2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/module.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/bootmem.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/rmap.h>
23 #include <linux/swap.h>
24 #include <linux/swapops.h>
25 #include <linux/page-isolation.h>
26 #include <linux/jhash.h>
29 #include <asm/pgtable.h>
33 #include <linux/hugetlb.h>
34 #include <linux/hugetlb_cgroup.h>
35 #include <linux/node.h>
38 int hugepages_treat_as_movable;
40 int hugetlb_max_hstate __read_mostly;
41 unsigned int default_hstate_idx;
42 struct hstate hstates[HUGE_MAX_HSTATE];
44 * Minimum page order among possible hugepage sizes, set to a proper value
47 static unsigned int minimum_order __read_mostly = UINT_MAX;
49 __initdata LIST_HEAD(huge_boot_pages);
51 /* for command line parsing */
52 static struct hstate * __initdata parsed_hstate;
53 static unsigned long __initdata default_hstate_max_huge_pages;
54 static unsigned long __initdata default_hstate_size;
57 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
58 * free_huge_pages, and surplus_huge_pages.
60 DEFINE_SPINLOCK(hugetlb_lock);
63 * Serializes faults on the same logical page. This is used to
64 * prevent spurious OOMs when the hugepage pool is fully utilized.
66 static int num_fault_mutexes;
67 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
69 /* Forward declaration */
70 static int hugetlb_acct_memory(struct hstate *h, long delta);
72 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
74 bool free = (spool->count == 0) && (spool->used_hpages == 0);
76 spin_unlock(&spool->lock);
78 /* If no pages are used, and no other handles to the subpool
79 * remain, give up any reservations mased on minimum size and
82 if (spool->min_hpages != -1)
83 hugetlb_acct_memory(spool->hstate,
89 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
92 struct hugepage_subpool *spool;
94 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
98 spin_lock_init(&spool->lock);
100 spool->max_hpages = max_hpages;
102 spool->min_hpages = min_hpages;
104 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
108 spool->rsv_hpages = min_hpages;
113 void hugepage_put_subpool(struct hugepage_subpool *spool)
115 spin_lock(&spool->lock);
116 BUG_ON(!spool->count);
118 unlock_or_release_subpool(spool);
122 * Subpool accounting for allocating and reserving pages.
123 * Return -ENOMEM if there are not enough resources to satisfy the
124 * the request. Otherwise, return the number of pages by which the
125 * global pools must be adjusted (upward). The returned value may
126 * only be different than the passed value (delta) in the case where
127 * a subpool minimum size must be manitained.
129 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
137 spin_lock(&spool->lock);
139 if (spool->max_hpages != -1) { /* maximum size accounting */
140 if ((spool->used_hpages + delta) <= spool->max_hpages)
141 spool->used_hpages += delta;
148 if (spool->min_hpages != -1) { /* minimum size accounting */
149 if (delta > spool->rsv_hpages) {
151 * Asking for more reserves than those already taken on
152 * behalf of subpool. Return difference.
154 ret = delta - spool->rsv_hpages;
155 spool->rsv_hpages = 0;
157 ret = 0; /* reserves already accounted for */
158 spool->rsv_hpages -= delta;
163 spin_unlock(&spool->lock);
168 * Subpool accounting for freeing and unreserving pages.
169 * Return the number of global page reservations that must be dropped.
170 * The return value may only be different than the passed value (delta)
171 * in the case where a subpool minimum size must be maintained.
173 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
181 spin_lock(&spool->lock);
183 if (spool->max_hpages != -1) /* maximum size accounting */
184 spool->used_hpages -= delta;
186 if (spool->min_hpages != -1) { /* minimum size accounting */
187 if (spool->rsv_hpages + delta <= spool->min_hpages)
190 ret = spool->rsv_hpages + delta - spool->min_hpages;
192 spool->rsv_hpages += delta;
193 if (spool->rsv_hpages > spool->min_hpages)
194 spool->rsv_hpages = spool->min_hpages;
198 * If hugetlbfs_put_super couldn't free spool due to an outstanding
199 * quota reference, free it now.
201 unlock_or_release_subpool(spool);
206 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
208 return HUGETLBFS_SB(inode->i_sb)->spool;
211 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
213 return subpool_inode(file_inode(vma->vm_file));
217 * Region tracking -- allows tracking of reservations and instantiated pages
218 * across the pages in a mapping.
220 * The region data structures are embedded into a resv_map and protected
221 * by a resv_map's lock. The set of regions within the resv_map represent
222 * reservations for huge pages, or huge pages that have already been
223 * instantiated within the map. The from and to elements are huge page
224 * indicies into the associated mapping. from indicates the starting index
225 * of the region. to represents the first index past the end of the region.
227 * For example, a file region structure with from == 0 and to == 4 represents
228 * four huge pages in a mapping. It is important to note that the to element
229 * represents the first element past the end of the region. This is used in
230 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
232 * Interval notation of the form [from, to) will be used to indicate that
233 * the endpoint from is inclusive and to is exclusive.
236 struct list_head link;
242 * Add the huge page range represented by [f, t) to the reserve
243 * map. In the normal case, existing regions will be expanded
244 * to accommodate the specified range. Sufficient regions should
245 * exist for expansion due to the previous call to region_chg
246 * with the same range. However, it is possible that region_del
247 * could have been called after region_chg and modifed the map
248 * in such a way that no region exists to be expanded. In this
249 * case, pull a region descriptor from the cache associated with
250 * the map and use that for the new range.
252 * Return the number of new huge pages added to the map. This
253 * number is greater than or equal to zero.
255 static long region_add(struct resv_map *resv, long f, long t)
257 struct list_head *head = &resv->regions;
258 struct file_region *rg, *nrg, *trg;
261 spin_lock(&resv->lock);
262 /* Locate the region we are either in or before. */
263 list_for_each_entry(rg, head, link)
268 * If no region exists which can be expanded to include the
269 * specified range, the list must have been modified by an
270 * interleving call to region_del(). Pull a region descriptor
271 * from the cache and use it for this range.
273 if (&rg->link == head || t < rg->from) {
274 VM_BUG_ON(resv->region_cache_count <= 0);
276 resv->region_cache_count--;
277 nrg = list_first_entry(&resv->region_cache, struct file_region,
279 list_del(&nrg->link);
283 list_add(&nrg->link, rg->link.prev);
289 /* Round our left edge to the current segment if it encloses us. */
293 /* Check for and consume any regions we now overlap with. */
295 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
296 if (&rg->link == head)
301 /* If this area reaches higher then extend our area to
302 * include it completely. If this is not the first area
303 * which we intend to reuse, free it. */
307 /* Decrement return value by the deleted range.
308 * Another range will span this area so that by
309 * end of routine add will be >= zero
311 add -= (rg->to - rg->from);
317 add += (nrg->from - f); /* Added to beginning of region */
319 add += t - nrg->to; /* Added to end of region */
323 resv->adds_in_progress--;
324 spin_unlock(&resv->lock);
330 * Examine the existing reserve map and determine how many
331 * huge pages in the specified range [f, t) are NOT currently
332 * represented. This routine is called before a subsequent
333 * call to region_add that will actually modify the reserve
334 * map to add the specified range [f, t). region_chg does
335 * not change the number of huge pages represented by the
336 * map. However, if the existing regions in the map can not
337 * be expanded to represent the new range, a new file_region
338 * structure is added to the map as a placeholder. This is
339 * so that the subsequent region_add call will have all the
340 * regions it needs and will not fail.
342 * Upon entry, region_chg will also examine the cache of region descriptors
343 * associated with the map. If there are not enough descriptors cached, one
344 * will be allocated for the in progress add operation.
346 * Returns the number of huge pages that need to be added to the existing
347 * reservation map for the range [f, t). This number is greater or equal to
348 * zero. -ENOMEM is returned if a new file_region structure or cache entry
349 * is needed and can not be allocated.
351 static long region_chg(struct resv_map *resv, long f, long t)
353 struct list_head *head = &resv->regions;
354 struct file_region *rg, *nrg = NULL;
358 spin_lock(&resv->lock);
360 resv->adds_in_progress++;
363 * Check for sufficient descriptors in the cache to accommodate
364 * the number of in progress add operations.
366 if (resv->adds_in_progress > resv->region_cache_count) {
367 struct file_region *trg;
369 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
370 /* Must drop lock to allocate a new descriptor. */
371 resv->adds_in_progress--;
372 spin_unlock(&resv->lock);
374 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
378 spin_lock(&resv->lock);
379 list_add(&trg->link, &resv->region_cache);
380 resv->region_cache_count++;
384 /* Locate the region we are before or in. */
385 list_for_each_entry(rg, head, link)
389 /* If we are below the current region then a new region is required.
390 * Subtle, allocate a new region at the position but make it zero
391 * size such that we can guarantee to record the reservation. */
392 if (&rg->link == head || t < rg->from) {
394 resv->adds_in_progress--;
395 spin_unlock(&resv->lock);
396 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
402 INIT_LIST_HEAD(&nrg->link);
406 list_add(&nrg->link, rg->link.prev);
411 /* Round our left edge to the current segment if it encloses us. */
416 /* Check for and consume any regions we now overlap with. */
417 list_for_each_entry(rg, rg->link.prev, link) {
418 if (&rg->link == head)
423 /* We overlap with this area, if it extends further than
424 * us then we must extend ourselves. Account for its
425 * existing reservation. */
430 chg -= rg->to - rg->from;
434 spin_unlock(&resv->lock);
435 /* We already know we raced and no longer need the new region */
439 spin_unlock(&resv->lock);
444 * Abort the in progress add operation. The adds_in_progress field
445 * of the resv_map keeps track of the operations in progress between
446 * calls to region_chg and region_add. Operations are sometimes
447 * aborted after the call to region_chg. In such cases, region_abort
448 * is called to decrement the adds_in_progress counter.
450 * NOTE: The range arguments [f, t) are not needed or used in this
451 * routine. They are kept to make reading the calling code easier as
452 * arguments will match the associated region_chg call.
454 static void region_abort(struct resv_map *resv, long f, long t)
456 spin_lock(&resv->lock);
457 VM_BUG_ON(!resv->region_cache_count);
458 resv->adds_in_progress--;
459 spin_unlock(&resv->lock);
463 * Delete the specified range [f, t) from the reserve map. If the
464 * t parameter is LONG_MAX, this indicates that ALL regions after f
465 * should be deleted. Locate the regions which intersect [f, t)
466 * and either trim, delete or split the existing regions.
468 * Returns the number of huge pages deleted from the reserve map.
469 * In the normal case, the return value is zero or more. In the
470 * case where a region must be split, a new region descriptor must
471 * be allocated. If the allocation fails, -ENOMEM will be returned.
472 * NOTE: If the parameter t == LONG_MAX, then we will never split
473 * a region and possibly return -ENOMEM. Callers specifying
474 * t == LONG_MAX do not need to check for -ENOMEM error.
476 static long region_del(struct resv_map *resv, long f, long t)
478 struct list_head *head = &resv->regions;
479 struct file_region *rg, *trg;
480 struct file_region *nrg = NULL;
484 spin_lock(&resv->lock);
485 list_for_each_entry_safe(rg, trg, head, link) {
491 if (f > rg->from && t < rg->to) { /* Must split region */
493 * Check for an entry in the cache before dropping
494 * lock and attempting allocation.
497 resv->region_cache_count > resv->adds_in_progress) {
498 nrg = list_first_entry(&resv->region_cache,
501 list_del(&nrg->link);
502 resv->region_cache_count--;
506 spin_unlock(&resv->lock);
507 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
515 /* New entry for end of split region */
518 INIT_LIST_HEAD(&nrg->link);
520 /* Original entry is trimmed */
523 list_add(&nrg->link, &rg->link);
528 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
529 del += rg->to - rg->from;
535 if (f <= rg->from) { /* Trim beginning of region */
538 } else { /* Trim end of region */
544 spin_unlock(&resv->lock);
550 * A rare out of memory error was encountered which prevented removal of
551 * the reserve map region for a page. The huge page itself was free'ed
552 * and removed from the page cache. This routine will adjust the subpool
553 * usage count, and the global reserve count if needed. By incrementing
554 * these counts, the reserve map entry which could not be deleted will
555 * appear as a "reserved" entry instead of simply dangling with incorrect
558 void hugetlb_fix_reserve_counts(struct inode *inode, bool restore_reserve)
560 struct hugepage_subpool *spool = subpool_inode(inode);
563 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
564 if (restore_reserve && rsv_adjust) {
565 struct hstate *h = hstate_inode(inode);
567 hugetlb_acct_memory(h, 1);
572 * Count and return the number of huge pages in the reserve map
573 * that intersect with the range [f, t).
575 static long region_count(struct resv_map *resv, long f, long t)
577 struct list_head *head = &resv->regions;
578 struct file_region *rg;
581 spin_lock(&resv->lock);
582 /* Locate each segment we overlap with, and count that overlap. */
583 list_for_each_entry(rg, head, link) {
592 seg_from = max(rg->from, f);
593 seg_to = min(rg->to, t);
595 chg += seg_to - seg_from;
597 spin_unlock(&resv->lock);
603 * Convert the address within this vma to the page offset within
604 * the mapping, in pagecache page units; huge pages here.
606 static pgoff_t vma_hugecache_offset(struct hstate *h,
607 struct vm_area_struct *vma, unsigned long address)
609 return ((address - vma->vm_start) >> huge_page_shift(h)) +
610 (vma->vm_pgoff >> huge_page_order(h));
613 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
614 unsigned long address)
616 return vma_hugecache_offset(hstate_vma(vma), vma, address);
620 * Return the size of the pages allocated when backing a VMA. In the majority
621 * cases this will be same size as used by the page table entries.
623 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
625 struct hstate *hstate;
627 if (!is_vm_hugetlb_page(vma))
630 hstate = hstate_vma(vma);
632 return 1UL << huge_page_shift(hstate);
634 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
637 * Return the page size being used by the MMU to back a VMA. In the majority
638 * of cases, the page size used by the kernel matches the MMU size. On
639 * architectures where it differs, an architecture-specific version of this
640 * function is required.
642 #ifndef vma_mmu_pagesize
643 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
645 return vma_kernel_pagesize(vma);
650 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
651 * bits of the reservation map pointer, which are always clear due to
654 #define HPAGE_RESV_OWNER (1UL << 0)
655 #define HPAGE_RESV_UNMAPPED (1UL << 1)
656 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
659 * These helpers are used to track how many pages are reserved for
660 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
661 * is guaranteed to have their future faults succeed.
663 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
664 * the reserve counters are updated with the hugetlb_lock held. It is safe
665 * to reset the VMA at fork() time as it is not in use yet and there is no
666 * chance of the global counters getting corrupted as a result of the values.
668 * The private mapping reservation is represented in a subtly different
669 * manner to a shared mapping. A shared mapping has a region map associated
670 * with the underlying file, this region map represents the backing file
671 * pages which have ever had a reservation assigned which this persists even
672 * after the page is instantiated. A private mapping has a region map
673 * associated with the original mmap which is attached to all VMAs which
674 * reference it, this region map represents those offsets which have consumed
675 * reservation ie. where pages have been instantiated.
677 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
679 return (unsigned long)vma->vm_private_data;
682 static void set_vma_private_data(struct vm_area_struct *vma,
685 vma->vm_private_data = (void *)value;
688 struct resv_map *resv_map_alloc(void)
690 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
691 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
693 if (!resv_map || !rg) {
699 kref_init(&resv_map->refs);
700 spin_lock_init(&resv_map->lock);
701 INIT_LIST_HEAD(&resv_map->regions);
703 resv_map->adds_in_progress = 0;
705 INIT_LIST_HEAD(&resv_map->region_cache);
706 list_add(&rg->link, &resv_map->region_cache);
707 resv_map->region_cache_count = 1;
712 void resv_map_release(struct kref *ref)
714 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
715 struct list_head *head = &resv_map->region_cache;
716 struct file_region *rg, *trg;
718 /* Clear out any active regions before we release the map. */
719 region_del(resv_map, 0, LONG_MAX);
721 /* ... and any entries left in the cache */
722 list_for_each_entry_safe(rg, trg, head, link) {
727 VM_BUG_ON(resv_map->adds_in_progress);
732 static inline struct resv_map *inode_resv_map(struct inode *inode)
734 return inode->i_mapping->private_data;
737 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
739 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
740 if (vma->vm_flags & VM_MAYSHARE) {
741 struct address_space *mapping = vma->vm_file->f_mapping;
742 struct inode *inode = mapping->host;
744 return inode_resv_map(inode);
747 return (struct resv_map *)(get_vma_private_data(vma) &
752 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
754 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
755 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
757 set_vma_private_data(vma, (get_vma_private_data(vma) &
758 HPAGE_RESV_MASK) | (unsigned long)map);
761 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
763 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
764 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
766 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
769 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
771 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
773 return (get_vma_private_data(vma) & flag) != 0;
776 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
777 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
779 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
780 if (!(vma->vm_flags & VM_MAYSHARE))
781 vma->vm_private_data = (void *)0;
784 /* Returns true if the VMA has associated reserve pages */
785 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
787 if (vma->vm_flags & VM_NORESERVE) {
789 * This address is already reserved by other process(chg == 0),
790 * so, we should decrement reserved count. Without decrementing,
791 * reserve count remains after releasing inode, because this
792 * allocated page will go into page cache and is regarded as
793 * coming from reserved pool in releasing step. Currently, we
794 * don't have any other solution to deal with this situation
795 * properly, so add work-around here.
797 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
803 /* Shared mappings always use reserves */
804 if (vma->vm_flags & VM_MAYSHARE) {
806 * We know VM_NORESERVE is not set. Therefore, there SHOULD
807 * be a region map for all pages. The only situation where
808 * there is no region map is if a hole was punched via
809 * fallocate. In this case, there really are no reverves to
810 * use. This situation is indicated if chg != 0.
819 * Only the process that called mmap() has reserves for
822 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
828 static void enqueue_huge_page(struct hstate *h, struct page *page)
830 int nid = page_to_nid(page);
831 list_move(&page->lru, &h->hugepage_freelists[nid]);
832 h->free_huge_pages++;
833 h->free_huge_pages_node[nid]++;
836 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
840 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
841 if (!is_migrate_isolate_page(page))
844 * if 'non-isolated free hugepage' not found on the list,
845 * the allocation fails.
847 if (&h->hugepage_freelists[nid] == &page->lru)
849 list_move(&page->lru, &h->hugepage_activelist);
850 set_page_refcounted(page);
851 h->free_huge_pages--;
852 h->free_huge_pages_node[nid]--;
856 /* Movability of hugepages depends on migration support. */
857 static inline gfp_t htlb_alloc_mask(struct hstate *h)
859 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
860 return GFP_HIGHUSER_MOVABLE;
865 static struct page *dequeue_huge_page_vma(struct hstate *h,
866 struct vm_area_struct *vma,
867 unsigned long address, int avoid_reserve,
870 struct page *page = NULL;
871 struct mempolicy *mpol;
872 nodemask_t *nodemask;
873 struct zonelist *zonelist;
876 unsigned int cpuset_mems_cookie;
879 * A child process with MAP_PRIVATE mappings created by their parent
880 * have no page reserves. This check ensures that reservations are
881 * not "stolen". The child may still get SIGKILLed
883 if (!vma_has_reserves(vma, chg) &&
884 h->free_huge_pages - h->resv_huge_pages == 0)
887 /* If reserves cannot be used, ensure enough pages are in the pool */
888 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
892 cpuset_mems_cookie = read_mems_allowed_begin();
893 zonelist = huge_zonelist(vma, address,
894 htlb_alloc_mask(h), &mpol, &nodemask);
896 for_each_zone_zonelist_nodemask(zone, z, zonelist,
897 MAX_NR_ZONES - 1, nodemask) {
898 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
899 page = dequeue_huge_page_node(h, zone_to_nid(zone));
903 if (!vma_has_reserves(vma, chg))
906 SetPagePrivate(page);
907 h->resv_huge_pages--;
914 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
923 * common helper functions for hstate_next_node_to_{alloc|free}.
924 * We may have allocated or freed a huge page based on a different
925 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
926 * be outside of *nodes_allowed. Ensure that we use an allowed
927 * node for alloc or free.
929 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
931 nid = next_node(nid, *nodes_allowed);
932 if (nid == MAX_NUMNODES)
933 nid = first_node(*nodes_allowed);
934 VM_BUG_ON(nid >= MAX_NUMNODES);
939 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
941 if (!node_isset(nid, *nodes_allowed))
942 nid = next_node_allowed(nid, nodes_allowed);
947 * returns the previously saved node ["this node"] from which to
948 * allocate a persistent huge page for the pool and advance the
949 * next node from which to allocate, handling wrap at end of node
952 static int hstate_next_node_to_alloc(struct hstate *h,
953 nodemask_t *nodes_allowed)
957 VM_BUG_ON(!nodes_allowed);
959 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
960 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
966 * helper for free_pool_huge_page() - return the previously saved
967 * node ["this node"] from which to free a huge page. Advance the
968 * next node id whether or not we find a free huge page to free so
969 * that the next attempt to free addresses the next node.
971 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
975 VM_BUG_ON(!nodes_allowed);
977 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
978 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
983 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
984 for (nr_nodes = nodes_weight(*mask); \
986 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
989 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
990 for (nr_nodes = nodes_weight(*mask); \
992 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
995 #if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
996 static void destroy_compound_gigantic_page(struct page *page,
1000 int nr_pages = 1 << order;
1001 struct page *p = page + 1;
1003 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1005 set_page_refcounted(p);
1006 p->first_page = NULL;
1009 set_compound_order(page, 0);
1010 __ClearPageHead(page);
1013 static void free_gigantic_page(struct page *page, unsigned order)
1015 free_contig_range(page_to_pfn(page), 1 << order);
1018 static int __alloc_gigantic_page(unsigned long start_pfn,
1019 unsigned long nr_pages)
1021 unsigned long end_pfn = start_pfn + nr_pages;
1022 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1025 static bool pfn_range_valid_gigantic(unsigned long start_pfn,
1026 unsigned long nr_pages)
1028 unsigned long i, end_pfn = start_pfn + nr_pages;
1031 for (i = start_pfn; i < end_pfn; i++) {
1035 page = pfn_to_page(i);
1037 if (PageReserved(page))
1040 if (page_count(page) > 0)
1050 static bool zone_spans_last_pfn(const struct zone *zone,
1051 unsigned long start_pfn, unsigned long nr_pages)
1053 unsigned long last_pfn = start_pfn + nr_pages - 1;
1054 return zone_spans_pfn(zone, last_pfn);
1057 static struct page *alloc_gigantic_page(int nid, unsigned order)
1059 unsigned long nr_pages = 1 << order;
1060 unsigned long ret, pfn, flags;
1063 z = NODE_DATA(nid)->node_zones;
1064 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
1065 spin_lock_irqsave(&z->lock, flags);
1067 pfn = ALIGN(z->zone_start_pfn, nr_pages);
1068 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
1069 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
1071 * We release the zone lock here because
1072 * alloc_contig_range() will also lock the zone
1073 * at some point. If there's an allocation
1074 * spinning on this lock, it may win the race
1075 * and cause alloc_contig_range() to fail...
1077 spin_unlock_irqrestore(&z->lock, flags);
1078 ret = __alloc_gigantic_page(pfn, nr_pages);
1080 return pfn_to_page(pfn);
1081 spin_lock_irqsave(&z->lock, flags);
1086 spin_unlock_irqrestore(&z->lock, flags);
1092 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1093 static void prep_compound_gigantic_page(struct page *page, unsigned long order);
1095 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1099 page = alloc_gigantic_page(nid, huge_page_order(h));
1101 prep_compound_gigantic_page(page, huge_page_order(h));
1102 prep_new_huge_page(h, page, nid);
1108 static int alloc_fresh_gigantic_page(struct hstate *h,
1109 nodemask_t *nodes_allowed)
1111 struct page *page = NULL;
1114 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1115 page = alloc_fresh_gigantic_page_node(h, node);
1123 static inline bool gigantic_page_supported(void) { return true; }
1125 static inline bool gigantic_page_supported(void) { return false; }
1126 static inline void free_gigantic_page(struct page *page, unsigned order) { }
1127 static inline void destroy_compound_gigantic_page(struct page *page,
1128 unsigned long order) { }
1129 static inline int alloc_fresh_gigantic_page(struct hstate *h,
1130 nodemask_t *nodes_allowed) { return 0; }
1133 static void update_and_free_page(struct hstate *h, struct page *page)
1137 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1141 h->nr_huge_pages_node[page_to_nid(page)]--;
1142 for (i = 0; i < pages_per_huge_page(h); i++) {
1143 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1144 1 << PG_referenced | 1 << PG_dirty |
1145 1 << PG_active | 1 << PG_private |
1148 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1149 set_compound_page_dtor(page, NULL);
1150 set_page_refcounted(page);
1151 if (hstate_is_gigantic(h)) {
1152 destroy_compound_gigantic_page(page, huge_page_order(h));
1153 free_gigantic_page(page, huge_page_order(h));
1155 __free_pages(page, huge_page_order(h));
1159 struct hstate *size_to_hstate(unsigned long size)
1163 for_each_hstate(h) {
1164 if (huge_page_size(h) == size)
1171 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1172 * to hstate->hugepage_activelist.)
1174 * This function can be called for tail pages, but never returns true for them.
1176 bool page_huge_active(struct page *page)
1178 VM_BUG_ON_PAGE(!PageHuge(page), page);
1179 return PageHead(page) && PagePrivate(&page[1]);
1182 /* never called for tail page */
1183 static void set_page_huge_active(struct page *page)
1185 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1186 SetPagePrivate(&page[1]);
1189 static void clear_page_huge_active(struct page *page)
1191 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1192 ClearPagePrivate(&page[1]);
1195 void free_huge_page(struct page *page)
1198 * Can't pass hstate in here because it is called from the
1199 * compound page destructor.
1201 struct hstate *h = page_hstate(page);
1202 int nid = page_to_nid(page);
1203 struct hugepage_subpool *spool =
1204 (struct hugepage_subpool *)page_private(page);
1205 bool restore_reserve;
1207 set_page_private(page, 0);
1208 page->mapping = NULL;
1209 BUG_ON(page_count(page));
1210 BUG_ON(page_mapcount(page));
1211 restore_reserve = PagePrivate(page);
1212 ClearPagePrivate(page);
1215 * A return code of zero implies that the subpool will be under its
1216 * minimum size if the reservation is not restored after page is free.
1217 * Therefore, force restore_reserve operation.
1219 if (hugepage_subpool_put_pages(spool, 1) == 0)
1220 restore_reserve = true;
1222 spin_lock(&hugetlb_lock);
1223 clear_page_huge_active(page);
1224 hugetlb_cgroup_uncharge_page(hstate_index(h),
1225 pages_per_huge_page(h), page);
1226 if (restore_reserve)
1227 h->resv_huge_pages++;
1229 if (h->surplus_huge_pages_node[nid]) {
1230 /* remove the page from active list */
1231 list_del(&page->lru);
1232 update_and_free_page(h, page);
1233 h->surplus_huge_pages--;
1234 h->surplus_huge_pages_node[nid]--;
1236 arch_clear_hugepage_flags(page);
1237 enqueue_huge_page(h, page);
1239 spin_unlock(&hugetlb_lock);
1242 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1244 INIT_LIST_HEAD(&page->lru);
1245 set_compound_page_dtor(page, free_huge_page);
1246 spin_lock(&hugetlb_lock);
1247 set_hugetlb_cgroup(page, NULL);
1249 h->nr_huge_pages_node[nid]++;
1250 spin_unlock(&hugetlb_lock);
1251 put_page(page); /* free it into the hugepage allocator */
1254 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
1257 int nr_pages = 1 << order;
1258 struct page *p = page + 1;
1260 /* we rely on prep_new_huge_page to set the destructor */
1261 set_compound_order(page, order);
1262 __SetPageHead(page);
1263 __ClearPageReserved(page);
1264 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1266 * For gigantic hugepages allocated through bootmem at
1267 * boot, it's safer to be consistent with the not-gigantic
1268 * hugepages and clear the PG_reserved bit from all tail pages
1269 * too. Otherwse drivers using get_user_pages() to access tail
1270 * pages may get the reference counting wrong if they see
1271 * PG_reserved set on a tail page (despite the head page not
1272 * having PG_reserved set). Enforcing this consistency between
1273 * head and tail pages allows drivers to optimize away a check
1274 * on the head page when they need know if put_page() is needed
1275 * after get_user_pages().
1277 __ClearPageReserved(p);
1278 set_page_count(p, 0);
1279 p->first_page = page;
1280 /* Make sure p->first_page is always valid for PageTail() */
1287 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1288 * transparent huge pages. See the PageTransHuge() documentation for more
1291 int PageHuge(struct page *page)
1293 if (!PageCompound(page))
1296 page = compound_head(page);
1297 return get_compound_page_dtor(page) == free_huge_page;
1299 EXPORT_SYMBOL_GPL(PageHuge);
1302 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1303 * normal or transparent huge pages.
1305 int PageHeadHuge(struct page *page_head)
1307 if (!PageHead(page_head))
1310 return get_compound_page_dtor(page_head) == free_huge_page;
1313 pgoff_t __basepage_index(struct page *page)
1315 struct page *page_head = compound_head(page);
1316 pgoff_t index = page_index(page_head);
1317 unsigned long compound_idx;
1319 if (!PageHuge(page_head))
1320 return page_index(page);
1322 if (compound_order(page_head) >= MAX_ORDER)
1323 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1325 compound_idx = page - page_head;
1327 return (index << compound_order(page_head)) + compound_idx;
1330 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1334 page = __alloc_pages_node(nid,
1335 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1336 __GFP_REPEAT|__GFP_NOWARN,
1337 huge_page_order(h));
1339 prep_new_huge_page(h, page, nid);
1345 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1351 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1352 page = alloc_fresh_huge_page_node(h, node);
1360 count_vm_event(HTLB_BUDDY_PGALLOC);
1362 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1368 * Free huge page from pool from next node to free.
1369 * Attempt to keep persistent huge pages more or less
1370 * balanced over allowed nodes.
1371 * Called with hugetlb_lock locked.
1373 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1379 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1381 * If we're returning unused surplus pages, only examine
1382 * nodes with surplus pages.
1384 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1385 !list_empty(&h->hugepage_freelists[node])) {
1387 list_entry(h->hugepage_freelists[node].next,
1389 list_del(&page->lru);
1390 h->free_huge_pages--;
1391 h->free_huge_pages_node[node]--;
1393 h->surplus_huge_pages--;
1394 h->surplus_huge_pages_node[node]--;
1396 update_and_free_page(h, page);
1406 * Dissolve a given free hugepage into free buddy pages. This function does
1407 * nothing for in-use (including surplus) hugepages.
1409 static void dissolve_free_huge_page(struct page *page)
1411 spin_lock(&hugetlb_lock);
1412 if (PageHuge(page) && !page_count(page)) {
1413 struct hstate *h = page_hstate(page);
1414 int nid = page_to_nid(page);
1415 list_del(&page->lru);
1416 h->free_huge_pages--;
1417 h->free_huge_pages_node[nid]--;
1418 update_and_free_page(h, page);
1420 spin_unlock(&hugetlb_lock);
1424 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1425 * make specified memory blocks removable from the system.
1426 * Note that start_pfn should aligned with (minimum) hugepage size.
1428 void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1432 if (!hugepages_supported())
1435 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order));
1436 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
1437 dissolve_free_huge_page(pfn_to_page(pfn));
1441 * There are 3 ways this can get called:
1442 * 1. With vma+addr: we use the VMA's memory policy
1443 * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge
1444 * page from any node, and let the buddy allocator itself figure
1446 * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page
1447 * strictly from 'nid'
1449 static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1450 struct vm_area_struct *vma, unsigned long addr, int nid)
1452 int order = huge_page_order(h);
1453 gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN;
1454 unsigned int cpuset_mems_cookie;
1457 * We need a VMA to get a memory policy. If we do not
1458 * have one, we use the 'nid' argument.
1460 * The mempolicy stuff below has some non-inlined bits
1461 * and calls ->vm_ops. That makes it hard to optimize at
1462 * compile-time, even when NUMA is off and it does
1463 * nothing. This helps the compiler optimize it out.
1465 if (!IS_ENABLED(CONFIG_NUMA) || !vma) {
1467 * If a specific node is requested, make sure to
1468 * get memory from there, but only when a node
1469 * is explicitly specified.
1471 if (nid != NUMA_NO_NODE)
1472 gfp |= __GFP_THISNODE;
1474 * Make sure to call something that can handle
1477 return alloc_pages_node(nid, gfp, order);
1481 * OK, so we have a VMA. Fetch the mempolicy and try to
1482 * allocate a huge page with it. We will only reach this
1483 * when CONFIG_NUMA=y.
1487 struct mempolicy *mpol;
1488 struct zonelist *zl;
1489 nodemask_t *nodemask;
1491 cpuset_mems_cookie = read_mems_allowed_begin();
1492 zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask);
1493 mpol_cond_put(mpol);
1494 page = __alloc_pages_nodemask(gfp, order, zl, nodemask);
1497 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1503 * There are two ways to allocate a huge page:
1504 * 1. When you have a VMA and an address (like a fault)
1505 * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1507 * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in
1508 * this case which signifies that the allocation should be done with
1509 * respect for the VMA's memory policy.
1511 * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1512 * implies that memory policies will not be taken in to account.
1514 static struct page *__alloc_buddy_huge_page(struct hstate *h,
1515 struct vm_area_struct *vma, unsigned long addr, int nid)
1520 if (hstate_is_gigantic(h))
1524 * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1525 * This makes sure the caller is picking _one_ of the modes with which
1526 * we can call this function, not both.
1528 if (vma || (addr != -1)) {
1529 VM_WARN_ON_ONCE(addr == -1);
1530 VM_WARN_ON_ONCE(nid != NUMA_NO_NODE);
1533 * Assume we will successfully allocate the surplus page to
1534 * prevent racing processes from causing the surplus to exceed
1537 * This however introduces a different race, where a process B
1538 * tries to grow the static hugepage pool while alloc_pages() is
1539 * called by process A. B will only examine the per-node
1540 * counters in determining if surplus huge pages can be
1541 * converted to normal huge pages in adjust_pool_surplus(). A
1542 * won't be able to increment the per-node counter, until the
1543 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1544 * no more huge pages can be converted from surplus to normal
1545 * state (and doesn't try to convert again). Thus, we have a
1546 * case where a surplus huge page exists, the pool is grown, and
1547 * the surplus huge page still exists after, even though it
1548 * should just have been converted to a normal huge page. This
1549 * does not leak memory, though, as the hugepage will be freed
1550 * once it is out of use. It also does not allow the counters to
1551 * go out of whack in adjust_pool_surplus() as we don't modify
1552 * the node values until we've gotten the hugepage and only the
1553 * per-node value is checked there.
1555 spin_lock(&hugetlb_lock);
1556 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1557 spin_unlock(&hugetlb_lock);
1561 h->surplus_huge_pages++;
1563 spin_unlock(&hugetlb_lock);
1565 page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid);
1567 spin_lock(&hugetlb_lock);
1569 INIT_LIST_HEAD(&page->lru);
1570 r_nid = page_to_nid(page);
1571 set_compound_page_dtor(page, free_huge_page);
1572 set_hugetlb_cgroup(page, NULL);
1574 * We incremented the global counters already
1576 h->nr_huge_pages_node[r_nid]++;
1577 h->surplus_huge_pages_node[r_nid]++;
1578 __count_vm_event(HTLB_BUDDY_PGALLOC);
1581 h->surplus_huge_pages--;
1582 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1584 spin_unlock(&hugetlb_lock);
1590 * Allocate a huge page from 'nid'. Note, 'nid' may be
1591 * NUMA_NO_NODE, which means that it may be allocated
1595 struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid)
1597 unsigned long addr = -1;
1599 return __alloc_buddy_huge_page(h, NULL, addr, nid);
1603 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1606 struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1607 struct vm_area_struct *vma, unsigned long addr)
1609 return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE);
1613 * This allocation function is useful in the context where vma is irrelevant.
1614 * E.g. soft-offlining uses this function because it only cares physical
1615 * address of error page.
1617 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1619 struct page *page = NULL;
1621 spin_lock(&hugetlb_lock);
1622 if (h->free_huge_pages - h->resv_huge_pages > 0)
1623 page = dequeue_huge_page_node(h, nid);
1624 spin_unlock(&hugetlb_lock);
1627 page = __alloc_buddy_huge_page_no_mpol(h, nid);
1633 * Increase the hugetlb pool such that it can accommodate a reservation
1636 static int gather_surplus_pages(struct hstate *h, int delta)
1638 struct list_head surplus_list;
1639 struct page *page, *tmp;
1641 int needed, allocated;
1642 bool alloc_ok = true;
1644 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1646 h->resv_huge_pages += delta;
1651 INIT_LIST_HEAD(&surplus_list);
1655 spin_unlock(&hugetlb_lock);
1656 for (i = 0; i < needed; i++) {
1657 page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE);
1662 list_add(&page->lru, &surplus_list);
1667 * After retaking hugetlb_lock, we need to recalculate 'needed'
1668 * because either resv_huge_pages or free_huge_pages may have changed.
1670 spin_lock(&hugetlb_lock);
1671 needed = (h->resv_huge_pages + delta) -
1672 (h->free_huge_pages + allocated);
1677 * We were not able to allocate enough pages to
1678 * satisfy the entire reservation so we free what
1679 * we've allocated so far.
1684 * The surplus_list now contains _at_least_ the number of extra pages
1685 * needed to accommodate the reservation. Add the appropriate number
1686 * of pages to the hugetlb pool and free the extras back to the buddy
1687 * allocator. Commit the entire reservation here to prevent another
1688 * process from stealing the pages as they are added to the pool but
1689 * before they are reserved.
1691 needed += allocated;
1692 h->resv_huge_pages += delta;
1695 /* Free the needed pages to the hugetlb pool */
1696 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1700 * This page is now managed by the hugetlb allocator and has
1701 * no users -- drop the buddy allocator's reference.
1703 put_page_testzero(page);
1704 VM_BUG_ON_PAGE(page_count(page), page);
1705 enqueue_huge_page(h, page);
1708 spin_unlock(&hugetlb_lock);
1710 /* Free unnecessary surplus pages to the buddy allocator */
1711 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1713 spin_lock(&hugetlb_lock);
1719 * When releasing a hugetlb pool reservation, any surplus pages that were
1720 * allocated to satisfy the reservation must be explicitly freed if they were
1722 * Called with hugetlb_lock held.
1724 static void return_unused_surplus_pages(struct hstate *h,
1725 unsigned long unused_resv_pages)
1727 unsigned long nr_pages;
1729 /* Uncommit the reservation */
1730 h->resv_huge_pages -= unused_resv_pages;
1732 /* Cannot return gigantic pages currently */
1733 if (hstate_is_gigantic(h))
1736 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1739 * We want to release as many surplus pages as possible, spread
1740 * evenly across all nodes with memory. Iterate across these nodes
1741 * until we can no longer free unreserved surplus pages. This occurs
1742 * when the nodes with surplus pages have no free pages.
1743 * free_pool_huge_page() will balance the the freed pages across the
1744 * on-line nodes with memory and will handle the hstate accounting.
1746 while (nr_pages--) {
1747 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1749 cond_resched_lock(&hugetlb_lock);
1755 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1756 * are used by the huge page allocation routines to manage reservations.
1758 * vma_needs_reservation is called to determine if the huge page at addr
1759 * within the vma has an associated reservation. If a reservation is
1760 * needed, the value 1 is returned. The caller is then responsible for
1761 * managing the global reservation and subpool usage counts. After
1762 * the huge page has been allocated, vma_commit_reservation is called
1763 * to add the page to the reservation map. If the page allocation fails,
1764 * the reservation must be ended instead of committed. vma_end_reservation
1765 * is called in such cases.
1767 * In the normal case, vma_commit_reservation returns the same value
1768 * as the preceding vma_needs_reservation call. The only time this
1769 * is not the case is if a reserve map was changed between calls. It
1770 * is the responsibility of the caller to notice the difference and
1771 * take appropriate action.
1773 enum vma_resv_mode {
1778 static long __vma_reservation_common(struct hstate *h,
1779 struct vm_area_struct *vma, unsigned long addr,
1780 enum vma_resv_mode mode)
1782 struct resv_map *resv;
1786 resv = vma_resv_map(vma);
1790 idx = vma_hugecache_offset(h, vma, addr);
1792 case VMA_NEEDS_RESV:
1793 ret = region_chg(resv, idx, idx + 1);
1795 case VMA_COMMIT_RESV:
1796 ret = region_add(resv, idx, idx + 1);
1799 region_abort(resv, idx, idx + 1);
1806 if (vma->vm_flags & VM_MAYSHARE)
1809 return ret < 0 ? ret : 0;
1812 static long vma_needs_reservation(struct hstate *h,
1813 struct vm_area_struct *vma, unsigned long addr)
1815 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1818 static long vma_commit_reservation(struct hstate *h,
1819 struct vm_area_struct *vma, unsigned long addr)
1821 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1824 static void vma_end_reservation(struct hstate *h,
1825 struct vm_area_struct *vma, unsigned long addr)
1827 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1830 struct page *alloc_huge_page(struct vm_area_struct *vma,
1831 unsigned long addr, int avoid_reserve)
1833 struct hugepage_subpool *spool = subpool_vma(vma);
1834 struct hstate *h = hstate_vma(vma);
1836 long map_chg, map_commit;
1839 struct hugetlb_cgroup *h_cg;
1841 idx = hstate_index(h);
1843 * Examine the region/reserve map to determine if the process
1844 * has a reservation for the page to be allocated. A return
1845 * code of zero indicates a reservation exists (no change).
1847 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
1849 return ERR_PTR(-ENOMEM);
1852 * Processes that did not create the mapping will have no
1853 * reserves as indicated by the region/reserve map. Check
1854 * that the allocation will not exceed the subpool limit.
1855 * Allocations for MAP_NORESERVE mappings also need to be
1856 * checked against any subpool limit.
1858 if (map_chg || avoid_reserve) {
1859 gbl_chg = hugepage_subpool_get_pages(spool, 1);
1861 vma_end_reservation(h, vma, addr);
1862 return ERR_PTR(-ENOSPC);
1866 * Even though there was no reservation in the region/reserve
1867 * map, there could be reservations associated with the
1868 * subpool that can be used. This would be indicated if the
1869 * return value of hugepage_subpool_get_pages() is zero.
1870 * However, if avoid_reserve is specified we still avoid even
1871 * the subpool reservations.
1877 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1879 goto out_subpool_put;
1881 spin_lock(&hugetlb_lock);
1883 * glb_chg is passed to indicate whether or not a page must be taken
1884 * from the global free pool (global change). gbl_chg == 0 indicates
1885 * a reservation exists for the allocation.
1887 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
1889 spin_unlock(&hugetlb_lock);
1890 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
1892 goto out_uncharge_cgroup;
1894 spin_lock(&hugetlb_lock);
1895 list_move(&page->lru, &h->hugepage_activelist);
1898 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1899 spin_unlock(&hugetlb_lock);
1901 set_page_private(page, (unsigned long)spool);
1903 map_commit = vma_commit_reservation(h, vma, addr);
1904 if (unlikely(map_chg > map_commit)) {
1906 * The page was added to the reservation map between
1907 * vma_needs_reservation and vma_commit_reservation.
1908 * This indicates a race with hugetlb_reserve_pages.
1909 * Adjust for the subpool count incremented above AND
1910 * in hugetlb_reserve_pages for the same page. Also,
1911 * the reservation count added in hugetlb_reserve_pages
1912 * no longer applies.
1916 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
1917 hugetlb_acct_memory(h, -rsv_adjust);
1921 out_uncharge_cgroup:
1922 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1924 if (map_chg || avoid_reserve)
1925 hugepage_subpool_put_pages(spool, 1);
1926 vma_end_reservation(h, vma, addr);
1927 return ERR_PTR(-ENOSPC);
1931 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1932 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1933 * where no ERR_VALUE is expected to be returned.
1935 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1936 unsigned long addr, int avoid_reserve)
1938 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1944 int __weak alloc_bootmem_huge_page(struct hstate *h)
1946 struct huge_bootmem_page *m;
1949 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1952 addr = memblock_virt_alloc_try_nid_nopanic(
1953 huge_page_size(h), huge_page_size(h),
1954 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1957 * Use the beginning of the huge page to store the
1958 * huge_bootmem_page struct (until gather_bootmem
1959 * puts them into the mem_map).
1968 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
1969 /* Put them into a private list first because mem_map is not up yet */
1970 list_add(&m->list, &huge_boot_pages);
1975 static void __init prep_compound_huge_page(struct page *page, int order)
1977 if (unlikely(order > (MAX_ORDER - 1)))
1978 prep_compound_gigantic_page(page, order);
1980 prep_compound_page(page, order);
1983 /* Put bootmem huge pages into the standard lists after mem_map is up */
1984 static void __init gather_bootmem_prealloc(void)
1986 struct huge_bootmem_page *m;
1988 list_for_each_entry(m, &huge_boot_pages, list) {
1989 struct hstate *h = m->hstate;
1992 #ifdef CONFIG_HIGHMEM
1993 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1994 memblock_free_late(__pa(m),
1995 sizeof(struct huge_bootmem_page));
1997 page = virt_to_page(m);
1999 WARN_ON(page_count(page) != 1);
2000 prep_compound_huge_page(page, h->order);
2001 WARN_ON(PageReserved(page));
2002 prep_new_huge_page(h, page, page_to_nid(page));
2004 * If we had gigantic hugepages allocated at boot time, we need
2005 * to restore the 'stolen' pages to totalram_pages in order to
2006 * fix confusing memory reports from free(1) and another
2007 * side-effects, like CommitLimit going negative.
2009 if (hstate_is_gigantic(h))
2010 adjust_managed_page_count(page, 1 << h->order);
2014 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2018 for (i = 0; i < h->max_huge_pages; ++i) {
2019 if (hstate_is_gigantic(h)) {
2020 if (!alloc_bootmem_huge_page(h))
2022 } else if (!alloc_fresh_huge_page(h,
2023 &node_states[N_MEMORY]))
2026 h->max_huge_pages = i;
2029 static void __init hugetlb_init_hstates(void)
2033 for_each_hstate(h) {
2034 if (minimum_order > huge_page_order(h))
2035 minimum_order = huge_page_order(h);
2037 /* oversize hugepages were init'ed in early boot */
2038 if (!hstate_is_gigantic(h))
2039 hugetlb_hstate_alloc_pages(h);
2041 VM_BUG_ON(minimum_order == UINT_MAX);
2044 static char * __init memfmt(char *buf, unsigned long n)
2046 if (n >= (1UL << 30))
2047 sprintf(buf, "%lu GB", n >> 30);
2048 else if (n >= (1UL << 20))
2049 sprintf(buf, "%lu MB", n >> 20);
2051 sprintf(buf, "%lu KB", n >> 10);
2055 static void __init report_hugepages(void)
2059 for_each_hstate(h) {
2061 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2062 memfmt(buf, huge_page_size(h)),
2063 h->free_huge_pages);
2067 #ifdef CONFIG_HIGHMEM
2068 static void try_to_free_low(struct hstate *h, unsigned long count,
2069 nodemask_t *nodes_allowed)
2073 if (hstate_is_gigantic(h))
2076 for_each_node_mask(i, *nodes_allowed) {
2077 struct page *page, *next;
2078 struct list_head *freel = &h->hugepage_freelists[i];
2079 list_for_each_entry_safe(page, next, freel, lru) {
2080 if (count >= h->nr_huge_pages)
2082 if (PageHighMem(page))
2084 list_del(&page->lru);
2085 update_and_free_page(h, page);
2086 h->free_huge_pages--;
2087 h->free_huge_pages_node[page_to_nid(page)]--;
2092 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2093 nodemask_t *nodes_allowed)
2099 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2100 * balanced by operating on them in a round-robin fashion.
2101 * Returns 1 if an adjustment was made.
2103 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2108 VM_BUG_ON(delta != -1 && delta != 1);
2111 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2112 if (h->surplus_huge_pages_node[node])
2116 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2117 if (h->surplus_huge_pages_node[node] <
2118 h->nr_huge_pages_node[node])
2125 h->surplus_huge_pages += delta;
2126 h->surplus_huge_pages_node[node] += delta;
2130 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2131 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2132 nodemask_t *nodes_allowed)
2134 unsigned long min_count, ret;
2136 if (hstate_is_gigantic(h) && !gigantic_page_supported())
2137 return h->max_huge_pages;
2140 * Increase the pool size
2141 * First take pages out of surplus state. Then make up the
2142 * remaining difference by allocating fresh huge pages.
2144 * We might race with alloc_buddy_huge_page() here and be unable
2145 * to convert a surplus huge page to a normal huge page. That is
2146 * not critical, though, it just means the overall size of the
2147 * pool might be one hugepage larger than it needs to be, but
2148 * within all the constraints specified by the sysctls.
2150 spin_lock(&hugetlb_lock);
2151 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2152 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2156 while (count > persistent_huge_pages(h)) {
2158 * If this allocation races such that we no longer need the
2159 * page, free_huge_page will handle it by freeing the page
2160 * and reducing the surplus.
2162 spin_unlock(&hugetlb_lock);
2163 if (hstate_is_gigantic(h))
2164 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2166 ret = alloc_fresh_huge_page(h, nodes_allowed);
2167 spin_lock(&hugetlb_lock);
2171 /* Bail for signals. Probably ctrl-c from user */
2172 if (signal_pending(current))
2177 * Decrease the pool size
2178 * First return free pages to the buddy allocator (being careful
2179 * to keep enough around to satisfy reservations). Then place
2180 * pages into surplus state as needed so the pool will shrink
2181 * to the desired size as pages become free.
2183 * By placing pages into the surplus state independent of the
2184 * overcommit value, we are allowing the surplus pool size to
2185 * exceed overcommit. There are few sane options here. Since
2186 * alloc_buddy_huge_page() is checking the global counter,
2187 * though, we'll note that we're not allowed to exceed surplus
2188 * and won't grow the pool anywhere else. Not until one of the
2189 * sysctls are changed, or the surplus pages go out of use.
2191 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2192 min_count = max(count, min_count);
2193 try_to_free_low(h, min_count, nodes_allowed);
2194 while (min_count < persistent_huge_pages(h)) {
2195 if (!free_pool_huge_page(h, nodes_allowed, 0))
2197 cond_resched_lock(&hugetlb_lock);
2199 while (count < persistent_huge_pages(h)) {
2200 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2204 ret = persistent_huge_pages(h);
2205 spin_unlock(&hugetlb_lock);
2209 #define HSTATE_ATTR_RO(_name) \
2210 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2212 #define HSTATE_ATTR(_name) \
2213 static struct kobj_attribute _name##_attr = \
2214 __ATTR(_name, 0644, _name##_show, _name##_store)
2216 static struct kobject *hugepages_kobj;
2217 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2219 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2221 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2225 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2226 if (hstate_kobjs[i] == kobj) {
2228 *nidp = NUMA_NO_NODE;
2232 return kobj_to_node_hstate(kobj, nidp);
2235 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2236 struct kobj_attribute *attr, char *buf)
2239 unsigned long nr_huge_pages;
2242 h = kobj_to_hstate(kobj, &nid);
2243 if (nid == NUMA_NO_NODE)
2244 nr_huge_pages = h->nr_huge_pages;
2246 nr_huge_pages = h->nr_huge_pages_node[nid];
2248 return sprintf(buf, "%lu\n", nr_huge_pages);
2251 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2252 struct hstate *h, int nid,
2253 unsigned long count, size_t len)
2256 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2258 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2263 if (nid == NUMA_NO_NODE) {
2265 * global hstate attribute
2267 if (!(obey_mempolicy &&
2268 init_nodemask_of_mempolicy(nodes_allowed))) {
2269 NODEMASK_FREE(nodes_allowed);
2270 nodes_allowed = &node_states[N_MEMORY];
2272 } else if (nodes_allowed) {
2274 * per node hstate attribute: adjust count to global,
2275 * but restrict alloc/free to the specified node.
2277 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2278 init_nodemask_of_node(nodes_allowed, nid);
2280 nodes_allowed = &node_states[N_MEMORY];
2282 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2284 if (nodes_allowed != &node_states[N_MEMORY])
2285 NODEMASK_FREE(nodes_allowed);
2289 NODEMASK_FREE(nodes_allowed);
2293 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2294 struct kobject *kobj, const char *buf,
2298 unsigned long count;
2302 err = kstrtoul(buf, 10, &count);
2306 h = kobj_to_hstate(kobj, &nid);
2307 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2310 static ssize_t nr_hugepages_show(struct kobject *kobj,
2311 struct kobj_attribute *attr, char *buf)
2313 return nr_hugepages_show_common(kobj, attr, buf);
2316 static ssize_t nr_hugepages_store(struct kobject *kobj,
2317 struct kobj_attribute *attr, const char *buf, size_t len)
2319 return nr_hugepages_store_common(false, kobj, buf, len);
2321 HSTATE_ATTR(nr_hugepages);
2326 * hstate attribute for optionally mempolicy-based constraint on persistent
2327 * huge page alloc/free.
2329 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2330 struct kobj_attribute *attr, char *buf)
2332 return nr_hugepages_show_common(kobj, attr, buf);
2335 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2336 struct kobj_attribute *attr, const char *buf, size_t len)
2338 return nr_hugepages_store_common(true, kobj, buf, len);
2340 HSTATE_ATTR(nr_hugepages_mempolicy);
2344 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2345 struct kobj_attribute *attr, char *buf)
2347 struct hstate *h = kobj_to_hstate(kobj, NULL);
2348 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2351 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2352 struct kobj_attribute *attr, const char *buf, size_t count)
2355 unsigned long input;
2356 struct hstate *h = kobj_to_hstate(kobj, NULL);
2358 if (hstate_is_gigantic(h))
2361 err = kstrtoul(buf, 10, &input);
2365 spin_lock(&hugetlb_lock);
2366 h->nr_overcommit_huge_pages = input;
2367 spin_unlock(&hugetlb_lock);
2371 HSTATE_ATTR(nr_overcommit_hugepages);
2373 static ssize_t free_hugepages_show(struct kobject *kobj,
2374 struct kobj_attribute *attr, char *buf)
2377 unsigned long free_huge_pages;
2380 h = kobj_to_hstate(kobj, &nid);
2381 if (nid == NUMA_NO_NODE)
2382 free_huge_pages = h->free_huge_pages;
2384 free_huge_pages = h->free_huge_pages_node[nid];
2386 return sprintf(buf, "%lu\n", free_huge_pages);
2388 HSTATE_ATTR_RO(free_hugepages);
2390 static ssize_t resv_hugepages_show(struct kobject *kobj,
2391 struct kobj_attribute *attr, char *buf)
2393 struct hstate *h = kobj_to_hstate(kobj, NULL);
2394 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2396 HSTATE_ATTR_RO(resv_hugepages);
2398 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2399 struct kobj_attribute *attr, char *buf)
2402 unsigned long surplus_huge_pages;
2405 h = kobj_to_hstate(kobj, &nid);
2406 if (nid == NUMA_NO_NODE)
2407 surplus_huge_pages = h->surplus_huge_pages;
2409 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2411 return sprintf(buf, "%lu\n", surplus_huge_pages);
2413 HSTATE_ATTR_RO(surplus_hugepages);
2415 static struct attribute *hstate_attrs[] = {
2416 &nr_hugepages_attr.attr,
2417 &nr_overcommit_hugepages_attr.attr,
2418 &free_hugepages_attr.attr,
2419 &resv_hugepages_attr.attr,
2420 &surplus_hugepages_attr.attr,
2422 &nr_hugepages_mempolicy_attr.attr,
2427 static struct attribute_group hstate_attr_group = {
2428 .attrs = hstate_attrs,
2431 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2432 struct kobject **hstate_kobjs,
2433 struct attribute_group *hstate_attr_group)
2436 int hi = hstate_index(h);
2438 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2439 if (!hstate_kobjs[hi])
2442 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2444 kobject_put(hstate_kobjs[hi]);
2449 static void __init hugetlb_sysfs_init(void)
2454 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2455 if (!hugepages_kobj)
2458 for_each_hstate(h) {
2459 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2460 hstate_kobjs, &hstate_attr_group);
2462 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2469 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2470 * with node devices in node_devices[] using a parallel array. The array
2471 * index of a node device or _hstate == node id.
2472 * This is here to avoid any static dependency of the node device driver, in
2473 * the base kernel, on the hugetlb module.
2475 struct node_hstate {
2476 struct kobject *hugepages_kobj;
2477 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2479 static struct node_hstate node_hstates[MAX_NUMNODES];
2482 * A subset of global hstate attributes for node devices
2484 static struct attribute *per_node_hstate_attrs[] = {
2485 &nr_hugepages_attr.attr,
2486 &free_hugepages_attr.attr,
2487 &surplus_hugepages_attr.attr,
2491 static struct attribute_group per_node_hstate_attr_group = {
2492 .attrs = per_node_hstate_attrs,
2496 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2497 * Returns node id via non-NULL nidp.
2499 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2503 for (nid = 0; nid < nr_node_ids; nid++) {
2504 struct node_hstate *nhs = &node_hstates[nid];
2506 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2507 if (nhs->hstate_kobjs[i] == kobj) {
2519 * Unregister hstate attributes from a single node device.
2520 * No-op if no hstate attributes attached.
2522 static void hugetlb_unregister_node(struct node *node)
2525 struct node_hstate *nhs = &node_hstates[node->dev.id];
2527 if (!nhs->hugepages_kobj)
2528 return; /* no hstate attributes */
2530 for_each_hstate(h) {
2531 int idx = hstate_index(h);
2532 if (nhs->hstate_kobjs[idx]) {
2533 kobject_put(nhs->hstate_kobjs[idx]);
2534 nhs->hstate_kobjs[idx] = NULL;
2538 kobject_put(nhs->hugepages_kobj);
2539 nhs->hugepages_kobj = NULL;
2543 * hugetlb module exit: unregister hstate attributes from node devices
2546 static void hugetlb_unregister_all_nodes(void)
2551 * disable node device registrations.
2553 register_hugetlbfs_with_node(NULL, NULL);
2556 * remove hstate attributes from any nodes that have them.
2558 for (nid = 0; nid < nr_node_ids; nid++)
2559 hugetlb_unregister_node(node_devices[nid]);
2563 * Register hstate attributes for a single node device.
2564 * No-op if attributes already registered.
2566 static void hugetlb_register_node(struct node *node)
2569 struct node_hstate *nhs = &node_hstates[node->dev.id];
2572 if (nhs->hugepages_kobj)
2573 return; /* already allocated */
2575 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2577 if (!nhs->hugepages_kobj)
2580 for_each_hstate(h) {
2581 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2583 &per_node_hstate_attr_group);
2585 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2586 h->name, node->dev.id);
2587 hugetlb_unregister_node(node);
2594 * hugetlb init time: register hstate attributes for all registered node
2595 * devices of nodes that have memory. All on-line nodes should have
2596 * registered their associated device by this time.
2598 static void __init hugetlb_register_all_nodes(void)
2602 for_each_node_state(nid, N_MEMORY) {
2603 struct node *node = node_devices[nid];
2604 if (node->dev.id == nid)
2605 hugetlb_register_node(node);
2609 * Let the node device driver know we're here so it can
2610 * [un]register hstate attributes on node hotplug.
2612 register_hugetlbfs_with_node(hugetlb_register_node,
2613 hugetlb_unregister_node);
2615 #else /* !CONFIG_NUMA */
2617 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2625 static void hugetlb_unregister_all_nodes(void) { }
2627 static void hugetlb_register_all_nodes(void) { }
2631 static void __exit hugetlb_exit(void)
2635 hugetlb_unregister_all_nodes();
2637 for_each_hstate(h) {
2638 kobject_put(hstate_kobjs[hstate_index(h)]);
2641 kobject_put(hugepages_kobj);
2642 kfree(hugetlb_fault_mutex_table);
2644 module_exit(hugetlb_exit);
2646 static int __init hugetlb_init(void)
2650 if (!hugepages_supported())
2653 if (!size_to_hstate(default_hstate_size)) {
2654 default_hstate_size = HPAGE_SIZE;
2655 if (!size_to_hstate(default_hstate_size))
2656 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2658 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2659 if (default_hstate_max_huge_pages)
2660 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2662 hugetlb_init_hstates();
2663 gather_bootmem_prealloc();
2666 hugetlb_sysfs_init();
2667 hugetlb_register_all_nodes();
2668 hugetlb_cgroup_file_init();
2671 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2673 num_fault_mutexes = 1;
2675 hugetlb_fault_mutex_table =
2676 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2677 BUG_ON(!hugetlb_fault_mutex_table);
2679 for (i = 0; i < num_fault_mutexes; i++)
2680 mutex_init(&hugetlb_fault_mutex_table[i]);
2683 module_init(hugetlb_init);
2685 /* Should be called on processing a hugepagesz=... option */
2686 void __init hugetlb_add_hstate(unsigned order)
2691 if (size_to_hstate(PAGE_SIZE << order)) {
2692 pr_warning("hugepagesz= specified twice, ignoring\n");
2695 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2697 h = &hstates[hugetlb_max_hstate++];
2699 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2700 h->nr_huge_pages = 0;
2701 h->free_huge_pages = 0;
2702 for (i = 0; i < MAX_NUMNODES; ++i)
2703 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2704 INIT_LIST_HEAD(&h->hugepage_activelist);
2705 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2706 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2707 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2708 huge_page_size(h)/1024);
2713 static int __init hugetlb_nrpages_setup(char *s)
2716 static unsigned long *last_mhp;
2719 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2720 * so this hugepages= parameter goes to the "default hstate".
2722 if (!hugetlb_max_hstate)
2723 mhp = &default_hstate_max_huge_pages;
2725 mhp = &parsed_hstate->max_huge_pages;
2727 if (mhp == last_mhp) {
2728 pr_warning("hugepages= specified twice without "
2729 "interleaving hugepagesz=, ignoring\n");
2733 if (sscanf(s, "%lu", mhp) <= 0)
2737 * Global state is always initialized later in hugetlb_init.
2738 * But we need to allocate >= MAX_ORDER hstates here early to still
2739 * use the bootmem allocator.
2741 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2742 hugetlb_hstate_alloc_pages(parsed_hstate);
2748 __setup("hugepages=", hugetlb_nrpages_setup);
2750 static int __init hugetlb_default_setup(char *s)
2752 default_hstate_size = memparse(s, &s);
2755 __setup("default_hugepagesz=", hugetlb_default_setup);
2757 static unsigned int cpuset_mems_nr(unsigned int *array)
2760 unsigned int nr = 0;
2762 for_each_node_mask(node, cpuset_current_mems_allowed)
2768 #ifdef CONFIG_SYSCTL
2769 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2770 struct ctl_table *table, int write,
2771 void __user *buffer, size_t *length, loff_t *ppos)
2773 struct hstate *h = &default_hstate;
2774 unsigned long tmp = h->max_huge_pages;
2777 if (!hugepages_supported())
2781 table->maxlen = sizeof(unsigned long);
2782 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2787 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2788 NUMA_NO_NODE, tmp, *length);
2793 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2794 void __user *buffer, size_t *length, loff_t *ppos)
2797 return hugetlb_sysctl_handler_common(false, table, write,
2798 buffer, length, ppos);
2802 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2803 void __user *buffer, size_t *length, loff_t *ppos)
2805 return hugetlb_sysctl_handler_common(true, table, write,
2806 buffer, length, ppos);
2808 #endif /* CONFIG_NUMA */
2810 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2811 void __user *buffer,
2812 size_t *length, loff_t *ppos)
2814 struct hstate *h = &default_hstate;
2818 if (!hugepages_supported())
2821 tmp = h->nr_overcommit_huge_pages;
2823 if (write && hstate_is_gigantic(h))
2827 table->maxlen = sizeof(unsigned long);
2828 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2833 spin_lock(&hugetlb_lock);
2834 h->nr_overcommit_huge_pages = tmp;
2835 spin_unlock(&hugetlb_lock);
2841 #endif /* CONFIG_SYSCTL */
2843 void hugetlb_report_meminfo(struct seq_file *m)
2845 struct hstate *h = &default_hstate;
2846 if (!hugepages_supported())
2849 "HugePages_Total: %5lu\n"
2850 "HugePages_Free: %5lu\n"
2851 "HugePages_Rsvd: %5lu\n"
2852 "HugePages_Surp: %5lu\n"
2853 "Hugepagesize: %8lu kB\n",
2857 h->surplus_huge_pages,
2858 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2861 int hugetlb_report_node_meminfo(int nid, char *buf)
2863 struct hstate *h = &default_hstate;
2864 if (!hugepages_supported())
2867 "Node %d HugePages_Total: %5u\n"
2868 "Node %d HugePages_Free: %5u\n"
2869 "Node %d HugePages_Surp: %5u\n",
2870 nid, h->nr_huge_pages_node[nid],
2871 nid, h->free_huge_pages_node[nid],
2872 nid, h->surplus_huge_pages_node[nid]);
2875 void hugetlb_show_meminfo(void)
2880 if (!hugepages_supported())
2883 for_each_node_state(nid, N_MEMORY)
2885 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2887 h->nr_huge_pages_node[nid],
2888 h->free_huge_pages_node[nid],
2889 h->surplus_huge_pages_node[nid],
2890 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2893 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
2895 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
2896 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
2899 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2900 unsigned long hugetlb_total_pages(void)
2903 unsigned long nr_total_pages = 0;
2906 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2907 return nr_total_pages;
2910 static int hugetlb_acct_memory(struct hstate *h, long delta)
2914 spin_lock(&hugetlb_lock);
2916 * When cpuset is configured, it breaks the strict hugetlb page
2917 * reservation as the accounting is done on a global variable. Such
2918 * reservation is completely rubbish in the presence of cpuset because
2919 * the reservation is not checked against page availability for the
2920 * current cpuset. Application can still potentially OOM'ed by kernel
2921 * with lack of free htlb page in cpuset that the task is in.
2922 * Attempt to enforce strict accounting with cpuset is almost
2923 * impossible (or too ugly) because cpuset is too fluid that
2924 * task or memory node can be dynamically moved between cpusets.
2926 * The change of semantics for shared hugetlb mapping with cpuset is
2927 * undesirable. However, in order to preserve some of the semantics,
2928 * we fall back to check against current free page availability as
2929 * a best attempt and hopefully to minimize the impact of changing
2930 * semantics that cpuset has.
2933 if (gather_surplus_pages(h, delta) < 0)
2936 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2937 return_unused_surplus_pages(h, delta);
2944 return_unused_surplus_pages(h, (unsigned long) -delta);
2947 spin_unlock(&hugetlb_lock);
2951 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2953 struct resv_map *resv = vma_resv_map(vma);
2956 * This new VMA should share its siblings reservation map if present.
2957 * The VMA will only ever have a valid reservation map pointer where
2958 * it is being copied for another still existing VMA. As that VMA
2959 * has a reference to the reservation map it cannot disappear until
2960 * after this open call completes. It is therefore safe to take a
2961 * new reference here without additional locking.
2963 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2964 kref_get(&resv->refs);
2967 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2969 struct hstate *h = hstate_vma(vma);
2970 struct resv_map *resv = vma_resv_map(vma);
2971 struct hugepage_subpool *spool = subpool_vma(vma);
2972 unsigned long reserve, start, end;
2975 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2978 start = vma_hugecache_offset(h, vma, vma->vm_start);
2979 end = vma_hugecache_offset(h, vma, vma->vm_end);
2981 reserve = (end - start) - region_count(resv, start, end);
2983 kref_put(&resv->refs, resv_map_release);
2987 * Decrement reserve counts. The global reserve count may be
2988 * adjusted if the subpool has a minimum size.
2990 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
2991 hugetlb_acct_memory(h, -gbl_reserve);
2996 * We cannot handle pagefaults against hugetlb pages at all. They cause
2997 * handle_mm_fault() to try to instantiate regular-sized pages in the
2998 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3001 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3007 const struct vm_operations_struct hugetlb_vm_ops = {
3008 .fault = hugetlb_vm_op_fault,
3009 .open = hugetlb_vm_op_open,
3010 .close = hugetlb_vm_op_close,
3013 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3019 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3020 vma->vm_page_prot)));
3022 entry = huge_pte_wrprotect(mk_huge_pte(page,
3023 vma->vm_page_prot));
3025 entry = pte_mkyoung(entry);
3026 entry = pte_mkhuge(entry);
3027 entry = arch_make_huge_pte(entry, vma, page, writable);
3032 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3033 unsigned long address, pte_t *ptep)
3037 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3038 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3039 update_mmu_cache(vma, address, ptep);
3042 static int is_hugetlb_entry_migration(pte_t pte)
3046 if (huge_pte_none(pte) || pte_present(pte))
3048 swp = pte_to_swp_entry(pte);
3049 if (non_swap_entry(swp) && is_migration_entry(swp))
3055 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3059 if (huge_pte_none(pte) || pte_present(pte))
3061 swp = pte_to_swp_entry(pte);
3062 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3068 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3069 struct vm_area_struct *vma)
3071 pte_t *src_pte, *dst_pte, entry;
3072 struct page *ptepage;
3075 struct hstate *h = hstate_vma(vma);
3076 unsigned long sz = huge_page_size(h);
3077 unsigned long mmun_start; /* For mmu_notifiers */
3078 unsigned long mmun_end; /* For mmu_notifiers */
3081 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3083 mmun_start = vma->vm_start;
3084 mmun_end = vma->vm_end;
3086 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3088 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3089 spinlock_t *src_ptl, *dst_ptl;
3090 src_pte = huge_pte_offset(src, addr);
3093 dst_pte = huge_pte_alloc(dst, addr, sz);
3099 /* If the pagetables are shared don't copy or take references */
3100 if (dst_pte == src_pte)
3103 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3104 src_ptl = huge_pte_lockptr(h, src, src_pte);
3105 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3106 entry = huge_ptep_get(src_pte);
3107 if (huge_pte_none(entry)) { /* skip none entry */
3109 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3110 is_hugetlb_entry_hwpoisoned(entry))) {
3111 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3113 if (is_write_migration_entry(swp_entry) && cow) {
3115 * COW mappings require pages in both
3116 * parent and child to be set to read.
3118 make_migration_entry_read(&swp_entry);
3119 entry = swp_entry_to_pte(swp_entry);
3120 set_huge_pte_at(src, addr, src_pte, entry);
3122 set_huge_pte_at(dst, addr, dst_pte, entry);
3125 huge_ptep_set_wrprotect(src, addr, src_pte);
3126 mmu_notifier_invalidate_range(src, mmun_start,
3129 entry = huge_ptep_get(src_pte);
3130 ptepage = pte_page(entry);
3132 page_dup_rmap(ptepage);
3133 set_huge_pte_at(dst, addr, dst_pte, entry);
3134 hugetlb_count_add(pages_per_huge_page(h), dst);
3136 spin_unlock(src_ptl);
3137 spin_unlock(dst_ptl);
3141 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3146 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3147 unsigned long start, unsigned long end,
3148 struct page *ref_page)
3150 int force_flush = 0;
3151 struct mm_struct *mm = vma->vm_mm;
3152 unsigned long address;
3157 struct hstate *h = hstate_vma(vma);
3158 unsigned long sz = huge_page_size(h);
3159 const unsigned long mmun_start = start; /* For mmu_notifiers */
3160 const unsigned long mmun_end = end; /* For mmu_notifiers */
3162 WARN_ON(!is_vm_hugetlb_page(vma));
3163 BUG_ON(start & ~huge_page_mask(h));
3164 BUG_ON(end & ~huge_page_mask(h));
3166 tlb_start_vma(tlb, vma);
3167 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3170 for (; address < end; address += sz) {
3171 ptep = huge_pte_offset(mm, address);
3175 ptl = huge_pte_lock(h, mm, ptep);
3176 if (huge_pmd_unshare(mm, &address, ptep))
3179 pte = huge_ptep_get(ptep);
3180 if (huge_pte_none(pte))
3184 * Migrating hugepage or HWPoisoned hugepage is already
3185 * unmapped and its refcount is dropped, so just clear pte here.
3187 if (unlikely(!pte_present(pte))) {
3188 huge_pte_clear(mm, address, ptep);
3192 page = pte_page(pte);
3194 * If a reference page is supplied, it is because a specific
3195 * page is being unmapped, not a range. Ensure the page we
3196 * are about to unmap is the actual page of interest.
3199 if (page != ref_page)
3203 * Mark the VMA as having unmapped its page so that
3204 * future faults in this VMA will fail rather than
3205 * looking like data was lost
3207 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3210 pte = huge_ptep_get_and_clear(mm, address, ptep);
3211 tlb_remove_tlb_entry(tlb, ptep, address);
3212 if (huge_pte_dirty(pte))
3213 set_page_dirty(page);
3215 hugetlb_count_sub(pages_per_huge_page(h), mm);
3216 page_remove_rmap(page);
3217 force_flush = !__tlb_remove_page(tlb, page);
3223 /* Bail out after unmapping reference page if supplied */
3232 * mmu_gather ran out of room to batch pages, we break out of
3233 * the PTE lock to avoid doing the potential expensive TLB invalidate
3234 * and page-free while holding it.
3239 if (address < end && !ref_page)
3242 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3243 tlb_end_vma(tlb, vma);
3246 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3247 struct vm_area_struct *vma, unsigned long start,
3248 unsigned long end, struct page *ref_page)
3250 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3253 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3254 * test will fail on a vma being torn down, and not grab a page table
3255 * on its way out. We're lucky that the flag has such an appropriate
3256 * name, and can in fact be safely cleared here. We could clear it
3257 * before the __unmap_hugepage_range above, but all that's necessary
3258 * is to clear it before releasing the i_mmap_rwsem. This works
3259 * because in the context this is called, the VMA is about to be
3260 * destroyed and the i_mmap_rwsem is held.
3262 vma->vm_flags &= ~VM_MAYSHARE;
3265 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3266 unsigned long end, struct page *ref_page)
3268 struct mm_struct *mm;
3269 struct mmu_gather tlb;
3273 tlb_gather_mmu(&tlb, mm, start, end);
3274 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3275 tlb_finish_mmu(&tlb, start, end);
3279 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3280 * mappping it owns the reserve page for. The intention is to unmap the page
3281 * from other VMAs and let the children be SIGKILLed if they are faulting the
3284 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3285 struct page *page, unsigned long address)
3287 struct hstate *h = hstate_vma(vma);
3288 struct vm_area_struct *iter_vma;
3289 struct address_space *mapping;
3293 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3294 * from page cache lookup which is in HPAGE_SIZE units.
3296 address = address & huge_page_mask(h);
3297 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3299 mapping = file_inode(vma->vm_file)->i_mapping;
3302 * Take the mapping lock for the duration of the table walk. As
3303 * this mapping should be shared between all the VMAs,
3304 * __unmap_hugepage_range() is called as the lock is already held
3306 i_mmap_lock_write(mapping);
3307 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3308 /* Do not unmap the current VMA */
3309 if (iter_vma == vma)
3313 * Shared VMAs have their own reserves and do not affect
3314 * MAP_PRIVATE accounting but it is possible that a shared
3315 * VMA is using the same page so check and skip such VMAs.
3317 if (iter_vma->vm_flags & VM_MAYSHARE)
3321 * Unmap the page from other VMAs without their own reserves.
3322 * They get marked to be SIGKILLed if they fault in these
3323 * areas. This is because a future no-page fault on this VMA
3324 * could insert a zeroed page instead of the data existing
3325 * from the time of fork. This would look like data corruption
3327 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3328 unmap_hugepage_range(iter_vma, address,
3329 address + huge_page_size(h), page);
3331 i_mmap_unlock_write(mapping);
3335 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3336 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3337 * cannot race with other handlers or page migration.
3338 * Keep the pte_same checks anyway to make transition from the mutex easier.
3340 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3341 unsigned long address, pte_t *ptep, pte_t pte,
3342 struct page *pagecache_page, spinlock_t *ptl)
3344 struct hstate *h = hstate_vma(vma);
3345 struct page *old_page, *new_page;
3346 int ret = 0, outside_reserve = 0;
3347 unsigned long mmun_start; /* For mmu_notifiers */
3348 unsigned long mmun_end; /* For mmu_notifiers */
3350 old_page = pte_page(pte);
3353 /* If no-one else is actually using this page, avoid the copy
3354 * and just make the page writable */
3355 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3356 page_move_anon_rmap(old_page, vma, address);
3357 set_huge_ptep_writable(vma, address, ptep);
3362 * If the process that created a MAP_PRIVATE mapping is about to
3363 * perform a COW due to a shared page count, attempt to satisfy
3364 * the allocation without using the existing reserves. The pagecache
3365 * page is used to determine if the reserve at this address was
3366 * consumed or not. If reserves were used, a partial faulted mapping
3367 * at the time of fork() could consume its reserves on COW instead
3368 * of the full address range.
3370 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3371 old_page != pagecache_page)
3372 outside_reserve = 1;
3374 page_cache_get(old_page);
3377 * Drop page table lock as buddy allocator may be called. It will
3378 * be acquired again before returning to the caller, as expected.
3381 new_page = alloc_huge_page(vma, address, outside_reserve);
3383 if (IS_ERR(new_page)) {
3385 * If a process owning a MAP_PRIVATE mapping fails to COW,
3386 * it is due to references held by a child and an insufficient
3387 * huge page pool. To guarantee the original mappers
3388 * reliability, unmap the page from child processes. The child
3389 * may get SIGKILLed if it later faults.
3391 if (outside_reserve) {
3392 page_cache_release(old_page);
3393 BUG_ON(huge_pte_none(pte));
3394 unmap_ref_private(mm, vma, old_page, address);
3395 BUG_ON(huge_pte_none(pte));
3397 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3399 pte_same(huge_ptep_get(ptep), pte)))
3400 goto retry_avoidcopy;
3402 * race occurs while re-acquiring page table
3403 * lock, and our job is done.
3408 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3409 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3410 goto out_release_old;
3414 * When the original hugepage is shared one, it does not have
3415 * anon_vma prepared.
3417 if (unlikely(anon_vma_prepare(vma))) {
3419 goto out_release_all;
3422 copy_user_huge_page(new_page, old_page, address, vma,
3423 pages_per_huge_page(h));
3424 __SetPageUptodate(new_page);
3425 set_page_huge_active(new_page);
3427 mmun_start = address & huge_page_mask(h);
3428 mmun_end = mmun_start + huge_page_size(h);
3429 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3432 * Retake the page table lock to check for racing updates
3433 * before the page tables are altered
3436 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3437 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3438 ClearPagePrivate(new_page);
3441 huge_ptep_clear_flush(vma, address, ptep);
3442 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3443 set_huge_pte_at(mm, address, ptep,
3444 make_huge_pte(vma, new_page, 1));
3445 page_remove_rmap(old_page);
3446 hugepage_add_new_anon_rmap(new_page, vma, address);
3447 /* Make the old page be freed below */
3448 new_page = old_page;
3451 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3453 page_cache_release(new_page);
3455 page_cache_release(old_page);
3457 spin_lock(ptl); /* Caller expects lock to be held */
3461 /* Return the pagecache page at a given address within a VMA */
3462 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3463 struct vm_area_struct *vma, unsigned long address)
3465 struct address_space *mapping;
3468 mapping = vma->vm_file->f_mapping;
3469 idx = vma_hugecache_offset(h, vma, address);
3471 return find_lock_page(mapping, idx);
3475 * Return whether there is a pagecache page to back given address within VMA.
3476 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3478 static bool hugetlbfs_pagecache_present(struct hstate *h,
3479 struct vm_area_struct *vma, unsigned long address)
3481 struct address_space *mapping;
3485 mapping = vma->vm_file->f_mapping;
3486 idx = vma_hugecache_offset(h, vma, address);
3488 page = find_get_page(mapping, idx);
3491 return page != NULL;
3494 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3497 struct inode *inode = mapping->host;
3498 struct hstate *h = hstate_inode(inode);
3499 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3503 ClearPagePrivate(page);
3505 spin_lock(&inode->i_lock);
3506 inode->i_blocks += blocks_per_huge_page(h);
3507 spin_unlock(&inode->i_lock);
3511 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3512 struct address_space *mapping, pgoff_t idx,
3513 unsigned long address, pte_t *ptep, unsigned int flags)
3515 struct hstate *h = hstate_vma(vma);
3516 int ret = VM_FAULT_SIGBUS;
3524 * Currently, we are forced to kill the process in the event the
3525 * original mapper has unmapped pages from the child due to a failed
3526 * COW. Warn that such a situation has occurred as it may not be obvious
3528 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3529 pr_warning("PID %d killed due to inadequate hugepage pool\n",
3535 * Use page lock to guard against racing truncation
3536 * before we get page_table_lock.
3539 page = find_lock_page(mapping, idx);
3541 size = i_size_read(mapping->host) >> huge_page_shift(h);
3544 page = alloc_huge_page(vma, address, 0);
3546 ret = PTR_ERR(page);
3550 ret = VM_FAULT_SIGBUS;
3553 clear_huge_page(page, address, pages_per_huge_page(h));
3554 __SetPageUptodate(page);
3555 set_page_huge_active(page);
3557 if (vma->vm_flags & VM_MAYSHARE) {
3558 int err = huge_add_to_page_cache(page, mapping, idx);
3567 if (unlikely(anon_vma_prepare(vma))) {
3569 goto backout_unlocked;
3575 * If memory error occurs between mmap() and fault, some process
3576 * don't have hwpoisoned swap entry for errored virtual address.
3577 * So we need to block hugepage fault by PG_hwpoison bit check.
3579 if (unlikely(PageHWPoison(page))) {
3580 ret = VM_FAULT_HWPOISON |
3581 VM_FAULT_SET_HINDEX(hstate_index(h));
3582 goto backout_unlocked;
3587 * If we are going to COW a private mapping later, we examine the
3588 * pending reservations for this page now. This will ensure that
3589 * any allocations necessary to record that reservation occur outside
3592 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3593 if (vma_needs_reservation(h, vma, address) < 0) {
3595 goto backout_unlocked;
3597 /* Just decrements count, does not deallocate */
3598 vma_end_reservation(h, vma, address);
3601 ptl = huge_pte_lockptr(h, mm, ptep);
3603 size = i_size_read(mapping->host) >> huge_page_shift(h);
3608 if (!huge_pte_none(huge_ptep_get(ptep)))
3612 ClearPagePrivate(page);
3613 hugepage_add_new_anon_rmap(page, vma, address);
3615 page_dup_rmap(page);
3616 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3617 && (vma->vm_flags & VM_SHARED)));
3618 set_huge_pte_at(mm, address, ptep, new_pte);
3620 hugetlb_count_add(pages_per_huge_page(h), mm);
3621 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3622 /* Optimization, do the COW without a second fault */
3623 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
3640 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3641 struct vm_area_struct *vma,
3642 struct address_space *mapping,
3643 pgoff_t idx, unsigned long address)
3645 unsigned long key[2];
3648 if (vma->vm_flags & VM_SHARED) {
3649 key[0] = (unsigned long) mapping;
3652 key[0] = (unsigned long) mm;
3653 key[1] = address >> huge_page_shift(h);
3656 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3658 return hash & (num_fault_mutexes - 1);
3662 * For uniprocesor systems we always use a single mutex, so just
3663 * return 0 and avoid the hashing overhead.
3665 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3666 struct vm_area_struct *vma,
3667 struct address_space *mapping,
3668 pgoff_t idx, unsigned long address)
3674 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3675 unsigned long address, unsigned int flags)
3682 struct page *page = NULL;
3683 struct page *pagecache_page = NULL;
3684 struct hstate *h = hstate_vma(vma);
3685 struct address_space *mapping;
3686 int need_wait_lock = 0;
3688 address &= huge_page_mask(h);
3690 ptep = huge_pte_offset(mm, address);
3692 entry = huge_ptep_get(ptep);
3693 if (unlikely(is_hugetlb_entry_migration(entry))) {
3694 migration_entry_wait_huge(vma, mm, ptep);
3696 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3697 return VM_FAULT_HWPOISON_LARGE |
3698 VM_FAULT_SET_HINDEX(hstate_index(h));
3701 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3703 return VM_FAULT_OOM;
3705 mapping = vma->vm_file->f_mapping;
3706 idx = vma_hugecache_offset(h, vma, address);
3709 * Serialize hugepage allocation and instantiation, so that we don't
3710 * get spurious allocation failures if two CPUs race to instantiate
3711 * the same page in the page cache.
3713 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3714 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3716 entry = huge_ptep_get(ptep);
3717 if (huge_pte_none(entry)) {
3718 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3725 * entry could be a migration/hwpoison entry at this point, so this
3726 * check prevents the kernel from going below assuming that we have
3727 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3728 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3731 if (!pte_present(entry))
3735 * If we are going to COW the mapping later, we examine the pending
3736 * reservations for this page now. This will ensure that any
3737 * allocations necessary to record that reservation occur outside the
3738 * spinlock. For private mappings, we also lookup the pagecache
3739 * page now as it is used to determine if a reservation has been
3742 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3743 if (vma_needs_reservation(h, vma, address) < 0) {
3747 /* Just decrements count, does not deallocate */
3748 vma_end_reservation(h, vma, address);
3750 if (!(vma->vm_flags & VM_MAYSHARE))
3751 pagecache_page = hugetlbfs_pagecache_page(h,
3755 ptl = huge_pte_lock(h, mm, ptep);
3757 /* Check for a racing update before calling hugetlb_cow */
3758 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3762 * hugetlb_cow() requires page locks of pte_page(entry) and
3763 * pagecache_page, so here we need take the former one
3764 * when page != pagecache_page or !pagecache_page.
3766 page = pte_page(entry);
3767 if (page != pagecache_page)
3768 if (!trylock_page(page)) {
3775 if (flags & FAULT_FLAG_WRITE) {
3776 if (!huge_pte_write(entry)) {
3777 ret = hugetlb_cow(mm, vma, address, ptep, entry,
3778 pagecache_page, ptl);
3781 entry = huge_pte_mkdirty(entry);
3783 entry = pte_mkyoung(entry);
3784 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3785 flags & FAULT_FLAG_WRITE))
3786 update_mmu_cache(vma, address, ptep);
3788 if (page != pagecache_page)
3794 if (pagecache_page) {
3795 unlock_page(pagecache_page);
3796 put_page(pagecache_page);
3799 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3801 * Generally it's safe to hold refcount during waiting page lock. But
3802 * here we just wait to defer the next page fault to avoid busy loop and
3803 * the page is not used after unlocked before returning from the current
3804 * page fault. So we are safe from accessing freed page, even if we wait
3805 * here without taking refcount.
3808 wait_on_page_locked(page);
3812 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3813 struct page **pages, struct vm_area_struct **vmas,
3814 unsigned long *position, unsigned long *nr_pages,
3815 long i, unsigned int flags)
3817 unsigned long pfn_offset;
3818 unsigned long vaddr = *position;
3819 unsigned long remainder = *nr_pages;
3820 struct hstate *h = hstate_vma(vma);
3822 while (vaddr < vma->vm_end && remainder) {
3824 spinlock_t *ptl = NULL;
3829 * If we have a pending SIGKILL, don't keep faulting pages and
3830 * potentially allocating memory.
3832 if (unlikely(fatal_signal_pending(current))) {
3838 * Some archs (sparc64, sh*) have multiple pte_ts to
3839 * each hugepage. We have to make sure we get the
3840 * first, for the page indexing below to work.
3842 * Note that page table lock is not held when pte is null.
3844 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3846 ptl = huge_pte_lock(h, mm, pte);
3847 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3850 * When coredumping, it suits get_dump_page if we just return
3851 * an error where there's an empty slot with no huge pagecache
3852 * to back it. This way, we avoid allocating a hugepage, and
3853 * the sparse dumpfile avoids allocating disk blocks, but its
3854 * huge holes still show up with zeroes where they need to be.
3856 if (absent && (flags & FOLL_DUMP) &&
3857 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3865 * We need call hugetlb_fault for both hugepages under migration
3866 * (in which case hugetlb_fault waits for the migration,) and
3867 * hwpoisoned hugepages (in which case we need to prevent the
3868 * caller from accessing to them.) In order to do this, we use
3869 * here is_swap_pte instead of is_hugetlb_entry_migration and
3870 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3871 * both cases, and because we can't follow correct pages
3872 * directly from any kind of swap entries.
3874 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3875 ((flags & FOLL_WRITE) &&
3876 !huge_pte_write(huge_ptep_get(pte)))) {
3881 ret = hugetlb_fault(mm, vma, vaddr,
3882 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3883 if (!(ret & VM_FAULT_ERROR))
3890 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3891 page = pte_page(huge_ptep_get(pte));
3894 pages[i] = mem_map_offset(page, pfn_offset);
3895 get_page_foll(pages[i]);
3905 if (vaddr < vma->vm_end && remainder &&
3906 pfn_offset < pages_per_huge_page(h)) {
3908 * We use pfn_offset to avoid touching the pageframes
3909 * of this compound page.
3915 *nr_pages = remainder;
3918 return i ? i : -EFAULT;
3921 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3922 unsigned long address, unsigned long end, pgprot_t newprot)
3924 struct mm_struct *mm = vma->vm_mm;
3925 unsigned long start = address;
3928 struct hstate *h = hstate_vma(vma);
3929 unsigned long pages = 0;
3931 BUG_ON(address >= end);
3932 flush_cache_range(vma, address, end);
3934 mmu_notifier_invalidate_range_start(mm, start, end);
3935 i_mmap_lock_write(vma->vm_file->f_mapping);
3936 for (; address < end; address += huge_page_size(h)) {
3938 ptep = huge_pte_offset(mm, address);
3941 ptl = huge_pte_lock(h, mm, ptep);
3942 if (huge_pmd_unshare(mm, &address, ptep)) {
3947 pte = huge_ptep_get(ptep);
3948 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3952 if (unlikely(is_hugetlb_entry_migration(pte))) {
3953 swp_entry_t entry = pte_to_swp_entry(pte);
3955 if (is_write_migration_entry(entry)) {
3958 make_migration_entry_read(&entry);
3959 newpte = swp_entry_to_pte(entry);
3960 set_huge_pte_at(mm, address, ptep, newpte);
3966 if (!huge_pte_none(pte)) {
3967 pte = huge_ptep_get_and_clear(mm, address, ptep);
3968 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3969 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3970 set_huge_pte_at(mm, address, ptep, pte);
3976 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
3977 * may have cleared our pud entry and done put_page on the page table:
3978 * once we release i_mmap_rwsem, another task can do the final put_page
3979 * and that page table be reused and filled with junk.
3981 flush_tlb_range(vma, start, end);
3982 mmu_notifier_invalidate_range(mm, start, end);
3983 i_mmap_unlock_write(vma->vm_file->f_mapping);
3984 mmu_notifier_invalidate_range_end(mm, start, end);
3986 return pages << h->order;
3989 int hugetlb_reserve_pages(struct inode *inode,
3991 struct vm_area_struct *vma,
3992 vm_flags_t vm_flags)
3995 struct hstate *h = hstate_inode(inode);
3996 struct hugepage_subpool *spool = subpool_inode(inode);
3997 struct resv_map *resv_map;
4001 * Only apply hugepage reservation if asked. At fault time, an
4002 * attempt will be made for VM_NORESERVE to allocate a page
4003 * without using reserves
4005 if (vm_flags & VM_NORESERVE)
4009 * Shared mappings base their reservation on the number of pages that
4010 * are already allocated on behalf of the file. Private mappings need
4011 * to reserve the full area even if read-only as mprotect() may be
4012 * called to make the mapping read-write. Assume !vma is a shm mapping
4014 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4015 resv_map = inode_resv_map(inode);
4017 chg = region_chg(resv_map, from, to);
4020 resv_map = resv_map_alloc();
4026 set_vma_resv_map(vma, resv_map);
4027 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4036 * There must be enough pages in the subpool for the mapping. If
4037 * the subpool has a minimum size, there may be some global
4038 * reservations already in place (gbl_reserve).
4040 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4041 if (gbl_reserve < 0) {
4047 * Check enough hugepages are available for the reservation.
4048 * Hand the pages back to the subpool if there are not
4050 ret = hugetlb_acct_memory(h, gbl_reserve);
4052 /* put back original number of pages, chg */
4053 (void)hugepage_subpool_put_pages(spool, chg);
4058 * Account for the reservations made. Shared mappings record regions
4059 * that have reservations as they are shared by multiple VMAs.
4060 * When the last VMA disappears, the region map says how much
4061 * the reservation was and the page cache tells how much of
4062 * the reservation was consumed. Private mappings are per-VMA and
4063 * only the consumed reservations are tracked. When the VMA
4064 * disappears, the original reservation is the VMA size and the
4065 * consumed reservations are stored in the map. Hence, nothing
4066 * else has to be done for private mappings here
4068 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4069 long add = region_add(resv_map, from, to);
4071 if (unlikely(chg > add)) {
4073 * pages in this range were added to the reserve
4074 * map between region_chg and region_add. This
4075 * indicates a race with alloc_huge_page. Adjust
4076 * the subpool and reserve counts modified above
4077 * based on the difference.
4081 rsv_adjust = hugepage_subpool_put_pages(spool,
4083 hugetlb_acct_memory(h, -rsv_adjust);
4088 if (!vma || vma->vm_flags & VM_MAYSHARE)
4089 region_abort(resv_map, from, to);
4090 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4091 kref_put(&resv_map->refs, resv_map_release);
4095 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4098 struct hstate *h = hstate_inode(inode);
4099 struct resv_map *resv_map = inode_resv_map(inode);
4101 struct hugepage_subpool *spool = subpool_inode(inode);
4105 chg = region_del(resv_map, start, end);
4107 * region_del() can fail in the rare case where a region
4108 * must be split and another region descriptor can not be
4109 * allocated. If end == LONG_MAX, it will not fail.
4115 spin_lock(&inode->i_lock);
4116 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4117 spin_unlock(&inode->i_lock);
4120 * If the subpool has a minimum size, the number of global
4121 * reservations to be released may be adjusted.
4123 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4124 hugetlb_acct_memory(h, -gbl_reserve);
4129 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4130 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4131 struct vm_area_struct *vma,
4132 unsigned long addr, pgoff_t idx)
4134 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4136 unsigned long sbase = saddr & PUD_MASK;
4137 unsigned long s_end = sbase + PUD_SIZE;
4139 /* Allow segments to share if only one is marked locked */
4140 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4141 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4144 * match the virtual addresses, permission and the alignment of the
4147 if (pmd_index(addr) != pmd_index(saddr) ||
4148 vm_flags != svm_flags ||
4149 sbase < svma->vm_start || svma->vm_end < s_end)
4155 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4157 unsigned long base = addr & PUD_MASK;
4158 unsigned long end = base + PUD_SIZE;
4161 * check on proper vm_flags and page table alignment
4163 if (vma->vm_flags & VM_MAYSHARE &&
4164 vma->vm_start <= base && end <= vma->vm_end)
4170 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4171 * and returns the corresponding pte. While this is not necessary for the
4172 * !shared pmd case because we can allocate the pmd later as well, it makes the
4173 * code much cleaner. pmd allocation is essential for the shared case because
4174 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4175 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4176 * bad pmd for sharing.
4178 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4180 struct vm_area_struct *vma = find_vma(mm, addr);
4181 struct address_space *mapping = vma->vm_file->f_mapping;
4182 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4184 struct vm_area_struct *svma;
4185 unsigned long saddr;
4190 if (!vma_shareable(vma, addr))
4191 return (pte_t *)pmd_alloc(mm, pud, addr);
4193 i_mmap_lock_write(mapping);
4194 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4198 saddr = page_table_shareable(svma, vma, addr, idx);
4200 spte = huge_pte_offset(svma->vm_mm, saddr);
4203 get_page(virt_to_page(spte));
4212 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
4214 if (pud_none(*pud)) {
4215 pud_populate(mm, pud,
4216 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4218 put_page(virt_to_page(spte));
4223 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4224 i_mmap_unlock_write(mapping);
4229 * unmap huge page backed by shared pte.
4231 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4232 * indicated by page_count > 1, unmap is achieved by clearing pud and
4233 * decrementing the ref count. If count == 1, the pte page is not shared.
4235 * called with page table lock held.
4237 * returns: 1 successfully unmapped a shared pte page
4238 * 0 the underlying pte page is not shared, or it is the last user
4240 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4242 pgd_t *pgd = pgd_offset(mm, *addr);
4243 pud_t *pud = pud_offset(pgd, *addr);
4245 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4246 if (page_count(virt_to_page(ptep)) == 1)
4250 put_page(virt_to_page(ptep));
4252 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4255 #define want_pmd_share() (1)
4256 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4257 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4262 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4266 #define want_pmd_share() (0)
4267 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4269 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4270 pte_t *huge_pte_alloc(struct mm_struct *mm,
4271 unsigned long addr, unsigned long sz)
4277 pgd = pgd_offset(mm, addr);
4278 pud = pud_alloc(mm, pgd, addr);
4280 if (sz == PUD_SIZE) {
4283 BUG_ON(sz != PMD_SIZE);
4284 if (want_pmd_share() && pud_none(*pud))
4285 pte = huge_pmd_share(mm, addr, pud);
4287 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4290 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
4295 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
4301 pgd = pgd_offset(mm, addr);
4302 if (pgd_present(*pgd)) {
4303 pud = pud_offset(pgd, addr);
4304 if (pud_present(*pud)) {
4306 return (pte_t *)pud;
4307 pmd = pmd_offset(pud, addr);
4310 return (pte_t *) pmd;
4313 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4316 * These functions are overwritable if your architecture needs its own
4319 struct page * __weak
4320 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4323 return ERR_PTR(-EINVAL);
4326 struct page * __weak
4327 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4328 pmd_t *pmd, int flags)
4330 struct page *page = NULL;
4333 ptl = pmd_lockptr(mm, pmd);
4336 * make sure that the address range covered by this pmd is not
4337 * unmapped from other threads.
4339 if (!pmd_huge(*pmd))
4341 if (pmd_present(*pmd)) {
4342 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4343 if (flags & FOLL_GET)
4346 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
4348 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4352 * hwpoisoned entry is treated as no_page_table in
4353 * follow_page_mask().
4361 struct page * __weak
4362 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4363 pud_t *pud, int flags)
4365 if (flags & FOLL_GET)
4368 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4371 #ifdef CONFIG_MEMORY_FAILURE
4374 * This function is called from memory failure code.
4375 * Assume the caller holds page lock of the head page.
4377 int dequeue_hwpoisoned_huge_page(struct page *hpage)
4379 struct hstate *h = page_hstate(hpage);
4380 int nid = page_to_nid(hpage);
4383 spin_lock(&hugetlb_lock);
4385 * Just checking !page_huge_active is not enough, because that could be
4386 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4388 if (!page_huge_active(hpage) && !page_count(hpage)) {
4390 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4391 * but dangling hpage->lru can trigger list-debug warnings
4392 * (this happens when we call unpoison_memory() on it),
4393 * so let it point to itself with list_del_init().
4395 list_del_init(&hpage->lru);
4396 set_page_refcounted(hpage);
4397 h->free_huge_pages--;
4398 h->free_huge_pages_node[nid]--;
4401 spin_unlock(&hugetlb_lock);
4406 bool isolate_huge_page(struct page *page, struct list_head *list)
4410 VM_BUG_ON_PAGE(!PageHead(page), page);
4411 spin_lock(&hugetlb_lock);
4412 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4416 clear_page_huge_active(page);
4417 list_move_tail(&page->lru, list);
4419 spin_unlock(&hugetlb_lock);
4423 void putback_active_hugepage(struct page *page)
4425 VM_BUG_ON_PAGE(!PageHead(page), page);
4426 spin_lock(&hugetlb_lock);
4427 set_page_huge_active(page);
4428 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4429 spin_unlock(&hugetlb_lock);