2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
8 #include <linux/seq_file.h>
9 #include <linux/sysctl.h>
10 #include <linux/highmem.h>
11 #include <linux/mmu_notifier.h>
12 #include <linux/nodemask.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/compiler.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/rmap.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
24 #include <linux/page-isolation.h>
25 #include <linux/jhash.h>
28 #include <asm/pgtable.h>
32 #include <linux/hugetlb.h>
33 #include <linux/hugetlb_cgroup.h>
34 #include <linux/node.h>
37 int hugepages_treat_as_movable;
39 int hugetlb_max_hstate __read_mostly;
40 unsigned int default_hstate_idx;
41 struct hstate hstates[HUGE_MAX_HSTATE];
43 * Minimum page order among possible hugepage sizes, set to a proper value
46 static unsigned int minimum_order __read_mostly = UINT_MAX;
48 __initdata LIST_HEAD(huge_boot_pages);
50 /* for command line parsing */
51 static struct hstate * __initdata parsed_hstate;
52 static unsigned long __initdata default_hstate_max_huge_pages;
53 static unsigned long __initdata default_hstate_size;
56 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
57 * free_huge_pages, and surplus_huge_pages.
59 DEFINE_SPINLOCK(hugetlb_lock);
62 * Serializes faults on the same logical page. This is used to
63 * prevent spurious OOMs when the hugepage pool is fully utilized.
65 static int num_fault_mutexes;
66 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
68 /* Forward declaration */
69 static int hugetlb_acct_memory(struct hstate *h, long delta);
71 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
73 bool free = (spool->count == 0) && (spool->used_hpages == 0);
75 spin_unlock(&spool->lock);
77 /* If no pages are used, and no other handles to the subpool
78 * remain, give up any reservations mased on minimum size and
81 if (spool->min_hpages != -1)
82 hugetlb_acct_memory(spool->hstate,
88 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
91 struct hugepage_subpool *spool;
93 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
97 spin_lock_init(&spool->lock);
99 spool->max_hpages = max_hpages;
101 spool->min_hpages = min_hpages;
103 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
107 spool->rsv_hpages = min_hpages;
112 void hugepage_put_subpool(struct hugepage_subpool *spool)
114 spin_lock(&spool->lock);
115 BUG_ON(!spool->count);
117 unlock_or_release_subpool(spool);
121 * Subpool accounting for allocating and reserving pages.
122 * Return -ENOMEM if there are not enough resources to satisfy the
123 * the request. Otherwise, return the number of pages by which the
124 * global pools must be adjusted (upward). The returned value may
125 * only be different than the passed value (delta) in the case where
126 * a subpool minimum size must be manitained.
128 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
136 spin_lock(&spool->lock);
138 if (spool->max_hpages != -1) { /* maximum size accounting */
139 if ((spool->used_hpages + delta) <= spool->max_hpages)
140 spool->used_hpages += delta;
147 if (spool->min_hpages != -1) { /* minimum size accounting */
148 if (delta > spool->rsv_hpages) {
150 * Asking for more reserves than those already taken on
151 * behalf of subpool. Return difference.
153 ret = delta - spool->rsv_hpages;
154 spool->rsv_hpages = 0;
156 ret = 0; /* reserves already accounted for */
157 spool->rsv_hpages -= delta;
162 spin_unlock(&spool->lock);
167 * Subpool accounting for freeing and unreserving pages.
168 * Return the number of global page reservations that must be dropped.
169 * The return value may only be different than the passed value (delta)
170 * in the case where a subpool minimum size must be maintained.
172 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
180 spin_lock(&spool->lock);
182 if (spool->max_hpages != -1) /* maximum size accounting */
183 spool->used_hpages -= delta;
185 if (spool->min_hpages != -1) { /* minimum size accounting */
186 if (spool->rsv_hpages + delta <= spool->min_hpages)
189 ret = spool->rsv_hpages + delta - spool->min_hpages;
191 spool->rsv_hpages += delta;
192 if (spool->rsv_hpages > spool->min_hpages)
193 spool->rsv_hpages = spool->min_hpages;
197 * If hugetlbfs_put_super couldn't free spool due to an outstanding
198 * quota reference, free it now.
200 unlock_or_release_subpool(spool);
205 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
207 return HUGETLBFS_SB(inode->i_sb)->spool;
210 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
212 return subpool_inode(file_inode(vma->vm_file));
216 * Region tracking -- allows tracking of reservations and instantiated pages
217 * across the pages in a mapping.
219 * The region data structures are embedded into a resv_map and protected
220 * by a resv_map's lock. The set of regions within the resv_map represent
221 * reservations for huge pages, or huge pages that have already been
222 * instantiated within the map. The from and to elements are huge page
223 * indicies into the associated mapping. from indicates the starting index
224 * of the region. to represents the first index past the end of the region.
226 * For example, a file region structure with from == 0 and to == 4 represents
227 * four huge pages in a mapping. It is important to note that the to element
228 * represents the first element past the end of the region. This is used in
229 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
231 * Interval notation of the form [from, to) will be used to indicate that
232 * the endpoint from is inclusive and to is exclusive.
235 struct list_head link;
241 * Add the huge page range represented by [f, t) to the reserve
242 * map. In the normal case, existing regions will be expanded
243 * to accommodate the specified range. Sufficient regions should
244 * exist for expansion due to the previous call to region_chg
245 * with the same range. However, it is possible that region_del
246 * could have been called after region_chg and modifed the map
247 * in such a way that no region exists to be expanded. In this
248 * case, pull a region descriptor from the cache associated with
249 * the map and use that for the new range.
251 * Return the number of new huge pages added to the map. This
252 * number is greater than or equal to zero.
254 static long region_add(struct resv_map *resv, long f, long t)
256 struct list_head *head = &resv->regions;
257 struct file_region *rg, *nrg, *trg;
260 spin_lock(&resv->lock);
261 /* Locate the region we are either in or before. */
262 list_for_each_entry(rg, head, link)
267 * If no region exists which can be expanded to include the
268 * specified range, the list must have been modified by an
269 * interleving call to region_del(). Pull a region descriptor
270 * from the cache and use it for this range.
272 if (&rg->link == head || t < rg->from) {
273 VM_BUG_ON(resv->region_cache_count <= 0);
275 resv->region_cache_count--;
276 nrg = list_first_entry(&resv->region_cache, struct file_region,
278 list_del(&nrg->link);
282 list_add(&nrg->link, rg->link.prev);
288 /* Round our left edge to the current segment if it encloses us. */
292 /* Check for and consume any regions we now overlap with. */
294 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
295 if (&rg->link == head)
300 /* If this area reaches higher then extend our area to
301 * include it completely. If this is not the first area
302 * which we intend to reuse, free it. */
306 /* Decrement return value by the deleted range.
307 * Another range will span this area so that by
308 * end of routine add will be >= zero
310 add -= (rg->to - rg->from);
316 add += (nrg->from - f); /* Added to beginning of region */
318 add += t - nrg->to; /* Added to end of region */
322 resv->adds_in_progress--;
323 spin_unlock(&resv->lock);
329 * Examine the existing reserve map and determine how many
330 * huge pages in the specified range [f, t) are NOT currently
331 * represented. This routine is called before a subsequent
332 * call to region_add that will actually modify the reserve
333 * map to add the specified range [f, t). region_chg does
334 * not change the number of huge pages represented by the
335 * map. However, if the existing regions in the map can not
336 * be expanded to represent the new range, a new file_region
337 * structure is added to the map as a placeholder. This is
338 * so that the subsequent region_add call will have all the
339 * regions it needs and will not fail.
341 * Upon entry, region_chg will also examine the cache of region descriptors
342 * associated with the map. If there are not enough descriptors cached, one
343 * will be allocated for the in progress add operation.
345 * Returns the number of huge pages that need to be added to the existing
346 * reservation map for the range [f, t). This number is greater or equal to
347 * zero. -ENOMEM is returned if a new file_region structure or cache entry
348 * is needed and can not be allocated.
350 static long region_chg(struct resv_map *resv, long f, long t)
352 struct list_head *head = &resv->regions;
353 struct file_region *rg, *nrg = NULL;
357 spin_lock(&resv->lock);
359 resv->adds_in_progress++;
362 * Check for sufficient descriptors in the cache to accommodate
363 * the number of in progress add operations.
365 if (resv->adds_in_progress > resv->region_cache_count) {
366 struct file_region *trg;
368 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
369 /* Must drop lock to allocate a new descriptor. */
370 resv->adds_in_progress--;
371 spin_unlock(&resv->lock);
373 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
379 spin_lock(&resv->lock);
380 list_add(&trg->link, &resv->region_cache);
381 resv->region_cache_count++;
385 /* Locate the region we are before or in. */
386 list_for_each_entry(rg, head, link)
390 /* If we are below the current region then a new region is required.
391 * Subtle, allocate a new region at the position but make it zero
392 * size such that we can guarantee to record the reservation. */
393 if (&rg->link == head || t < rg->from) {
395 resv->adds_in_progress--;
396 spin_unlock(&resv->lock);
397 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
403 INIT_LIST_HEAD(&nrg->link);
407 list_add(&nrg->link, rg->link.prev);
412 /* Round our left edge to the current segment if it encloses us. */
417 /* Check for and consume any regions we now overlap with. */
418 list_for_each_entry(rg, rg->link.prev, link) {
419 if (&rg->link == head)
424 /* We overlap with this area, if it extends further than
425 * us then we must extend ourselves. Account for its
426 * existing reservation. */
431 chg -= rg->to - rg->from;
435 spin_unlock(&resv->lock);
436 /* We already know we raced and no longer need the new region */
440 spin_unlock(&resv->lock);
445 * Abort the in progress add operation. The adds_in_progress field
446 * of the resv_map keeps track of the operations in progress between
447 * calls to region_chg and region_add. Operations are sometimes
448 * aborted after the call to region_chg. In such cases, region_abort
449 * is called to decrement the adds_in_progress counter.
451 * NOTE: The range arguments [f, t) are not needed or used in this
452 * routine. They are kept to make reading the calling code easier as
453 * arguments will match the associated region_chg call.
455 static void region_abort(struct resv_map *resv, long f, long t)
457 spin_lock(&resv->lock);
458 VM_BUG_ON(!resv->region_cache_count);
459 resv->adds_in_progress--;
460 spin_unlock(&resv->lock);
464 * Delete the specified range [f, t) from the reserve map. If the
465 * t parameter is LONG_MAX, this indicates that ALL regions after f
466 * should be deleted. Locate the regions which intersect [f, t)
467 * and either trim, delete or split the existing regions.
469 * Returns the number of huge pages deleted from the reserve map.
470 * In the normal case, the return value is zero or more. In the
471 * case where a region must be split, a new region descriptor must
472 * be allocated. If the allocation fails, -ENOMEM will be returned.
473 * NOTE: If the parameter t == LONG_MAX, then we will never split
474 * a region and possibly return -ENOMEM. Callers specifying
475 * t == LONG_MAX do not need to check for -ENOMEM error.
477 static long region_del(struct resv_map *resv, long f, long t)
479 struct list_head *head = &resv->regions;
480 struct file_region *rg, *trg;
481 struct file_region *nrg = NULL;
485 spin_lock(&resv->lock);
486 list_for_each_entry_safe(rg, trg, head, link) {
488 * Skip regions before the range to be deleted. file_region
489 * ranges are normally of the form [from, to). However, there
490 * may be a "placeholder" entry in the map which is of the form
491 * (from, to) with from == to. Check for placeholder entries
492 * at the beginning of the range to be deleted.
494 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
500 if (f > rg->from && t < rg->to) { /* Must split region */
502 * Check for an entry in the cache before dropping
503 * lock and attempting allocation.
506 resv->region_cache_count > resv->adds_in_progress) {
507 nrg = list_first_entry(&resv->region_cache,
510 list_del(&nrg->link);
511 resv->region_cache_count--;
515 spin_unlock(&resv->lock);
516 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
524 /* New entry for end of split region */
527 INIT_LIST_HEAD(&nrg->link);
529 /* Original entry is trimmed */
532 list_add(&nrg->link, &rg->link);
537 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
538 del += rg->to - rg->from;
544 if (f <= rg->from) { /* Trim beginning of region */
547 } else { /* Trim end of region */
553 spin_unlock(&resv->lock);
559 * A rare out of memory error was encountered which prevented removal of
560 * the reserve map region for a page. The huge page itself was free'ed
561 * and removed from the page cache. This routine will adjust the subpool
562 * usage count, and the global reserve count if needed. By incrementing
563 * these counts, the reserve map entry which could not be deleted will
564 * appear as a "reserved" entry instead of simply dangling with incorrect
567 void hugetlb_fix_reserve_counts(struct inode *inode, bool restore_reserve)
569 struct hugepage_subpool *spool = subpool_inode(inode);
572 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
573 if (restore_reserve && rsv_adjust) {
574 struct hstate *h = hstate_inode(inode);
576 hugetlb_acct_memory(h, 1);
581 * Count and return the number of huge pages in the reserve map
582 * that intersect with the range [f, t).
584 static long region_count(struct resv_map *resv, long f, long t)
586 struct list_head *head = &resv->regions;
587 struct file_region *rg;
590 spin_lock(&resv->lock);
591 /* Locate each segment we overlap with, and count that overlap. */
592 list_for_each_entry(rg, head, link) {
601 seg_from = max(rg->from, f);
602 seg_to = min(rg->to, t);
604 chg += seg_to - seg_from;
606 spin_unlock(&resv->lock);
612 * Convert the address within this vma to the page offset within
613 * the mapping, in pagecache page units; huge pages here.
615 static pgoff_t vma_hugecache_offset(struct hstate *h,
616 struct vm_area_struct *vma, unsigned long address)
618 return ((address - vma->vm_start) >> huge_page_shift(h)) +
619 (vma->vm_pgoff >> huge_page_order(h));
622 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
623 unsigned long address)
625 return vma_hugecache_offset(hstate_vma(vma), vma, address);
629 * Return the size of the pages allocated when backing a VMA. In the majority
630 * cases this will be same size as used by the page table entries.
632 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
634 struct hstate *hstate;
636 if (!is_vm_hugetlb_page(vma))
639 hstate = hstate_vma(vma);
641 return 1UL << huge_page_shift(hstate);
643 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
646 * Return the page size being used by the MMU to back a VMA. In the majority
647 * of cases, the page size used by the kernel matches the MMU size. On
648 * architectures where it differs, an architecture-specific version of this
649 * function is required.
651 #ifndef vma_mmu_pagesize
652 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
654 return vma_kernel_pagesize(vma);
659 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
660 * bits of the reservation map pointer, which are always clear due to
663 #define HPAGE_RESV_OWNER (1UL << 0)
664 #define HPAGE_RESV_UNMAPPED (1UL << 1)
665 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
668 * These helpers are used to track how many pages are reserved for
669 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
670 * is guaranteed to have their future faults succeed.
672 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
673 * the reserve counters are updated with the hugetlb_lock held. It is safe
674 * to reset the VMA at fork() time as it is not in use yet and there is no
675 * chance of the global counters getting corrupted as a result of the values.
677 * The private mapping reservation is represented in a subtly different
678 * manner to a shared mapping. A shared mapping has a region map associated
679 * with the underlying file, this region map represents the backing file
680 * pages which have ever had a reservation assigned which this persists even
681 * after the page is instantiated. A private mapping has a region map
682 * associated with the original mmap which is attached to all VMAs which
683 * reference it, this region map represents those offsets which have consumed
684 * reservation ie. where pages have been instantiated.
686 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
688 return (unsigned long)vma->vm_private_data;
691 static void set_vma_private_data(struct vm_area_struct *vma,
694 vma->vm_private_data = (void *)value;
697 struct resv_map *resv_map_alloc(void)
699 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
700 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
702 if (!resv_map || !rg) {
708 kref_init(&resv_map->refs);
709 spin_lock_init(&resv_map->lock);
710 INIT_LIST_HEAD(&resv_map->regions);
712 resv_map->adds_in_progress = 0;
714 INIT_LIST_HEAD(&resv_map->region_cache);
715 list_add(&rg->link, &resv_map->region_cache);
716 resv_map->region_cache_count = 1;
721 void resv_map_release(struct kref *ref)
723 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
724 struct list_head *head = &resv_map->region_cache;
725 struct file_region *rg, *trg;
727 /* Clear out any active regions before we release the map. */
728 region_del(resv_map, 0, LONG_MAX);
730 /* ... and any entries left in the cache */
731 list_for_each_entry_safe(rg, trg, head, link) {
736 VM_BUG_ON(resv_map->adds_in_progress);
741 static inline struct resv_map *inode_resv_map(struct inode *inode)
743 return inode->i_mapping->private_data;
746 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
748 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
749 if (vma->vm_flags & VM_MAYSHARE) {
750 struct address_space *mapping = vma->vm_file->f_mapping;
751 struct inode *inode = mapping->host;
753 return inode_resv_map(inode);
756 return (struct resv_map *)(get_vma_private_data(vma) &
761 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
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) &
767 HPAGE_RESV_MASK) | (unsigned long)map);
770 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
772 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
773 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
775 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
778 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
780 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
782 return (get_vma_private_data(vma) & flag) != 0;
785 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
786 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
788 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
789 if (!(vma->vm_flags & VM_MAYSHARE))
790 vma->vm_private_data = (void *)0;
793 /* Returns true if the VMA has associated reserve pages */
794 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
796 if (vma->vm_flags & VM_NORESERVE) {
798 * This address is already reserved by other process(chg == 0),
799 * so, we should decrement reserved count. Without decrementing,
800 * reserve count remains after releasing inode, because this
801 * allocated page will go into page cache and is regarded as
802 * coming from reserved pool in releasing step. Currently, we
803 * don't have any other solution to deal with this situation
804 * properly, so add work-around here.
806 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
812 /* Shared mappings always use reserves */
813 if (vma->vm_flags & VM_MAYSHARE) {
815 * We know VM_NORESERVE is not set. Therefore, there SHOULD
816 * be a region map for all pages. The only situation where
817 * there is no region map is if a hole was punched via
818 * fallocate. In this case, there really are no reverves to
819 * use. This situation is indicated if chg != 0.
828 * Only the process that called mmap() has reserves for
831 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
837 static void enqueue_huge_page(struct hstate *h, struct page *page)
839 int nid = page_to_nid(page);
840 list_move(&page->lru, &h->hugepage_freelists[nid]);
841 h->free_huge_pages++;
842 h->free_huge_pages_node[nid]++;
845 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
849 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
850 if (!is_migrate_isolate_page(page))
853 * if 'non-isolated free hugepage' not found on the list,
854 * the allocation fails.
856 if (&h->hugepage_freelists[nid] == &page->lru)
858 list_move(&page->lru, &h->hugepage_activelist);
859 set_page_refcounted(page);
860 h->free_huge_pages--;
861 h->free_huge_pages_node[nid]--;
865 /* Movability of hugepages depends on migration support. */
866 static inline gfp_t htlb_alloc_mask(struct hstate *h)
868 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
869 return GFP_HIGHUSER_MOVABLE;
874 static struct page *dequeue_huge_page_vma(struct hstate *h,
875 struct vm_area_struct *vma,
876 unsigned long address, int avoid_reserve,
879 struct page *page = NULL;
880 struct mempolicy *mpol;
881 nodemask_t *nodemask;
882 struct zonelist *zonelist;
885 unsigned int cpuset_mems_cookie;
888 * A child process with MAP_PRIVATE mappings created by their parent
889 * have no page reserves. This check ensures that reservations are
890 * not "stolen". The child may still get SIGKILLed
892 if (!vma_has_reserves(vma, chg) &&
893 h->free_huge_pages - h->resv_huge_pages == 0)
896 /* If reserves cannot be used, ensure enough pages are in the pool */
897 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
901 cpuset_mems_cookie = read_mems_allowed_begin();
902 zonelist = huge_zonelist(vma, address,
903 htlb_alloc_mask(h), &mpol, &nodemask);
905 for_each_zone_zonelist_nodemask(zone, z, zonelist,
906 MAX_NR_ZONES - 1, nodemask) {
907 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
908 page = dequeue_huge_page_node(h, zone_to_nid(zone));
912 if (!vma_has_reserves(vma, chg))
915 SetPagePrivate(page);
916 h->resv_huge_pages--;
923 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
932 * common helper functions for hstate_next_node_to_{alloc|free}.
933 * We may have allocated or freed a huge page based on a different
934 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
935 * be outside of *nodes_allowed. Ensure that we use an allowed
936 * node for alloc or free.
938 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
940 nid = next_node(nid, *nodes_allowed);
941 if (nid == MAX_NUMNODES)
942 nid = first_node(*nodes_allowed);
943 VM_BUG_ON(nid >= MAX_NUMNODES);
948 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
950 if (!node_isset(nid, *nodes_allowed))
951 nid = next_node_allowed(nid, nodes_allowed);
956 * returns the previously saved node ["this node"] from which to
957 * allocate a persistent huge page for the pool and advance the
958 * next node from which to allocate, handling wrap at end of node
961 static int hstate_next_node_to_alloc(struct hstate *h,
962 nodemask_t *nodes_allowed)
966 VM_BUG_ON(!nodes_allowed);
968 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
969 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
975 * helper for free_pool_huge_page() - return the previously saved
976 * node ["this node"] from which to free a huge page. Advance the
977 * next node id whether or not we find a free huge page to free so
978 * that the next attempt to free addresses the next node.
980 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
984 VM_BUG_ON(!nodes_allowed);
986 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
987 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
992 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
993 for (nr_nodes = nodes_weight(*mask); \
995 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
998 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
999 for (nr_nodes = nodes_weight(*mask); \
1001 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1004 #if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
1005 static void destroy_compound_gigantic_page(struct page *page,
1009 int nr_pages = 1 << order;
1010 struct page *p = page + 1;
1012 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1013 clear_compound_head(p);
1014 set_page_refcounted(p);
1017 set_compound_order(page, 0);
1018 __ClearPageHead(page);
1021 static void free_gigantic_page(struct page *page, unsigned int order)
1023 free_contig_range(page_to_pfn(page), 1 << order);
1026 static int __alloc_gigantic_page(unsigned long start_pfn,
1027 unsigned long nr_pages)
1029 unsigned long end_pfn = start_pfn + nr_pages;
1030 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1033 static bool pfn_range_valid_gigantic(unsigned long start_pfn,
1034 unsigned long nr_pages)
1036 unsigned long i, end_pfn = start_pfn + nr_pages;
1039 for (i = start_pfn; i < end_pfn; i++) {
1043 page = pfn_to_page(i);
1045 if (PageReserved(page))
1048 if (page_count(page) > 0)
1058 static bool zone_spans_last_pfn(const struct zone *zone,
1059 unsigned long start_pfn, unsigned long nr_pages)
1061 unsigned long last_pfn = start_pfn + nr_pages - 1;
1062 return zone_spans_pfn(zone, last_pfn);
1065 static struct page *alloc_gigantic_page(int nid, unsigned int order)
1067 unsigned long nr_pages = 1 << order;
1068 unsigned long ret, pfn, flags;
1071 z = NODE_DATA(nid)->node_zones;
1072 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
1073 spin_lock_irqsave(&z->lock, flags);
1075 pfn = ALIGN(z->zone_start_pfn, nr_pages);
1076 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
1077 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
1079 * We release the zone lock here because
1080 * alloc_contig_range() will also lock the zone
1081 * at some point. If there's an allocation
1082 * spinning on this lock, it may win the race
1083 * and cause alloc_contig_range() to fail...
1085 spin_unlock_irqrestore(&z->lock, flags);
1086 ret = __alloc_gigantic_page(pfn, nr_pages);
1088 return pfn_to_page(pfn);
1089 spin_lock_irqsave(&z->lock, flags);
1094 spin_unlock_irqrestore(&z->lock, flags);
1100 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1101 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1103 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1107 page = alloc_gigantic_page(nid, huge_page_order(h));
1109 prep_compound_gigantic_page(page, huge_page_order(h));
1110 prep_new_huge_page(h, page, nid);
1116 static int alloc_fresh_gigantic_page(struct hstate *h,
1117 nodemask_t *nodes_allowed)
1119 struct page *page = NULL;
1122 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1123 page = alloc_fresh_gigantic_page_node(h, node);
1131 static inline bool gigantic_page_supported(void) { return true; }
1133 static inline bool gigantic_page_supported(void) { return false; }
1134 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1135 static inline void destroy_compound_gigantic_page(struct page *page,
1136 unsigned int order) { }
1137 static inline int alloc_fresh_gigantic_page(struct hstate *h,
1138 nodemask_t *nodes_allowed) { return 0; }
1141 static void update_and_free_page(struct hstate *h, struct page *page)
1145 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1149 h->nr_huge_pages_node[page_to_nid(page)]--;
1150 for (i = 0; i < pages_per_huge_page(h); i++) {
1151 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1152 1 << PG_referenced | 1 << PG_dirty |
1153 1 << PG_active | 1 << PG_private |
1156 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1157 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1158 set_page_refcounted(page);
1159 if (hstate_is_gigantic(h)) {
1160 destroy_compound_gigantic_page(page, huge_page_order(h));
1161 free_gigantic_page(page, huge_page_order(h));
1163 __free_pages(page, huge_page_order(h));
1167 struct hstate *size_to_hstate(unsigned long size)
1171 for_each_hstate(h) {
1172 if (huge_page_size(h) == size)
1179 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1180 * to hstate->hugepage_activelist.)
1182 * This function can be called for tail pages, but never returns true for them.
1184 bool page_huge_active(struct page *page)
1186 VM_BUG_ON_PAGE(!PageHuge(page), page);
1187 return PageHead(page) && PagePrivate(&page[1]);
1190 /* never called for tail page */
1191 static void set_page_huge_active(struct page *page)
1193 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1194 SetPagePrivate(&page[1]);
1197 static void clear_page_huge_active(struct page *page)
1199 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1200 ClearPagePrivate(&page[1]);
1203 void free_huge_page(struct page *page)
1206 * Can't pass hstate in here because it is called from the
1207 * compound page destructor.
1209 struct hstate *h = page_hstate(page);
1210 int nid = page_to_nid(page);
1211 struct hugepage_subpool *spool =
1212 (struct hugepage_subpool *)page_private(page);
1213 bool restore_reserve;
1215 set_page_private(page, 0);
1216 page->mapping = NULL;
1217 BUG_ON(page_count(page));
1218 BUG_ON(page_mapcount(page));
1219 restore_reserve = PagePrivate(page);
1220 ClearPagePrivate(page);
1223 * A return code of zero implies that the subpool will be under its
1224 * minimum size if the reservation is not restored after page is free.
1225 * Therefore, force restore_reserve operation.
1227 if (hugepage_subpool_put_pages(spool, 1) == 0)
1228 restore_reserve = true;
1230 spin_lock(&hugetlb_lock);
1231 clear_page_huge_active(page);
1232 hugetlb_cgroup_uncharge_page(hstate_index(h),
1233 pages_per_huge_page(h), page);
1234 if (restore_reserve)
1235 h->resv_huge_pages++;
1237 if (h->surplus_huge_pages_node[nid]) {
1238 /* remove the page from active list */
1239 list_del(&page->lru);
1240 update_and_free_page(h, page);
1241 h->surplus_huge_pages--;
1242 h->surplus_huge_pages_node[nid]--;
1244 arch_clear_hugepage_flags(page);
1245 enqueue_huge_page(h, page);
1247 spin_unlock(&hugetlb_lock);
1250 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1252 INIT_LIST_HEAD(&page->lru);
1253 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1254 spin_lock(&hugetlb_lock);
1255 set_hugetlb_cgroup(page, NULL);
1257 h->nr_huge_pages_node[nid]++;
1258 spin_unlock(&hugetlb_lock);
1259 put_page(page); /* free it into the hugepage allocator */
1262 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1265 int nr_pages = 1 << order;
1266 struct page *p = page + 1;
1268 /* we rely on prep_new_huge_page to set the destructor */
1269 set_compound_order(page, order);
1270 __SetPageHead(page);
1271 __ClearPageReserved(page);
1272 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1274 * For gigantic hugepages allocated through bootmem at
1275 * boot, it's safer to be consistent with the not-gigantic
1276 * hugepages and clear the PG_reserved bit from all tail pages
1277 * too. Otherwse drivers using get_user_pages() to access tail
1278 * pages may get the reference counting wrong if they see
1279 * PG_reserved set on a tail page (despite the head page not
1280 * having PG_reserved set). Enforcing this consistency between
1281 * head and tail pages allows drivers to optimize away a check
1282 * on the head page when they need know if put_page() is needed
1283 * after get_user_pages().
1285 __ClearPageReserved(p);
1286 set_page_count(p, 0);
1287 set_compound_head(p, page);
1292 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1293 * transparent huge pages. See the PageTransHuge() documentation for more
1296 int PageHuge(struct page *page)
1298 if (!PageCompound(page))
1301 page = compound_head(page);
1302 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1304 EXPORT_SYMBOL_GPL(PageHuge);
1307 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1308 * normal or transparent huge pages.
1310 int PageHeadHuge(struct page *page_head)
1312 if (!PageHead(page_head))
1315 return get_compound_page_dtor(page_head) == free_huge_page;
1318 pgoff_t __basepage_index(struct page *page)
1320 struct page *page_head = compound_head(page);
1321 pgoff_t index = page_index(page_head);
1322 unsigned long compound_idx;
1324 if (!PageHuge(page_head))
1325 return page_index(page);
1327 if (compound_order(page_head) >= MAX_ORDER)
1328 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1330 compound_idx = page - page_head;
1332 return (index << compound_order(page_head)) + compound_idx;
1335 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1339 page = __alloc_pages_node(nid,
1340 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1341 __GFP_REPEAT|__GFP_NOWARN,
1342 huge_page_order(h));
1344 prep_new_huge_page(h, page, nid);
1350 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1356 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1357 page = alloc_fresh_huge_page_node(h, node);
1365 count_vm_event(HTLB_BUDDY_PGALLOC);
1367 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1373 * Free huge page from pool from next node to free.
1374 * Attempt to keep persistent huge pages more or less
1375 * balanced over allowed nodes.
1376 * Called with hugetlb_lock locked.
1378 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1384 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1386 * If we're returning unused surplus pages, only examine
1387 * nodes with surplus pages.
1389 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1390 !list_empty(&h->hugepage_freelists[node])) {
1392 list_entry(h->hugepage_freelists[node].next,
1394 list_del(&page->lru);
1395 h->free_huge_pages--;
1396 h->free_huge_pages_node[node]--;
1398 h->surplus_huge_pages--;
1399 h->surplus_huge_pages_node[node]--;
1401 update_and_free_page(h, page);
1411 * Dissolve a given free hugepage into free buddy pages. This function does
1412 * nothing for in-use (including surplus) hugepages.
1414 static void dissolve_free_huge_page(struct page *page)
1416 spin_lock(&hugetlb_lock);
1417 if (PageHuge(page) && !page_count(page)) {
1418 struct hstate *h = page_hstate(page);
1419 int nid = page_to_nid(page);
1420 list_del(&page->lru);
1421 h->free_huge_pages--;
1422 h->free_huge_pages_node[nid]--;
1423 update_and_free_page(h, page);
1425 spin_unlock(&hugetlb_lock);
1429 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1430 * make specified memory blocks removable from the system.
1431 * Note that start_pfn should aligned with (minimum) hugepage size.
1433 void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1437 if (!hugepages_supported())
1440 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order));
1441 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
1442 dissolve_free_huge_page(pfn_to_page(pfn));
1446 * There are 3 ways this can get called:
1447 * 1. With vma+addr: we use the VMA's memory policy
1448 * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge
1449 * page from any node, and let the buddy allocator itself figure
1451 * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page
1452 * strictly from 'nid'
1454 static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1455 struct vm_area_struct *vma, unsigned long addr, int nid)
1457 int order = huge_page_order(h);
1458 gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN;
1459 unsigned int cpuset_mems_cookie;
1462 * We need a VMA to get a memory policy. If we do not
1463 * have one, we use the 'nid' argument.
1465 * The mempolicy stuff below has some non-inlined bits
1466 * and calls ->vm_ops. That makes it hard to optimize at
1467 * compile-time, even when NUMA is off and it does
1468 * nothing. This helps the compiler optimize it out.
1470 if (!IS_ENABLED(CONFIG_NUMA) || !vma) {
1472 * If a specific node is requested, make sure to
1473 * get memory from there, but only when a node
1474 * is explicitly specified.
1476 if (nid != NUMA_NO_NODE)
1477 gfp |= __GFP_THISNODE;
1479 * Make sure to call something that can handle
1482 return alloc_pages_node(nid, gfp, order);
1486 * OK, so we have a VMA. Fetch the mempolicy and try to
1487 * allocate a huge page with it. We will only reach this
1488 * when CONFIG_NUMA=y.
1492 struct mempolicy *mpol;
1493 struct zonelist *zl;
1494 nodemask_t *nodemask;
1496 cpuset_mems_cookie = read_mems_allowed_begin();
1497 zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask);
1498 mpol_cond_put(mpol);
1499 page = __alloc_pages_nodemask(gfp, order, zl, nodemask);
1502 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1508 * There are two ways to allocate a huge page:
1509 * 1. When you have a VMA and an address (like a fault)
1510 * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1512 * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in
1513 * this case which signifies that the allocation should be done with
1514 * respect for the VMA's memory policy.
1516 * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1517 * implies that memory policies will not be taken in to account.
1519 static struct page *__alloc_buddy_huge_page(struct hstate *h,
1520 struct vm_area_struct *vma, unsigned long addr, int nid)
1525 if (hstate_is_gigantic(h))
1529 * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1530 * This makes sure the caller is picking _one_ of the modes with which
1531 * we can call this function, not both.
1533 if (vma || (addr != -1)) {
1534 VM_WARN_ON_ONCE(addr == -1);
1535 VM_WARN_ON_ONCE(nid != NUMA_NO_NODE);
1538 * Assume we will successfully allocate the surplus page to
1539 * prevent racing processes from causing the surplus to exceed
1542 * This however introduces a different race, where a process B
1543 * tries to grow the static hugepage pool while alloc_pages() is
1544 * called by process A. B will only examine the per-node
1545 * counters in determining if surplus huge pages can be
1546 * converted to normal huge pages in adjust_pool_surplus(). A
1547 * won't be able to increment the per-node counter, until the
1548 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1549 * no more huge pages can be converted from surplus to normal
1550 * state (and doesn't try to convert again). Thus, we have a
1551 * case where a surplus huge page exists, the pool is grown, and
1552 * the surplus huge page still exists after, even though it
1553 * should just have been converted to a normal huge page. This
1554 * does not leak memory, though, as the hugepage will be freed
1555 * once it is out of use. It also does not allow the counters to
1556 * go out of whack in adjust_pool_surplus() as we don't modify
1557 * the node values until we've gotten the hugepage and only the
1558 * per-node value is checked there.
1560 spin_lock(&hugetlb_lock);
1561 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1562 spin_unlock(&hugetlb_lock);
1566 h->surplus_huge_pages++;
1568 spin_unlock(&hugetlb_lock);
1570 page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid);
1572 spin_lock(&hugetlb_lock);
1574 INIT_LIST_HEAD(&page->lru);
1575 r_nid = page_to_nid(page);
1576 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1577 set_hugetlb_cgroup(page, NULL);
1579 * We incremented the global counters already
1581 h->nr_huge_pages_node[r_nid]++;
1582 h->surplus_huge_pages_node[r_nid]++;
1583 __count_vm_event(HTLB_BUDDY_PGALLOC);
1586 h->surplus_huge_pages--;
1587 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1589 spin_unlock(&hugetlb_lock);
1595 * Allocate a huge page from 'nid'. Note, 'nid' may be
1596 * NUMA_NO_NODE, which means that it may be allocated
1600 struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid)
1602 unsigned long addr = -1;
1604 return __alloc_buddy_huge_page(h, NULL, addr, nid);
1608 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1611 struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1612 struct vm_area_struct *vma, unsigned long addr)
1614 return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE);
1618 * This allocation function is useful in the context where vma is irrelevant.
1619 * E.g. soft-offlining uses this function because it only cares physical
1620 * address of error page.
1622 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1624 struct page *page = NULL;
1626 spin_lock(&hugetlb_lock);
1627 if (h->free_huge_pages - h->resv_huge_pages > 0)
1628 page = dequeue_huge_page_node(h, nid);
1629 spin_unlock(&hugetlb_lock);
1632 page = __alloc_buddy_huge_page_no_mpol(h, nid);
1638 * Increase the hugetlb pool such that it can accommodate a reservation
1641 static int gather_surplus_pages(struct hstate *h, int delta)
1643 struct list_head surplus_list;
1644 struct page *page, *tmp;
1646 int needed, allocated;
1647 bool alloc_ok = true;
1649 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1651 h->resv_huge_pages += delta;
1656 INIT_LIST_HEAD(&surplus_list);
1660 spin_unlock(&hugetlb_lock);
1661 for (i = 0; i < needed; i++) {
1662 page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE);
1667 list_add(&page->lru, &surplus_list);
1672 * After retaking hugetlb_lock, we need to recalculate 'needed'
1673 * because either resv_huge_pages or free_huge_pages may have changed.
1675 spin_lock(&hugetlb_lock);
1676 needed = (h->resv_huge_pages + delta) -
1677 (h->free_huge_pages + allocated);
1682 * We were not able to allocate enough pages to
1683 * satisfy the entire reservation so we free what
1684 * we've allocated so far.
1689 * The surplus_list now contains _at_least_ the number of extra pages
1690 * needed to accommodate the reservation. Add the appropriate number
1691 * of pages to the hugetlb pool and free the extras back to the buddy
1692 * allocator. Commit the entire reservation here to prevent another
1693 * process from stealing the pages as they are added to the pool but
1694 * before they are reserved.
1696 needed += allocated;
1697 h->resv_huge_pages += delta;
1700 /* Free the needed pages to the hugetlb pool */
1701 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1705 * This page is now managed by the hugetlb allocator and has
1706 * no users -- drop the buddy allocator's reference.
1708 put_page_testzero(page);
1709 VM_BUG_ON_PAGE(page_count(page), page);
1710 enqueue_huge_page(h, page);
1713 spin_unlock(&hugetlb_lock);
1715 /* Free unnecessary surplus pages to the buddy allocator */
1716 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1718 spin_lock(&hugetlb_lock);
1724 * When releasing a hugetlb pool reservation, any surplus pages that were
1725 * allocated to satisfy the reservation must be explicitly freed if they were
1727 * Called with hugetlb_lock held.
1729 static void return_unused_surplus_pages(struct hstate *h,
1730 unsigned long unused_resv_pages)
1732 unsigned long nr_pages;
1734 /* Uncommit the reservation */
1735 h->resv_huge_pages -= unused_resv_pages;
1737 /* Cannot return gigantic pages currently */
1738 if (hstate_is_gigantic(h))
1741 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1744 * We want to release as many surplus pages as possible, spread
1745 * evenly across all nodes with memory. Iterate across these nodes
1746 * until we can no longer free unreserved surplus pages. This occurs
1747 * when the nodes with surplus pages have no free pages.
1748 * free_pool_huge_page() will balance the the freed pages across the
1749 * on-line nodes with memory and will handle the hstate accounting.
1751 while (nr_pages--) {
1752 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1754 cond_resched_lock(&hugetlb_lock);
1760 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1761 * are used by the huge page allocation routines to manage reservations.
1763 * vma_needs_reservation is called to determine if the huge page at addr
1764 * within the vma has an associated reservation. If a reservation is
1765 * needed, the value 1 is returned. The caller is then responsible for
1766 * managing the global reservation and subpool usage counts. After
1767 * the huge page has been allocated, vma_commit_reservation is called
1768 * to add the page to the reservation map. If the page allocation fails,
1769 * the reservation must be ended instead of committed. vma_end_reservation
1770 * is called in such cases.
1772 * In the normal case, vma_commit_reservation returns the same value
1773 * as the preceding vma_needs_reservation call. The only time this
1774 * is not the case is if a reserve map was changed between calls. It
1775 * is the responsibility of the caller to notice the difference and
1776 * take appropriate action.
1778 enum vma_resv_mode {
1783 static long __vma_reservation_common(struct hstate *h,
1784 struct vm_area_struct *vma, unsigned long addr,
1785 enum vma_resv_mode mode)
1787 struct resv_map *resv;
1791 resv = vma_resv_map(vma);
1795 idx = vma_hugecache_offset(h, vma, addr);
1797 case VMA_NEEDS_RESV:
1798 ret = region_chg(resv, idx, idx + 1);
1800 case VMA_COMMIT_RESV:
1801 ret = region_add(resv, idx, idx + 1);
1804 region_abort(resv, idx, idx + 1);
1811 if (vma->vm_flags & VM_MAYSHARE)
1814 return ret < 0 ? ret : 0;
1817 static long vma_needs_reservation(struct hstate *h,
1818 struct vm_area_struct *vma, unsigned long addr)
1820 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1823 static long vma_commit_reservation(struct hstate *h,
1824 struct vm_area_struct *vma, unsigned long addr)
1826 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1829 static void vma_end_reservation(struct hstate *h,
1830 struct vm_area_struct *vma, unsigned long addr)
1832 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1835 struct page *alloc_huge_page(struct vm_area_struct *vma,
1836 unsigned long addr, int avoid_reserve)
1838 struct hugepage_subpool *spool = subpool_vma(vma);
1839 struct hstate *h = hstate_vma(vma);
1841 long map_chg, map_commit;
1844 struct hugetlb_cgroup *h_cg;
1846 idx = hstate_index(h);
1848 * Examine the region/reserve map to determine if the process
1849 * has a reservation for the page to be allocated. A return
1850 * code of zero indicates a reservation exists (no change).
1852 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
1854 return ERR_PTR(-ENOMEM);
1857 * Processes that did not create the mapping will have no
1858 * reserves as indicated by the region/reserve map. Check
1859 * that the allocation will not exceed the subpool limit.
1860 * Allocations for MAP_NORESERVE mappings also need to be
1861 * checked against any subpool limit.
1863 if (map_chg || avoid_reserve) {
1864 gbl_chg = hugepage_subpool_get_pages(spool, 1);
1866 vma_end_reservation(h, vma, addr);
1867 return ERR_PTR(-ENOSPC);
1871 * Even though there was no reservation in the region/reserve
1872 * map, there could be reservations associated with the
1873 * subpool that can be used. This would be indicated if the
1874 * return value of hugepage_subpool_get_pages() is zero.
1875 * However, if avoid_reserve is specified we still avoid even
1876 * the subpool reservations.
1882 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1884 goto out_subpool_put;
1886 spin_lock(&hugetlb_lock);
1888 * glb_chg is passed to indicate whether or not a page must be taken
1889 * from the global free pool (global change). gbl_chg == 0 indicates
1890 * a reservation exists for the allocation.
1892 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
1894 spin_unlock(&hugetlb_lock);
1895 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
1897 goto out_uncharge_cgroup;
1898 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
1899 SetPagePrivate(page);
1900 h->resv_huge_pages--;
1902 spin_lock(&hugetlb_lock);
1903 list_move(&page->lru, &h->hugepage_activelist);
1906 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1907 spin_unlock(&hugetlb_lock);
1909 set_page_private(page, (unsigned long)spool);
1911 map_commit = vma_commit_reservation(h, vma, addr);
1912 if (unlikely(map_chg > map_commit)) {
1914 * The page was added to the reservation map between
1915 * vma_needs_reservation and vma_commit_reservation.
1916 * This indicates a race with hugetlb_reserve_pages.
1917 * Adjust for the subpool count incremented above AND
1918 * in hugetlb_reserve_pages for the same page. Also,
1919 * the reservation count added in hugetlb_reserve_pages
1920 * no longer applies.
1924 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
1925 hugetlb_acct_memory(h, -rsv_adjust);
1929 out_uncharge_cgroup:
1930 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1932 if (map_chg || avoid_reserve)
1933 hugepage_subpool_put_pages(spool, 1);
1934 vma_end_reservation(h, vma, addr);
1935 return ERR_PTR(-ENOSPC);
1939 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1940 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1941 * where no ERR_VALUE is expected to be returned.
1943 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1944 unsigned long addr, int avoid_reserve)
1946 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1952 int __weak alloc_bootmem_huge_page(struct hstate *h)
1954 struct huge_bootmem_page *m;
1957 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1960 addr = memblock_virt_alloc_try_nid_nopanic(
1961 huge_page_size(h), huge_page_size(h),
1962 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1965 * Use the beginning of the huge page to store the
1966 * huge_bootmem_page struct (until gather_bootmem
1967 * puts them into the mem_map).
1976 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
1977 /* Put them into a private list first because mem_map is not up yet */
1978 list_add(&m->list, &huge_boot_pages);
1983 static void __init prep_compound_huge_page(struct page *page,
1986 if (unlikely(order > (MAX_ORDER - 1)))
1987 prep_compound_gigantic_page(page, order);
1989 prep_compound_page(page, order);
1992 /* Put bootmem huge pages into the standard lists after mem_map is up */
1993 static void __init gather_bootmem_prealloc(void)
1995 struct huge_bootmem_page *m;
1997 list_for_each_entry(m, &huge_boot_pages, list) {
1998 struct hstate *h = m->hstate;
2001 #ifdef CONFIG_HIGHMEM
2002 page = pfn_to_page(m->phys >> PAGE_SHIFT);
2003 memblock_free_late(__pa(m),
2004 sizeof(struct huge_bootmem_page));
2006 page = virt_to_page(m);
2008 WARN_ON(page_count(page) != 1);
2009 prep_compound_huge_page(page, h->order);
2010 WARN_ON(PageReserved(page));
2011 prep_new_huge_page(h, page, page_to_nid(page));
2013 * If we had gigantic hugepages allocated at boot time, we need
2014 * to restore the 'stolen' pages to totalram_pages in order to
2015 * fix confusing memory reports from free(1) and another
2016 * side-effects, like CommitLimit going negative.
2018 if (hstate_is_gigantic(h))
2019 adjust_managed_page_count(page, 1 << h->order);
2023 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2027 for (i = 0; i < h->max_huge_pages; ++i) {
2028 if (hstate_is_gigantic(h)) {
2029 if (!alloc_bootmem_huge_page(h))
2031 } else if (!alloc_fresh_huge_page(h,
2032 &node_states[N_MEMORY]))
2035 h->max_huge_pages = i;
2038 static void __init hugetlb_init_hstates(void)
2042 for_each_hstate(h) {
2043 if (minimum_order > huge_page_order(h))
2044 minimum_order = huge_page_order(h);
2046 /* oversize hugepages were init'ed in early boot */
2047 if (!hstate_is_gigantic(h))
2048 hugetlb_hstate_alloc_pages(h);
2050 VM_BUG_ON(minimum_order == UINT_MAX);
2053 static char * __init memfmt(char *buf, unsigned long n)
2055 if (n >= (1UL << 30))
2056 sprintf(buf, "%lu GB", n >> 30);
2057 else if (n >= (1UL << 20))
2058 sprintf(buf, "%lu MB", n >> 20);
2060 sprintf(buf, "%lu KB", n >> 10);
2064 static void __init report_hugepages(void)
2068 for_each_hstate(h) {
2070 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2071 memfmt(buf, huge_page_size(h)),
2072 h->free_huge_pages);
2076 #ifdef CONFIG_HIGHMEM
2077 static void try_to_free_low(struct hstate *h, unsigned long count,
2078 nodemask_t *nodes_allowed)
2082 if (hstate_is_gigantic(h))
2085 for_each_node_mask(i, *nodes_allowed) {
2086 struct page *page, *next;
2087 struct list_head *freel = &h->hugepage_freelists[i];
2088 list_for_each_entry_safe(page, next, freel, lru) {
2089 if (count >= h->nr_huge_pages)
2091 if (PageHighMem(page))
2093 list_del(&page->lru);
2094 update_and_free_page(h, page);
2095 h->free_huge_pages--;
2096 h->free_huge_pages_node[page_to_nid(page)]--;
2101 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2102 nodemask_t *nodes_allowed)
2108 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2109 * balanced by operating on them in a round-robin fashion.
2110 * Returns 1 if an adjustment was made.
2112 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2117 VM_BUG_ON(delta != -1 && delta != 1);
2120 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2121 if (h->surplus_huge_pages_node[node])
2125 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2126 if (h->surplus_huge_pages_node[node] <
2127 h->nr_huge_pages_node[node])
2134 h->surplus_huge_pages += delta;
2135 h->surplus_huge_pages_node[node] += delta;
2139 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2140 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2141 nodemask_t *nodes_allowed)
2143 unsigned long min_count, ret;
2145 if (hstate_is_gigantic(h) && !gigantic_page_supported())
2146 return h->max_huge_pages;
2149 * Increase the pool size
2150 * First take pages out of surplus state. Then make up the
2151 * remaining difference by allocating fresh huge pages.
2153 * We might race with __alloc_buddy_huge_page() here and be unable
2154 * to convert a surplus huge page to a normal huge page. That is
2155 * not critical, though, it just means the overall size of the
2156 * pool might be one hugepage larger than it needs to be, but
2157 * within all the constraints specified by the sysctls.
2159 spin_lock(&hugetlb_lock);
2160 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2161 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2165 while (count > persistent_huge_pages(h)) {
2167 * If this allocation races such that we no longer need the
2168 * page, free_huge_page will handle it by freeing the page
2169 * and reducing the surplus.
2171 spin_unlock(&hugetlb_lock);
2172 if (hstate_is_gigantic(h))
2173 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2175 ret = alloc_fresh_huge_page(h, nodes_allowed);
2176 spin_lock(&hugetlb_lock);
2180 /* Bail for signals. Probably ctrl-c from user */
2181 if (signal_pending(current))
2186 * Decrease the pool size
2187 * First return free pages to the buddy allocator (being careful
2188 * to keep enough around to satisfy reservations). Then place
2189 * pages into surplus state as needed so the pool will shrink
2190 * to the desired size as pages become free.
2192 * By placing pages into the surplus state independent of the
2193 * overcommit value, we are allowing the surplus pool size to
2194 * exceed overcommit. There are few sane options here. Since
2195 * __alloc_buddy_huge_page() is checking the global counter,
2196 * though, we'll note that we're not allowed to exceed surplus
2197 * and won't grow the pool anywhere else. Not until one of the
2198 * sysctls are changed, or the surplus pages go out of use.
2200 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2201 min_count = max(count, min_count);
2202 try_to_free_low(h, min_count, nodes_allowed);
2203 while (min_count < persistent_huge_pages(h)) {
2204 if (!free_pool_huge_page(h, nodes_allowed, 0))
2206 cond_resched_lock(&hugetlb_lock);
2208 while (count < persistent_huge_pages(h)) {
2209 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2213 ret = persistent_huge_pages(h);
2214 spin_unlock(&hugetlb_lock);
2218 #define HSTATE_ATTR_RO(_name) \
2219 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2221 #define HSTATE_ATTR(_name) \
2222 static struct kobj_attribute _name##_attr = \
2223 __ATTR(_name, 0644, _name##_show, _name##_store)
2225 static struct kobject *hugepages_kobj;
2226 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2228 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2230 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2234 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2235 if (hstate_kobjs[i] == kobj) {
2237 *nidp = NUMA_NO_NODE;
2241 return kobj_to_node_hstate(kobj, nidp);
2244 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2245 struct kobj_attribute *attr, char *buf)
2248 unsigned long nr_huge_pages;
2251 h = kobj_to_hstate(kobj, &nid);
2252 if (nid == NUMA_NO_NODE)
2253 nr_huge_pages = h->nr_huge_pages;
2255 nr_huge_pages = h->nr_huge_pages_node[nid];
2257 return sprintf(buf, "%lu\n", nr_huge_pages);
2260 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2261 struct hstate *h, int nid,
2262 unsigned long count, size_t len)
2265 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2267 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2272 if (nid == NUMA_NO_NODE) {
2274 * global hstate attribute
2276 if (!(obey_mempolicy &&
2277 init_nodemask_of_mempolicy(nodes_allowed))) {
2278 NODEMASK_FREE(nodes_allowed);
2279 nodes_allowed = &node_states[N_MEMORY];
2281 } else if (nodes_allowed) {
2283 * per node hstate attribute: adjust count to global,
2284 * but restrict alloc/free to the specified node.
2286 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2287 init_nodemask_of_node(nodes_allowed, nid);
2289 nodes_allowed = &node_states[N_MEMORY];
2291 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2293 if (nodes_allowed != &node_states[N_MEMORY])
2294 NODEMASK_FREE(nodes_allowed);
2298 NODEMASK_FREE(nodes_allowed);
2302 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2303 struct kobject *kobj, const char *buf,
2307 unsigned long count;
2311 err = kstrtoul(buf, 10, &count);
2315 h = kobj_to_hstate(kobj, &nid);
2316 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2319 static ssize_t nr_hugepages_show(struct kobject *kobj,
2320 struct kobj_attribute *attr, char *buf)
2322 return nr_hugepages_show_common(kobj, attr, buf);
2325 static ssize_t nr_hugepages_store(struct kobject *kobj,
2326 struct kobj_attribute *attr, const char *buf, size_t len)
2328 return nr_hugepages_store_common(false, kobj, buf, len);
2330 HSTATE_ATTR(nr_hugepages);
2335 * hstate attribute for optionally mempolicy-based constraint on persistent
2336 * huge page alloc/free.
2338 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2339 struct kobj_attribute *attr, char *buf)
2341 return nr_hugepages_show_common(kobj, attr, buf);
2344 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2345 struct kobj_attribute *attr, const char *buf, size_t len)
2347 return nr_hugepages_store_common(true, kobj, buf, len);
2349 HSTATE_ATTR(nr_hugepages_mempolicy);
2353 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2354 struct kobj_attribute *attr, char *buf)
2356 struct hstate *h = kobj_to_hstate(kobj, NULL);
2357 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2360 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2361 struct kobj_attribute *attr, const char *buf, size_t count)
2364 unsigned long input;
2365 struct hstate *h = kobj_to_hstate(kobj, NULL);
2367 if (hstate_is_gigantic(h))
2370 err = kstrtoul(buf, 10, &input);
2374 spin_lock(&hugetlb_lock);
2375 h->nr_overcommit_huge_pages = input;
2376 spin_unlock(&hugetlb_lock);
2380 HSTATE_ATTR(nr_overcommit_hugepages);
2382 static ssize_t free_hugepages_show(struct kobject *kobj,
2383 struct kobj_attribute *attr, char *buf)
2386 unsigned long free_huge_pages;
2389 h = kobj_to_hstate(kobj, &nid);
2390 if (nid == NUMA_NO_NODE)
2391 free_huge_pages = h->free_huge_pages;
2393 free_huge_pages = h->free_huge_pages_node[nid];
2395 return sprintf(buf, "%lu\n", free_huge_pages);
2397 HSTATE_ATTR_RO(free_hugepages);
2399 static ssize_t resv_hugepages_show(struct kobject *kobj,
2400 struct kobj_attribute *attr, char *buf)
2402 struct hstate *h = kobj_to_hstate(kobj, NULL);
2403 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2405 HSTATE_ATTR_RO(resv_hugepages);
2407 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2408 struct kobj_attribute *attr, char *buf)
2411 unsigned long surplus_huge_pages;
2414 h = kobj_to_hstate(kobj, &nid);
2415 if (nid == NUMA_NO_NODE)
2416 surplus_huge_pages = h->surplus_huge_pages;
2418 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2420 return sprintf(buf, "%lu\n", surplus_huge_pages);
2422 HSTATE_ATTR_RO(surplus_hugepages);
2424 static struct attribute *hstate_attrs[] = {
2425 &nr_hugepages_attr.attr,
2426 &nr_overcommit_hugepages_attr.attr,
2427 &free_hugepages_attr.attr,
2428 &resv_hugepages_attr.attr,
2429 &surplus_hugepages_attr.attr,
2431 &nr_hugepages_mempolicy_attr.attr,
2436 static struct attribute_group hstate_attr_group = {
2437 .attrs = hstate_attrs,
2440 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2441 struct kobject **hstate_kobjs,
2442 struct attribute_group *hstate_attr_group)
2445 int hi = hstate_index(h);
2447 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2448 if (!hstate_kobjs[hi])
2451 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2453 kobject_put(hstate_kobjs[hi]);
2458 static void __init hugetlb_sysfs_init(void)
2463 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2464 if (!hugepages_kobj)
2467 for_each_hstate(h) {
2468 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2469 hstate_kobjs, &hstate_attr_group);
2471 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2478 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2479 * with node devices in node_devices[] using a parallel array. The array
2480 * index of a node device or _hstate == node id.
2481 * This is here to avoid any static dependency of the node device driver, in
2482 * the base kernel, on the hugetlb module.
2484 struct node_hstate {
2485 struct kobject *hugepages_kobj;
2486 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2488 static struct node_hstate node_hstates[MAX_NUMNODES];
2491 * A subset of global hstate attributes for node devices
2493 static struct attribute *per_node_hstate_attrs[] = {
2494 &nr_hugepages_attr.attr,
2495 &free_hugepages_attr.attr,
2496 &surplus_hugepages_attr.attr,
2500 static struct attribute_group per_node_hstate_attr_group = {
2501 .attrs = per_node_hstate_attrs,
2505 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2506 * Returns node id via non-NULL nidp.
2508 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2512 for (nid = 0; nid < nr_node_ids; nid++) {
2513 struct node_hstate *nhs = &node_hstates[nid];
2515 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2516 if (nhs->hstate_kobjs[i] == kobj) {
2528 * Unregister hstate attributes from a single node device.
2529 * No-op if no hstate attributes attached.
2531 static void hugetlb_unregister_node(struct node *node)
2534 struct node_hstate *nhs = &node_hstates[node->dev.id];
2536 if (!nhs->hugepages_kobj)
2537 return; /* no hstate attributes */
2539 for_each_hstate(h) {
2540 int idx = hstate_index(h);
2541 if (nhs->hstate_kobjs[idx]) {
2542 kobject_put(nhs->hstate_kobjs[idx]);
2543 nhs->hstate_kobjs[idx] = NULL;
2547 kobject_put(nhs->hugepages_kobj);
2548 nhs->hugepages_kobj = NULL;
2553 * Register hstate attributes for a single node device.
2554 * No-op if attributes already registered.
2556 static void hugetlb_register_node(struct node *node)
2559 struct node_hstate *nhs = &node_hstates[node->dev.id];
2562 if (nhs->hugepages_kobj)
2563 return; /* already allocated */
2565 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2567 if (!nhs->hugepages_kobj)
2570 for_each_hstate(h) {
2571 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2573 &per_node_hstate_attr_group);
2575 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2576 h->name, node->dev.id);
2577 hugetlb_unregister_node(node);
2584 * hugetlb init time: register hstate attributes for all registered node
2585 * devices of nodes that have memory. All on-line nodes should have
2586 * registered their associated device by this time.
2588 static void __init hugetlb_register_all_nodes(void)
2592 for_each_node_state(nid, N_MEMORY) {
2593 struct node *node = node_devices[nid];
2594 if (node->dev.id == nid)
2595 hugetlb_register_node(node);
2599 * Let the node device driver know we're here so it can
2600 * [un]register hstate attributes on node hotplug.
2602 register_hugetlbfs_with_node(hugetlb_register_node,
2603 hugetlb_unregister_node);
2605 #else /* !CONFIG_NUMA */
2607 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2615 static void hugetlb_register_all_nodes(void) { }
2619 static int __init hugetlb_init(void)
2623 if (!hugepages_supported())
2626 if (!size_to_hstate(default_hstate_size)) {
2627 default_hstate_size = HPAGE_SIZE;
2628 if (!size_to_hstate(default_hstate_size))
2629 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2631 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2632 if (default_hstate_max_huge_pages)
2633 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2635 hugetlb_init_hstates();
2636 gather_bootmem_prealloc();
2639 hugetlb_sysfs_init();
2640 hugetlb_register_all_nodes();
2641 hugetlb_cgroup_file_init();
2644 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2646 num_fault_mutexes = 1;
2648 hugetlb_fault_mutex_table =
2649 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2650 BUG_ON(!hugetlb_fault_mutex_table);
2652 for (i = 0; i < num_fault_mutexes; i++)
2653 mutex_init(&hugetlb_fault_mutex_table[i]);
2656 subsys_initcall(hugetlb_init);
2658 /* Should be called on processing a hugepagesz=... option */
2659 void __init hugetlb_add_hstate(unsigned int order)
2664 if (size_to_hstate(PAGE_SIZE << order)) {
2665 pr_warning("hugepagesz= specified twice, ignoring\n");
2668 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2670 h = &hstates[hugetlb_max_hstate++];
2672 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2673 h->nr_huge_pages = 0;
2674 h->free_huge_pages = 0;
2675 for (i = 0; i < MAX_NUMNODES; ++i)
2676 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2677 INIT_LIST_HEAD(&h->hugepage_activelist);
2678 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2679 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2680 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2681 huge_page_size(h)/1024);
2686 static int __init hugetlb_nrpages_setup(char *s)
2689 static unsigned long *last_mhp;
2692 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2693 * so this hugepages= parameter goes to the "default hstate".
2695 if (!hugetlb_max_hstate)
2696 mhp = &default_hstate_max_huge_pages;
2698 mhp = &parsed_hstate->max_huge_pages;
2700 if (mhp == last_mhp) {
2701 pr_warning("hugepages= specified twice without "
2702 "interleaving hugepagesz=, ignoring\n");
2706 if (sscanf(s, "%lu", mhp) <= 0)
2710 * Global state is always initialized later in hugetlb_init.
2711 * But we need to allocate >= MAX_ORDER hstates here early to still
2712 * use the bootmem allocator.
2714 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2715 hugetlb_hstate_alloc_pages(parsed_hstate);
2721 __setup("hugepages=", hugetlb_nrpages_setup);
2723 static int __init hugetlb_default_setup(char *s)
2725 default_hstate_size = memparse(s, &s);
2728 __setup("default_hugepagesz=", hugetlb_default_setup);
2730 static unsigned int cpuset_mems_nr(unsigned int *array)
2733 unsigned int nr = 0;
2735 for_each_node_mask(node, cpuset_current_mems_allowed)
2741 #ifdef CONFIG_SYSCTL
2742 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2743 struct ctl_table *table, int write,
2744 void __user *buffer, size_t *length, loff_t *ppos)
2746 struct hstate *h = &default_hstate;
2747 unsigned long tmp = h->max_huge_pages;
2750 if (!hugepages_supported())
2754 table->maxlen = sizeof(unsigned long);
2755 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2760 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2761 NUMA_NO_NODE, tmp, *length);
2766 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2767 void __user *buffer, size_t *length, loff_t *ppos)
2770 return hugetlb_sysctl_handler_common(false, table, write,
2771 buffer, length, ppos);
2775 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2776 void __user *buffer, size_t *length, loff_t *ppos)
2778 return hugetlb_sysctl_handler_common(true, table, write,
2779 buffer, length, ppos);
2781 #endif /* CONFIG_NUMA */
2783 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2784 void __user *buffer,
2785 size_t *length, loff_t *ppos)
2787 struct hstate *h = &default_hstate;
2791 if (!hugepages_supported())
2794 tmp = h->nr_overcommit_huge_pages;
2796 if (write && hstate_is_gigantic(h))
2800 table->maxlen = sizeof(unsigned long);
2801 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2806 spin_lock(&hugetlb_lock);
2807 h->nr_overcommit_huge_pages = tmp;
2808 spin_unlock(&hugetlb_lock);
2814 #endif /* CONFIG_SYSCTL */
2816 void hugetlb_report_meminfo(struct seq_file *m)
2818 struct hstate *h = &default_hstate;
2819 if (!hugepages_supported())
2822 "HugePages_Total: %5lu\n"
2823 "HugePages_Free: %5lu\n"
2824 "HugePages_Rsvd: %5lu\n"
2825 "HugePages_Surp: %5lu\n"
2826 "Hugepagesize: %8lu kB\n",
2830 h->surplus_huge_pages,
2831 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2834 int hugetlb_report_node_meminfo(int nid, char *buf)
2836 struct hstate *h = &default_hstate;
2837 if (!hugepages_supported())
2840 "Node %d HugePages_Total: %5u\n"
2841 "Node %d HugePages_Free: %5u\n"
2842 "Node %d HugePages_Surp: %5u\n",
2843 nid, h->nr_huge_pages_node[nid],
2844 nid, h->free_huge_pages_node[nid],
2845 nid, h->surplus_huge_pages_node[nid]);
2848 void hugetlb_show_meminfo(void)
2853 if (!hugepages_supported())
2856 for_each_node_state(nid, N_MEMORY)
2858 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2860 h->nr_huge_pages_node[nid],
2861 h->free_huge_pages_node[nid],
2862 h->surplus_huge_pages_node[nid],
2863 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2866 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
2868 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
2869 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
2872 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2873 unsigned long hugetlb_total_pages(void)
2876 unsigned long nr_total_pages = 0;
2879 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2880 return nr_total_pages;
2883 static int hugetlb_acct_memory(struct hstate *h, long delta)
2887 spin_lock(&hugetlb_lock);
2889 * When cpuset is configured, it breaks the strict hugetlb page
2890 * reservation as the accounting is done on a global variable. Such
2891 * reservation is completely rubbish in the presence of cpuset because
2892 * the reservation is not checked against page availability for the
2893 * current cpuset. Application can still potentially OOM'ed by kernel
2894 * with lack of free htlb page in cpuset that the task is in.
2895 * Attempt to enforce strict accounting with cpuset is almost
2896 * impossible (or too ugly) because cpuset is too fluid that
2897 * task or memory node can be dynamically moved between cpusets.
2899 * The change of semantics for shared hugetlb mapping with cpuset is
2900 * undesirable. However, in order to preserve some of the semantics,
2901 * we fall back to check against current free page availability as
2902 * a best attempt and hopefully to minimize the impact of changing
2903 * semantics that cpuset has.
2906 if (gather_surplus_pages(h, delta) < 0)
2909 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2910 return_unused_surplus_pages(h, delta);
2917 return_unused_surplus_pages(h, (unsigned long) -delta);
2920 spin_unlock(&hugetlb_lock);
2924 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2926 struct resv_map *resv = vma_resv_map(vma);
2929 * This new VMA should share its siblings reservation map if present.
2930 * The VMA will only ever have a valid reservation map pointer where
2931 * it is being copied for another still existing VMA. As that VMA
2932 * has a reference to the reservation map it cannot disappear until
2933 * after this open call completes. It is therefore safe to take a
2934 * new reference here without additional locking.
2936 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2937 kref_get(&resv->refs);
2940 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2942 struct hstate *h = hstate_vma(vma);
2943 struct resv_map *resv = vma_resv_map(vma);
2944 struct hugepage_subpool *spool = subpool_vma(vma);
2945 unsigned long reserve, start, end;
2948 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2951 start = vma_hugecache_offset(h, vma, vma->vm_start);
2952 end = vma_hugecache_offset(h, vma, vma->vm_end);
2954 reserve = (end - start) - region_count(resv, start, end);
2956 kref_put(&resv->refs, resv_map_release);
2960 * Decrement reserve counts. The global reserve count may be
2961 * adjusted if the subpool has a minimum size.
2963 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
2964 hugetlb_acct_memory(h, -gbl_reserve);
2969 * We cannot handle pagefaults against hugetlb pages at all. They cause
2970 * handle_mm_fault() to try to instantiate regular-sized pages in the
2971 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2974 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2980 const struct vm_operations_struct hugetlb_vm_ops = {
2981 .fault = hugetlb_vm_op_fault,
2982 .open = hugetlb_vm_op_open,
2983 .close = hugetlb_vm_op_close,
2986 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2992 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2993 vma->vm_page_prot)));
2995 entry = huge_pte_wrprotect(mk_huge_pte(page,
2996 vma->vm_page_prot));
2998 entry = pte_mkyoung(entry);
2999 entry = pte_mkhuge(entry);
3000 entry = arch_make_huge_pte(entry, vma, page, writable);
3005 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3006 unsigned long address, pte_t *ptep)
3010 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3011 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3012 update_mmu_cache(vma, address, ptep);
3015 static int is_hugetlb_entry_migration(pte_t pte)
3019 if (huge_pte_none(pte) || pte_present(pte))
3021 swp = pte_to_swp_entry(pte);
3022 if (non_swap_entry(swp) && is_migration_entry(swp))
3028 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3032 if (huge_pte_none(pte) || pte_present(pte))
3034 swp = pte_to_swp_entry(pte);
3035 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3041 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3042 struct vm_area_struct *vma)
3044 pte_t *src_pte, *dst_pte, entry;
3045 struct page *ptepage;
3048 struct hstate *h = hstate_vma(vma);
3049 unsigned long sz = huge_page_size(h);
3050 unsigned long mmun_start; /* For mmu_notifiers */
3051 unsigned long mmun_end; /* For mmu_notifiers */
3054 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3056 mmun_start = vma->vm_start;
3057 mmun_end = vma->vm_end;
3059 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3061 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3062 spinlock_t *src_ptl, *dst_ptl;
3063 src_pte = huge_pte_offset(src, addr);
3066 dst_pte = huge_pte_alloc(dst, addr, sz);
3072 /* If the pagetables are shared don't copy or take references */
3073 if (dst_pte == src_pte)
3076 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3077 src_ptl = huge_pte_lockptr(h, src, src_pte);
3078 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3079 entry = huge_ptep_get(src_pte);
3080 if (huge_pte_none(entry)) { /* skip none entry */
3082 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3083 is_hugetlb_entry_hwpoisoned(entry))) {
3084 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3086 if (is_write_migration_entry(swp_entry) && cow) {
3088 * COW mappings require pages in both
3089 * parent and child to be set to read.
3091 make_migration_entry_read(&swp_entry);
3092 entry = swp_entry_to_pte(swp_entry);
3093 set_huge_pte_at(src, addr, src_pte, entry);
3095 set_huge_pte_at(dst, addr, dst_pte, entry);
3098 huge_ptep_set_wrprotect(src, addr, src_pte);
3099 mmu_notifier_invalidate_range(src, mmun_start,
3102 entry = huge_ptep_get(src_pte);
3103 ptepage = pte_page(entry);
3105 page_dup_rmap(ptepage);
3106 set_huge_pte_at(dst, addr, dst_pte, entry);
3107 hugetlb_count_add(pages_per_huge_page(h), dst);
3109 spin_unlock(src_ptl);
3110 spin_unlock(dst_ptl);
3114 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3119 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3120 unsigned long start, unsigned long end,
3121 struct page *ref_page)
3123 int force_flush = 0;
3124 struct mm_struct *mm = vma->vm_mm;
3125 unsigned long address;
3130 struct hstate *h = hstate_vma(vma);
3131 unsigned long sz = huge_page_size(h);
3132 const unsigned long mmun_start = start; /* For mmu_notifiers */
3133 const unsigned long mmun_end = end; /* For mmu_notifiers */
3135 WARN_ON(!is_vm_hugetlb_page(vma));
3136 BUG_ON(start & ~huge_page_mask(h));
3137 BUG_ON(end & ~huge_page_mask(h));
3139 tlb_start_vma(tlb, vma);
3140 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3143 for (; address < end; address += sz) {
3144 ptep = huge_pte_offset(mm, address);
3148 ptl = huge_pte_lock(h, mm, ptep);
3149 if (huge_pmd_unshare(mm, &address, ptep))
3152 pte = huge_ptep_get(ptep);
3153 if (huge_pte_none(pte))
3157 * Migrating hugepage or HWPoisoned hugepage is already
3158 * unmapped and its refcount is dropped, so just clear pte here.
3160 if (unlikely(!pte_present(pte))) {
3161 huge_pte_clear(mm, address, ptep);
3165 page = pte_page(pte);
3167 * If a reference page is supplied, it is because a specific
3168 * page is being unmapped, not a range. Ensure the page we
3169 * are about to unmap is the actual page of interest.
3172 if (page != ref_page)
3176 * Mark the VMA as having unmapped its page so that
3177 * future faults in this VMA will fail rather than
3178 * looking like data was lost
3180 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3183 pte = huge_ptep_get_and_clear(mm, address, ptep);
3184 tlb_remove_tlb_entry(tlb, ptep, address);
3185 if (huge_pte_dirty(pte))
3186 set_page_dirty(page);
3188 hugetlb_count_sub(pages_per_huge_page(h), mm);
3189 page_remove_rmap(page);
3190 force_flush = !__tlb_remove_page(tlb, page);
3196 /* Bail out after unmapping reference page if supplied */
3205 * mmu_gather ran out of room to batch pages, we break out of
3206 * the PTE lock to avoid doing the potential expensive TLB invalidate
3207 * and page-free while holding it.
3212 if (address < end && !ref_page)
3215 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3216 tlb_end_vma(tlb, vma);
3219 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3220 struct vm_area_struct *vma, unsigned long start,
3221 unsigned long end, struct page *ref_page)
3223 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3226 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3227 * test will fail on a vma being torn down, and not grab a page table
3228 * on its way out. We're lucky that the flag has such an appropriate
3229 * name, and can in fact be safely cleared here. We could clear it
3230 * before the __unmap_hugepage_range above, but all that's necessary
3231 * is to clear it before releasing the i_mmap_rwsem. This works
3232 * because in the context this is called, the VMA is about to be
3233 * destroyed and the i_mmap_rwsem is held.
3235 vma->vm_flags &= ~VM_MAYSHARE;
3238 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3239 unsigned long end, struct page *ref_page)
3241 struct mm_struct *mm;
3242 struct mmu_gather tlb;
3246 tlb_gather_mmu(&tlb, mm, start, end);
3247 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3248 tlb_finish_mmu(&tlb, start, end);
3252 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3253 * mappping it owns the reserve page for. The intention is to unmap the page
3254 * from other VMAs and let the children be SIGKILLed if they are faulting the
3257 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3258 struct page *page, unsigned long address)
3260 struct hstate *h = hstate_vma(vma);
3261 struct vm_area_struct *iter_vma;
3262 struct address_space *mapping;
3266 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3267 * from page cache lookup which is in HPAGE_SIZE units.
3269 address = address & huge_page_mask(h);
3270 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3272 mapping = file_inode(vma->vm_file)->i_mapping;
3275 * Take the mapping lock for the duration of the table walk. As
3276 * this mapping should be shared between all the VMAs,
3277 * __unmap_hugepage_range() is called as the lock is already held
3279 i_mmap_lock_write(mapping);
3280 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3281 /* Do not unmap the current VMA */
3282 if (iter_vma == vma)
3286 * Shared VMAs have their own reserves and do not affect
3287 * MAP_PRIVATE accounting but it is possible that a shared
3288 * VMA is using the same page so check and skip such VMAs.
3290 if (iter_vma->vm_flags & VM_MAYSHARE)
3294 * Unmap the page from other VMAs without their own reserves.
3295 * They get marked to be SIGKILLed if they fault in these
3296 * areas. This is because a future no-page fault on this VMA
3297 * could insert a zeroed page instead of the data existing
3298 * from the time of fork. This would look like data corruption
3300 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3301 unmap_hugepage_range(iter_vma, address,
3302 address + huge_page_size(h), page);
3304 i_mmap_unlock_write(mapping);
3308 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3309 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3310 * cannot race with other handlers or page migration.
3311 * Keep the pte_same checks anyway to make transition from the mutex easier.
3313 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3314 unsigned long address, pte_t *ptep, pte_t pte,
3315 struct page *pagecache_page, spinlock_t *ptl)
3317 struct hstate *h = hstate_vma(vma);
3318 struct page *old_page, *new_page;
3319 int ret = 0, outside_reserve = 0;
3320 unsigned long mmun_start; /* For mmu_notifiers */
3321 unsigned long mmun_end; /* For mmu_notifiers */
3323 old_page = pte_page(pte);
3326 /* If no-one else is actually using this page, avoid the copy
3327 * and just make the page writable */
3328 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3329 page_move_anon_rmap(old_page, vma, address);
3330 set_huge_ptep_writable(vma, address, ptep);
3335 * If the process that created a MAP_PRIVATE mapping is about to
3336 * perform a COW due to a shared page count, attempt to satisfy
3337 * the allocation without using the existing reserves. The pagecache
3338 * page is used to determine if the reserve at this address was
3339 * consumed or not. If reserves were used, a partial faulted mapping
3340 * at the time of fork() could consume its reserves on COW instead
3341 * of the full address range.
3343 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3344 old_page != pagecache_page)
3345 outside_reserve = 1;
3347 page_cache_get(old_page);
3350 * Drop page table lock as buddy allocator may be called. It will
3351 * be acquired again before returning to the caller, as expected.
3354 new_page = alloc_huge_page(vma, address, outside_reserve);
3356 if (IS_ERR(new_page)) {
3358 * If a process owning a MAP_PRIVATE mapping fails to COW,
3359 * it is due to references held by a child and an insufficient
3360 * huge page pool. To guarantee the original mappers
3361 * reliability, unmap the page from child processes. The child
3362 * may get SIGKILLed if it later faults.
3364 if (outside_reserve) {
3365 page_cache_release(old_page);
3366 BUG_ON(huge_pte_none(pte));
3367 unmap_ref_private(mm, vma, old_page, address);
3368 BUG_ON(huge_pte_none(pte));
3370 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3372 pte_same(huge_ptep_get(ptep), pte)))
3373 goto retry_avoidcopy;
3375 * race occurs while re-acquiring page table
3376 * lock, and our job is done.
3381 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3382 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3383 goto out_release_old;
3387 * When the original hugepage is shared one, it does not have
3388 * anon_vma prepared.
3390 if (unlikely(anon_vma_prepare(vma))) {
3392 goto out_release_all;
3395 copy_user_huge_page(new_page, old_page, address, vma,
3396 pages_per_huge_page(h));
3397 __SetPageUptodate(new_page);
3398 set_page_huge_active(new_page);
3400 mmun_start = address & huge_page_mask(h);
3401 mmun_end = mmun_start + huge_page_size(h);
3402 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3405 * Retake the page table lock to check for racing updates
3406 * before the page tables are altered
3409 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3410 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3411 ClearPagePrivate(new_page);
3414 huge_ptep_clear_flush(vma, address, ptep);
3415 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3416 set_huge_pte_at(mm, address, ptep,
3417 make_huge_pte(vma, new_page, 1));
3418 page_remove_rmap(old_page);
3419 hugepage_add_new_anon_rmap(new_page, vma, address);
3420 /* Make the old page be freed below */
3421 new_page = old_page;
3424 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3426 page_cache_release(new_page);
3428 page_cache_release(old_page);
3430 spin_lock(ptl); /* Caller expects lock to be held */
3434 /* Return the pagecache page at a given address within a VMA */
3435 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3436 struct vm_area_struct *vma, unsigned long address)
3438 struct address_space *mapping;
3441 mapping = vma->vm_file->f_mapping;
3442 idx = vma_hugecache_offset(h, vma, address);
3444 return find_lock_page(mapping, idx);
3448 * Return whether there is a pagecache page to back given address within VMA.
3449 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3451 static bool hugetlbfs_pagecache_present(struct hstate *h,
3452 struct vm_area_struct *vma, unsigned long address)
3454 struct address_space *mapping;
3458 mapping = vma->vm_file->f_mapping;
3459 idx = vma_hugecache_offset(h, vma, address);
3461 page = find_get_page(mapping, idx);
3464 return page != NULL;
3467 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3470 struct inode *inode = mapping->host;
3471 struct hstate *h = hstate_inode(inode);
3472 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3476 ClearPagePrivate(page);
3478 spin_lock(&inode->i_lock);
3479 inode->i_blocks += blocks_per_huge_page(h);
3480 spin_unlock(&inode->i_lock);
3484 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3485 struct address_space *mapping, pgoff_t idx,
3486 unsigned long address, pte_t *ptep, unsigned int flags)
3488 struct hstate *h = hstate_vma(vma);
3489 int ret = VM_FAULT_SIGBUS;
3497 * Currently, we are forced to kill the process in the event the
3498 * original mapper has unmapped pages from the child due to a failed
3499 * COW. Warn that such a situation has occurred as it may not be obvious
3501 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3502 pr_warning("PID %d killed due to inadequate hugepage pool\n",
3508 * Use page lock to guard against racing truncation
3509 * before we get page_table_lock.
3512 page = find_lock_page(mapping, idx);
3514 size = i_size_read(mapping->host) >> huge_page_shift(h);
3517 page = alloc_huge_page(vma, address, 0);
3519 ret = PTR_ERR(page);
3523 ret = VM_FAULT_SIGBUS;
3526 clear_huge_page(page, address, pages_per_huge_page(h));
3527 __SetPageUptodate(page);
3528 set_page_huge_active(page);
3530 if (vma->vm_flags & VM_MAYSHARE) {
3531 int err = huge_add_to_page_cache(page, mapping, idx);
3540 if (unlikely(anon_vma_prepare(vma))) {
3542 goto backout_unlocked;
3548 * If memory error occurs between mmap() and fault, some process
3549 * don't have hwpoisoned swap entry for errored virtual address.
3550 * So we need to block hugepage fault by PG_hwpoison bit check.
3552 if (unlikely(PageHWPoison(page))) {
3553 ret = VM_FAULT_HWPOISON |
3554 VM_FAULT_SET_HINDEX(hstate_index(h));
3555 goto backout_unlocked;
3560 * If we are going to COW a private mapping later, we examine the
3561 * pending reservations for this page now. This will ensure that
3562 * any allocations necessary to record that reservation occur outside
3565 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3566 if (vma_needs_reservation(h, vma, address) < 0) {
3568 goto backout_unlocked;
3570 /* Just decrements count, does not deallocate */
3571 vma_end_reservation(h, vma, address);
3574 ptl = huge_pte_lockptr(h, mm, ptep);
3576 size = i_size_read(mapping->host) >> huge_page_shift(h);
3581 if (!huge_pte_none(huge_ptep_get(ptep)))
3585 ClearPagePrivate(page);
3586 hugepage_add_new_anon_rmap(page, vma, address);
3588 page_dup_rmap(page);
3589 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3590 && (vma->vm_flags & VM_SHARED)));
3591 set_huge_pte_at(mm, address, ptep, new_pte);
3593 hugetlb_count_add(pages_per_huge_page(h), mm);
3594 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3595 /* Optimization, do the COW without a second fault */
3596 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
3613 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3614 struct vm_area_struct *vma,
3615 struct address_space *mapping,
3616 pgoff_t idx, unsigned long address)
3618 unsigned long key[2];
3621 if (vma->vm_flags & VM_SHARED) {
3622 key[0] = (unsigned long) mapping;
3625 key[0] = (unsigned long) mm;
3626 key[1] = address >> huge_page_shift(h);
3629 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3631 return hash & (num_fault_mutexes - 1);
3635 * For uniprocesor systems we always use a single mutex, so just
3636 * return 0 and avoid the hashing overhead.
3638 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3639 struct vm_area_struct *vma,
3640 struct address_space *mapping,
3641 pgoff_t idx, unsigned long address)
3647 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3648 unsigned long address, unsigned int flags)
3655 struct page *page = NULL;
3656 struct page *pagecache_page = NULL;
3657 struct hstate *h = hstate_vma(vma);
3658 struct address_space *mapping;
3659 int need_wait_lock = 0;
3661 address &= huge_page_mask(h);
3663 ptep = huge_pte_offset(mm, address);
3665 entry = huge_ptep_get(ptep);
3666 if (unlikely(is_hugetlb_entry_migration(entry))) {
3667 migration_entry_wait_huge(vma, mm, ptep);
3669 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3670 return VM_FAULT_HWPOISON_LARGE |
3671 VM_FAULT_SET_HINDEX(hstate_index(h));
3673 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3675 return VM_FAULT_OOM;
3678 mapping = vma->vm_file->f_mapping;
3679 idx = vma_hugecache_offset(h, vma, address);
3682 * Serialize hugepage allocation and instantiation, so that we don't
3683 * get spurious allocation failures if two CPUs race to instantiate
3684 * the same page in the page cache.
3686 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3687 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3689 entry = huge_ptep_get(ptep);
3690 if (huge_pte_none(entry)) {
3691 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3698 * entry could be a migration/hwpoison entry at this point, so this
3699 * check prevents the kernel from going below assuming that we have
3700 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3701 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3704 if (!pte_present(entry))
3708 * If we are going to COW the mapping later, we examine the pending
3709 * reservations for this page now. This will ensure that any
3710 * allocations necessary to record that reservation occur outside the
3711 * spinlock. For private mappings, we also lookup the pagecache
3712 * page now as it is used to determine if a reservation has been
3715 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3716 if (vma_needs_reservation(h, vma, address) < 0) {
3720 /* Just decrements count, does not deallocate */
3721 vma_end_reservation(h, vma, address);
3723 if (!(vma->vm_flags & VM_MAYSHARE))
3724 pagecache_page = hugetlbfs_pagecache_page(h,
3728 ptl = huge_pte_lock(h, mm, ptep);
3730 /* Check for a racing update before calling hugetlb_cow */
3731 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3735 * hugetlb_cow() requires page locks of pte_page(entry) and
3736 * pagecache_page, so here we need take the former one
3737 * when page != pagecache_page or !pagecache_page.
3739 page = pte_page(entry);
3740 if (page != pagecache_page)
3741 if (!trylock_page(page)) {
3748 if (flags & FAULT_FLAG_WRITE) {
3749 if (!huge_pte_write(entry)) {
3750 ret = hugetlb_cow(mm, vma, address, ptep, entry,
3751 pagecache_page, ptl);
3754 entry = huge_pte_mkdirty(entry);
3756 entry = pte_mkyoung(entry);
3757 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3758 flags & FAULT_FLAG_WRITE))
3759 update_mmu_cache(vma, address, ptep);
3761 if (page != pagecache_page)
3767 if (pagecache_page) {
3768 unlock_page(pagecache_page);
3769 put_page(pagecache_page);
3772 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3774 * Generally it's safe to hold refcount during waiting page lock. But
3775 * here we just wait to defer the next page fault to avoid busy loop and
3776 * the page is not used after unlocked before returning from the current
3777 * page fault. So we are safe from accessing freed page, even if we wait
3778 * here without taking refcount.
3781 wait_on_page_locked(page);
3785 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3786 struct page **pages, struct vm_area_struct **vmas,
3787 unsigned long *position, unsigned long *nr_pages,
3788 long i, unsigned int flags)
3790 unsigned long pfn_offset;
3791 unsigned long vaddr = *position;
3792 unsigned long remainder = *nr_pages;
3793 struct hstate *h = hstate_vma(vma);
3795 while (vaddr < vma->vm_end && remainder) {
3797 spinlock_t *ptl = NULL;
3802 * If we have a pending SIGKILL, don't keep faulting pages and
3803 * potentially allocating memory.
3805 if (unlikely(fatal_signal_pending(current))) {
3811 * Some archs (sparc64, sh*) have multiple pte_ts to
3812 * each hugepage. We have to make sure we get the
3813 * first, for the page indexing below to work.
3815 * Note that page table lock is not held when pte is null.
3817 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3819 ptl = huge_pte_lock(h, mm, pte);
3820 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3823 * When coredumping, it suits get_dump_page if we just return
3824 * an error where there's an empty slot with no huge pagecache
3825 * to back it. This way, we avoid allocating a hugepage, and
3826 * the sparse dumpfile avoids allocating disk blocks, but its
3827 * huge holes still show up with zeroes where they need to be.
3829 if (absent && (flags & FOLL_DUMP) &&
3830 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3838 * We need call hugetlb_fault for both hugepages under migration
3839 * (in which case hugetlb_fault waits for the migration,) and
3840 * hwpoisoned hugepages (in which case we need to prevent the
3841 * caller from accessing to them.) In order to do this, we use
3842 * here is_swap_pte instead of is_hugetlb_entry_migration and
3843 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3844 * both cases, and because we can't follow correct pages
3845 * directly from any kind of swap entries.
3847 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3848 ((flags & FOLL_WRITE) &&
3849 !huge_pte_write(huge_ptep_get(pte)))) {
3854 ret = hugetlb_fault(mm, vma, vaddr,
3855 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3856 if (!(ret & VM_FAULT_ERROR))
3863 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3864 page = pte_page(huge_ptep_get(pte));
3867 pages[i] = mem_map_offset(page, pfn_offset);
3868 get_page_foll(pages[i]);
3878 if (vaddr < vma->vm_end && remainder &&
3879 pfn_offset < pages_per_huge_page(h)) {
3881 * We use pfn_offset to avoid touching the pageframes
3882 * of this compound page.
3888 *nr_pages = remainder;
3891 return i ? i : -EFAULT;
3894 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3895 unsigned long address, unsigned long end, pgprot_t newprot)
3897 struct mm_struct *mm = vma->vm_mm;
3898 unsigned long start = address;
3901 struct hstate *h = hstate_vma(vma);
3902 unsigned long pages = 0;
3904 BUG_ON(address >= end);
3905 flush_cache_range(vma, address, end);
3907 mmu_notifier_invalidate_range_start(mm, start, end);
3908 i_mmap_lock_write(vma->vm_file->f_mapping);
3909 for (; address < end; address += huge_page_size(h)) {
3911 ptep = huge_pte_offset(mm, address);
3914 ptl = huge_pte_lock(h, mm, ptep);
3915 if (huge_pmd_unshare(mm, &address, ptep)) {
3920 pte = huge_ptep_get(ptep);
3921 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3925 if (unlikely(is_hugetlb_entry_migration(pte))) {
3926 swp_entry_t entry = pte_to_swp_entry(pte);
3928 if (is_write_migration_entry(entry)) {
3931 make_migration_entry_read(&entry);
3932 newpte = swp_entry_to_pte(entry);
3933 set_huge_pte_at(mm, address, ptep, newpte);
3939 if (!huge_pte_none(pte)) {
3940 pte = huge_ptep_get_and_clear(mm, address, ptep);
3941 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3942 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3943 set_huge_pte_at(mm, address, ptep, pte);
3949 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
3950 * may have cleared our pud entry and done put_page on the page table:
3951 * once we release i_mmap_rwsem, another task can do the final put_page
3952 * and that page table be reused and filled with junk.
3954 flush_tlb_range(vma, start, end);
3955 mmu_notifier_invalidate_range(mm, start, end);
3956 i_mmap_unlock_write(vma->vm_file->f_mapping);
3957 mmu_notifier_invalidate_range_end(mm, start, end);
3959 return pages << h->order;
3962 int hugetlb_reserve_pages(struct inode *inode,
3964 struct vm_area_struct *vma,
3965 vm_flags_t vm_flags)
3968 struct hstate *h = hstate_inode(inode);
3969 struct hugepage_subpool *spool = subpool_inode(inode);
3970 struct resv_map *resv_map;
3974 * Only apply hugepage reservation if asked. At fault time, an
3975 * attempt will be made for VM_NORESERVE to allocate a page
3976 * without using reserves
3978 if (vm_flags & VM_NORESERVE)
3982 * Shared mappings base their reservation on the number of pages that
3983 * are already allocated on behalf of the file. Private mappings need
3984 * to reserve the full area even if read-only as mprotect() may be
3985 * called to make the mapping read-write. Assume !vma is a shm mapping
3987 if (!vma || vma->vm_flags & VM_MAYSHARE) {
3988 resv_map = inode_resv_map(inode);
3990 chg = region_chg(resv_map, from, to);
3993 resv_map = resv_map_alloc();
3999 set_vma_resv_map(vma, resv_map);
4000 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4009 * There must be enough pages in the subpool for the mapping. If
4010 * the subpool has a minimum size, there may be some global
4011 * reservations already in place (gbl_reserve).
4013 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4014 if (gbl_reserve < 0) {
4020 * Check enough hugepages are available for the reservation.
4021 * Hand the pages back to the subpool if there are not
4023 ret = hugetlb_acct_memory(h, gbl_reserve);
4025 /* put back original number of pages, chg */
4026 (void)hugepage_subpool_put_pages(spool, chg);
4031 * Account for the reservations made. Shared mappings record regions
4032 * that have reservations as they are shared by multiple VMAs.
4033 * When the last VMA disappears, the region map says how much
4034 * the reservation was and the page cache tells how much of
4035 * the reservation was consumed. Private mappings are per-VMA and
4036 * only the consumed reservations are tracked. When the VMA
4037 * disappears, the original reservation is the VMA size and the
4038 * consumed reservations are stored in the map. Hence, nothing
4039 * else has to be done for private mappings here
4041 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4042 long add = region_add(resv_map, from, to);
4044 if (unlikely(chg > add)) {
4046 * pages in this range were added to the reserve
4047 * map between region_chg and region_add. This
4048 * indicates a race with alloc_huge_page. Adjust
4049 * the subpool and reserve counts modified above
4050 * based on the difference.
4054 rsv_adjust = hugepage_subpool_put_pages(spool,
4056 hugetlb_acct_memory(h, -rsv_adjust);
4061 if (!vma || vma->vm_flags & VM_MAYSHARE)
4062 region_abort(resv_map, from, to);
4063 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4064 kref_put(&resv_map->refs, resv_map_release);
4068 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4071 struct hstate *h = hstate_inode(inode);
4072 struct resv_map *resv_map = inode_resv_map(inode);
4074 struct hugepage_subpool *spool = subpool_inode(inode);
4078 chg = region_del(resv_map, start, end);
4080 * region_del() can fail in the rare case where a region
4081 * must be split and another region descriptor can not be
4082 * allocated. If end == LONG_MAX, it will not fail.
4088 spin_lock(&inode->i_lock);
4089 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4090 spin_unlock(&inode->i_lock);
4093 * If the subpool has a minimum size, the number of global
4094 * reservations to be released may be adjusted.
4096 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4097 hugetlb_acct_memory(h, -gbl_reserve);
4102 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4103 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4104 struct vm_area_struct *vma,
4105 unsigned long addr, pgoff_t idx)
4107 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4109 unsigned long sbase = saddr & PUD_MASK;
4110 unsigned long s_end = sbase + PUD_SIZE;
4112 /* Allow segments to share if only one is marked locked */
4113 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4114 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4117 * match the virtual addresses, permission and the alignment of the
4120 if (pmd_index(addr) != pmd_index(saddr) ||
4121 vm_flags != svm_flags ||
4122 sbase < svma->vm_start || svma->vm_end < s_end)
4128 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4130 unsigned long base = addr & PUD_MASK;
4131 unsigned long end = base + PUD_SIZE;
4134 * check on proper vm_flags and page table alignment
4136 if (vma->vm_flags & VM_MAYSHARE &&
4137 vma->vm_start <= base && end <= vma->vm_end)
4143 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4144 * and returns the corresponding pte. While this is not necessary for the
4145 * !shared pmd case because we can allocate the pmd later as well, it makes the
4146 * code much cleaner. pmd allocation is essential for the shared case because
4147 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4148 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4149 * bad pmd for sharing.
4151 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4153 struct vm_area_struct *vma = find_vma(mm, addr);
4154 struct address_space *mapping = vma->vm_file->f_mapping;
4155 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4157 struct vm_area_struct *svma;
4158 unsigned long saddr;
4163 if (!vma_shareable(vma, addr))
4164 return (pte_t *)pmd_alloc(mm, pud, addr);
4166 i_mmap_lock_write(mapping);
4167 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4171 saddr = page_table_shareable(svma, vma, addr, idx);
4173 spte = huge_pte_offset(svma->vm_mm, saddr);
4176 get_page(virt_to_page(spte));
4185 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
4187 if (pud_none(*pud)) {
4188 pud_populate(mm, pud,
4189 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4191 put_page(virt_to_page(spte));
4196 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4197 i_mmap_unlock_write(mapping);
4202 * unmap huge page backed by shared pte.
4204 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4205 * indicated by page_count > 1, unmap is achieved by clearing pud and
4206 * decrementing the ref count. If count == 1, the pte page is not shared.
4208 * called with page table lock held.
4210 * returns: 1 successfully unmapped a shared pte page
4211 * 0 the underlying pte page is not shared, or it is the last user
4213 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4215 pgd_t *pgd = pgd_offset(mm, *addr);
4216 pud_t *pud = pud_offset(pgd, *addr);
4218 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4219 if (page_count(virt_to_page(ptep)) == 1)
4223 put_page(virt_to_page(ptep));
4225 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4228 #define want_pmd_share() (1)
4229 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4230 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4235 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4239 #define want_pmd_share() (0)
4240 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4242 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4243 pte_t *huge_pte_alloc(struct mm_struct *mm,
4244 unsigned long addr, unsigned long sz)
4250 pgd = pgd_offset(mm, addr);
4251 pud = pud_alloc(mm, pgd, addr);
4253 if (sz == PUD_SIZE) {
4256 BUG_ON(sz != PMD_SIZE);
4257 if (want_pmd_share() && pud_none(*pud))
4258 pte = huge_pmd_share(mm, addr, pud);
4260 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4263 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
4268 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
4274 pgd = pgd_offset(mm, addr);
4275 if (pgd_present(*pgd)) {
4276 pud = pud_offset(pgd, addr);
4277 if (pud_present(*pud)) {
4279 return (pte_t *)pud;
4280 pmd = pmd_offset(pud, addr);
4283 return (pte_t *) pmd;
4286 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4289 * These functions are overwritable if your architecture needs its own
4292 struct page * __weak
4293 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4296 return ERR_PTR(-EINVAL);
4299 struct page * __weak
4300 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4301 pmd_t *pmd, int flags)
4303 struct page *page = NULL;
4306 ptl = pmd_lockptr(mm, pmd);
4309 * make sure that the address range covered by this pmd is not
4310 * unmapped from other threads.
4312 if (!pmd_huge(*pmd))
4314 if (pmd_present(*pmd)) {
4315 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4316 if (flags & FOLL_GET)
4319 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
4321 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4325 * hwpoisoned entry is treated as no_page_table in
4326 * follow_page_mask().
4334 struct page * __weak
4335 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4336 pud_t *pud, int flags)
4338 if (flags & FOLL_GET)
4341 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4344 #ifdef CONFIG_MEMORY_FAILURE
4347 * This function is called from memory failure code.
4348 * Assume the caller holds page lock of the head page.
4350 int dequeue_hwpoisoned_huge_page(struct page *hpage)
4352 struct hstate *h = page_hstate(hpage);
4353 int nid = page_to_nid(hpage);
4356 spin_lock(&hugetlb_lock);
4358 * Just checking !page_huge_active is not enough, because that could be
4359 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4361 if (!page_huge_active(hpage) && !page_count(hpage)) {
4363 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4364 * but dangling hpage->lru can trigger list-debug warnings
4365 * (this happens when we call unpoison_memory() on it),
4366 * so let it point to itself with list_del_init().
4368 list_del_init(&hpage->lru);
4369 set_page_refcounted(hpage);
4370 h->free_huge_pages--;
4371 h->free_huge_pages_node[nid]--;
4374 spin_unlock(&hugetlb_lock);
4379 bool isolate_huge_page(struct page *page, struct list_head *list)
4383 VM_BUG_ON_PAGE(!PageHead(page), page);
4384 spin_lock(&hugetlb_lock);
4385 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4389 clear_page_huge_active(page);
4390 list_move_tail(&page->lru, list);
4392 spin_unlock(&hugetlb_lock);
4396 void putback_active_hugepage(struct page *page)
4398 VM_BUG_ON_PAGE(!PageHead(page), page);
4399 spin_lock(&hugetlb_lock);
4400 set_page_huge_active(page);
4401 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4402 spin_unlock(&hugetlb_lock);