem_page *m;
2018 list_for_each_entry(m, &huge_boot_pages, list) {
2019 struct hstate *h = m->hstate;
2022 #ifdef CONFIG_HIGHMEM
2023 page = pfn_to_page(m->phys >> PAGE_SHIFT);
2024 memblock_free_late(__pa(m),
2025 sizeof(struct huge_bootmem_page));
2027 page = virt_to_page(m);
2029 WARN_ON(page_count(page) != 1);
2030 prep_compound_huge_page(page, h->order);
2031 WARN_ON(PageReserved(page));
2032 prep_new_huge_page(h, page, page_to_nid(page));
2034 * If we had gigantic hugepages allocated at boot time, we need
2035 * to restore the 'stolen' pages to totalram_pages in order to
2036 * fix confusing memory reports from free(1) and another
2037 * side-effects, like CommitLimit going negative.
2039 if (hstate_is_gigantic(h))
2040 adjust_managed_page_count(page, 1 << h->order);
2045 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2049 for (i = 0; i < h->max_huge_pages; ++i) {
2050 if (hstate_is_gigantic(h)) {
2051 if (!alloc_bootmem_huge_page(h))
2053 } else if (!alloc_fresh_huge_page(h,
2054 &node_states[N_MEMORY]))
2057 h->max_huge_pages = i;
2060 static void __init hugetlb_init_hstates(void)
2064 for_each_hstate(h) {
2065 if (minimum_order > huge_page_order(h))
2066 minimum_order = huge_page_order(h);
2068 /* oversize hugepages were init'ed in early boot */
2069 if (!hstate_is_gigantic(h))
2070 hugetlb_hstate_alloc_pages(h);
2072 VM_BUG_ON(minimum_order == UINT_MAX);
2075 static char * __init memfmt(char *buf, unsigned long n)
2077 if (n >= (1UL << 30))
2078 sprintf(buf, "%lu GB", n >> 30);
2079 else if (n >= (1UL << 20))
2080 sprintf(buf, "%lu MB", n >> 20);
2082 sprintf(buf, "%lu KB", n >> 10);
2086 static void __init report_hugepages(void)
2090 for_each_hstate(h) {
2092 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2093 memfmt(buf, huge_page_size(h)),
2094 h->free_huge_pages);
2098 #ifdef CONFIG_HIGHMEM
2099 static void try_to_free_low(struct hstate *h, unsigned long count,
2100 nodemask_t *nodes_allowed)
2104 if (hstate_is_gigantic(h))
2107 for_each_node_mask(i, *nodes_allowed) {
2108 struct page *page, *next;
2109 struct list_head *freel = &h->hugepage_freelists[i];
2110 list_for_each_entry_safe(page, next, freel, lru) {
2111 if (count >= h->nr_huge_pages)
2113 if (PageHighMem(page))
2115 list_del(&page->lru);
2116 update_and_free_page(h, page);
2117 h->free_huge_pages--;
2118 h->free_huge_pages_node[page_to_nid(page)]--;
2123 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2124 nodemask_t *nodes_allowed)
2130 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2131 * balanced by operating on them in a round-robin fashion.
2132 * Returns 1 if an adjustment was made.
2134 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2139 VM_BUG_ON(delta != -1 && delta != 1);
2142 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2143 if (h->surplus_huge_pages_node[node])
2147 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2148 if (h->surplus_huge_pages_node[node] <
2149 h->nr_huge_pages_node[node])
2156 h->surplus_huge_pages += delta;
2157 h->surplus_huge_pages_node[node] += delta;
2161 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2162 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2163 nodemask_t *nodes_allowed)
2165 unsigned long min_count, ret;
2167 if (hstate_is_gigantic(h) && !gigantic_page_supported())
2168 return h->max_huge_pages;
2171 * Increase the pool size
2172 * First take pages out of surplus state. Then make up the
2173 * remaining difference by allocating fresh huge pages.
2175 * We might race with __alloc_buddy_huge_page() here and be unable
2176 * to convert a surplus huge page to a normal huge page. That is
2177 * not critical, though, it just means the overall size of the
2178 * pool might be one hugepage larger than it needs to be, but
2179 * within all the constraints specified by the sysctls.
2181 spin_lock(&hugetlb_lock);
2182 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2183 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2187 while (count > persistent_huge_pages(h)) {
2189 * If this allocation races such that we no longer need the
2190 * page, free_huge_page will handle it by freeing the page
2191 * and reducing the surplus.
2193 spin_unlock(&hugetlb_lock);
2195 /* yield cpu to avoid soft lockup */
2198 if (hstate_is_gigantic(h))
2199 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2201 ret = alloc_fresh_huge_page(h, nodes_allowed);
2202 spin_lock(&hugetlb_lock);
2206 /* Bail for signals. Probably ctrl-c from user */
2207 if (signal_pending(current))
2212 * Decrease the pool size
2213 * First return free pages to the buddy allocator (being careful
2214 * to keep enough around to satisfy reservations). Then place
2215 * pages into surplus state as needed so the pool will shrink
2216 * to the desired size as pages become free.
2218 * By placing pages into the surplus state independent of the
2219 * overcommit value, we are allowing the surplus pool size to
2220 * exceed overcommit. There are few sane options here. Since
2221 * __alloc_buddy_huge_page() is checking the global counter,
2222 * though, we'll note that we're not allowed to exceed surplus
2223 * and won't grow the pool anywhere else. Not until one of the
2224 * sysctls are changed, or the surplus pages go out of use.
2226 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2227 min_count = max(count, min_count);
2228 try_to_free_low(h, min_count, nodes_allowed);
2229 while (min_count < persistent_huge_pages(h)) {
2230 if (!free_pool_huge_page(h, nodes_allowed, 0))
2232 cond_resched_lock(&hugetlb_lock);
2234 while (count < persistent_huge_pages(h)) {
2235 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2239 ret = persistent_huge_pages(h);
2240 spin_unlock(&hugetlb_lock);
2244 #define HSTATE_ATTR_RO(_name) \
2245 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2247 #define HSTATE_ATTR(_name) \
2248 static struct kobj_attribute _name##_attr = \
2249 __ATTR(_name, 0644, _name##_show, _name##_store)
2251 static struct kobject *hugepages_kobj;
2252 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2254 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2256 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2260 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2261 if (hstate_kobjs[i] == kobj) {
2263 *nidp = NUMA_NO_NODE;
2267 return kobj_to_node_hstate(kobj, nidp);
2270 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2271 struct kobj_attribute *attr, char *buf)
2274 unsigned long nr_huge_pages;
2277 h = kobj_to_hstate(kobj, &nid);
2278 if (nid == NUMA_NO_NODE)
2279 nr_huge_pages = h->nr_huge_pages;
2281 nr_huge_pages = h->nr_huge_pages_node[nid];
2283 return sprintf(buf, "%lu\n", nr_huge_pages);
2286 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2287 struct hstate *h, int nid,
2288 unsigned long count, size_t len)
2291 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2293 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2298 if (nid == NUMA_NO_NODE) {
2300 * global hstate attribute
2302 if (!(obey_mempolicy &&
2303 init_nodemask_of_mempolicy(nodes_allowed))) {
2304 NODEMASK_FREE(nodes_allowed);
2305 nodes_allowed = &node_states[N_MEMORY];
2307 } else if (nodes_allowed) {
2309 * per node hstate attribute: adjust count to global,
2310 * but restrict alloc/free to the specified node.
2312 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2313 init_nodemask_of_node(nodes_allowed, nid);
2315 nodes_allowed = &node_states[N_MEMORY];
2317 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2319 if (nodes_allowed != &node_states[N_MEMORY])
2320 NODEMASK_FREE(nodes_allowed);
2324 NODEMASK_FREE(nodes_allowed);
2328 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2329 struct kobject *kobj, const char *buf,
2333 unsigned long count;
2337 err = kstrtoul(buf, 10, &count);
2341 h = kobj_to_hstate(kobj, &nid);
2342 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2345 static ssize_t nr_hugepages_show(struct kobject *kobj,
2346 struct kobj_attribute *attr, char *buf)
2348 return nr_hugepages_show_common(kobj, attr, buf);
2351 static ssize_t nr_hugepages_store(struct kobject *kobj,
2352 struct kobj_attribute *attr, const char *buf, size_t len)
2354 return nr_hugepages_store_common(false, kobj, buf, len);
2356 HSTATE_ATTR(nr_hugepages);
2361 * hstate attribute for optionally mempolicy-based constraint on persistent
2362 * huge page alloc/free.
2364 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2365 struct kobj_attribute *attr, char *buf)
2367 return nr_hugepages_show_common(kobj, attr, buf);
2370 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2371 struct kobj_attribute *attr, const char *buf, size_t len)
2373 return nr_hugepages_store_common(true, kobj, buf, len);
2375 HSTATE_ATTR(nr_hugepages_mempolicy);
2379 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2380 struct kobj_attribute *attr, char *buf)
2382 struct hstate *h = kobj_to_hstate(kobj, NULL);
2383 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2386 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2387 struct kobj_attribute *attr, const char *buf, size_t count)
2390 unsigned long input;
2391 struct hstate *h = kobj_to_hstate(kobj, NULL);
2393 if (hstate_is_gigantic(h))
2396 err = kstrtoul(buf, 10, &input);
2400 spin_lock(&hugetlb_lock);
2401 h->nr_overcommit_huge_pages = input;
2402 spin_unlock(&hugetlb_lock);
2406 HSTATE_ATTR(nr_overcommit_hugepages);
2408 static ssize_t free_hugepages_show(struct kobject *kobj,
2409 struct kobj_attribute *attr, char *buf)
2412 unsigned long free_huge_pages;
2415 h = kobj_to_hstate(kobj, &nid);
2416 if (nid == NUMA_NO_NODE)
2417 free_huge_pages = h->free_huge_pages;
2419 free_huge_pages = h->free_huge_pages_node[nid];
2421 return sprintf(buf, "%lu\n", free_huge_pages);
2423 HSTATE_ATTR_RO(free_hugepages);
2425 static ssize_t resv_hugepages_show(struct kobject *kobj,
2426 struct kobj_attribute *attr, char *buf)
2428 struct hstate *h = kobj_to_hstate(kobj, NULL);
2429 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2431 HSTATE_ATTR_RO(resv_hugepages);
2433 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2434 struct kobj_attribute *attr, char *buf)
2437 unsigned long surplus_huge_pages;
2440 h = kobj_to_hstate(kobj, &nid);
2441 if (nid == NUMA_NO_NODE)
2442 surplus_huge_pages = h->surplus_huge_pages;
2444 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2446 return sprintf(buf, "%lu\n", surplus_huge_pages);
2448 HSTATE_ATTR_RO(surplus_hugepages);
2450 static struct attribute *hstate_attrs[] = {
2451 &nr_hugepages_attr.attr,
2452 &nr_overcommit_hugepages_attr.attr,
2453 &free_hugepages_attr.attr,
2454 &resv_hugepages_attr.attr,
2455 &surplus_hugepages_attr.attr,
2457 &nr_hugepages_mempolicy_attr.attr,
2462 static struct attribute_group hstate_attr_group = {
2463 .attrs = hstate_attrs,
2466 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2467 struct kobject **hstate_kobjs,
2468 struct attribute_group *hstate_attr_group)
2471 int hi = hstate_index(h);
2473 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2474 if (!hstate_kobjs[hi])
2477 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2479 kobject_put(hstate_kobjs[hi]);
2484 static void __init hugetlb_sysfs_init(void)
2489 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2490 if (!hugepages_kobj)
2493 for_each_hstate(h) {
2494 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2495 hstate_kobjs, &hstate_attr_group);
2497 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2504 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2505 * with node devices in node_devices[] using a parallel array. The array
2506 * index of a node device or _hstate == node id.
2507 * This is here to avoid any static dependency of the node device driver, in
2508 * the base kernel, on the hugetlb module.
2510 struct node_hstate {
2511 struct kobject *hugepages_kobj;
2512 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2514 static struct node_hstate node_hstates[MAX_NUMNODES];
2517 * A subset of global hstate attributes for node devices
2519 static struct attribute *per_node_hstate_attrs[] = {
2520 &nr_hugepages_attr.attr,
2521 &free_hugepages_attr.attr,
2522 &surplus_hugepages_attr.attr,
2526 static struct attribute_group per_node_hstate_attr_group = {
2527 .attrs = per_node_hstate_attrs,
2531 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2532 * Returns node id via non-NULL nidp.
2534 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2538 for (nid = 0; nid < nr_node_ids; nid++) {
2539 struct node_hstate *nhs = &node_hstates[nid];
2541 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2542 if (nhs->hstate_kobjs[i] == kobj) {
2554 * Unregister hstate attributes from a single node device.
2555 * No-op if no hstate attributes attached.
2557 static void hugetlb_unregister_node(struct node *node)
2560 struct node_hstate *nhs = &node_hstates[node->dev.id];
2562 if (!nhs->hugepages_kobj)
2563 return; /* no hstate attributes */
2565 for_each_hstate(h) {
2566 int idx = hstate_index(h);
2567 if (nhs->hstate_kobjs[idx]) {
2568 kobject_put(nhs->hstate_kobjs[idx]);
2569 nhs->hstate_kobjs[idx] = NULL;
2573 kobject_put(nhs->hugepages_kobj);
2574 nhs->hugepages_kobj = NULL;
2578 * hugetlb module exit: unregister hstate attributes from node devices
2581 static void hugetlb_unregister_all_nodes(void)
2586 * disable node device registrations.
2588 register_hugetlbfs_with_node(NULL, NULL);
2591 * remove hstate attributes from any nodes that have them.
2593 for (nid = 0; nid < nr_node_ids; nid++)
2594 hugetlb_unregister_node(node_devices[nid]);
2598 * Register hstate attributes for a single node device.
2599 * No-op if attributes already registered.
2601 static void hugetlb_register_node(struct node *node)
2604 struct node_hstate *nhs = &node_hstates[node->dev.id];
2607 if (nhs->hugepages_kobj)
2608 return; /* already allocated */
2610 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2612 if (!nhs->hugepages_kobj)
2615 for_each_hstate(h) {
2616 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2618 &per_node_hstate_attr_group);
2620 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2621 h->name, node->dev.id);
2622 hugetlb_unregister_node(node);
2629 * hugetlb init time: register hstate attributes for all registered node
2630 * devices of nodes that have memory. All on-line nodes should have
2631 * registered their associated device by this time.
2633 static void __init hugetlb_register_all_nodes(void)
2637 for_each_node_state(nid, N_MEMORY) {
2638 struct node *node = node_devices[nid];
2639 if (node->dev.id == nid)
2640 hugetlb_register_node(node);
2644 * Let the node device driver know we're here so it can
2645 * [un]register hstate attributes on node hotplug.
2647 register_hugetlbfs_with_node(hugetlb_register_node,
2648 hugetlb_unregister_node);
2650 #else /* !CONFIG_NUMA */
2652 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2660 static void hugetlb_unregister_all_nodes(void) { }
2662 static void hugetlb_register_all_nodes(void) { }
2666 static void __exit hugetlb_exit(void)
2670 hugetlb_unregister_all_nodes();
2672 for_each_hstate(h) {
2673 kobject_put(hstate_kobjs[hstate_index(h)]);
2676 kobject_put(hugepages_kobj);
2677 kfree(hugetlb_fault_mutex_table);
2679 module_exit(hugetlb_exit);
2681 static int __init hugetlb_init(void)
2685 if (!hugepages_supported())
2688 if (!size_to_hstate(default_hstate_size)) {
2689 default_hstate_size = HPAGE_SIZE;
2690 if (!size_to_hstate(default_hstate_size))
2691 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2693 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2694 if (default_hstate_max_huge_pages)
2695 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2697 hugetlb_init_hstates();
2698 gather_bootmem_prealloc();
2701 hugetlb_sysfs_init();
2702 hugetlb_register_all_nodes();
2703 hugetlb_cgroup_file_init();
2706 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2708 num_fault_mutexes = 1;
2710 hugetlb_fault_mutex_table =
2711 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2712 BUG_ON(!hugetlb_fault_mutex_table);
2714 for (i = 0; i < num_fault_mutexes; i++)
2715 mutex_init(&hugetlb_fault_mutex_table[i]);
2718 module_init(hugetlb_init);
2720 /* Should be called on processing a hugepagesz=... option */
2721 void __init hugetlb_add_hstate(unsigned int order)
2726 if (size_to_hstate(PAGE_SIZE << order)) {
2727 pr_warning("hugepagesz= specified twice, ignoring\n");
2730 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2732 h = &hstates[hugetlb_max_hstate++];
2734 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2735 h->nr_huge_pages = 0;
2736 h->free_huge_pages = 0;
2737 for (i = 0; i < MAX_NUMNODES; ++i)
2738 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2739 INIT_LIST_HEAD(&h->hugepage_activelist);
2740 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2741 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2742 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2743 huge_page_size(h)/1024);
2748 static int __init hugetlb_nrpages_setup(char *s)
2751 static unsigned long *last_mhp;
2754 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2755 * so this hugepages= parameter goes to the "default hstate".
2757 if (!hugetlb_max_hstate)
2758 mhp = &default_hstate_max_huge_pages;
2760 mhp = &parsed_hstate->max_huge_pages;
2762 if (mhp == last_mhp) {
2763 pr_warning("hugepages= specified twice without "
2764 "interleaving hugepagesz=, ignoring\n");
2768 if (sscanf(s, "%lu", mhp) <= 0)
2772 * Global state is always initialized later in hugetlb_init.
2773 * But we need to allocate >= MAX_ORDER hstates here early to still
2774 * use the bootmem allocator.
2776 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2777 hugetlb_hstate_alloc_pages(parsed_hstate);
2783 __setup("hugepages=", hugetlb_nrpages_setup);
2785 static int __init hugetlb_default_setup(char *s)
2787 default_hstate_size = memparse(s, &s);
2790 __setup("default_hugepagesz=", hugetlb_default_setup);
2792 static unsigned int cpuset_mems_nr(unsigned int *array)
2795 unsigned int nr = 0;
2797 for_each_node_mask(node, cpuset_current_mems_allowed)
2803 #ifdef CONFIG_SYSCTL
2804 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2805 struct ctl_table *table, int write,
2806 void __user *buffer, size_t *length, loff_t *ppos)
2808 struct hstate *h = &default_hstate;
2809 unsigned long tmp = h->max_huge_pages;
2812 if (!hugepages_supported())
2816 table->maxlen = sizeof(unsigned long);
2817 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2822 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2823 NUMA_NO_NODE, tmp, *length);
2828 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2829 void __user *buffer, size_t *length, loff_t *ppos)
2832 return hugetlb_sysctl_handler_common(false, table, write,
2833 buffer, length, ppos);
2837 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2838 void __user *buffer, size_t *length, loff_t *ppos)
2840 return hugetlb_sysctl_handler_common(true, table, write,
2841 buffer, length, ppos);
2843 #endif /* CONFIG_NUMA */
2845 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2846 void __user *buffer,
2847 size_t *length, loff_t *ppos)
2849 struct hstate *h = &default_hstate;
2853 if (!hugepages_supported())
2856 tmp = h->nr_overcommit_huge_pages;
2858 if (write && hstate_is_gigantic(h))
2862 table->maxlen = sizeof(unsigned long);
2863 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2868 spin_lock(&hugetlb_lock);
2869 h->nr_overcommit_huge_pages = tmp;
2870 spin_unlock(&hugetlb_lock);
2876 #endif /* CONFIG_SYSCTL */
2878 void hugetlb_report_meminfo(struct seq_file *m)
2880 struct hstate *h = &default_hstate;
2881 if (!hugepages_supported())
2884 "HugePages_Total: %5lu\n"
2885 "HugePages_Free: %5lu\n"
2886 "HugePages_Rsvd: %5lu\n"
2887 "HugePages_Surp: %5lu\n"
2888 "Hugepagesize: %8lu kB\n",
2892 h->surplus_huge_pages,
2893 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2896 int hugetlb_report_node_meminfo(int nid, char *buf)
2898 struct hstate *h = &default_hstate;
2899 if (!hugepages_supported())
2902 "Node %d HugePages_Total: %5u\n"
2903 "Node %d HugePages_Free: %5u\n"
2904 "Node %d HugePages_Surp: %5u\n",
2905 nid, h->nr_huge_pages_node[nid],
2906 nid, h->free_huge_pages_node[nid],
2907 nid, h->surplus_huge_pages_node[nid]);
2910 void hugetlb_show_meminfo(void)
2915 if (!hugepages_supported())
2918 for_each_node_state(nid, N_MEMORY)
2920 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2922 h->nr_huge_pages_node[nid],
2923 h->free_huge_pages_node[nid],
2924 h->surplus_huge_pages_node[nid],
2925 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2928 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
2930 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
2931 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
2934 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2935 unsigned long hugetlb_total_pages(void)
2938 unsigned long nr_total_pages = 0;
2941 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2942 return nr_total_pages;
2945 static int hugetlb_acct_memory(struct hstate *h, long delta)
2949 spin_lock(&hugetlb_lock);
2951 * When cpuset is configured, it breaks the strict hugetlb page
2952 * reservation as the accounting is done on a global variable. Such
2953 * reservation is completely rubbish in the presence of cpuset because
2954 * the reservation is not checked against page availability for the
2955 * current cpuset. Application can still potentially OOM'ed by kernel
2956 * with lack of free htlb page in cpuset that the task is in.
2957 * Attempt to enforce strict accounting with cpuset is almost
2958 * impossible (or too ugly) because cpuset is too fluid that
2959 * task or memory node can be dynamically moved between cpusets.
2961 * The change of semantics for shared hugetlb mapping with cpuset is
2962 * undesirable. However, in order to preserve some of the semantics,
2963 * we fall back to check against current free page availability as
2964 * a best attempt and hopefully to minimize the impact of changing
2965 * semantics that cpuset has.
2968 if (gather_surplus_pages(h, delta) < 0)
2971 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2972 return_unused_surplus_pages(h, delta);
2979 return_unused_surplus_pages(h, (unsigned long) -delta);
2982 spin_unlock(&hugetlb_lock);
2986 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2988 struct resv_map *resv = vma_resv_map(vma);
2991 * This new VMA should share its siblings reservation map if present.
2992 * The VMA will only ever have a valid reservation map pointer where
2993 * it is being copied for another still existing VMA. As that VMA
2994 * has a reference to the reservation map it cannot disappear until
2995 * after this open call completes. It is therefore safe to take a
2996 * new reference here without additional locking.
2998 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2999 kref_get(&resv->refs);
3002 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3004 struct hstate *h = hstate_vma(vma);
3005 struct resv_map *resv = vma_resv_map(vma);
3006 struct hugepage_subpool *spool = subpool_vma(vma);
3007 unsigned long reserve, start, end;
3010 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3013 start = vma_hugecache_offset(h, vma, vma->vm_start);
3014 end = vma_hugecache_offset(h, vma, vma->vm_end);
3016 reserve = (end - start) - region_count(resv, start, end);
3018 kref_put(&resv->refs, resv_map_release);
3022 * Decrement reserve counts. The global reserve count may be
3023 * adjusted if the subpool has a minimum size.
3025 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3026 hugetlb_acct_memory(h, -gbl_reserve);
3031 * We cannot handle pagefaults against hugetlb pages at all. They cause
3032 * handle_mm_fault() to try to instantiate regular-sized pages in the
3033 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3036 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3042 const struct vm_operations_struct hugetlb_vm_ops = {
3043 .fault = hugetlb_vm_op_fault,
3044 .open = hugetlb_vm_op_open,
3045 .close = hugetlb_vm_op_close,
3048 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3054 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3055 vma->vm_page_prot)));
3057 entry = huge_pte_wrprotect(mk_huge_pte(page,
3058 vma->vm_page_prot));
3060 entry = pte_mkyoung(entry);
3061 entry = pte_mkhuge(entry);
3062 entry = arch_make_huge_pte(entry, vma, page, writable);
3067 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3068 unsigned long address, pte_t *ptep)
3072 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3073 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3074 update_mmu_cache(vma, address, ptep);
3077 static int is_hugetlb_entry_migration(pte_t pte)
3081 if (huge_pte_none(pte) || pte_present(pte))
3083 swp = pte_to_swp_entry(pte);
3084 if (non_swap_entry(swp) && is_migration_entry(swp))
3090 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3094 if (huge_pte_none(pte) || pte_present(pte))
3096 swp = pte_to_swp_entry(pte);
3097 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3103 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3104 struct vm_area_struct *vma)
3106 pte_t *src_pte, *dst_pte, entry, dst_entry;
3107 struct page *ptepage;
3110 struct hstate *h = hstate_vma(vma);
3111 unsigned long sz = huge_page_size(h);
3112 unsigned long mmun_start; /* For mmu_notifiers */
3113 unsigned long mmun_end; /* For mmu_notifiers */
3116 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3118 mmun_start = vma->vm_start;
3119 mmun_end = vma->vm_end;
3121 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3123 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3124 spinlock_t *src_ptl, *dst_ptl;
3125 src_pte = huge_pte_offset(src, addr);
3128 dst_pte = huge_pte_alloc(dst, addr, sz);
3135 * If the pagetables are shared don't copy or take references.
3136 * dst_pte == src_pte is the common case of src/dest sharing.
3138 * However, src could have 'unshared' and dst shares with
3139 * another vma. If dst_pte !none, this implies sharing.
3140 * Check here before taking page table lock, and once again
3141 * after taking the lock below.
3143 dst_entry = huge_ptep_get(dst_pte);
3144 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3147 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3148 src_ptl = huge_pte_lockptr(h, src, src_pte);
3149 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3150 entry = huge_ptep_get(src_pte);
3151 dst_entry = huge_ptep_get(dst_pte);
3152 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3154 * Skip if src entry none. Also, skip in the
3155 * unlikely case dst entry !none as this implies
3156 * sharing with another vma.
3159 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3160 is_hugetlb_entry_hwpoisoned(entry))) {
3161 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3163 if (is_write_migration_entry(swp_entry) && cow) {
3165 * COW mappings require pages in both
3166 * parent and child to be set to read.
3168 make_migration_entry_read(&swp_entry);
3169 entry = swp_entry_to_pte(swp_entry);
3170 set_huge_pte_at(src, addr, src_pte, entry);
3172 set_huge_pte_at(dst, addr, dst_pte, entry);
3175 huge_ptep_set_wrprotect(src, addr, src_pte);
3176 mmu_notifier_invalidate_range(src, mmun_start,
3179 entry = huge_ptep_get(src_pte);
3180 ptepage = pte_page(entry);
3182 page_dup_rmap(ptepage);
3183 set_huge_pte_at(dst, addr, dst_pte, entry);
3184 hugetlb_count_add(pages_per_huge_page(h), dst);
3186 spin_unlock(src_ptl);
3187 spin_unlock(dst_ptl);
3191 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3196 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3197 unsigned long start, unsigned long end,
3198 struct page *ref_page)
3200 int force_flush = 0;
3201 struct mm_struct *mm = vma->vm_mm;
3202 unsigned long address;
3207 struct hstate *h = hstate_vma(vma);
3208 unsigned long sz = huge_page_size(h);
3209 const unsigned long mmun_start = start; /* For mmu_notifiers */
3210 const unsigned long mmun_end = end; /* For mmu_notifiers */
3212 WARN_ON(!is_vm_hugetlb_page(vma));
3213 BUG_ON(start & ~huge_page_mask(h));
3214 BUG_ON(end & ~huge_page_mask(h));
3216 tlb_start_vma(tlb, vma);
3217 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3220 for (; address < end; address += sz) {
3221 ptep = huge_pte_offset(mm, address);
3225 ptl = huge_pte_lock(h, mm, ptep);
3226 if (huge_pmd_unshare(mm, &address, ptep))
3229 pte = huge_ptep_get(ptep);
3230 if (huge_pte_none(pte))
3234 * Migrating hugepage or HWPoisoned hugepage is already
3235 * unmapped and its refcount is dropped, so just clear pte here.
3237 if (unlikely(!pte_present(pte))) {
3238 huge_pte_clear(mm, address, ptep);
3242 page = pte_page(pte);
3244 * If a reference page is supplied, it is because a specific
3245 * page is being unmapped, not a range. Ensure the page we
3246 * are about to unmap is the actual page of interest.
3249 if (page != ref_page)
3253 * Mark the VMA as having unmapped its page so that
3254 * future faults in this VMA will fail rather than
3255 * looking like data was lost
3257 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3260 pte = huge_ptep_get_and_clear(mm, address, ptep);
3261 tlb_remove_tlb_entry(tlb, ptep, address);
3262 if (huge_pte_dirty(pte))
3263 set_page_dirty(page);
3265 hugetlb_count_sub(pages_per_huge_page(h), mm);
3266 page_remove_rmap(page);
3267 force_flush = !__tlb_remove_page(tlb, page);
3273 /* Bail out after unmapping reference page if supplied */
3282 * mmu_gather ran out of room to batch pages, we break out of
3283 * the PTE lock to avoid doing the potential expensive TLB invalidate
3284 * and page-free while holding it.
3289 if (address < end && !ref_page)
3292 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3293 tlb_end_vma(tlb, vma);
3296 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3297 struct vm_area_struct *vma, unsigned long start,
3298 unsigned long end, struct page *ref_page)
3300 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3303 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3304 * test will fail on a vma being torn down, and not grab a page table
3305 * on its way out. We're lucky that the flag has such an appropriate
3306 * name, and can in fact be safely cleared here. We could clear it
3307 * before the __unmap_hugepage_range above, but all that's necessary
3308 * is to clear it before releasing the i_mmap_rwsem. This works
3309 * because in the context this is called, the VMA is about to be
3310 * destroyed and the i_mmap_rwsem is held.
3312 vma->vm_flags &= ~VM_MAYSHARE;
3315 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3316 unsigned long end, struct page *ref_page)
3318 struct mm_struct *mm;
3319 struct mmu_gather tlb;
3323 tlb_gather_mmu(&tlb, mm, start, end);
3324 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3325 tlb_finish_mmu(&tlb, start, end);
3329 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3330 * mappping it owns the reserve page for. The intention is to unmap the page
3331 * from other VMAs and let the children be SIGKILLed if they are faulting the
3334 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3335 struct page *page, unsigned long address)
3337 struct hstate *h = hstate_vma(vma);
3338 struct vm_area_struct *iter_vma;
3339 struct address_space *mapping;
3343 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3344 * from page cache lookup which is in HPAGE_SIZE units.
3346 address = address & huge_page_mask(h);
3347 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3349 mapping = file_inode(vma->vm_file)->i_mapping;
3352 * Take the mapping lock for the duration of the table walk. As
3353 * this mapping should be shared between all the VMAs,
3354 * __unmap_hugepage_range() is called as the lock is already held
3356 i_mmap_lock_write(mapping);
3357 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3358 /* Do not unmap the current VMA */
3359 if (iter_vma == vma)
3363 * Shared VMAs have their own reserves and do not affect
3364 * MAP_PRIVATE accounting but it is possible that a shared
3365 * VMA is using the same page so check and skip such VMAs.
3367 if (iter_vma->vm_flags & VM_MAYSHARE)
3371 * Unmap the page from other VMAs without their own reserves.
3372 * They get marked to be SIGKILLed if they fault in these
3373 * areas. This is because a future no-page fault on this VMA
3374 * could insert a zeroed page instead of the data existing
3375 * from the time of fork. This would look like data corruption
3377 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3378 unmap_hugepage_range(iter_vma, address,
3379 address + huge_page_size(h), page);
3381 i_mmap_unlock_write(mapping);
3385 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3386 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3387 * cannot race with other handlers or page migration.
3388 * Keep the pte_same checks anyway to make transition from the mutex easier.
3390 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3391 unsigned long address, pte_t *ptep, pte_t pte,
3392 struct page *pagecache_page, spinlock_t *ptl)
3394 struct hstate *h = hstate_vma(vma);
3395 struct page *old_page, *new_page;
3396 int ret = 0, outside_reserve = 0;
3397 unsigned long mmun_start; /* For mmu_notifiers */
3398 unsigned long mmun_end; /* For mmu_notifiers */
3400 old_page = pte_page(pte);
3403 /* If no-one else is actually using this page, avoid the copy
3404 * and just make the page writable */
3405 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3406 page_move_anon_rmap(old_page, vma, address);
3407 set_huge_ptep_writable(vma, address, ptep);
3412 * If the process that created a MAP_PRIVATE mapping is about to
3413 * perform a COW due to a shared page count, attempt to satisfy
3414 * the allocation without using the existing reserves. The pagecache
3415 * page is used to determine if the reserve at this address was
3416 * consumed or not. If reserves were used, a partial faulted mapping
3417 * at the time of fork() could consume its reserves on COW instead
3418 * of the full address range.
3420 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3421 old_page != pagecache_page)
3422 outside_reserve = 1;
3424 page_cache_get(old_page);
3427 * Drop page table lock as buddy allocator may be called. It will
3428 * be acquired again before returning to the caller, as expected.
3431 new_page = alloc_huge_page(vma, address, outside_reserve);
3433 if (IS_ERR(new_page)) {
3435 * If a process owning a MAP_PRIVATE mapping fails to COW,
3436 * it is due to references held by a child and an insufficient
3437 * huge page pool. To guarantee the original mappers
3438 * reliability, unmap the page from child processes. The child
3439 * may get SIGKILLed if it later faults.
3441 if (outside_reserve) {
3442 page_cache_release(old_page);
3443 BUG_ON(huge_pte_none(pte));
3444 unmap_ref_private(mm, vma, old_page, address);
3445 BUG_ON(huge_pte_none(pte));
3447 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3449 pte_same(huge_ptep_get(ptep), pte)))
3450 goto retry_avoidcopy;
3452 * race occurs while re-acquiring page table
3453 * lock, and our job is done.
3458 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3459 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3460 goto out_release_old;
3464 * When the original hugepage is shared one, it does not have
3465 * anon_vma prepared.
3467 if (unlikely(anon_vma_prepare(vma))) {
3469 goto out_release_all;
3472 copy_user_huge_page(new_page, old_page, address, vma,
3473 pages_per_huge_page(h));
3474 __SetPageUptodate(new_page);
3475 set_page_huge_active(new_page);
3477 mmun_start = address & huge_page_mask(h);
3478 mmun_end = mmun_start + huge_page_size(h);
3479 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3482 * Retake the page table lock to check for racing updates
3483 * before the page tables are altered
3486 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3487 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3488 ClearPagePrivate(new_page);
3491 huge_ptep_clear_flush(vma, address, ptep);
3492 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3493 set_huge_pte_at(mm, address, ptep,
3494 make_huge_pte(vma, new_page, 1));
3495 page_remove_rmap(old_page);
3496 hugepage_add_new_anon_rmap(new_page, vma, address);
3497 /* Make the old page be freed below */
3498 new_page = old_page;
3501 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3503 page_cache_release(new_page);
3505 page_cache_release(old_page);
3507 spin_lock(ptl); /* Caller expects lock to be held */
3511 /* Return the pagecache page at a given address within a VMA */
3512 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3513 struct vm_area_struct *vma, unsigned long address)
3515 struct address_space *mapping;
3518 mapping = vma->vm_file->f_mapping;
3519 idx = vma_hugecache_offset(h, vma, address);
3521 return find_lock_page(mapping, idx);
3525 * Return whether there is a pagecache page to back given address within VMA.
3526 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3528 static bool hugetlbfs_pagecache_present(struct hstate *h,
3529 struct vm_area_struct *vma, unsigned long address)
3531 struct address_space *mapping;
3535 mapping = vma->vm_file->f_mapping;
3536 idx = vma_hugecache_offset(h, vma, address);
3538 page = find_get_page(mapping, idx);
3541 return page != NULL;
3544 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3547 struct inode *inode = mapping->host;
3548 struct hstate *h = hstate_inode(inode);
3549 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3553 ClearPagePrivate(page);
3556 * set page dirty so that it will not be removed from cache/file
3557 * by non-hugetlbfs specific code paths.
3559 set_page_dirty(page);
3561 spin_lock(&inode->i_lock);
3562 inode->i_blocks += blocks_per_huge_page(h);
3563 spin_unlock(&inode->i_lock);
3567 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3568 struct address_space *mapping, pgoff_t idx,
3569 unsigned long address, pte_t *ptep, unsigned int flags)
3571 struct hstate *h = hstate_vma(vma);
3572 int ret = VM_FAULT_SIGBUS;
3580 * Currently, we are forced to kill the process in the event the
3581 * original mapper has unmapped pages from the child due to a failed
3582 * COW. Warn that such a situation has occurred as it may not be obvious
3584 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3585 pr_warning("PID %d killed due to inadequate hugepage pool\n",
3591 * Use page lock to guard against racing truncation
3592 * before we get page_table_lock.
3595 page = find_lock_page(mapping, idx);
3597 size = i_size_read(mapping->host) >> huge_page_shift(h);
3600 page = alloc_huge_page(vma, address, 0);
3602 ret = PTR_ERR(page);
3606 ret = VM_FAULT_SIGBUS;
3609 clear_huge_page(page, address, pages_per_huge_page(h));
3610 __SetPageUptodate(page);
3611 set_page_huge_active(page);
3613 if (vma->vm_flags & VM_MAYSHARE) {
3614 int err = huge_add_to_page_cache(page, mapping, idx);
3623 if (unlikely(anon_vma_prepare(vma))) {
3625 goto backout_unlocked;
3631 * If memory error occurs between mmap() and fault, some process
3632 * don't have hwpoisoned swap entry for errored virtual address.
3633 * So we need to block hugepage fault by PG_hwpoison bit check.
3635 if (unlikely(PageHWPoison(page))) {
3636 ret = VM_FAULT_HWPOISON |
3637 VM_FAULT_SET_HINDEX(hstate_index(h));
3638 goto backout_unlocked;
3643 * If we are going to COW a private mapping later, we examine the
3644 * pending reservations for this page now. This will ensure that
3645 * any allocations necessary to record that reservation occur outside
3648 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3649 if (vma_needs_reservation(h, vma, address) < 0) {
3651 goto backout_unlocked;
3653 /* Just decrements count, does not deallocate */
3654 vma_end_reservation(h, vma, address);
3657 ptl = huge_pte_lockptr(h, mm, ptep);
3659 size = i_size_read(mapping->host) >> huge_page_shift(h);
3664 if (!huge_pte_none(huge_ptep_get(ptep)))
3668 ClearPagePrivate(page);
3669 hugepage_add_new_anon_rmap(page, vma, address);
3671 page_dup_rmap(page);
3672 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3673 && (vma->vm_flags & VM_SHARED)));
3674 set_huge_pte_at(mm, address, ptep, new_pte);
3676 hugetlb_count_add(pages_per_huge_page(h), mm);
3677 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3678 /* Optimization, do the COW without a second fault */
3679 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
3696 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3697 struct vm_area_struct *vma,
3698 struct address_space *mapping,
3699 pgoff_t idx, unsigned long address)
3701 unsigned long key[2];
3704 if (vma->vm_flags & VM_SHARED) {
3705 key[0] = (unsigned long) mapping;
3708 key[0] = (unsigned long) mm;
3709 key[1] = address >> huge_page_shift(h);
3712 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3714 return hash & (num_fault_mutexes - 1);
3718 * For uniprocesor systems we always use a single mutex, so just
3719 * return 0 and avoid the hashing overhead.
3721 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3722 struct vm_area_struct *vma,
3723 struct address_space *mapping,
3724 pgoff_t idx, unsigned long address)
3730 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3731 unsigned long address, unsigned int flags)
3738 struct page *page = NULL;
3739 struct page *pagecache_page = NULL;
3740 struct hstate *h = hstate_vma(vma);
3741 struct address_space *mapping;
3742 int need_wait_lock = 0;
3744 address &= huge_page_mask(h);
3746 ptep = huge_pte_offset(mm, address);
3748 entry = huge_ptep_get(ptep);
3749 if (unlikely(is_hugetlb_entry_migration(entry))) {
3750 migration_entry_wait_huge(vma, mm, ptep);
3752 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3753 return VM_FAULT_HWPOISON_LARGE |
3754 VM_FAULT_SET_HINDEX(hstate_index(h));
3756 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3758 return VM_FAULT_OOM;
3761 mapping = vma->vm_file->f_mapping;
3762 idx = vma_hugecache_offset(h, vma, address);
3765 * Serialize hugepage allocation and instantiation, so that we don't
3766 * get spurious allocation failures if two CPUs race to instantiate
3767 * the same page in the page cache.
3769 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3770 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3772 entry = huge_ptep_get(ptep);
3773 if (huge_pte_none(entry)) {
3774 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3781 * entry could be a migration/hwpoison entry at this point, so this
3782 * check prevents the kernel from going below assuming that we have
3783 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3784 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3787 if (!pte_present(entry))
3791 * If we are going to COW the mapping later, we examine the pending
3792 * reservations for this page now. This will ensure that any
3793 * allocations necessary to record that reservation occur outside the
3794 * spinlock. For private mappings, we also lookup the pagecache
3795 * page now as it is used to determine if a reservation has been
3798 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3799 if (vma_needs_reservation(h, vma, address) < 0) {
3803 /* Just decrements count, does not deallocate */
3804 vma_end_reservation(h, vma, address);
3806 if (!(vma->vm_flags & VM_MAYSHARE))
3807 pagecache_page = hugetlbfs_pagecache_page(h,
3811 ptl = huge_pte_lock(h, mm, ptep);
3813 /* Check for a racing update before calling hugetlb_cow */
3814 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3818 * hugetlb_cow() requires page locks of pte_page(entry) and
3819 * pagecache_page, so here we need take the former one
3820 * when page != pagecache_page or !pagecache_page.
3822 page = pte_page(entry);
3823 if (page != pagecache_page)
3824 if (!trylock_page(page)) {
3831 if (flags & FAULT_FLAG_WRITE) {
3832 if (!huge_pte_write(entry)) {
3833 ret = hugetlb_cow(mm, vma, address, ptep, entry,
3834 pagecache_page, ptl);
3837 entry = huge_pte_mkdirty(entry);
3839 entry = pte_mkyoung(entry);
3840 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3841 flags & FAULT_FLAG_WRITE))
3842 update_mmu_cache(vma, address, ptep);
3844 if (page != pagecache_page)
3850 if (pagecache_page) {
3851 unlock_page(pagecache_page);
3852 put_page(pagecache_page);
3855 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3857 * Generally it's safe to hold refcount during waiting page lock. But
3858 * here we just wait to defer the next page fault to avoid busy loop and
3859 * the page is not used after unlocked before returning from the current
3860 * page fault. So we are safe from accessing freed page, even if we wait
3861 * here without taking refcount.
3864 wait_on_page_locked(page);
3868 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3869 struct page **pages, struct vm_area_struct **vmas,
3870 unsigned long *position, unsigned long *nr_pages,
3871 long i, unsigned int flags)
3873 unsigned long pfn_offset;
3874 unsigned long vaddr = *position;
3875 unsigned long remainder = *nr_pages;
3876 struct hstate *h = hstate_vma(vma);
3878 while (vaddr < vma->vm_end && remainder) {
3880 spinlock_t *ptl = NULL;
3885 * If we have a pending SIGKILL, don't keep faulting pages and
3886 * potentially allocating memory.
3888 if (unlikely(fatal_signal_pending(current))) {
3894 * Some archs (sparc64, sh*) have multiple pte_ts to
3895 * each hugepage. We have to make sure we get the
3896 * first, for the page indexing below to work.
3898 * Note that page table lock is not held when pte is null.
3900 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3902 ptl = huge_pte_lock(h, mm, pte);
3903 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3906 * When coredumping, it suits get_dump_page if we just return
3907 * an error where there's an empty slot with no huge pagecache
3908 * to back it. This way, we avoid allocating a hugepage, and
3909 * the sparse dumpfile avoids allocating disk blocks, but its
3910 * huge holes still show up with zeroes where they need to be.
3912 if (absent && (flags & FOLL_DUMP) &&
3913 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3921 * We need call hugetlb_fault for both hugepages under migration
3922 * (in which case hugetlb_fault waits for the migration,) and
3923 * hwpoisoned hugepages (in which case we need to prevent the
3924 * caller from accessing to them.) In order to do this, we use
3925 * here is_swap_pte instead of is_hugetlb_entry_migration and
3926 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3927 * both cases, and because we can't follow correct pages
3928 * directly from any kind of swap entries.
3930 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3931 ((flags & FOLL_WRITE) &&
3932 !huge_pte_write(huge_ptep_get(pte)))) {
3937 ret = hugetlb_fault(mm, vma, vaddr,
3938 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3939 if (!(ret & VM_FAULT_ERROR))
3946 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3947 page = pte_page(huge_ptep_get(pte));
3950 pages[i] = mem_map_offset(page, pfn_offset);
3951 get_page_foll(pages[i]);
3961 if (vaddr < vma->vm_end && remainder &&
3962 pfn_offset < pages_per_huge_page(h)) {
3964 * We use pfn_offset to avoid touching the pageframes
3965 * of this compound page.
3971 *nr_pages = remainder;
3974 return i ? i : -EFAULT;
3977 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3978 unsigned long address, unsigned long end, pgprot_t newprot)
3980 struct mm_struct *mm = vma->vm_mm;
3981 unsigned long start = address;
3984 struct hstate *h = hstate_vma(vma);
3985 unsigned long pages = 0;
3987 BUG_ON(address >= end);
3988 flush_cache_range(vma, address, end);
3990 mmu_notifier_invalidate_range_start(mm, start, end);
3991 i_mmap_lock_write(vma->vm_file->f_mapping);
3992 for (; address < end; address += huge_page_size(h)) {
3994 ptep = huge_pte_offset(mm, address);
3997 ptl = huge_pte_lock(h, mm, ptep);
3998 if (huge_pmd_unshare(mm, &address, ptep)) {
4003 pte = huge_ptep_get(ptep);
4004 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4008 if (unlikely(is_hugetlb_entry_migration(pte))) {
4009 swp_entry_t entry = pte_to_swp_entry(pte);
4011 if (is_write_migration_entry(entry)) {
4014 make_migration_entry_read(&entry);
4015 newpte = swp_entry_to_pte(entry);
4016 set_huge_pte_at(mm, address, ptep, newpte);
4022 if (!huge_pte_none(pte)) {
4023 pte = huge_ptep_get_and_clear(mm, address, ptep);
4024 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
4025 pte = arch_make_huge_pte(pte, vma, NULL, 0);
4026 set_huge_pte_at(mm, address, ptep, pte);
4032 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4033 * may have cleared our pud entry and done put_page on the page table:
4034 * once we release i_mmap_rwsem, another task can do the final put_page
4035 * and that page table be reused and filled with junk.
4037 flush_tlb_range(vma, start, end);
4038 mmu_notifier_invalidate_range(mm, start, end);
4039 i_mmap_unlock_write(vma->vm_file->f_mapping);
4040 mmu_notifier_invalidate_range_end(mm, start, end);
4042 return pages << h->order;
4045 int hugetlb_reserve_pages(struct inode *inode,
4047 struct vm_area_struct *vma,
4048 vm_flags_t vm_flags)
4051 struct hstate *h = hstate_inode(inode);
4052 struct hugepage_subpool *spool = subpool_inode(inode);
4053 struct resv_map *resv_map;
4057 * Only apply hugepage reservation if asked. At fault time, an
4058 * attempt will be made for VM_NORESERVE to allocate a page
4059 * without using reserves
4061 if (vm_flags & VM_NORESERVE)
4065 * Shared mappings base their reservation on the number of pages that
4066 * are already allocated on behalf of the file. Private mappings need
4067 * to reserve the full area even if read-only as mprotect() may be
4068 * called to make the mapping read-write. Assume !vma is a shm mapping
4070 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4071 resv_map = inode_resv_map(inode);
4073 chg = region_chg(resv_map, from, to);
4076 resv_map = resv_map_alloc();
4082 set_vma_resv_map(vma, resv_map);
4083 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4092 * There must be enough pages in the subpool for the mapping. If
4093 * the subpool has a minimum size, there may be some global
4094 * reservations already in place (gbl_reserve).
4096 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4097 if (gbl_reserve < 0) {
4103 * Check enough hugepages are available for the reservation.
4104 * Hand the pages back to the subpool if there are not
4106 ret = hugetlb_acct_memory(h, gbl_reserve);
4108 /* put back original number of pages, chg */
4109 (void)hugepage_subpool_put_pages(spool, chg);
4114 * Account for the reservations made. Shared mappings record regions
4115 * that have reservations as they are shared by multiple VMAs.
4116 * When the last VMA disappears, the region map says how much
4117 * the reservation was and the page cache tells how much of
4118 * the reservation was consumed. Private mappings are per-VMA and
4119 * only the consumed reservations are tracked. When the VMA
4120 * disappears, the original reservation is the VMA size and the
4121 * consumed reservations are stored in the map. Hence, nothing
4122 * else has to be done for private mappings here
4124 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4125 long add = region_add(resv_map, from, to);
4127 if (unlikely(chg > add)) {
4129 * pages in this range were added to the reserve
4130 * map between region_chg and region_add. This
4131 * indicates a race with alloc_huge_page. Adjust
4132 * the subpool and reserve counts modified above
4133 * based on the difference.
4137 rsv_adjust = hugepage_subpool_put_pages(spool,
4139 hugetlb_acct_memory(h, -rsv_adjust);
4144 if (!vma || vma->vm_flags & VM_MAYSHARE)
4145 /* Don't call region_abort if region_chg failed */
4147 region_abort(resv_map, from, to);
4148 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4149 kref_put(&resv_map->refs, resv_map_release);
4153 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4156 struct hstate *h = hstate_inode(inode);
4157 struct resv_map *resv_map = inode_resv_map(inode);
4159 struct hugepage_subpool *spool = subpool_inode(inode);
4163 chg = region_del(resv_map, start, end);
4165 * region_del() can fail in the rare case where a region
4166 * must be split and another region descriptor can not be
4167 * allocated. If end == LONG_MAX, it will not fail.
4173 spin_lock(&inode->i_lock);
4174 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4175 spin_unlock(&inode->i_lock);
4178 * If the subpool has a minimum size, the number of global
4179 * reservations to be released may be adjusted.
4181 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4182 hugetlb_acct_memory(h, -gbl_reserve);
4187 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4188 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4189 struct vm_area_struct *vma,
4190 unsigned long addr, pgoff_t idx)
4192 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4194 unsigned long sbase = saddr & PUD_MASK;
4195 unsigned long s_end = sbase + PUD_SIZE;
4197 /* Allow segments to share if only one is marked locked */
4198 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4199 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4202 * match the virtual addresses, permission and the alignment of the
4205 if (pmd_index(addr) != pmd_index(saddr) ||
4206 vm_flags != svm_flags ||
4207 sbase < svma->vm_start || svma->vm_end < s_end)
4213 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4215 unsigned long base = addr & PUD_MASK;
4216 unsigned long end = base + PUD_SIZE;
4219 * check on proper vm_flags and page table alignment
4221 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
4227 * Determine if start,end range within vma could be mapped by shared pmd.
4228 * If yes, adjust start and end to cover range associated with possible
4229 * shared pmd mappings.
4231 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4232 unsigned long *start, unsigned long *end)
4234 unsigned long check_addr = *start;
4236 if (!(vma->vm_flags & VM_MAYSHARE))
4239 for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
4240 unsigned long a_start = check_addr & PUD_MASK;
4241 unsigned long a_end = a_start + PUD_SIZE;
4244 * If sharing is possible, adjust start/end if necessary.
4246 if (range_in_vma(vma, a_start, a_end)) {
4247 if (a_start < *start)
4256 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4257 * and returns the corresponding pte. While this is not necessary for the
4258 * !shared pmd case because we can allocate the pmd later as well, it makes the
4259 * code much cleaner. pmd allocation is essential for the shared case because
4260 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4261 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4262 * bad pmd for sharing.
4264 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4266 struct vm_area_struct *vma = find_vma(mm, addr);
4267 struct address_space *mapping = vma->vm_file->f_mapping;
4268 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4270 struct vm_area_struct *svma;
4271 unsigned long saddr;
4276 if (!vma_shareable(vma, addr))
4277 return (pte_t *)pmd_alloc(mm, pud, addr);
4279 i_mmap_lock_write(mapping);
4280 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4284 saddr = page_table_shareable(svma, vma, addr, idx);
4286 spte = huge_pte_offset(svma->vm_mm, saddr);
4288 get_page(virt_to_page(spte));
4297 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
4299 if (pud_none(*pud)) {
4300 pud_populate(mm, pud,
4301 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4304 put_page(virt_to_page(spte));
4308 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4309 i_mmap_unlock_write(mapping);
4314 * unmap huge page backed by shared pte.
4316 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4317 * indicated by page_count > 1, unmap is achieved by clearing pud and
4318 * decrementing the ref count. If count == 1, the pte page is not shared.
4320 * called with page table lock held.
4322 * returns: 1 successfully unmapped a shared pte page
4323 * 0 the underlying pte page is not shared, or it is the last user
4325 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4327 pgd_t *pgd = pgd_offset(mm, *addr);
4328 pud_t *pud = pud_offset(pgd, *addr);
4330 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4331 if (page_count(virt_to_page(ptep)) == 1)
4335 put_page(virt_to_page(ptep));
4337 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4340 #define want_pmd_share() (1)
4341 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4342 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4347 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4352 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4353 unsigned long *start, unsigned long *end)
4356 #define want_pmd_share() (0)
4357 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4359 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4360 pte_t *huge_pte_alloc(struct mm_struct *mm,
4361 unsigned long addr, unsigned long sz)
4367 pgd = pgd_offset(mm, addr);
4368 pud = pud_alloc(mm, pgd, addr);
4370 if (sz == PUD_SIZE) {
4373 BUG_ON(sz != PMD_SIZE);
4374 if (want_pmd_share() && pud_none(*pud))
4375 pte = huge_pmd_share(mm, addr, pud);
4377 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4380 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
4385 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
4391 pgd = pgd_offset(mm, addr);
4392 if (pgd_present(*pgd)) {
4393 pud = pud_offset(pgd, addr);
4394 if (pud_present(*pud)) {
4396 return (pte_t *)pud;
4397 pmd = pmd_offset(pud, addr);
4400 return (pte_t *) pmd;
4403 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4406 * These functions are overwritable if your architecture needs its own
4409 struct page * __weak
4410 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4413 return ERR_PTR(-EINVAL);
4416 struct page * __weak
4417 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4418 pmd_t *pmd, int flags)
4420 struct page *page = NULL;
4424 ptl = pmd_lockptr(mm, pmd);
4427 * make sure that the address range covered by this pmd is not
4428 * unmapped from other threads.
4430 if (!pmd_huge(*pmd))
4432 pte = huge_ptep_get((pte_t *)pmd);
4433 if (pte_present(pte)) {
4434 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4435 if (flags & FOLL_GET)
4438 if (is_hugetlb_entry_migration(pte)) {
4440 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4444 * hwpoisoned entry is treated as no_page_table in
4445 * follow_page_mask().
4453 struct page * __weak
4454 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4455 pud_t *pud, int flags)
4457 if (flags & FOLL_GET)
4460 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4463 #ifdef CONFIG_MEMORY_FAILURE
4466 * This function is called from memory failure code.
4467 * Assume the caller holds page lock of the head page.
4469 int dequeue_hwpoisoned_huge_page(struct page *hpage)
4471 struct hstate *h = page_hstate(hpage);
4472 int nid = page_to_nid(hpage);
4475 spin_lock(&hugetlb_lock);
4477 * Just checking !page_huge_active is not enough, because that could be
4478 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4480 if (!page_huge_active(hpage) && !page_count(hpage)) {
4482 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4483 * but dangling hpage->lru can trigger list-debug warnings
4484 * (this happens when we call unpoison_memory() on it),
4485 * so let it point to itself with list_del_init().
4487 list_del_init(&hpage->lru);
4488 set_page_refcounted(hpage);
4489 h->free_huge_pages--;
4490 h->free_huge_pages_node[nid]--;
4493 spin_unlock(&hugetlb_lock);
4498 bool isolate_huge_page(struct page *page, struct list_head *list)
4502 VM_BUG_ON_PAGE(!PageHead(page), page);
4503 spin_lock(&hugetlb_lock);
4504 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4508 clear_page_huge_active(page);
4509 list_move_tail(&page->lru, list);
4511 spin_unlock(&hugetlb_lock);
4515 void putback_active_hugepage(struct page *page)
4517 VM_BUG_ON_PAGE(!PageHead(page), page);
4518 spin_lock(&hugetlb_lock);
4519 set_page_huge_active(page);
4520 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4521 spin_unlock(&hugetlb_lock);