2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/module.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/bootmem.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/rmap.h>
23 #include <linux/swap.h>
24 #include <linux/swapops.h>
25 #include <linux/page-isolation.h>
26 #include <linux/jhash.h>
29 #include <asm/pgtable.h>
33 #include <linux/hugetlb.h>
34 #include <linux/hugetlb_cgroup.h>
35 #include <linux/node.h>
38 int hugepages_treat_as_movable;
40 int hugetlb_max_hstate __read_mostly;
41 unsigned int default_hstate_idx;
42 struct hstate hstates[HUGE_MAX_HSTATE];
44 * Minimum page order among possible hugepage sizes, set to a proper value
47 static unsigned int minimum_order __read_mostly = UINT_MAX;
49 __initdata LIST_HEAD(huge_boot_pages);
51 /* for command line parsing */
52 static struct hstate * __initdata parsed_hstate;
53 static unsigned long __initdata default_hstate_max_huge_pages;
54 static unsigned long __initdata default_hstate_size;
57 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
58 * free_huge_pages, and surplus_huge_pages.
60 DEFINE_SPINLOCK(hugetlb_lock);
63 * Serializes faults on the same logical page. This is used to
64 * prevent spurious OOMs when the hugepage pool is fully utilized.
66 static int num_fault_mutexes;
67 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
69 /* Forward declaration */
70 static int hugetlb_acct_memory(struct hstate *h, long delta);
72 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
74 bool free = (spool->count == 0) && (spool->used_hpages == 0);
76 spin_unlock(&spool->lock);
78 /* If no pages are used, and no other handles to the subpool
79 * remain, give up any reservations mased on minimum size and
82 if (spool->min_hpages != -1)
83 hugetlb_acct_memory(spool->hstate,
89 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
92 struct hugepage_subpool *spool;
94 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
98 spin_lock_init(&spool->lock);
100 spool->max_hpages = max_hpages;
102 spool->min_hpages = min_hpages;
104 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
108 spool->rsv_hpages = min_hpages;
113 void hugepage_put_subpool(struct hugepage_subpool *spool)
115 spin_lock(&spool->lock);
116 BUG_ON(!spool->count);
118 unlock_or_release_subpool(spool);
122 * Subpool accounting for allocating and reserving pages.
123 * Return -ENOMEM if there are not enough resources to satisfy the
124 * the request. Otherwise, return the number of pages by which the
125 * global pools must be adjusted (upward). The returned value may
126 * only be different than the passed value (delta) in the case where
127 * a subpool minimum size must be manitained.
129 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
137 spin_lock(&spool->lock);
139 if (spool->max_hpages != -1) { /* maximum size accounting */
140 if ((spool->used_hpages + delta) <= spool->max_hpages)
141 spool->used_hpages += delta;
148 if (spool->min_hpages != -1) { /* minimum size accounting */
149 if (delta > spool->rsv_hpages) {
151 * Asking for more reserves than those already taken on
152 * behalf of subpool. Return difference.
154 ret = delta - spool->rsv_hpages;
155 spool->rsv_hpages = 0;
157 ret = 0; /* reserves already accounted for */
158 spool->rsv_hpages -= delta;
163 spin_unlock(&spool->lock);
168 * Subpool accounting for freeing and unreserving pages.
169 * Return the number of global page reservations that must be dropped.
170 * The return value may only be different than the passed value (delta)
171 * in the case where a subpool minimum size must be maintained.
173 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
181 spin_lock(&spool->lock);
183 if (spool->max_hpages != -1) /* maximum size accounting */
184 spool->used_hpages -= delta;
186 if (spool->min_hpages != -1) { /* minimum size accounting */
187 if (spool->rsv_hpages + delta <= spool->min_hpages)
190 ret = spool->rsv_hpages + delta - spool->min_hpages;
192 spool->rsv_hpages += delta;
193 if (spool->rsv_hpages > spool->min_hpages)
194 spool->rsv_hpages = spool->min_hpages;
198 * If hugetlbfs_put_super couldn't free spool due to an outstanding
199 * quota reference, free it now.
201 unlock_or_release_subpool(spool);
206 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
208 return HUGETLBFS_SB(inode->i_sb)->spool;
211 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
213 return subpool_inode(file_inode(vma->vm_file));
217 * Region tracking -- allows tracking of reservations and instantiated pages
218 * across the pages in a mapping.
220 * The region data structures are embedded into a resv_map and protected
221 * by a resv_map's lock. The set of regions within the resv_map represent
222 * reservations for huge pages, or huge pages that have already been
223 * instantiated within the map. The from and to elements are huge page
224 * indicies into the associated mapping. from indicates the starting index
225 * of the region. to represents the first index past the end of the region.
227 * For example, a file region structure with from == 0 and to == 4 represents
228 * four huge pages in a mapping. It is important to note that the to element
229 * represents the first element past the end of the region. This is used in
230 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
232 * Interval notation of the form [from, to) will be used to indicate that
233 * the endpoint from is inclusive and to is exclusive.
236 struct list_head link;
242 * Add the huge page range represented by [f, t) to the reserve
243 * map. In the normal case, existing regions will be expanded
244 * to accommodate the specified range. Sufficient regions should
245 * exist for expansion due to the previous call to region_chg
246 * with the same range. However, it is possible that region_del
247 * could have been called after region_chg and modifed the map
248 * in such a way that no region exists to be expanded. In this
249 * case, pull a region descriptor from the cache associated with
250 * the map and use that for the new range.
252 * Return the number of new huge pages added to the map. This
253 * number is greater than or equal to zero.
255 static long region_add(struct resv_map *resv, long f, long t)
257 struct list_head *head = &resv->regions;
258 struct file_region *rg, *nrg, *trg;
261 spin_lock(&resv->lock);
262 /* Locate the region we are either in or before. */
263 list_for_each_entry(rg, head, link)
268 * If no region exists which can be expanded to include the
269 * specified range, the list must have been modified by an
270 * interleving call to region_del(). Pull a region descriptor
271 * from the cache and use it for this range.
273 if (&rg->link == head || t < rg->from) {
274 VM_BUG_ON(resv->region_cache_count <= 0);
276 resv->region_cache_count--;
277 nrg = list_first_entry(&resv->region_cache, struct file_region,
279 list_del(&nrg->link);
283 list_add(&nrg->link, rg->link.prev);
289 /* Round our left edge to the current segment if it encloses us. */
293 /* Check for and consume any regions we now overlap with. */
295 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
296 if (&rg->link == head)
301 /* If this area reaches higher then extend our area to
302 * include it completely. If this is not the first area
303 * which we intend to reuse, free it. */
307 /* Decrement return value by the deleted range.
308 * Another range will span this area so that by
309 * end of routine add will be >= zero
311 add -= (rg->to - rg->from);
317 add += (nrg->from - f); /* Added to beginning of region */
319 add += t - nrg->to; /* Added to end of region */
323 resv->adds_in_progress--;
324 spin_unlock(&resv->lock);
330 * Examine the existing reserve map and determine how many
331 * huge pages in the specified range [f, t) are NOT currently
332 * represented. This routine is called before a subsequent
333 * call to region_add that will actually modify the reserve
334 * map to add the specified range [f, t). region_chg does
335 * not change the number of huge pages represented by the
336 * map. However, if the existing regions in the map can not
337 * be expanded to represent the new range, a new file_region
338 * structure is added to the map as a placeholder. This is
339 * so that the subsequent region_add call will have all the
340 * regions it needs and will not fail.
342 * Upon entry, region_chg will also examine the cache of region descriptors
343 * associated with the map. If there are not enough descriptors cached, one
344 * will be allocated for the in progress add operation.
346 * Returns the number of huge pages that need to be added to the existing
347 * reservation map for the range [f, t). This number is greater or equal to
348 * zero. -ENOMEM is returned if a new file_region structure or cache entry
349 * is needed and can not be allocated.
351 static long region_chg(struct resv_map *resv, long f, long t)
353 struct list_head *head = &resv->regions;
354 struct file_region *rg, *nrg = NULL;
358 spin_lock(&resv->lock);
360 resv->adds_in_progress++;
363 * Check for sufficient descriptors in the cache to accommodate
364 * the number of in progress add operations.
366 if (resv->adds_in_progress > resv->region_cache_count) {
367 struct file_region *trg;
369 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
370 /* Must drop lock to allocate a new descriptor. */
371 resv->adds_in_progress--;
372 spin_unlock(&resv->lock);
374 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
378 spin_lock(&resv->lock);
379 list_add(&trg->link, &resv->region_cache);
380 resv->region_cache_count++;
384 /* Locate the region we are before or in. */
385 list_for_each_entry(rg, head, link)
389 /* If we are below the current region then a new region is required.
390 * Subtle, allocate a new region at the position but make it zero
391 * size such that we can guarantee to record the reservation. */
392 if (&rg->link == head || t < rg->from) {
394 resv->adds_in_progress--;
395 spin_unlock(&resv->lock);
396 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
402 INIT_LIST_HEAD(&nrg->link);
406 list_add(&nrg->link, rg->link.prev);
411 /* Round our left edge to the current segment if it encloses us. */
416 /* Check for and consume any regions we now overlap with. */
417 list_for_each_entry(rg, rg->link.prev, link) {
418 if (&rg->link == head)
423 /* We overlap with this area, if it extends further than
424 * us then we must extend ourselves. Account for its
425 * existing reservation. */
430 chg -= rg->to - rg->from;
434 spin_unlock(&resv->lock);
435 /* We already know we raced and no longer need the new region */
439 spin_unlock(&resv->lock);
444 * Abort the in progress add operation. The adds_in_progress field
445 * of the resv_map keeps track of the operations in progress between
446 * calls to region_chg and region_add. Operations are sometimes
447 * aborted after the call to region_chg. In such cases, region_abort
448 * is called to decrement the adds_in_progress counter.
450 * NOTE: The range arguments [f, t) are not needed or used in this
451 * routine. They are kept to make reading the calling code easier as
452 * arguments will match the associated region_chg call.
454 static void region_abort(struct resv_map *resv, long f, long t)
456 spin_lock(&resv->lock);
457 VM_BUG_ON(!resv->region_cache_count);
458 resv->adds_in_progress--;
459 spin_unlock(&resv->lock);
463 * Delete the specified range [f, t) from the reserve map. If the
464 * t parameter is LONG_MAX, this indicates that ALL regions after f
465 * should be deleted. Locate the regions which intersect [f, t)
466 * and either trim, delete or split the existing regions.
468 * Returns the number of huge pages deleted from the reserve map.
469 * In the normal case, the return value is zero or more. In the
470 * case where a region must be split, a new region descriptor must
471 * be allocated. If the allocation fails, -ENOMEM will be returned.
472 * NOTE: If the parameter t == LONG_MAX, then we will never split
473 * a region and possibly return -ENOMEM. Callers specifying
474 * t == LONG_MAX do not need to check for -ENOMEM error.
476 static long region_del(struct resv_map *resv, long f, long t)
478 struct list_head *head = &resv->regions;
479 struct file_region *rg, *trg;
480 struct file_region *nrg = NULL;
484 spin_lock(&resv->lock);
485 list_for_each_entry_safe(rg, trg, head, link) {
491 if (f > rg->from && t < rg->to) { /* Must split region */
493 * Check for an entry in the cache before dropping
494 * lock and attempting allocation.
497 resv->region_cache_count > resv->adds_in_progress) {
498 nrg = list_first_entry(&resv->region_cache,
501 list_del(&nrg->link);
502 resv->region_cache_count--;
506 spin_unlock(&resv->lock);
507 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
515 /* New entry for end of split region */
518 INIT_LIST_HEAD(&nrg->link);
520 /* Original entry is trimmed */
523 list_add(&nrg->link, &rg->link);
528 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
529 del += rg->to - rg->from;
535 if (f <= rg->from) { /* Trim beginning of region */
538 } else { /* Trim end of region */
544 spin_unlock(&resv->lock);
550 * A rare out of memory error was encountered which prevented removal of
551 * the reserve map region for a page. The huge page itself was free'ed
552 * and removed from the page cache. This routine will adjust the subpool
553 * usage count, and the global reserve count if needed. By incrementing
554 * these counts, the reserve map entry which could not be deleted will
555 * appear as a "reserved" entry instead of simply dangling with incorrect
558 void hugetlb_fix_reserve_counts(struct inode *inode, bool restore_reserve)
560 struct hugepage_subpool *spool = subpool_inode(inode);
563 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
564 if (restore_reserve && rsv_adjust) {
565 struct hstate *h = hstate_inode(inode);
567 hugetlb_acct_memory(h, 1);
572 * Count and return the number of huge pages in the reserve map
573 * that intersect with the range [f, t).
575 static long region_count(struct resv_map *resv, long f, long t)
577 struct list_head *head = &resv->regions;
578 struct file_region *rg;
581 spin_lock(&resv->lock);
582 /* Locate each segment we overlap with, and count that overlap. */
583 list_for_each_entry(rg, head, link) {
592 seg_from = max(rg->from, f);
593 seg_to = min(rg->to, t);
595 chg += seg_to - seg_from;
597 spin_unlock(&resv->lock);
603 * Convert the address within this vma to the page offset within
604 * the mapping, in pagecache page units; huge pages here.
606 static pgoff_t vma_hugecache_offset(struct hstate *h,
607 struct vm_area_struct *vma, unsigned long address)
609 return ((address - vma->vm_start) >> huge_page_shift(h)) +
610 (vma->vm_pgoff >> huge_page_order(h));
613 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
614 unsigned long address)
616 return vma_hugecache_offset(hstate_vma(vma), vma, address);
620 * Return the size of the pages allocated when backing a VMA. In the majority
621 * cases this will be same size as used by the page table entries.
623 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
625 struct hstate *hstate;
627 if (!is_vm_hugetlb_page(vma))
630 hstate = hstate_vma(vma);
632 return 1UL << huge_page_shift(hstate);
634 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
637 * Return the page size being used by the MMU to back a VMA. In the majority
638 * of cases, the page size used by the kernel matches the MMU size. On
639 * architectures where it differs, an architecture-specific version of this
640 * function is required.
642 #ifndef vma_mmu_pagesize
643 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
645 return vma_kernel_pagesize(vma);
650 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
651 * bits of the reservation map pointer, which are always clear due to
654 #define HPAGE_RESV_OWNER (1UL << 0)
655 #define HPAGE_RESV_UNMAPPED (1UL << 1)
656 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
659 * These helpers are used to track how many pages are reserved for
660 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
661 * is guaranteed to have their future faults succeed.
663 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
664 * the reserve counters are updated with the hugetlb_lock held. It is safe
665 * to reset the VMA at fork() time as it is not in use yet and there is no
666 * chance of the global counters getting corrupted as a result of the values.
668 * The private mapping reservation is represented in a subtly different
669 * manner to a shared mapping. A shared mapping has a region map associated
670 * with the underlying file, this region map represents the backing file
671 * pages which have ever had a reservation assigned which this persists even
672 * after the page is instantiated. A private mapping has a region map
673 * associated with the original mmap which is attached to all VMAs which
674 * reference it, this region map represents those offsets which have consumed
675 * reservation ie. where pages have been instantiated.
677 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
679 return (unsigned long)vma->vm_private_data;
682 static void set_vma_private_data(struct vm_area_struct *vma,
685 vma->vm_private_data = (void *)value;
688 struct resv_map *resv_map_alloc(void)
690 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
691 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
693 if (!resv_map || !rg) {
699 kref_init(&resv_map->refs);
700 spin_lock_init(&resv_map->lock);
701 INIT_LIST_HEAD(&resv_map->regions);
703 resv_map->adds_in_progress = 0;
705 INIT_LIST_HEAD(&resv_map->region_cache);
706 list_add(&rg->link, &resv_map->region_cache);
707 resv_map->region_cache_count = 1;
712 void resv_map_release(struct kref *ref)
714 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
715 struct list_head *head = &resv_map->region_cache;
716 struct file_region *rg, *trg;
718 /* Clear out any active regions before we release the map. */
719 region_del(resv_map, 0, LONG_MAX);
721 /* ... and any entries left in the cache */
722 list_for_each_entry_safe(rg, trg, head, link) {
727 VM_BUG_ON(resv_map->adds_in_progress);
732 static inline struct resv_map *inode_resv_map(struct inode *inode)
734 return inode->i_mapping->private_data;
737 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
739 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
740 if (vma->vm_flags & VM_MAYSHARE) {
741 struct address_space *mapping = vma->vm_file->f_mapping;
742 struct inode *inode = mapping->host;
744 return inode_resv_map(inode);
747 return (struct resv_map *)(get_vma_private_data(vma) &
752 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
754 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
755 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
757 set_vma_private_data(vma, (get_vma_private_data(vma) &
758 HPAGE_RESV_MASK) | (unsigned long)map);
761 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
763 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
764 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
766 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
769 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
771 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
773 return (get_vma_private_data(vma) & flag) != 0;
776 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
777 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
779 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
780 if (!(vma->vm_flags & VM_MAYSHARE))
781 vma->vm_private_data = (void *)0;
784 /* Returns true if the VMA has associated reserve pages */
785 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
787 if (vma->vm_flags & VM_NORESERVE) {
789 * This address is already reserved by other process(chg == 0),
790 * so, we should decrement reserved count. Without decrementing,
791 * reserve count remains after releasing inode, because this
792 * allocated page will go into page cache and is regarded as
793 * coming from reserved pool in releasing step. Currently, we
794 * don't have any other solution to deal with this situation
795 * properly, so add work-around here.
797 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
803 /* Shared mappings always use reserves */
804 if (vma->vm_flags & VM_MAYSHARE) {
806 * We know VM_NORESERVE is not set. Therefore, there SHOULD
807 * be a region map for all pages. The only situation where
808 * there is no region map is if a hole was punched via
809 * fallocate. In this case, there really are no reverves to
810 * use. This situation is indicated if chg != 0.
819 * Only the process that called mmap() has reserves for
822 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
828 static void enqueue_huge_page(struct hstate *h, struct page *page)
830 int nid = page_to_nid(page);
831 list_move(&page->lru, &h->hugepage_freelists[nid]);
832 h->free_huge_pages++;
833 h->free_huge_pages_node[nid]++;
836 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
840 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
841 if (!is_migrate_isolate_page(page))
844 * if 'non-isolated free hugepage' not found on the list,
845 * the allocation fails.
847 if (&h->hugepage_freelists[nid] == &page->lru)
849 list_move(&page->lru, &h->hugepage_activelist);
850 set_page_refcounted(page);
851 h->free_huge_pages--;
852 h->free_huge_pages_node[nid]--;
856 /* Movability of hugepages depends on migration support. */
857 static inline gfp_t htlb_alloc_mask(struct hstate *h)
859 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
860 return GFP_HIGHUSER_MOVABLE;
865 static struct page *dequeue_huge_page_vma(struct hstate *h,
866 struct vm_area_struct *vma,
867 unsigned long address, int avoid_reserve,
870 struct page *page = NULL;
871 struct mempolicy *mpol;
872 nodemask_t *nodemask;
873 struct zonelist *zonelist;
876 unsigned int cpuset_mems_cookie;
879 * A child process with MAP_PRIVATE mappings created by their parent
880 * have no page reserves. This check ensures that reservations are
881 * not "stolen". The child may still get SIGKILLed
883 if (!vma_has_reserves(vma, chg) &&
884 h->free_huge_pages - h->resv_huge_pages == 0)
887 /* If reserves cannot be used, ensure enough pages are in the pool */
888 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
892 cpuset_mems_cookie = read_mems_allowed_begin();
893 zonelist = huge_zonelist(vma, address,
894 htlb_alloc_mask(h), &mpol, &nodemask);
896 for_each_zone_zonelist_nodemask(zone, z, zonelist,
897 MAX_NR_ZONES - 1, nodemask) {
898 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
899 page = dequeue_huge_page_node(h, zone_to_nid(zone));
903 if (!vma_has_reserves(vma, chg))
906 SetPagePrivate(page);
907 h->resv_huge_pages--;
914 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
923 * common helper functions for hstate_next_node_to_{alloc|free}.
924 * We may have allocated or freed a huge page based on a different
925 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
926 * be outside of *nodes_allowed. Ensure that we use an allowed
927 * node for alloc or free.
929 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
931 nid = next_node(nid, *nodes_allowed);
932 if (nid == MAX_NUMNODES)
933 nid = first_node(*nodes_allowed);
934 VM_BUG_ON(nid >= MAX_NUMNODES);
939 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
941 if (!node_isset(nid, *nodes_allowed))
942 nid = next_node_allowed(nid, nodes_allowed);
947 * returns the previously saved node ["this node"] from which to
948 * allocate a persistent huge page for the pool and advance the
949 * next node from which to allocate, handling wrap at end of node
952 static int hstate_next_node_to_alloc(struct hstate *h,
953 nodemask_t *nodes_allowed)
957 VM_BUG_ON(!nodes_allowed);
959 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
960 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
966 * helper for free_pool_huge_page() - return the previously saved
967 * node ["this node"] from which to free a huge page. Advance the
968 * next node id whether or not we find a free huge page to free so
969 * that the next attempt to free addresses the next node.
971 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
975 VM_BUG_ON(!nodes_allowed);
977 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
978 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
983 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
984 for (nr_nodes = nodes_weight(*mask); \
986 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
989 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
990 for (nr_nodes = nodes_weight(*mask); \
992 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
995 #if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
996 static void destroy_compound_gigantic_page(struct page *page,
1000 int nr_pages = 1 << order;
1001 struct page *p = page + 1;
1003 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1004 clear_compound_head(p);
1005 set_page_refcounted(p);
1008 set_compound_order(page, 0);
1009 __ClearPageHead(page);
1012 static void free_gigantic_page(struct page *page, unsigned int order)
1014 free_contig_range(page_to_pfn(page), 1 << order);
1017 static int __alloc_gigantic_page(unsigned long start_pfn,
1018 unsigned long nr_pages)
1020 unsigned long end_pfn = start_pfn + nr_pages;
1021 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1024 static bool pfn_range_valid_gigantic(unsigned long start_pfn,
1025 unsigned long nr_pages)
1027 unsigned long i, end_pfn = start_pfn + nr_pages;
1030 for (i = start_pfn; i < end_pfn; i++) {
1034 page = pfn_to_page(i);
1036 if (PageReserved(page))
1039 if (page_count(page) > 0)
1049 static bool zone_spans_last_pfn(const struct zone *zone,
1050 unsigned long start_pfn, unsigned long nr_pages)
1052 unsigned long last_pfn = start_pfn + nr_pages - 1;
1053 return zone_spans_pfn(zone, last_pfn);
1056 static struct page *alloc_gigantic_page(int nid, unsigned int order)
1058 unsigned long nr_pages = 1 << order;
1059 unsigned long ret, pfn, flags;
1062 z = NODE_DATA(nid)->node_zones;
1063 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
1064 spin_lock_irqsave(&z->lock, flags);
1066 pfn = ALIGN(z->zone_start_pfn, nr_pages);
1067 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
1068 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
1070 * We release the zone lock here because
1071 * alloc_contig_range() will also lock the zone
1072 * at some point. If there's an allocation
1073 * spinning on this lock, it may win the race
1074 * and cause alloc_contig_range() to fail...
1076 spin_unlock_irqrestore(&z->lock, flags);
1077 ret = __alloc_gigantic_page(pfn, nr_pages);
1079 return pfn_to_page(pfn);
1080 spin_lock_irqsave(&z->lock, flags);
1085 spin_unlock_irqrestore(&z->lock, flags);
1091 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1092 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1094 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1098 page = alloc_gigantic_page(nid, huge_page_order(h));
1100 prep_compound_gigantic_page(page, huge_page_order(h));
1101 prep_new_huge_page(h, page, nid);
1107 static int alloc_fresh_gigantic_page(struct hstate *h,
1108 nodemask_t *nodes_allowed)
1110 struct page *page = NULL;
1113 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1114 page = alloc_fresh_gigantic_page_node(h, node);
1122 static inline bool gigantic_page_supported(void) { return true; }
1124 static inline bool gigantic_page_supported(void) { return false; }
1125 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1126 static inline void destroy_compound_gigantic_page(struct page *page,
1127 unsigned int order) { }
1128 static inline int alloc_fresh_gigantic_page(struct hstate *h,
1129 nodemask_t *nodes_allowed) { return 0; }
1132 static void update_and_free_page(struct hstate *h, struct page *page)
1136 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1140 h->nr_huge_pages_node[page_to_nid(page)]--;
1141 for (i = 0; i < pages_per_huge_page(h); i++) {
1142 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1143 1 << PG_referenced | 1 << PG_dirty |
1144 1 << PG_active | 1 << PG_private |
1147 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1148 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1149 set_page_refcounted(page);
1150 if (hstate_is_gigantic(h)) {
1151 destroy_compound_gigantic_page(page, huge_page_order(h));
1152 free_gigantic_page(page, huge_page_order(h));
1154 __free_pages(page, huge_page_order(h));
1158 struct hstate *size_to_hstate(unsigned long size)
1162 for_each_hstate(h) {
1163 if (huge_page_size(h) == size)
1170 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1171 * to hstate->hugepage_activelist.)
1173 * This function can be called for tail pages, but never returns true for them.
1175 bool page_huge_active(struct page *page)
1177 VM_BUG_ON_PAGE(!PageHuge(page), page);
1178 return PageHead(page) && PagePrivate(&page[1]);
1181 /* never called for tail page */
1182 static void set_page_huge_active(struct page *page)
1184 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1185 SetPagePrivate(&page[1]);
1188 static void clear_page_huge_active(struct page *page)
1190 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1191 ClearPagePrivate(&page[1]);
1194 void free_huge_page(struct page *page)
1197 * Can't pass hstate in here because it is called from the
1198 * compound page destructor.
1200 struct hstate *h = page_hstate(page);
1201 int nid = page_to_nid(page);
1202 struct hugepage_subpool *spool =
1203 (struct hugepage_subpool *)page_private(page);
1204 bool restore_reserve;
1206 set_page_private(page, 0);
1207 page->mapping = NULL;
1208 BUG_ON(page_count(page));
1209 BUG_ON(page_mapcount(page));
1210 restore_reserve = PagePrivate(page);
1211 ClearPagePrivate(page);
1214 * A return code of zero implies that the subpool will be under its
1215 * minimum size if the reservation is not restored after page is free.
1216 * Therefore, force restore_reserve operation.
1218 if (hugepage_subpool_put_pages(spool, 1) == 0)
1219 restore_reserve = true;
1221 spin_lock(&hugetlb_lock);
1222 clear_page_huge_active(page);
1223 hugetlb_cgroup_uncharge_page(hstate_index(h),
1224 pages_per_huge_page(h), page);
1225 if (restore_reserve)
1226 h->resv_huge_pages++;
1228 if (h->surplus_huge_pages_node[nid]) {
1229 /* remove the page from active list */
1230 list_del(&page->lru);
1231 update_and_free_page(h, page);
1232 h->surplus_huge_pages--;
1233 h->surplus_huge_pages_node[nid]--;
1235 arch_clear_hugepage_flags(page);
1236 enqueue_huge_page(h, page);
1238 spin_unlock(&hugetlb_lock);
1241 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1243 INIT_LIST_HEAD(&page->lru);
1244 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1245 spin_lock(&hugetlb_lock);
1246 set_hugetlb_cgroup(page, NULL);
1248 h->nr_huge_pages_node[nid]++;
1249 spin_unlock(&hugetlb_lock);
1250 put_page(page); /* free it into the hugepage allocator */
1253 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1256 int nr_pages = 1 << order;
1257 struct page *p = page + 1;
1259 /* we rely on prep_new_huge_page to set the destructor */
1260 set_compound_order(page, order);
1261 __ClearPageReserved(page);
1262 __SetPageHead(page);
1263 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1265 * For gigantic hugepages allocated through bootmem at
1266 * boot, it's safer to be consistent with the not-gigantic
1267 * hugepages and clear the PG_reserved bit from all tail pages
1268 * too. Otherwse drivers using get_user_pages() to access tail
1269 * pages may get the reference counting wrong if they see
1270 * PG_reserved set on a tail page (despite the head page not
1271 * having PG_reserved set). Enforcing this consistency between
1272 * head and tail pages allows drivers to optimize away a check
1273 * on the head page when they need know if put_page() is needed
1274 * after get_user_pages().
1276 __ClearPageReserved(p);
1277 set_page_count(p, 0);
1278 set_compound_head(p, page);
1283 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1284 * transparent huge pages. See the PageTransHuge() documentation for more
1287 int PageHuge(struct page *page)
1289 if (!PageCompound(page))
1292 page = compound_head(page);
1293 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1295 EXPORT_SYMBOL_GPL(PageHuge);
1298 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1299 * normal or transparent huge pages.
1301 int PageHeadHuge(struct page *page_head)
1303 if (!PageHead(page_head))
1306 return get_compound_page_dtor(page_head) == free_huge_page;
1309 pgoff_t __basepage_index(struct page *page)
1311 struct page *page_head = compound_head(page);
1312 pgoff_t index = page_index(page_head);
1313 unsigned long compound_idx;
1315 if (!PageHuge(page_head))
1316 return page_index(page);
1318 if (compound_order(page_head) >= MAX_ORDER)
1319 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1321 compound_idx = page - page_head;
1323 return (index << compound_order(page_head)) + compound_idx;
1326 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1330 page = __alloc_pages_node(nid,
1331 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1332 __GFP_REPEAT|__GFP_NOWARN,
1333 huge_page_order(h));
1335 prep_new_huge_page(h, page, nid);
1341 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1347 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1348 page = alloc_fresh_huge_page_node(h, node);
1356 count_vm_event(HTLB_BUDDY_PGALLOC);
1358 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1364 * Free huge page from pool from next node to free.
1365 * Attempt to keep persistent huge pages more or less
1366 * balanced over allowed nodes.
1367 * Called with hugetlb_lock locked.
1369 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1375 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1377 * If we're returning unused surplus pages, only examine
1378 * nodes with surplus pages.
1380 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1381 !list_empty(&h->hugepage_freelists[node])) {
1383 list_entry(h->hugepage_freelists[node].next,
1385 list_del(&page->lru);
1386 h->free_huge_pages--;
1387 h->free_huge_pages_node[node]--;
1389 h->surplus_huge_pages--;
1390 h->surplus_huge_pages_node[node]--;
1392 update_and_free_page(h, page);
1402 * Dissolve a given free hugepage into free buddy pages. This function does
1403 * nothing for in-use (including surplus) hugepages.
1405 static void dissolve_free_huge_page(struct page *page)
1407 spin_lock(&hugetlb_lock);
1408 if (PageHuge(page) && !page_count(page)) {
1409 struct hstate *h = page_hstate(page);
1410 int nid = page_to_nid(page);
1411 list_del(&page->lru);
1412 h->free_huge_pages--;
1413 h->free_huge_pages_node[nid]--;
1414 update_and_free_page(h, page);
1416 spin_unlock(&hugetlb_lock);
1420 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1421 * make specified memory blocks removable from the system.
1422 * Note that start_pfn should aligned with (minimum) hugepage size.
1424 void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1428 if (!hugepages_supported())
1431 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order));
1432 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
1433 dissolve_free_huge_page(pfn_to_page(pfn));
1437 * There are 3 ways this can get called:
1438 * 1. With vma+addr: we use the VMA's memory policy
1439 * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge
1440 * page from any node, and let the buddy allocator itself figure
1442 * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page
1443 * strictly from 'nid'
1445 static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1446 struct vm_area_struct *vma, unsigned long addr, int nid)
1448 int order = huge_page_order(h);
1449 gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN;
1450 unsigned int cpuset_mems_cookie;
1453 * We need a VMA to get a memory policy. If we do not
1454 * have one, we use the 'nid' argument.
1456 * The mempolicy stuff below has some non-inlined bits
1457 * and calls ->vm_ops. That makes it hard to optimize at
1458 * compile-time, even when NUMA is off and it does
1459 * nothing. This helps the compiler optimize it out.
1461 if (!IS_ENABLED(CONFIG_NUMA) || !vma) {
1463 * If a specific node is requested, make sure to
1464 * get memory from there, but only when a node
1465 * is explicitly specified.
1467 if (nid != NUMA_NO_NODE)
1468 gfp |= __GFP_THISNODE;
1470 * Make sure to call something that can handle
1473 return alloc_pages_node(nid, gfp, order);
1477 * OK, so we have a VMA. Fetch the mempolicy and try to
1478 * allocate a huge page with it. We will only reach this
1479 * when CONFIG_NUMA=y.
1483 struct mempolicy *mpol;
1484 struct zonelist *zl;
1485 nodemask_t *nodemask;
1487 cpuset_mems_cookie = read_mems_allowed_begin();
1488 zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask);
1489 mpol_cond_put(mpol);
1490 page = __alloc_pages_nodemask(gfp, order, zl, nodemask);
1493 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1499 * There are two ways to allocate a huge page:
1500 * 1. When you have a VMA and an address (like a fault)
1501 * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1503 * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in
1504 * this case which signifies that the allocation should be done with
1505 * respect for the VMA's memory policy.
1507 * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1508 * implies that memory policies will not be taken in to account.
1510 static struct page *__alloc_buddy_huge_page(struct hstate *h,
1511 struct vm_area_struct *vma, unsigned long addr, int nid)
1516 if (hstate_is_gigantic(h))
1520 VM_WARN_ON_ONCE(!addr || addr == -1);
1521 VM_WARN_ON_ONCE(nid != NUMA_NO_NODE);
1524 * Assume we will successfully allocate the surplus page to
1525 * prevent racing processes from causing the surplus to exceed
1528 * This however introduces a different race, where a process B
1529 * tries to grow the static hugepage pool while alloc_pages() is
1530 * called by process A. B will only examine the per-node
1531 * counters in determining if surplus huge pages can be
1532 * converted to normal huge pages in adjust_pool_surplus(). A
1533 * won't be able to increment the per-node counter, until the
1534 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1535 * no more huge pages can be converted from surplus to normal
1536 * state (and doesn't try to convert again). Thus, we have a
1537 * case where a surplus huge page exists, the pool is grown, and
1538 * the surplus huge page still exists after, even though it
1539 * should just have been converted to a normal huge page. This
1540 * does not leak memory, though, as the hugepage will be freed
1541 * once it is out of use. It also does not allow the counters to
1542 * go out of whack in adjust_pool_surplus() as we don't modify
1543 * the node values until we've gotten the hugepage and only the
1544 * per-node value is checked there.
1546 spin_lock(&hugetlb_lock);
1547 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1548 spin_unlock(&hugetlb_lock);
1552 h->surplus_huge_pages++;
1554 spin_unlock(&hugetlb_lock);
1556 page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid);
1558 spin_lock(&hugetlb_lock);
1560 INIT_LIST_HEAD(&page->lru);
1561 r_nid = page_to_nid(page);
1562 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1563 set_hugetlb_cgroup(page, NULL);
1565 * We incremented the global counters already
1567 h->nr_huge_pages_node[r_nid]++;
1568 h->surplus_huge_pages_node[r_nid]++;
1569 __count_vm_event(HTLB_BUDDY_PGALLOC);
1572 h->surplus_huge_pages--;
1573 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1575 spin_unlock(&hugetlb_lock);
1581 * Allocate a huge page from 'nid'. Note, 'nid' may be
1582 * NUMA_NO_NODE, which means that it may be allocated
1586 struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid)
1588 unsigned long addr = -1;
1590 return __alloc_buddy_huge_page(h, NULL, addr, nid);
1594 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1597 struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1598 struct vm_area_struct *vma, unsigned long addr)
1600 return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE);
1604 * This allocation function is useful in the context where vma is irrelevant.
1605 * E.g. soft-offlining uses this function because it only cares physical
1606 * address of error page.
1608 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1610 struct page *page = NULL;
1612 spin_lock(&hugetlb_lock);
1613 if (h->free_huge_pages - h->resv_huge_pages > 0)
1614 page = dequeue_huge_page_node(h, nid);
1615 spin_unlock(&hugetlb_lock);
1618 page = __alloc_buddy_huge_page_no_mpol(h, nid);
1624 * Increase the hugetlb pool such that it can accommodate a reservation
1627 static int gather_surplus_pages(struct hstate *h, int delta)
1629 struct list_head surplus_list;
1630 struct page *page, *tmp;
1632 int needed, allocated;
1633 bool alloc_ok = true;
1635 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1637 h->resv_huge_pages += delta;
1642 INIT_LIST_HEAD(&surplus_list);
1646 spin_unlock(&hugetlb_lock);
1647 for (i = 0; i < needed; i++) {
1648 page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE);
1653 list_add(&page->lru, &surplus_list);
1658 * After retaking hugetlb_lock, we need to recalculate 'needed'
1659 * because either resv_huge_pages or free_huge_pages may have changed.
1661 spin_lock(&hugetlb_lock);
1662 needed = (h->resv_huge_pages + delta) -
1663 (h->free_huge_pages + allocated);
1668 * We were not able to allocate enough pages to
1669 * satisfy the entire reservation so we free what
1670 * we've allocated so far.
1675 * The surplus_list now contains _at_least_ the number of extra pages
1676 * needed to accommodate the reservation. Add the appropriate number
1677 * of pages to the hugetlb pool and free the extras back to the buddy
1678 * allocator. Commit the entire reservation here to prevent another
1679 * process from stealing the pages as they are added to the pool but
1680 * before they are reserved.
1682 needed += allocated;
1683 h->resv_huge_pages += delta;
1686 /* Free the needed pages to the hugetlb pool */
1687 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1691 * This page is now managed by the hugetlb allocator and has
1692 * no users -- drop the buddy allocator's reference.
1694 put_page_testzero(page);
1695 VM_BUG_ON_PAGE(page_count(page), page);
1696 enqueue_huge_page(h, page);
1699 spin_unlock(&hugetlb_lock);
1701 /* Free unnecessary surplus pages to the buddy allocator */
1702 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1704 spin_lock(&hugetlb_lock);
1710 * When releasing a hugetlb pool reservation, any surplus pages that were
1711 * allocated to satisfy the reservation must be explicitly freed if they were
1713 * Called with hugetlb_lock held.
1715 static void return_unused_surplus_pages(struct hstate *h,
1716 unsigned long unused_resv_pages)
1718 unsigned long nr_pages;
1720 /* Uncommit the reservation */
1721 h->resv_huge_pages -= unused_resv_pages;
1723 /* Cannot return gigantic pages currently */
1724 if (hstate_is_gigantic(h))
1727 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1730 * We want to release as many surplus pages as possible, spread
1731 * evenly across all nodes with memory. Iterate across these nodes
1732 * until we can no longer free unreserved surplus pages. This occurs
1733 * when the nodes with surplus pages have no free pages.
1734 * free_pool_huge_page() will balance the the freed pages across the
1735 * on-line nodes with memory and will handle the hstate accounting.
1737 while (nr_pages--) {
1738 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1740 cond_resched_lock(&hugetlb_lock);
1746 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1747 * are used by the huge page allocation routines to manage reservations.
1749 * vma_needs_reservation is called to determine if the huge page at addr
1750 * within the vma has an associated reservation. If a reservation is
1751 * needed, the value 1 is returned. The caller is then responsible for
1752 * managing the global reservation and subpool usage counts. After
1753 * the huge page has been allocated, vma_commit_reservation is called
1754 * to add the page to the reservation map. If the page allocation fails,
1755 * the reservation must be ended instead of committed. vma_end_reservation
1756 * is called in such cases.
1758 * In the normal case, vma_commit_reservation returns the same value
1759 * as the preceding vma_needs_reservation call. The only time this
1760 * is not the case is if a reserve map was changed between calls. It
1761 * is the responsibility of the caller to notice the difference and
1762 * take appropriate action.
1764 enum vma_resv_mode {
1769 static long __vma_reservation_common(struct hstate *h,
1770 struct vm_area_struct *vma, unsigned long addr,
1771 enum vma_resv_mode mode)
1773 struct resv_map *resv;
1777 resv = vma_resv_map(vma);
1781 idx = vma_hugecache_offset(h, vma, addr);
1783 case VMA_NEEDS_RESV:
1784 ret = region_chg(resv, idx, idx + 1);
1786 case VMA_COMMIT_RESV:
1787 ret = region_add(resv, idx, idx + 1);
1790 region_abort(resv, idx, idx + 1);
1797 if (vma->vm_flags & VM_MAYSHARE)
1800 return ret < 0 ? ret : 0;
1803 static long vma_needs_reservation(struct hstate *h,
1804 struct vm_area_struct *vma, unsigned long addr)
1806 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1809 static long vma_commit_reservation(struct hstate *h,
1810 struct vm_area_struct *vma, unsigned long addr)
1812 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1815 static void vma_end_reservation(struct hstate *h,
1816 struct vm_area_struct *vma, unsigned long addr)
1818 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1821 struct page *alloc_huge_page(struct vm_area_struct *vma,
1822 unsigned long addr, int avoid_reserve)
1824 struct hugepage_subpool *spool = subpool_vma(vma);
1825 struct hstate *h = hstate_vma(vma);
1827 long map_chg, map_commit;
1830 struct hugetlb_cgroup *h_cg;
1832 idx = hstate_index(h);
1834 * Examine the region/reserve map to determine if the process
1835 * has a reservation for the page to be allocated. A return
1836 * code of zero indicates a reservation exists (no change).
1838 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
1840 return ERR_PTR(-ENOMEM);
1843 * Processes that did not create the mapping will have no
1844 * reserves as indicated by the region/reserve map. Check
1845 * that the allocation will not exceed the subpool limit.
1846 * Allocations for MAP_NORESERVE mappings also need to be
1847 * checked against any subpool limit.
1849 if (map_chg || avoid_reserve) {
1850 gbl_chg = hugepage_subpool_get_pages(spool, 1);
1852 vma_end_reservation(h, vma, addr);
1853 return ERR_PTR(-ENOSPC);
1857 * Even though there was no reservation in the region/reserve
1858 * map, there could be reservations associated with the
1859 * subpool that can be used. This would be indicated if the
1860 * return value of hugepage_subpool_get_pages() is zero.
1861 * However, if avoid_reserve is specified we still avoid even
1862 * the subpool reservations.
1868 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1870 goto out_subpool_put;
1872 spin_lock(&hugetlb_lock);
1874 * glb_chg is passed to indicate whether or not a page must be taken
1875 * from the global free pool (global change). gbl_chg == 0 indicates
1876 * a reservation exists for the allocation.
1878 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
1880 spin_unlock(&hugetlb_lock);
1881 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
1883 goto out_uncharge_cgroup;
1885 spin_lock(&hugetlb_lock);
1886 list_move(&page->lru, &h->hugepage_activelist);
1889 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1890 spin_unlock(&hugetlb_lock);
1892 set_page_private(page, (unsigned long)spool);
1894 map_commit = vma_commit_reservation(h, vma, addr);
1895 if (unlikely(map_chg > map_commit)) {
1897 * The page was added to the reservation map between
1898 * vma_needs_reservation and vma_commit_reservation.
1899 * This indicates a race with hugetlb_reserve_pages.
1900 * Adjust for the subpool count incremented above AND
1901 * in hugetlb_reserve_pages for the same page. Also,
1902 * the reservation count added in hugetlb_reserve_pages
1903 * no longer applies.
1907 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
1908 hugetlb_acct_memory(h, -rsv_adjust);
1912 out_uncharge_cgroup:
1913 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1915 if (map_chg || avoid_reserve)
1916 hugepage_subpool_put_pages(spool, 1);
1917 vma_end_reservation(h, vma, addr);
1918 return ERR_PTR(-ENOSPC);
1922 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1923 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1924 * where no ERR_VALUE is expected to be returned.
1926 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1927 unsigned long addr, int avoid_reserve)
1929 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1935 int __weak alloc_bootmem_huge_page(struct hstate *h)
1937 struct huge_bootmem_page *m;
1940 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1943 addr = memblock_virt_alloc_try_nid_nopanic(
1944 huge_page_size(h), huge_page_size(h),
1945 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1948 * Use the beginning of the huge page to store the
1949 * huge_bootmem_page struct (until gather_bootmem
1950 * puts them into the mem_map).
1959 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
1960 /* Put them into a private list first because mem_map is not up yet */
1961 list_add(&m->list, &huge_boot_pages);
1966 static void __init prep_compound_huge_page(struct page *page,
1969 if (unlikely(order > (MAX_ORDER - 1)))
1970 prep_compound_gigantic_page(page, order);
1972 prep_compound_page(page, order);
1975 /* Put bootmem huge pages into the standard lists after mem_map is up */
1976 static void __init gather_bootmem_prealloc(void)
1978 struct huge_bootmem_page *m;
1980 list_for_each_entry(m, &huge_boot_pages, list) {
1981 struct hstate *h = m->hstate;
1984 #ifdef CONFIG_HIGHMEM
1985 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1986 memblock_free_late(__pa(m),
1987 sizeof(struct huge_bootmem_page));
1989 page = virt_to_page(m);
1991 WARN_ON(page_count(page) != 1);
1992 prep_compound_huge_page(page, h->order);
1993 WARN_ON(PageReserved(page));
1994 prep_new_huge_page(h, page, page_to_nid(page));
1996 * If we had gigantic hugepages allocated at boot time, we need
1997 * to restore the 'stolen' pages to totalram_pages in order to
1998 * fix confusing memory reports from free(1) and another
1999 * side-effects, like CommitLimit going negative.
2001 if (hstate_is_gigantic(h))
2002 adjust_managed_page_count(page, 1 << h->order);
2006 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2010 for (i = 0; i < h->max_huge_pages; ++i) {
2011 if (hstate_is_gigantic(h)) {
2012 if (!alloc_bootmem_huge_page(h))
2014 } else if (!alloc_fresh_huge_page(h,
2015 &node_states[N_MEMORY]))
2018 h->max_huge_pages = i;
2021 static void __init hugetlb_init_hstates(void)
2025 for_each_hstate(h) {
2026 if (minimum_order > huge_page_order(h))
2027 minimum_order = huge_page_order(h);
2029 /* oversize hugepages were init'ed in early boot */
2030 if (!hstate_is_gigantic(h))
2031 hugetlb_hstate_alloc_pages(h);
2033 VM_BUG_ON(minimum_order == UINT_MAX);
2036 static char * __init memfmt(char *buf, unsigned long n)
2038 if (n >= (1UL << 30))
2039 sprintf(buf, "%lu GB", n >> 30);
2040 else if (n >= (1UL << 20))
2041 sprintf(buf, "%lu MB", n >> 20);
2043 sprintf(buf, "%lu KB", n >> 10);
2047 static void __init report_hugepages(void)
2051 for_each_hstate(h) {
2053 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2054 memfmt(buf, huge_page_size(h)),
2055 h->free_huge_pages);
2059 #ifdef CONFIG_HIGHMEM
2060 static void try_to_free_low(struct hstate *h, unsigned long count,
2061 nodemask_t *nodes_allowed)
2065 if (hstate_is_gigantic(h))
2068 for_each_node_mask(i, *nodes_allowed) {
2069 struct page *page, *next;
2070 struct list_head *freel = &h->hugepage_freelists[i];
2071 list_for_each_entry_safe(page, next, freel, lru) {
2072 if (count >= h->nr_huge_pages)
2074 if (PageHighMem(page))
2076 list_del(&page->lru);
2077 update_and_free_page(h, page);
2078 h->free_huge_pages--;
2079 h->free_huge_pages_node[page_to_nid(page)]--;
2084 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2085 nodemask_t *nodes_allowed)
2091 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2092 * balanced by operating on them in a round-robin fashion.
2093 * Returns 1 if an adjustment was made.
2095 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2100 VM_BUG_ON(delta != -1 && delta != 1);
2103 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2104 if (h->surplus_huge_pages_node[node])
2108 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2109 if (h->surplus_huge_pages_node[node] <
2110 h->nr_huge_pages_node[node])
2117 h->surplus_huge_pages += delta;
2118 h->surplus_huge_pages_node[node] += delta;
2122 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2123 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2124 nodemask_t *nodes_allowed)
2126 unsigned long min_count, ret;
2128 if (hstate_is_gigantic(h) && !gigantic_page_supported())
2129 return h->max_huge_pages;
2132 * Increase the pool size
2133 * First take pages out of surplus state. Then make up the
2134 * remaining difference by allocating fresh huge pages.
2136 * We might race with alloc_buddy_huge_page() here and be unable
2137 * to convert a surplus huge page to a normal huge page. That is
2138 * not critical, though, it just means the overall size of the
2139 * pool might be one hugepage larger than it needs to be, but
2140 * within all the constraints specified by the sysctls.
2142 spin_lock(&hugetlb_lock);
2143 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2144 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2148 while (count > persistent_huge_pages(h)) {
2150 * If this allocation races such that we no longer need the
2151 * page, free_huge_page will handle it by freeing the page
2152 * and reducing the surplus.
2154 spin_unlock(&hugetlb_lock);
2155 if (hstate_is_gigantic(h))
2156 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2158 ret = alloc_fresh_huge_page(h, nodes_allowed);
2159 spin_lock(&hugetlb_lock);
2163 /* Bail for signals. Probably ctrl-c from user */
2164 if (signal_pending(current))
2169 * Decrease the pool size
2170 * First return free pages to the buddy allocator (being careful
2171 * to keep enough around to satisfy reservations). Then place
2172 * pages into surplus state as needed so the pool will shrink
2173 * to the desired size as pages become free.
2175 * By placing pages into the surplus state independent of the
2176 * overcommit value, we are allowing the surplus pool size to
2177 * exceed overcommit. There are few sane options here. Since
2178 * alloc_buddy_huge_page() is checking the global counter,
2179 * though, we'll note that we're not allowed to exceed surplus
2180 * and won't grow the pool anywhere else. Not until one of the
2181 * sysctls are changed, or the surplus pages go out of use.
2183 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2184 min_count = max(count, min_count);
2185 try_to_free_low(h, min_count, nodes_allowed);
2186 while (min_count < persistent_huge_pages(h)) {
2187 if (!free_pool_huge_page(h, nodes_allowed, 0))
2189 cond_resched_lock(&hugetlb_lock);
2191 while (count < persistent_huge_pages(h)) {
2192 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2196 ret = persistent_huge_pages(h);
2197 spin_unlock(&hugetlb_lock);
2201 #define HSTATE_ATTR_RO(_name) \
2202 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2204 #define HSTATE_ATTR(_name) \
2205 static struct kobj_attribute _name##_attr = \
2206 __ATTR(_name, 0644, _name##_show, _name##_store)
2208 static struct kobject *hugepages_kobj;
2209 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2211 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2213 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2217 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2218 if (hstate_kobjs[i] == kobj) {
2220 *nidp = NUMA_NO_NODE;
2224 return kobj_to_node_hstate(kobj, nidp);
2227 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2228 struct kobj_attribute *attr, char *buf)
2231 unsigned long nr_huge_pages;
2234 h = kobj_to_hstate(kobj, &nid);
2235 if (nid == NUMA_NO_NODE)
2236 nr_huge_pages = h->nr_huge_pages;
2238 nr_huge_pages = h->nr_huge_pages_node[nid];
2240 return sprintf(buf, "%lu\n", nr_huge_pages);
2243 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2244 struct hstate *h, int nid,
2245 unsigned long count, size_t len)
2248 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2250 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2255 if (nid == NUMA_NO_NODE) {
2257 * global hstate attribute
2259 if (!(obey_mempolicy &&
2260 init_nodemask_of_mempolicy(nodes_allowed))) {
2261 NODEMASK_FREE(nodes_allowed);
2262 nodes_allowed = &node_states[N_MEMORY];
2264 } else if (nodes_allowed) {
2266 * per node hstate attribute: adjust count to global,
2267 * but restrict alloc/free to the specified node.
2269 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2270 init_nodemask_of_node(nodes_allowed, nid);
2272 nodes_allowed = &node_states[N_MEMORY];
2274 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2276 if (nodes_allowed != &node_states[N_MEMORY])
2277 NODEMASK_FREE(nodes_allowed);
2281 NODEMASK_FREE(nodes_allowed);
2285 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2286 struct kobject *kobj, const char *buf,
2290 unsigned long count;
2294 err = kstrtoul(buf, 10, &count);
2298 h = kobj_to_hstate(kobj, &nid);
2299 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2302 static ssize_t nr_hugepages_show(struct kobject *kobj,
2303 struct kobj_attribute *attr, char *buf)
2305 return nr_hugepages_show_common(kobj, attr, buf);
2308 static ssize_t nr_hugepages_store(struct kobject *kobj,
2309 struct kobj_attribute *attr, const char *buf, size_t len)
2311 return nr_hugepages_store_common(false, kobj, buf, len);
2313 HSTATE_ATTR(nr_hugepages);
2318 * hstate attribute for optionally mempolicy-based constraint on persistent
2319 * huge page alloc/free.
2321 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2322 struct kobj_attribute *attr, char *buf)
2324 return nr_hugepages_show_common(kobj, attr, buf);
2327 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2328 struct kobj_attribute *attr, const char *buf, size_t len)
2330 return nr_hugepages_store_common(true, kobj, buf, len);
2332 HSTATE_ATTR(nr_hugepages_mempolicy);
2336 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2337 struct kobj_attribute *attr, char *buf)
2339 struct hstate *h = kobj_to_hstate(kobj, NULL);
2340 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2343 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2344 struct kobj_attribute *attr, const char *buf, size_t count)
2347 unsigned long input;
2348 struct hstate *h = kobj_to_hstate(kobj, NULL);
2350 if (hstate_is_gigantic(h))
2353 err = kstrtoul(buf, 10, &input);
2357 spin_lock(&hugetlb_lock);
2358 h->nr_overcommit_huge_pages = input;
2359 spin_unlock(&hugetlb_lock);
2363 HSTATE_ATTR(nr_overcommit_hugepages);
2365 static ssize_t free_hugepages_show(struct kobject *kobj,
2366 struct kobj_attribute *attr, char *buf)
2369 unsigned long free_huge_pages;
2372 h = kobj_to_hstate(kobj, &nid);
2373 if (nid == NUMA_NO_NODE)
2374 free_huge_pages = h->free_huge_pages;
2376 free_huge_pages = h->free_huge_pages_node[nid];
2378 return sprintf(buf, "%lu\n", free_huge_pages);
2380 HSTATE_ATTR_RO(free_hugepages);
2382 static ssize_t resv_hugepages_show(struct kobject *kobj,
2383 struct kobj_attribute *attr, char *buf)
2385 struct hstate *h = kobj_to_hstate(kobj, NULL);
2386 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2388 HSTATE_ATTR_RO(resv_hugepages);
2390 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2391 struct kobj_attribute *attr, char *buf)
2394 unsigned long surplus_huge_pages;
2397 h = kobj_to_hstate(kobj, &nid);
2398 if (nid == NUMA_NO_NODE)
2399 surplus_huge_pages = h->surplus_huge_pages;
2401 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2403 return sprintf(buf, "%lu\n", surplus_huge_pages);
2405 HSTATE_ATTR_RO(surplus_hugepages);
2407 static struct attribute *hstate_attrs[] = {
2408 &nr_hugepages_attr.attr,
2409 &nr_overcommit_hugepages_attr.attr,
2410 &free_hugepages_attr.attr,
2411 &resv_hugepages_attr.attr,
2412 &surplus_hugepages_attr.attr,
2414 &nr_hugepages_mempolicy_attr.attr,
2419 static struct attribute_group hstate_attr_group = {
2420 .attrs = hstate_attrs,
2423 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2424 struct kobject **hstate_kobjs,
2425 struct attribute_group *hstate_attr_group)
2428 int hi = hstate_index(h);
2430 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2431 if (!hstate_kobjs[hi])
2434 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2436 kobject_put(hstate_kobjs[hi]);
2441 static void __init hugetlb_sysfs_init(void)
2446 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2447 if (!hugepages_kobj)
2450 for_each_hstate(h) {
2451 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2452 hstate_kobjs, &hstate_attr_group);
2454 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2461 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2462 * with node devices in node_devices[] using a parallel array. The array
2463 * index of a node device or _hstate == node id.
2464 * This is here to avoid any static dependency of the node device driver, in
2465 * the base kernel, on the hugetlb module.
2467 struct node_hstate {
2468 struct kobject *hugepages_kobj;
2469 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2471 static struct node_hstate node_hstates[MAX_NUMNODES];
2474 * A subset of global hstate attributes for node devices
2476 static struct attribute *per_node_hstate_attrs[] = {
2477 &nr_hugepages_attr.attr,
2478 &free_hugepages_attr.attr,
2479 &surplus_hugepages_attr.attr,
2483 static struct attribute_group per_node_hstate_attr_group = {
2484 .attrs = per_node_hstate_attrs,
2488 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2489 * Returns node id via non-NULL nidp.
2491 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2495 for (nid = 0; nid < nr_node_ids; nid++) {
2496 struct node_hstate *nhs = &node_hstates[nid];
2498 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2499 if (nhs->hstate_kobjs[i] == kobj) {
2511 * Unregister hstate attributes from a single node device.
2512 * No-op if no hstate attributes attached.
2514 static void hugetlb_unregister_node(struct node *node)
2517 struct node_hstate *nhs = &node_hstates[node->dev.id];
2519 if (!nhs->hugepages_kobj)
2520 return; /* no hstate attributes */
2522 for_each_hstate(h) {
2523 int idx = hstate_index(h);
2524 if (nhs->hstate_kobjs[idx]) {
2525 kobject_put(nhs->hstate_kobjs[idx]);
2526 nhs->hstate_kobjs[idx] = NULL;
2530 kobject_put(nhs->hugepages_kobj);
2531 nhs->hugepages_kobj = NULL;
2535 * hugetlb module exit: unregister hstate attributes from node devices
2538 static void hugetlb_unregister_all_nodes(void)
2543 * disable node device registrations.
2545 register_hugetlbfs_with_node(NULL, NULL);
2548 * remove hstate attributes from any nodes that have them.
2550 for (nid = 0; nid < nr_node_ids; nid++)
2551 hugetlb_unregister_node(node_devices[nid]);
2555 * Register hstate attributes for a single node device.
2556 * No-op if attributes already registered.
2558 static void hugetlb_register_node(struct node *node)
2561 struct node_hstate *nhs = &node_hstates[node->dev.id];
2564 if (nhs->hugepages_kobj)
2565 return; /* already allocated */
2567 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2569 if (!nhs->hugepages_kobj)
2572 for_each_hstate(h) {
2573 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2575 &per_node_hstate_attr_group);
2577 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2578 h->name, node->dev.id);
2579 hugetlb_unregister_node(node);
2586 * hugetlb init time: register hstate attributes for all registered node
2587 * devices of nodes that have memory. All on-line nodes should have
2588 * registered their associated device by this time.
2590 static void __init hugetlb_register_all_nodes(void)
2594 for_each_node_state(nid, N_MEMORY) {
2595 struct node *node = node_devices[nid];
2596 if (node->dev.id == nid)
2597 hugetlb_register_node(node);
2601 * Let the node device driver know we're here so it can
2602 * [un]register hstate attributes on node hotplug.
2604 register_hugetlbfs_with_node(hugetlb_register_node,
2605 hugetlb_unregister_node);
2607 #else /* !CONFIG_NUMA */
2609 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2617 static void hugetlb_unregister_all_nodes(void) { }
2619 static void hugetlb_register_all_nodes(void) { }
2623 static void __exit hugetlb_exit(void)
2627 hugetlb_unregister_all_nodes();
2629 for_each_hstate(h) {
2630 kobject_put(hstate_kobjs[hstate_index(h)]);
2633 kobject_put(hugepages_kobj);
2634 kfree(hugetlb_fault_mutex_table);
2636 module_exit(hugetlb_exit);
2638 static int __init hugetlb_init(void)
2642 if (!hugepages_supported())
2645 if (!size_to_hstate(default_hstate_size)) {
2646 default_hstate_size = HPAGE_SIZE;
2647 if (!size_to_hstate(default_hstate_size))
2648 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2650 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2651 if (default_hstate_max_huge_pages)
2652 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2654 hugetlb_init_hstates();
2655 gather_bootmem_prealloc();
2658 hugetlb_sysfs_init();
2659 hugetlb_register_all_nodes();
2660 hugetlb_cgroup_file_init();
2663 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2665 num_fault_mutexes = 1;
2667 hugetlb_fault_mutex_table =
2668 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2669 BUG_ON(!hugetlb_fault_mutex_table);
2671 for (i = 0; i < num_fault_mutexes; i++)
2672 mutex_init(&hugetlb_fault_mutex_table[i]);
2675 module_init(hugetlb_init);
2677 /* Should be called on processing a hugepagesz=... option */
2678 void __init hugetlb_add_hstate(unsigned int order)
2683 if (size_to_hstate(PAGE_SIZE << order)) {
2684 pr_warning("hugepagesz= specified twice, ignoring\n");
2687 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2689 h = &hstates[hugetlb_max_hstate++];
2691 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2692 h->nr_huge_pages = 0;
2693 h->free_huge_pages = 0;
2694 for (i = 0; i < MAX_NUMNODES; ++i)
2695 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2696 INIT_LIST_HEAD(&h->hugepage_activelist);
2697 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2698 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2699 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2700 huge_page_size(h)/1024);
2705 static int __init hugetlb_nrpages_setup(char *s)
2708 static unsigned long *last_mhp;
2711 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2712 * so this hugepages= parameter goes to the "default hstate".
2714 if (!hugetlb_max_hstate)
2715 mhp = &default_hstate_max_huge_pages;
2717 mhp = &parsed_hstate->max_huge_pages;
2719 if (mhp == last_mhp) {
2720 pr_warning("hugepages= specified twice without "
2721 "interleaving hugepagesz=, ignoring\n");
2725 if (sscanf(s, "%lu", mhp) <= 0)
2729 * Global state is always initialized later in hugetlb_init.
2730 * But we need to allocate >= MAX_ORDER hstates here early to still
2731 * use the bootmem allocator.
2733 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2734 hugetlb_hstate_alloc_pages(parsed_hstate);
2740 __setup("hugepages=", hugetlb_nrpages_setup);
2742 static int __init hugetlb_default_setup(char *s)
2744 default_hstate_size = memparse(s, &s);
2747 __setup("default_hugepagesz=", hugetlb_default_setup);
2749 static unsigned int cpuset_mems_nr(unsigned int *array)
2752 unsigned int nr = 0;
2754 for_each_node_mask(node, cpuset_current_mems_allowed)
2760 #ifdef CONFIG_SYSCTL
2761 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2762 struct ctl_table *table, int write,
2763 void __user *buffer, size_t *length, loff_t *ppos)
2765 struct hstate *h = &default_hstate;
2766 unsigned long tmp = h->max_huge_pages;
2769 if (!hugepages_supported())
2773 table->maxlen = sizeof(unsigned long);
2774 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2779 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2780 NUMA_NO_NODE, tmp, *length);
2785 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2786 void __user *buffer, size_t *length, loff_t *ppos)
2789 return hugetlb_sysctl_handler_common(false, table, write,
2790 buffer, length, ppos);
2794 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2795 void __user *buffer, size_t *length, loff_t *ppos)
2797 return hugetlb_sysctl_handler_common(true, table, write,
2798 buffer, length, ppos);
2800 #endif /* CONFIG_NUMA */
2802 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2803 void __user *buffer,
2804 size_t *length, loff_t *ppos)
2806 struct hstate *h = &default_hstate;
2810 if (!hugepages_supported())
2813 tmp = h->nr_overcommit_huge_pages;
2815 if (write && hstate_is_gigantic(h))
2819 table->maxlen = sizeof(unsigned long);
2820 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2825 spin_lock(&hugetlb_lock);
2826 h->nr_overcommit_huge_pages = tmp;
2827 spin_unlock(&hugetlb_lock);
2833 #endif /* CONFIG_SYSCTL */
2835 void hugetlb_report_meminfo(struct seq_file *m)
2837 struct hstate *h = &default_hstate;
2838 if (!hugepages_supported())
2841 "HugePages_Total: %5lu\n"
2842 "HugePages_Free: %5lu\n"
2843 "HugePages_Rsvd: %5lu\n"
2844 "HugePages_Surp: %5lu\n"
2845 "Hugepagesize: %8lu kB\n",
2849 h->surplus_huge_pages,
2850 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2853 int hugetlb_report_node_meminfo(int nid, char *buf)
2855 struct hstate *h = &default_hstate;
2856 if (!hugepages_supported())
2859 "Node %d HugePages_Total: %5u\n"
2860 "Node %d HugePages_Free: %5u\n"
2861 "Node %d HugePages_Surp: %5u\n",
2862 nid, h->nr_huge_pages_node[nid],
2863 nid, h->free_huge_pages_node[nid],
2864 nid, h->surplus_huge_pages_node[nid]);
2867 void hugetlb_show_meminfo(void)
2872 if (!hugepages_supported())
2875 for_each_node_state(nid, N_MEMORY)
2877 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2879 h->nr_huge_pages_node[nid],
2880 h->free_huge_pages_node[nid],
2881 h->surplus_huge_pages_node[nid],
2882 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2885 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
2887 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
2888 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
2891 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2892 unsigned long hugetlb_total_pages(void)
2895 unsigned long nr_total_pages = 0;
2898 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2899 return nr_total_pages;
2902 static int hugetlb_acct_memory(struct hstate *h, long delta)
2906 spin_lock(&hugetlb_lock);
2908 * When cpuset is configured, it breaks the strict hugetlb page
2909 * reservation as the accounting is done on a global variable. Such
2910 * reservation is completely rubbish in the presence of cpuset because
2911 * the reservation is not checked against page availability for the
2912 * current cpuset. Application can still potentially OOM'ed by kernel
2913 * with lack of free htlb page in cpuset that the task is in.
2914 * Attempt to enforce strict accounting with cpuset is almost
2915 * impossible (or too ugly) because cpuset is too fluid that
2916 * task or memory node can be dynamically moved between cpusets.
2918 * The change of semantics for shared hugetlb mapping with cpuset is
2919 * undesirable. However, in order to preserve some of the semantics,
2920 * we fall back to check against current free page availability as
2921 * a best attempt and hopefully to minimize the impact of changing
2922 * semantics that cpuset has.
2925 if (gather_surplus_pages(h, delta) < 0)
2928 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2929 return_unused_surplus_pages(h, delta);
2936 return_unused_surplus_pages(h, (unsigned long) -delta);
2939 spin_unlock(&hugetlb_lock);
2943 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2945 struct resv_map *resv = vma_resv_map(vma);
2948 * This new VMA should share its siblings reservation map if present.
2949 * The VMA will only ever have a valid reservation map pointer where
2950 * it is being copied for another still existing VMA. As that VMA
2951 * has a reference to the reservation map it cannot disappear until
2952 * after this open call completes. It is therefore safe to take a
2953 * new reference here without additional locking.
2955 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2956 kref_get(&resv->refs);
2959 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2961 struct hstate *h = hstate_vma(vma);
2962 struct resv_map *resv = vma_resv_map(vma);
2963 struct hugepage_subpool *spool = subpool_vma(vma);
2964 unsigned long reserve, start, end;
2967 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2970 start = vma_hugecache_offset(h, vma, vma->vm_start);
2971 end = vma_hugecache_offset(h, vma, vma->vm_end);
2973 reserve = (end - start) - region_count(resv, start, end);
2975 kref_put(&resv->refs, resv_map_release);
2979 * Decrement reserve counts. The global reserve count may be
2980 * adjusted if the subpool has a minimum size.
2982 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
2983 hugetlb_acct_memory(h, -gbl_reserve);
2988 * We cannot handle pagefaults against hugetlb pages at all. They cause
2989 * handle_mm_fault() to try to instantiate regular-sized pages in the
2990 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2993 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2999 const struct vm_operations_struct hugetlb_vm_ops = {
3000 .fault = hugetlb_vm_op_fault,
3001 .open = hugetlb_vm_op_open,
3002 .close = hugetlb_vm_op_close,
3005 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3011 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3012 vma->vm_page_prot)));
3014 entry = huge_pte_wrprotect(mk_huge_pte(page,
3015 vma->vm_page_prot));
3017 entry = pte_mkyoung(entry);
3018 entry = pte_mkhuge(entry);
3019 entry = arch_make_huge_pte(entry, vma, page, writable);
3024 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3025 unsigned long address, pte_t *ptep)
3029 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3030 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3031 update_mmu_cache(vma, address, ptep);
3034 static int is_hugetlb_entry_migration(pte_t pte)
3038 if (huge_pte_none(pte) || pte_present(pte))
3040 swp = pte_to_swp_entry(pte);
3041 if (non_swap_entry(swp) && is_migration_entry(swp))
3047 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3051 if (huge_pte_none(pte) || pte_present(pte))
3053 swp = pte_to_swp_entry(pte);
3054 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3060 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3061 struct vm_area_struct *vma)
3063 pte_t *src_pte, *dst_pte, entry;
3064 struct page *ptepage;
3067 struct hstate *h = hstate_vma(vma);
3068 unsigned long sz = huge_page_size(h);
3069 unsigned long mmun_start; /* For mmu_notifiers */
3070 unsigned long mmun_end; /* For mmu_notifiers */
3073 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3075 mmun_start = vma->vm_start;
3076 mmun_end = vma->vm_end;
3078 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3080 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3081 spinlock_t *src_ptl, *dst_ptl;
3082 src_pte = huge_pte_offset(src, addr);
3085 dst_pte = huge_pte_alloc(dst, addr, sz);
3091 /* If the pagetables are shared don't copy or take references */
3092 if (dst_pte == src_pte)
3095 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3096 src_ptl = huge_pte_lockptr(h, src, src_pte);
3097 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3098 entry = huge_ptep_get(src_pte);
3099 if (huge_pte_none(entry)) { /* skip none entry */
3101 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3102 is_hugetlb_entry_hwpoisoned(entry))) {
3103 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3105 if (is_write_migration_entry(swp_entry) && cow) {
3107 * COW mappings require pages in both
3108 * parent and child to be set to read.
3110 make_migration_entry_read(&swp_entry);
3111 entry = swp_entry_to_pte(swp_entry);
3112 set_huge_pte_at(src, addr, src_pte, entry);
3114 set_huge_pte_at(dst, addr, dst_pte, entry);
3117 huge_ptep_set_wrprotect(src, addr, src_pte);
3118 mmu_notifier_invalidate_range(src, mmun_start,
3121 entry = huge_ptep_get(src_pte);
3122 ptepage = pte_page(entry);
3124 page_dup_rmap(ptepage, true);
3125 set_huge_pte_at(dst, addr, dst_pte, entry);
3126 hugetlb_count_add(pages_per_huge_page(h), dst);
3128 spin_unlock(src_ptl);
3129 spin_unlock(dst_ptl);
3133 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3138 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3139 unsigned long start, unsigned long end,
3140 struct page *ref_page)
3142 int force_flush = 0;
3143 struct mm_struct *mm = vma->vm_mm;
3144 unsigned long address;
3149 struct hstate *h = hstate_vma(vma);
3150 unsigned long sz = huge_page_size(h);
3151 const unsigned long mmun_start = start; /* For mmu_notifiers */
3152 const unsigned long mmun_end = end; /* For mmu_notifiers */
3154 WARN_ON(!is_vm_hugetlb_page(vma));
3155 BUG_ON(start & ~huge_page_mask(h));
3156 BUG_ON(end & ~huge_page_mask(h));
3158 tlb_start_vma(tlb, vma);
3159 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3162 for (; address < end; address += sz) {
3163 ptep = huge_pte_offset(mm, address);
3167 ptl = huge_pte_lock(h, mm, ptep);
3168 if (huge_pmd_unshare(mm, &address, ptep))
3171 pte = huge_ptep_get(ptep);
3172 if (huge_pte_none(pte))
3176 * Migrating hugepage or HWPoisoned hugepage is already
3177 * unmapped and its refcount is dropped, so just clear pte here.
3179 if (unlikely(!pte_present(pte))) {
3180 huge_pte_clear(mm, address, ptep);
3184 page = pte_page(pte);
3186 * If a reference page is supplied, it is because a specific
3187 * page is being unmapped, not a range. Ensure the page we
3188 * are about to unmap is the actual page of interest.
3191 if (page != ref_page)
3195 * Mark the VMA as having unmapped its page so that
3196 * future faults in this VMA will fail rather than
3197 * looking like data was lost
3199 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3202 pte = huge_ptep_get_and_clear(mm, address, ptep);
3203 tlb_remove_tlb_entry(tlb, ptep, address);
3204 if (huge_pte_dirty(pte))
3205 set_page_dirty(page);
3207 hugetlb_count_sub(pages_per_huge_page(h), mm);
3208 page_remove_rmap(page, true);
3209 force_flush = !__tlb_remove_page(tlb, page);
3215 /* Bail out after unmapping reference page if supplied */
3224 * mmu_gather ran out of room to batch pages, we break out of
3225 * the PTE lock to avoid doing the potential expensive TLB invalidate
3226 * and page-free while holding it.
3231 if (address < end && !ref_page)
3234 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3235 tlb_end_vma(tlb, vma);
3238 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3239 struct vm_area_struct *vma, unsigned long start,
3240 unsigned long end, struct page *ref_page)
3242 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3245 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3246 * test will fail on a vma being torn down, and not grab a page table
3247 * on its way out. We're lucky that the flag has such an appropriate
3248 * name, and can in fact be safely cleared here. We could clear it
3249 * before the __unmap_hugepage_range above, but all that's necessary
3250 * is to clear it before releasing the i_mmap_rwsem. This works
3251 * because in the context this is called, the VMA is about to be
3252 * destroyed and the i_mmap_rwsem is held.
3254 vma->vm_flags &= ~VM_MAYSHARE;
3257 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3258 unsigned long end, struct page *ref_page)
3260 struct mm_struct *mm;
3261 struct mmu_gather tlb;
3265 tlb_gather_mmu(&tlb, mm, start, end);
3266 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3267 tlb_finish_mmu(&tlb, start, end);
3271 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3272 * mappping it owns the reserve page for. The intention is to unmap the page
3273 * from other VMAs and let the children be SIGKILLed if they are faulting the
3276 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3277 struct page *page, unsigned long address)
3279 struct hstate *h = hstate_vma(vma);
3280 struct vm_area_struct *iter_vma;
3281 struct address_space *mapping;
3285 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3286 * from page cache lookup which is in HPAGE_SIZE units.
3288 address = address & huge_page_mask(h);
3289 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3291 mapping = file_inode(vma->vm_file)->i_mapping;
3294 * Take the mapping lock for the duration of the table walk. As
3295 * this mapping should be shared between all the VMAs,
3296 * __unmap_hugepage_range() is called as the lock is already held
3298 i_mmap_lock_write(mapping);
3299 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3300 /* Do not unmap the current VMA */
3301 if (iter_vma == vma)
3305 * Shared VMAs have their own reserves and do not affect
3306 * MAP_PRIVATE accounting but it is possible that a shared
3307 * VMA is using the same page so check and skip such VMAs.
3309 if (iter_vma->vm_flags & VM_MAYSHARE)
3313 * Unmap the page from other VMAs without their own reserves.
3314 * They get marked to be SIGKILLed if they fault in these
3315 * areas. This is because a future no-page fault on this VMA
3316 * could insert a zeroed page instead of the data existing
3317 * from the time of fork. This would look like data corruption
3319 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3320 unmap_hugepage_range(iter_vma, address,
3321 address + huge_page_size(h), page);
3323 i_mmap_unlock_write(mapping);
3327 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3328 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3329 * cannot race with other handlers or page migration.
3330 * Keep the pte_same checks anyway to make transition from the mutex easier.
3332 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3333 unsigned long address, pte_t *ptep, pte_t pte,
3334 struct page *pagecache_page, spinlock_t *ptl)
3336 struct hstate *h = hstate_vma(vma);
3337 struct page *old_page, *new_page;
3338 int ret = 0, outside_reserve = 0;
3339 unsigned long mmun_start; /* For mmu_notifiers */
3340 unsigned long mmun_end; /* For mmu_notifiers */
3342 old_page = pte_page(pte);
3345 /* If no-one else is actually using this page, avoid the copy
3346 * and just make the page writable */
3347 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3348 page_move_anon_rmap(old_page, vma, address);
3349 set_huge_ptep_writable(vma, address, ptep);
3354 * If the process that created a MAP_PRIVATE mapping is about to
3355 * perform a COW due to a shared page count, attempt to satisfy
3356 * the allocation without using the existing reserves. The pagecache
3357 * page is used to determine if the reserve at this address was
3358 * consumed or not. If reserves were used, a partial faulted mapping
3359 * at the time of fork() could consume its reserves on COW instead
3360 * of the full address range.
3362 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3363 old_page != pagecache_page)
3364 outside_reserve = 1;
3366 page_cache_get(old_page);
3369 * Drop page table lock as buddy allocator may be called. It will
3370 * be acquired again before returning to the caller, as expected.
3373 new_page = alloc_huge_page(vma, address, outside_reserve);
3375 if (IS_ERR(new_page)) {
3377 * If a process owning a MAP_PRIVATE mapping fails to COW,
3378 * it is due to references held by a child and an insufficient
3379 * huge page pool. To guarantee the original mappers
3380 * reliability, unmap the page from child processes. The child
3381 * may get SIGKILLed if it later faults.
3383 if (outside_reserve) {
3384 page_cache_release(old_page);
3385 BUG_ON(huge_pte_none(pte));
3386 unmap_ref_private(mm, vma, old_page, address);
3387 BUG_ON(huge_pte_none(pte));
3389 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3391 pte_same(huge_ptep_get(ptep), pte)))
3392 goto retry_avoidcopy;
3394 * race occurs while re-acquiring page table
3395 * lock, and our job is done.
3400 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3401 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3402 goto out_release_old;
3406 * When the original hugepage is shared one, it does not have
3407 * anon_vma prepared.
3409 if (unlikely(anon_vma_prepare(vma))) {
3411 goto out_release_all;
3414 copy_user_huge_page(new_page, old_page, address, vma,
3415 pages_per_huge_page(h));
3416 __SetPageUptodate(new_page);
3417 set_page_huge_active(new_page);
3419 mmun_start = address & huge_page_mask(h);
3420 mmun_end = mmun_start + huge_page_size(h);
3421 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3424 * Retake the page table lock to check for racing updates
3425 * before the page tables are altered
3428 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3429 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3430 ClearPagePrivate(new_page);
3433 huge_ptep_clear_flush(vma, address, ptep);
3434 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3435 set_huge_pte_at(mm, address, ptep,
3436 make_huge_pte(vma, new_page, 1));
3437 page_remove_rmap(old_page, true);
3438 hugepage_add_new_anon_rmap(new_page, vma, address);
3439 /* Make the old page be freed below */
3440 new_page = old_page;
3443 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3445 page_cache_release(new_page);
3447 page_cache_release(old_page);
3449 spin_lock(ptl); /* Caller expects lock to be held */
3453 /* Return the pagecache page at a given address within a VMA */
3454 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3455 struct vm_area_struct *vma, unsigned long address)
3457 struct address_space *mapping;
3460 mapping = vma->vm_file->f_mapping;
3461 idx = vma_hugecache_offset(h, vma, address);
3463 return find_lock_page(mapping, idx);
3467 * Return whether there is a pagecache page to back given address within VMA.
3468 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3470 static bool hugetlbfs_pagecache_present(struct hstate *h,
3471 struct vm_area_struct *vma, unsigned long address)
3473 struct address_space *mapping;
3477 mapping = vma->vm_file->f_mapping;
3478 idx = vma_hugecache_offset(h, vma, address);
3480 page = find_get_page(mapping, idx);
3483 return page != NULL;
3486 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3489 struct inode *inode = mapping->host;
3490 struct hstate *h = hstate_inode(inode);
3491 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3495 ClearPagePrivate(page);
3497 spin_lock(&inode->i_lock);
3498 inode->i_blocks += blocks_per_huge_page(h);
3499 spin_unlock(&inode->i_lock);
3503 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3504 struct address_space *mapping, pgoff_t idx,
3505 unsigned long address, pte_t *ptep, unsigned int flags)
3507 struct hstate *h = hstate_vma(vma);
3508 int ret = VM_FAULT_SIGBUS;
3516 * Currently, we are forced to kill the process in the event the
3517 * original mapper has unmapped pages from the child due to a failed
3518 * COW. Warn that such a situation has occurred as it may not be obvious
3520 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3521 pr_warning("PID %d killed due to inadequate hugepage pool\n",
3527 * Use page lock to guard against racing truncation
3528 * before we get page_table_lock.
3531 page = find_lock_page(mapping, idx);
3533 size = i_size_read(mapping->host) >> huge_page_shift(h);
3536 page = alloc_huge_page(vma, address, 0);
3538 ret = PTR_ERR(page);
3542 ret = VM_FAULT_SIGBUS;
3545 clear_huge_page(page, address, pages_per_huge_page(h));
3546 __SetPageUptodate(page);
3547 set_page_huge_active(page);
3549 if (vma->vm_flags & VM_MAYSHARE) {
3550 int err = huge_add_to_page_cache(page, mapping, idx);
3559 if (unlikely(anon_vma_prepare(vma))) {
3561 goto backout_unlocked;
3567 * If memory error occurs between mmap() and fault, some process
3568 * don't have hwpoisoned swap entry for errored virtual address.
3569 * So we need to block hugepage fault by PG_hwpoison bit check.
3571 if (unlikely(PageHWPoison(page))) {
3572 ret = VM_FAULT_HWPOISON |
3573 VM_FAULT_SET_HINDEX(hstate_index(h));
3574 goto backout_unlocked;
3579 * If we are going to COW a private mapping later, we examine the
3580 * pending reservations for this page now. This will ensure that
3581 * any allocations necessary to record that reservation occur outside
3584 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3585 if (vma_needs_reservation(h, vma, address) < 0) {
3587 goto backout_unlocked;
3589 /* Just decrements count, does not deallocate */
3590 vma_end_reservation(h, vma, address);
3593 ptl = huge_pte_lockptr(h, mm, ptep);
3595 size = i_size_read(mapping->host) >> huge_page_shift(h);
3600 if (!huge_pte_none(huge_ptep_get(ptep)))
3604 ClearPagePrivate(page);
3605 hugepage_add_new_anon_rmap(page, vma, address);
3607 page_dup_rmap(page, true);
3608 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3609 && (vma->vm_flags & VM_SHARED)));
3610 set_huge_pte_at(mm, address, ptep, new_pte);
3612 hugetlb_count_add(pages_per_huge_page(h), mm);
3613 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3614 /* Optimization, do the COW without a second fault */
3615 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
3632 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3633 struct vm_area_struct *vma,
3634 struct address_space *mapping,
3635 pgoff_t idx, unsigned long address)
3637 unsigned long key[2];
3640 if (vma->vm_flags & VM_SHARED) {
3641 key[0] = (unsigned long) mapping;
3644 key[0] = (unsigned long) mm;
3645 key[1] = address >> huge_page_shift(h);
3648 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3650 return hash & (num_fault_mutexes - 1);
3654 * For uniprocesor systems we always use a single mutex, so just
3655 * return 0 and avoid the hashing overhead.
3657 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3658 struct vm_area_struct *vma,
3659 struct address_space *mapping,
3660 pgoff_t idx, unsigned long address)
3666 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3667 unsigned long address, unsigned int flags)
3674 struct page *page = NULL;
3675 struct page *pagecache_page = NULL;
3676 struct hstate *h = hstate_vma(vma);
3677 struct address_space *mapping;
3678 struct inode *inode = file_inode(vma->vm_file);
3679 int need_wait_lock = 0;
3681 address &= huge_page_mask(h);
3683 ptep = huge_pte_offset(mm, address);
3685 entry = huge_ptep_get(ptep);
3686 if (unlikely(is_hugetlb_entry_migration(entry))) {
3687 migration_entry_wait_huge(vma, mm, ptep);
3689 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3690 return VM_FAULT_HWPOISON_LARGE |
3691 VM_FAULT_SET_HINDEX(hstate_index(h));
3694 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3696 return VM_FAULT_OOM;
3698 mapping = vma->vm_file->f_mapping;
3699 idx = vma_hugecache_offset(h, vma, address);
3702 * page faults could race with fallocate hole punch. If a page
3703 * is faulted between unmap and deallocation, it will still remain
3704 * in the punched hole. During hole punch operations, a hugetlb_falloc
3705 * structure will be pointed to by i_private. If this fault is for
3706 * a page in a hole being punched, wait for the operation to finish
3707 * before proceeding.
3709 * Even with this strategy, it is still possible for a page fault to
3710 * race with hole punch. In this case, remove_inode_hugepages() will
3711 * unmap the page and then remove. Checking i_private as below should
3712 * catch most of these races as we want to minimize unmapping a page
3715 if (unlikely(inode->i_private)) {
3716 struct hugetlb_falloc *hugetlb_falloc;
3718 spin_lock(&inode->i_lock);
3719 hugetlb_falloc = inode->i_private;
3720 if (hugetlb_falloc && hugetlb_falloc->waitq &&
3721 idx >= hugetlb_falloc->start &&
3722 idx <= hugetlb_falloc->end) {
3723 wait_queue_head_t *hugetlb_falloc_waitq;
3724 DEFINE_WAIT(hugetlb_fault_wait);
3726 hugetlb_falloc_waitq = hugetlb_falloc->waitq;
3727 prepare_to_wait(hugetlb_falloc_waitq,
3728 &hugetlb_fault_wait,
3729 TASK_UNINTERRUPTIBLE);
3730 spin_unlock(&inode->i_lock);
3733 spin_lock(&inode->i_lock);
3734 finish_wait(hugetlb_falloc_waitq, &hugetlb_fault_wait);
3736 spin_unlock(&inode->i_lock);
3740 * Serialize hugepage allocation and instantiation, so that we don't
3741 * get spurious allocation failures if two CPUs race to instantiate
3742 * the same page in the page cache.
3744 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3745 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3747 entry = huge_ptep_get(ptep);
3748 if (huge_pte_none(entry)) {
3749 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3756 * entry could be a migration/hwpoison entry at this point, so this
3757 * check prevents the kernel from going below assuming that we have
3758 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3759 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3762 if (!pte_present(entry))
3766 * If we are going to COW the mapping later, we examine the pending
3767 * reservations for this page now. This will ensure that any
3768 * allocations necessary to record that reservation occur outside the
3769 * spinlock. For private mappings, we also lookup the pagecache
3770 * page now as it is used to determine if a reservation has been
3773 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3774 if (vma_needs_reservation(h, vma, address) < 0) {
3778 /* Just decrements count, does not deallocate */
3779 vma_end_reservation(h, vma, address);
3781 if (!(vma->vm_flags & VM_MAYSHARE))
3782 pagecache_page = hugetlbfs_pagecache_page(h,
3786 ptl = huge_pte_lock(h, mm, ptep);
3788 /* Check for a racing update before calling hugetlb_cow */
3789 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3793 * hugetlb_cow() requires page locks of pte_page(entry) and
3794 * pagecache_page, so here we need take the former one
3795 * when page != pagecache_page or !pagecache_page.
3797 page = pte_page(entry);
3798 if (page != pagecache_page)
3799 if (!trylock_page(page)) {
3806 if (flags & FAULT_FLAG_WRITE) {
3807 if (!huge_pte_write(entry)) {
3808 ret = hugetlb_cow(mm, vma, address, ptep, entry,
3809 pagecache_page, ptl);
3812 entry = huge_pte_mkdirty(entry);
3814 entry = pte_mkyoung(entry);
3815 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3816 flags & FAULT_FLAG_WRITE))
3817 update_mmu_cache(vma, address, ptep);
3819 if (page != pagecache_page)
3825 if (pagecache_page) {
3826 unlock_page(pagecache_page);
3827 put_page(pagecache_page);
3830 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3832 * Generally it's safe to hold refcount during waiting page lock. But
3833 * here we just wait to defer the next page fault to avoid busy loop and
3834 * the page is not used after unlocked before returning from the current
3835 * page fault. So we are safe from accessing freed page, even if we wait
3836 * here without taking refcount.
3839 wait_on_page_locked(page);
3843 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3844 struct page **pages, struct vm_area_struct **vmas,
3845 unsigned long *position, unsigned long *nr_pages,
3846 long i, unsigned int flags)
3848 unsigned long pfn_offset;
3849 unsigned long vaddr = *position;
3850 unsigned long remainder = *nr_pages;
3851 struct hstate *h = hstate_vma(vma);
3853 while (vaddr < vma->vm_end && remainder) {
3855 spinlock_t *ptl = NULL;
3860 * If we have a pending SIGKILL, don't keep faulting pages and
3861 * potentially allocating memory.
3863 if (unlikely(fatal_signal_pending(current))) {
3869 * Some archs (sparc64, sh*) have multiple pte_ts to
3870 * each hugepage. We have to make sure we get the
3871 * first, for the page indexing below to work.
3873 * Note that page table lock is not held when pte is null.
3875 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3877 ptl = huge_pte_lock(h, mm, pte);
3878 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3881 * When coredumping, it suits get_dump_page if we just return
3882 * an error where there's an empty slot with no huge pagecache
3883 * to back it. This way, we avoid allocating a hugepage, and
3884 * the sparse dumpfile avoids allocating disk blocks, but its
3885 * huge holes still show up with zeroes where they need to be.
3887 if (absent && (flags & FOLL_DUMP) &&
3888 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3896 * We need call hugetlb_fault for both hugepages under migration
3897 * (in which case hugetlb_fault waits for the migration,) and
3898 * hwpoisoned hugepages (in which case we need to prevent the
3899 * caller from accessing to them.) In order to do this, we use
3900 * here is_swap_pte instead of is_hugetlb_entry_migration and
3901 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3902 * both cases, and because we can't follow correct pages
3903 * directly from any kind of swap entries.
3905 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3906 ((flags & FOLL_WRITE) &&
3907 !huge_pte_write(huge_ptep_get(pte)))) {
3912 ret = hugetlb_fault(mm, vma, vaddr,
3913 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3914 if (!(ret & VM_FAULT_ERROR))
3921 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3922 page = pte_page(huge_ptep_get(pte));
3925 pages[i] = mem_map_offset(page, pfn_offset);
3936 if (vaddr < vma->vm_end && remainder &&
3937 pfn_offset < pages_per_huge_page(h)) {
3939 * We use pfn_offset to avoid touching the pageframes
3940 * of this compound page.
3946 *nr_pages = remainder;
3949 return i ? i : -EFAULT;
3952 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3953 unsigned long address, unsigned long end, pgprot_t newprot)
3955 struct mm_struct *mm = vma->vm_mm;
3956 unsigned long start = address;
3959 struct hstate *h = hstate_vma(vma);
3960 unsigned long pages = 0;
3962 BUG_ON(address >= end);
3963 flush_cache_range(vma, address, end);
3965 mmu_notifier_invalidate_range_start(mm, start, end);
3966 i_mmap_lock_write(vma->vm_file->f_mapping);
3967 for (; address < end; address += huge_page_size(h)) {
3969 ptep = huge_pte_offset(mm, address);
3972 ptl = huge_pte_lock(h, mm, ptep);
3973 if (huge_pmd_unshare(mm, &address, ptep)) {
3978 pte = huge_ptep_get(ptep);
3979 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3983 if (unlikely(is_hugetlb_entry_migration(pte))) {
3984 swp_entry_t entry = pte_to_swp_entry(pte);
3986 if (is_write_migration_entry(entry)) {
3989 make_migration_entry_read(&entry);
3990 newpte = swp_entry_to_pte(entry);
3991 set_huge_pte_at(mm, address, ptep, newpte);
3997 if (!huge_pte_none(pte)) {
3998 pte = huge_ptep_get_and_clear(mm, address, ptep);
3999 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
4000 pte = arch_make_huge_pte(pte, vma, NULL, 0);
4001 set_huge_pte_at(mm, address, ptep, pte);
4007 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4008 * may have cleared our pud entry and done put_page on the page table:
4009 * once we release i_mmap_rwsem, another task can do the final put_page
4010 * and that page table be reused and filled with junk.
4012 flush_tlb_range(vma, start, end);
4013 mmu_notifier_invalidate_range(mm, start, end);
4014 i_mmap_unlock_write(vma->vm_file->f_mapping);
4015 mmu_notifier_invalidate_range_end(mm, start, end);
4017 return pages << h->order;
4020 int hugetlb_reserve_pages(struct inode *inode,
4022 struct vm_area_struct *vma,
4023 vm_flags_t vm_flags)
4026 struct hstate *h = hstate_inode(inode);
4027 struct hugepage_subpool *spool = subpool_inode(inode);
4028 struct resv_map *resv_map;
4032 * Only apply hugepage reservation if asked. At fault time, an
4033 * attempt will be made for VM_NORESERVE to allocate a page
4034 * without using reserves
4036 if (vm_flags & VM_NORESERVE)
4040 * Shared mappings base their reservation on the number of pages that
4041 * are already allocated on behalf of the file. Private mappings need
4042 * to reserve the full area even if read-only as mprotect() may be
4043 * called to make the mapping read-write. Assume !vma is a shm mapping
4045 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4046 resv_map = inode_resv_map(inode);
4048 chg = region_chg(resv_map, from, to);
4051 resv_map = resv_map_alloc();
4057 set_vma_resv_map(vma, resv_map);
4058 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4067 * There must be enough pages in the subpool for the mapping. If
4068 * the subpool has a minimum size, there may be some global
4069 * reservations already in place (gbl_reserve).
4071 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4072 if (gbl_reserve < 0) {
4078 * Check enough hugepages are available for the reservation.
4079 * Hand the pages back to the subpool if there are not
4081 ret = hugetlb_acct_memory(h, gbl_reserve);
4083 /* put back original number of pages, chg */
4084 (void)hugepage_subpool_put_pages(spool, chg);
4089 * Account for the reservations made. Shared mappings record regions
4090 * that have reservations as they are shared by multiple VMAs.
4091 * When the last VMA disappears, the region map says how much
4092 * the reservation was and the page cache tells how much of
4093 * the reservation was consumed. Private mappings are per-VMA and
4094 * only the consumed reservations are tracked. When the VMA
4095 * disappears, the original reservation is the VMA size and the
4096 * consumed reservations are stored in the map. Hence, nothing
4097 * else has to be done for private mappings here
4099 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4100 long add = region_add(resv_map, from, to);
4102 if (unlikely(chg > add)) {
4104 * pages in this range were added to the reserve
4105 * map between region_chg and region_add. This
4106 * indicates a race with alloc_huge_page. Adjust
4107 * the subpool and reserve counts modified above
4108 * based on the difference.
4112 rsv_adjust = hugepage_subpool_put_pages(spool,
4114 hugetlb_acct_memory(h, -rsv_adjust);
4119 if (!vma || vma->vm_flags & VM_MAYSHARE)
4120 region_abort(resv_map, from, to);
4121 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4122 kref_put(&resv_map->refs, resv_map_release);
4126 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4129 struct hstate *h = hstate_inode(inode);
4130 struct resv_map *resv_map = inode_resv_map(inode);
4132 struct hugepage_subpool *spool = subpool_inode(inode);
4136 chg = region_del(resv_map, start, end);
4138 * region_del() can fail in the rare case where a region
4139 * must be split and another region descriptor can not be
4140 * allocated. If end == LONG_MAX, it will not fail.
4146 spin_lock(&inode->i_lock);
4147 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4148 spin_unlock(&inode->i_lock);
4151 * If the subpool has a minimum size, the number of global
4152 * reservations to be released may be adjusted.
4154 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4155 hugetlb_acct_memory(h, -gbl_reserve);
4160 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4161 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4162 struct vm_area_struct *vma,
4163 unsigned long addr, pgoff_t idx)
4165 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4167 unsigned long sbase = saddr & PUD_MASK;
4168 unsigned long s_end = sbase + PUD_SIZE;
4170 /* Allow segments to share if only one is marked locked */
4171 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4172 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4175 * match the virtual addresses, permission and the alignment of the
4178 if (pmd_index(addr) != pmd_index(saddr) ||
4179 vm_flags != svm_flags ||
4180 sbase < svma->vm_start || svma->vm_end < s_end)
4186 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4188 unsigned long base = addr & PUD_MASK;
4189 unsigned long end = base + PUD_SIZE;
4192 * check on proper vm_flags and page table alignment
4194 if (vma->vm_flags & VM_MAYSHARE &&
4195 vma->vm_start <= base && end <= vma->vm_end)
4201 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4202 * and returns the corresponding pte. While this is not necessary for the
4203 * !shared pmd case because we can allocate the pmd later as well, it makes the
4204 * code much cleaner. pmd allocation is essential for the shared case because
4205 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4206 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4207 * bad pmd for sharing.
4209 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4211 struct vm_area_struct *vma = find_vma(mm, addr);
4212 struct address_space *mapping = vma->vm_file->f_mapping;
4213 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4215 struct vm_area_struct *svma;
4216 unsigned long saddr;
4221 if (!vma_shareable(vma, addr))
4222 return (pte_t *)pmd_alloc(mm, pud, addr);
4224 i_mmap_lock_write(mapping);
4225 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4229 saddr = page_table_shareable(svma, vma, addr, idx);
4231 spte = huge_pte_offset(svma->vm_mm, saddr);
4234 get_page(virt_to_page(spte));
4243 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
4245 if (pud_none(*pud)) {
4246 pud_populate(mm, pud,
4247 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4249 put_page(virt_to_page(spte));
4254 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4255 i_mmap_unlock_write(mapping);
4260 * unmap huge page backed by shared pte.
4262 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4263 * indicated by page_count > 1, unmap is achieved by clearing pud and
4264 * decrementing the ref count. If count == 1, the pte page is not shared.
4266 * called with page table lock held.
4268 * returns: 1 successfully unmapped a shared pte page
4269 * 0 the underlying pte page is not shared, or it is the last user
4271 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4273 pgd_t *pgd = pgd_offset(mm, *addr);
4274 pud_t *pud = pud_offset(pgd, *addr);
4276 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4277 if (page_count(virt_to_page(ptep)) == 1)
4281 put_page(virt_to_page(ptep));
4283 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4286 #define want_pmd_share() (1)
4287 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4288 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4293 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4297 #define want_pmd_share() (0)
4298 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4300 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4301 pte_t *huge_pte_alloc(struct mm_struct *mm,
4302 unsigned long addr, unsigned long sz)
4308 pgd = pgd_offset(mm, addr);
4309 pud = pud_alloc(mm, pgd, addr);
4311 if (sz == PUD_SIZE) {
4314 BUG_ON(sz != PMD_SIZE);
4315 if (want_pmd_share() && pud_none(*pud))
4316 pte = huge_pmd_share(mm, addr, pud);
4318 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4321 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
4326 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
4332 pgd = pgd_offset(mm, addr);
4333 if (pgd_present(*pgd)) {
4334 pud = pud_offset(pgd, addr);
4335 if (pud_present(*pud)) {
4337 return (pte_t *)pud;
4338 pmd = pmd_offset(pud, addr);
4341 return (pte_t *) pmd;
4344 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4347 * These functions are overwritable if your architecture needs its own
4350 struct page * __weak
4351 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4354 return ERR_PTR(-EINVAL);
4357 struct page * __weak
4358 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4359 pmd_t *pmd, int flags)
4361 struct page *page = NULL;
4364 ptl = pmd_lockptr(mm, pmd);
4367 * make sure that the address range covered by this pmd is not
4368 * unmapped from other threads.
4370 if (!pmd_huge(*pmd))
4372 if (pmd_present(*pmd)) {
4373 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4374 if (flags & FOLL_GET)
4377 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
4379 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4383 * hwpoisoned entry is treated as no_page_table in
4384 * follow_page_mask().
4392 struct page * __weak
4393 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4394 pud_t *pud, int flags)
4396 if (flags & FOLL_GET)
4399 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4402 #ifdef CONFIG_MEMORY_FAILURE
4405 * This function is called from memory failure code.
4406 * Assume the caller holds page lock of the head page.
4408 int dequeue_hwpoisoned_huge_page(struct page *hpage)
4410 struct hstate *h = page_hstate(hpage);
4411 int nid = page_to_nid(hpage);
4414 spin_lock(&hugetlb_lock);
4416 * Just checking !page_huge_active is not enough, because that could be
4417 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4419 if (!page_huge_active(hpage) && !page_count(hpage)) {
4421 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4422 * but dangling hpage->lru can trigger list-debug warnings
4423 * (this happens when we call unpoison_memory() on it),
4424 * so let it point to itself with list_del_init().
4426 list_del_init(&hpage->lru);
4427 set_page_refcounted(hpage);
4428 h->free_huge_pages--;
4429 h->free_huge_pages_node[nid]--;
4432 spin_unlock(&hugetlb_lock);
4437 bool isolate_huge_page(struct page *page, struct list_head *list)
4441 VM_BUG_ON_PAGE(!PageHead(page), page);
4442 spin_lock(&hugetlb_lock);
4443 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4447 clear_page_huge_active(page);
4448 list_move_tail(&page->lru, list);
4450 spin_unlock(&hugetlb_lock);
4454 void putback_active_hugepage(struct page *page)
4456 VM_BUG_ON_PAGE(!PageHead(page), page);
4457 spin_lock(&hugetlb_lock);
4458 set_page_huge_active(page);
4459 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4460 spin_unlock(&hugetlb_lock);