1 // SPDX-License-Identifier: GPL-2.0-only
3 * Generic hugetlb support.
4 * (C) Nadia Yvette Chambers, April 2004
6 #include <linux/list.h>
7 #include <linux/init.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/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/mmdebug.h>
23 #include <linux/sched/signal.h>
24 #include <linux/rmap.h>
25 #include <linux/string_helpers.h>
26 #include <linux/swap.h>
27 #include <linux/swapops.h>
28 #include <linux/jhash.h>
29 #include <linux/numa.h>
30 #include <linux/llist.h>
33 #include <asm/pgtable.h>
37 #include <linux/hugetlb.h>
38 #include <linux/hugetlb_cgroup.h>
39 #include <linux/node.h>
40 #include <linux/userfaultfd_k.h>
41 #include <linux/page_owner.h>
44 int hugetlb_max_hstate __read_mostly;
45 unsigned int default_hstate_idx;
46 struct hstate hstates[HUGE_MAX_HSTATE];
48 * Minimum page order among possible hugepage sizes, set to a proper value
51 static unsigned int minimum_order __read_mostly = UINT_MAX;
53 __initdata LIST_HEAD(huge_boot_pages);
55 /* for command line parsing */
56 static struct hstate * __initdata parsed_hstate;
57 static unsigned long __initdata default_hstate_max_huge_pages;
58 static unsigned long __initdata default_hstate_size;
59 static bool __initdata parsed_valid_hugepagesz = true;
62 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
63 * free_huge_pages, and surplus_huge_pages.
65 DEFINE_SPINLOCK(hugetlb_lock);
68 * Serializes faults on the same logical page. This is used to
69 * prevent spurious OOMs when the hugepage pool is fully utilized.
71 static int num_fault_mutexes;
72 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
74 static inline bool PageHugeFreed(struct page *head)
76 return page_private(head + 4) == -1UL;
79 static inline void SetPageHugeFreed(struct page *head)
81 set_page_private(head + 4, -1UL);
84 static inline void ClearPageHugeFreed(struct page *head)
86 set_page_private(head + 4, 0);
89 /* Forward declaration */
90 static int hugetlb_acct_memory(struct hstate *h, long delta);
92 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
94 bool free = (spool->count == 0) && (spool->used_hpages == 0);
96 spin_unlock(&spool->lock);
98 /* If no pages are used, and no other handles to the subpool
99 * remain, give up any reservations mased on minimum size and
100 * free the subpool */
102 if (spool->min_hpages != -1)
103 hugetlb_acct_memory(spool->hstate,
109 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
112 struct hugepage_subpool *spool;
114 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
118 spin_lock_init(&spool->lock);
120 spool->max_hpages = max_hpages;
122 spool->min_hpages = min_hpages;
124 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
128 spool->rsv_hpages = min_hpages;
133 void hugepage_put_subpool(struct hugepage_subpool *spool)
135 spin_lock(&spool->lock);
136 BUG_ON(!spool->count);
138 unlock_or_release_subpool(spool);
142 * Subpool accounting for allocating and reserving pages.
143 * Return -ENOMEM if there are not enough resources to satisfy the
144 * the request. Otherwise, return the number of pages by which the
145 * global pools must be adjusted (upward). The returned value may
146 * only be different than the passed value (delta) in the case where
147 * a subpool minimum size must be manitained.
149 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
157 spin_lock(&spool->lock);
159 if (spool->max_hpages != -1) { /* maximum size accounting */
160 if ((spool->used_hpages + delta) <= spool->max_hpages)
161 spool->used_hpages += delta;
168 /* minimum size accounting */
169 if (spool->min_hpages != -1 && spool->rsv_hpages) {
170 if (delta > spool->rsv_hpages) {
172 * Asking for more reserves than those already taken on
173 * behalf of subpool. Return difference.
175 ret = delta - spool->rsv_hpages;
176 spool->rsv_hpages = 0;
178 ret = 0; /* reserves already accounted for */
179 spool->rsv_hpages -= delta;
184 spin_unlock(&spool->lock);
189 * Subpool accounting for freeing and unreserving pages.
190 * Return the number of global page reservations that must be dropped.
191 * The return value may only be different than the passed value (delta)
192 * in the case where a subpool minimum size must be maintained.
194 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
202 spin_lock(&spool->lock);
204 if (spool->max_hpages != -1) /* maximum size accounting */
205 spool->used_hpages -= delta;
207 /* minimum size accounting */
208 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
209 if (spool->rsv_hpages + delta <= spool->min_hpages)
212 ret = spool->rsv_hpages + delta - spool->min_hpages;
214 spool->rsv_hpages += delta;
215 if (spool->rsv_hpages > spool->min_hpages)
216 spool->rsv_hpages = spool->min_hpages;
220 * If hugetlbfs_put_super couldn't free spool due to an outstanding
221 * quota reference, free it now.
223 unlock_or_release_subpool(spool);
228 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
230 return HUGETLBFS_SB(inode->i_sb)->spool;
233 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
235 return subpool_inode(file_inode(vma->vm_file));
239 * Region tracking -- allows tracking of reservations and instantiated pages
240 * across the pages in a mapping.
242 * The region data structures are embedded into a resv_map and protected
243 * by a resv_map's lock. The set of regions within the resv_map represent
244 * reservations for huge pages, or huge pages that have already been
245 * instantiated within the map. The from and to elements are huge page
246 * indicies into the associated mapping. from indicates the starting index
247 * of the region. to represents the first index past the end of the region.
249 * For example, a file region structure with from == 0 and to == 4 represents
250 * four huge pages in a mapping. It is important to note that the to element
251 * represents the first element past the end of the region. This is used in
252 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
254 * Interval notation of the form [from, to) will be used to indicate that
255 * the endpoint from is inclusive and to is exclusive.
258 struct list_head link;
264 * Add the huge page range represented by [f, t) to the reserve
265 * map. In the normal case, existing regions will be expanded
266 * to accommodate the specified range. Sufficient regions should
267 * exist for expansion due to the previous call to region_chg
268 * with the same range. However, it is possible that region_del
269 * could have been called after region_chg and modifed the map
270 * in such a way that no region exists to be expanded. In this
271 * case, pull a region descriptor from the cache associated with
272 * the map and use that for the new range.
274 * Return the number of new huge pages added to the map. This
275 * number is greater than or equal to zero.
277 static long region_add(struct resv_map *resv, long f, long t)
279 struct list_head *head = &resv->regions;
280 struct file_region *rg, *nrg, *trg;
283 spin_lock(&resv->lock);
284 /* Locate the region we are either in or before. */
285 list_for_each_entry(rg, head, link)
290 * If no region exists which can be expanded to include the
291 * specified range, the list must have been modified by an
292 * interleving call to region_del(). Pull a region descriptor
293 * from the cache and use it for this range.
295 if (&rg->link == head || t < rg->from) {
296 VM_BUG_ON(resv->region_cache_count <= 0);
298 resv->region_cache_count--;
299 nrg = list_first_entry(&resv->region_cache, struct file_region,
301 list_del(&nrg->link);
305 list_add(&nrg->link, rg->link.prev);
311 /* Round our left edge to the current segment if it encloses us. */
315 /* Check for and consume any regions we now overlap with. */
317 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
318 if (&rg->link == head)
323 /* If this area reaches higher then extend our area to
324 * include it completely. If this is not the first area
325 * which we intend to reuse, free it. */
329 /* Decrement return value by the deleted range.
330 * Another range will span this area so that by
331 * end of routine add will be >= zero
333 add -= (rg->to - rg->from);
339 add += (nrg->from - f); /* Added to beginning of region */
341 add += t - nrg->to; /* Added to end of region */
345 resv->adds_in_progress--;
346 spin_unlock(&resv->lock);
352 * Examine the existing reserve map and determine how many
353 * huge pages in the specified range [f, t) are NOT currently
354 * represented. This routine is called before a subsequent
355 * call to region_add that will actually modify the reserve
356 * map to add the specified range [f, t). region_chg does
357 * not change the number of huge pages represented by the
358 * map. However, if the existing regions in the map can not
359 * be expanded to represent the new range, a new file_region
360 * structure is added to the map as a placeholder. This is
361 * so that the subsequent region_add call will have all the
362 * regions it needs and will not fail.
364 * Upon entry, region_chg will also examine the cache of region descriptors
365 * associated with the map. If there are not enough descriptors cached, one
366 * will be allocated for the in progress add operation.
368 * Returns the number of huge pages that need to be added to the existing
369 * reservation map for the range [f, t). This number is greater or equal to
370 * zero. -ENOMEM is returned if a new file_region structure or cache entry
371 * is needed and can not be allocated.
373 static long region_chg(struct resv_map *resv, long f, long t)
375 struct list_head *head = &resv->regions;
376 struct file_region *rg, *nrg = NULL;
380 spin_lock(&resv->lock);
382 resv->adds_in_progress++;
385 * Check for sufficient descriptors in the cache to accommodate
386 * the number of in progress add operations.
388 if (resv->adds_in_progress > resv->region_cache_count) {
389 struct file_region *trg;
391 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
392 /* Must drop lock to allocate a new descriptor. */
393 resv->adds_in_progress--;
394 spin_unlock(&resv->lock);
396 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
402 spin_lock(&resv->lock);
403 list_add(&trg->link, &resv->region_cache);
404 resv->region_cache_count++;
408 /* Locate the region we are before or in. */
409 list_for_each_entry(rg, head, link)
413 /* If we are below the current region then a new region is required.
414 * Subtle, allocate a new region at the position but make it zero
415 * size such that we can guarantee to record the reservation. */
416 if (&rg->link == head || t < rg->from) {
418 resv->adds_in_progress--;
419 spin_unlock(&resv->lock);
420 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
426 INIT_LIST_HEAD(&nrg->link);
430 list_add(&nrg->link, rg->link.prev);
435 /* Round our left edge to the current segment if it encloses us. */
440 /* Check for and consume any regions we now overlap with. */
441 list_for_each_entry(rg, rg->link.prev, link) {
442 if (&rg->link == head)
447 /* We overlap with this area, if it extends further than
448 * us then we must extend ourselves. Account for its
449 * existing reservation. */
454 chg -= rg->to - rg->from;
458 spin_unlock(&resv->lock);
459 /* We already know we raced and no longer need the new region */
463 spin_unlock(&resv->lock);
468 * Abort the in progress add operation. The adds_in_progress field
469 * of the resv_map keeps track of the operations in progress between
470 * calls to region_chg and region_add. Operations are sometimes
471 * aborted after the call to region_chg. In such cases, region_abort
472 * is called to decrement the adds_in_progress counter.
474 * NOTE: The range arguments [f, t) are not needed or used in this
475 * routine. They are kept to make reading the calling code easier as
476 * arguments will match the associated region_chg call.
478 static void region_abort(struct resv_map *resv, long f, long t)
480 spin_lock(&resv->lock);
481 VM_BUG_ON(!resv->region_cache_count);
482 resv->adds_in_progress--;
483 spin_unlock(&resv->lock);
487 * Delete the specified range [f, t) from the reserve map. If the
488 * t parameter is LONG_MAX, this indicates that ALL regions after f
489 * should be deleted. Locate the regions which intersect [f, t)
490 * and either trim, delete or split the existing regions.
492 * Returns the number of huge pages deleted from the reserve map.
493 * In the normal case, the return value is zero or more. In the
494 * case where a region must be split, a new region descriptor must
495 * be allocated. If the allocation fails, -ENOMEM will be returned.
496 * NOTE: If the parameter t == LONG_MAX, then we will never split
497 * a region and possibly return -ENOMEM. Callers specifying
498 * t == LONG_MAX do not need to check for -ENOMEM error.
500 static long region_del(struct resv_map *resv, long f, long t)
502 struct list_head *head = &resv->regions;
503 struct file_region *rg, *trg;
504 struct file_region *nrg = NULL;
508 spin_lock(&resv->lock);
509 list_for_each_entry_safe(rg, trg, head, link) {
511 * Skip regions before the range to be deleted. file_region
512 * ranges are normally of the form [from, to). However, there
513 * may be a "placeholder" entry in the map which is of the form
514 * (from, to) with from == to. Check for placeholder entries
515 * at the beginning of the range to be deleted.
517 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
523 if (f > rg->from && t < rg->to) { /* Must split region */
525 * Check for an entry in the cache before dropping
526 * lock and attempting allocation.
529 resv->region_cache_count > resv->adds_in_progress) {
530 nrg = list_first_entry(&resv->region_cache,
533 list_del(&nrg->link);
534 resv->region_cache_count--;
538 spin_unlock(&resv->lock);
539 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
547 /* New entry for end of split region */
550 INIT_LIST_HEAD(&nrg->link);
552 /* Original entry is trimmed */
555 list_add(&nrg->link, &rg->link);
560 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
561 del += rg->to - rg->from;
567 if (f <= rg->from) { /* Trim beginning of region */
570 } else { /* Trim end of region */
576 spin_unlock(&resv->lock);
582 * A rare out of memory error was encountered which prevented removal of
583 * the reserve map region for a page. The huge page itself was free'ed
584 * and removed from the page cache. This routine will adjust the subpool
585 * usage count, and the global reserve count if needed. By incrementing
586 * these counts, the reserve map entry which could not be deleted will
587 * appear as a "reserved" entry instead of simply dangling with incorrect
590 void hugetlb_fix_reserve_counts(struct inode *inode)
592 struct hugepage_subpool *spool = subpool_inode(inode);
594 bool reserved = false;
596 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
597 if (rsv_adjust > 0) {
598 struct hstate *h = hstate_inode(inode);
600 if (!hugetlb_acct_memory(h, 1))
602 } else if (!rsv_adjust) {
607 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
611 * Count and return the number of huge pages in the reserve map
612 * that intersect with the range [f, t).
614 static long region_count(struct resv_map *resv, long f, long t)
616 struct list_head *head = &resv->regions;
617 struct file_region *rg;
620 spin_lock(&resv->lock);
621 /* Locate each segment we overlap with, and count that overlap. */
622 list_for_each_entry(rg, head, link) {
631 seg_from = max(rg->from, f);
632 seg_to = min(rg->to, t);
634 chg += seg_to - seg_from;
636 spin_unlock(&resv->lock);
642 * Convert the address within this vma to the page offset within
643 * the mapping, in pagecache page units; huge pages here.
645 static pgoff_t vma_hugecache_offset(struct hstate *h,
646 struct vm_area_struct *vma, unsigned long address)
648 return ((address - vma->vm_start) >> huge_page_shift(h)) +
649 (vma->vm_pgoff >> huge_page_order(h));
652 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
653 unsigned long address)
655 return vma_hugecache_offset(hstate_vma(vma), vma, address);
657 EXPORT_SYMBOL_GPL(linear_hugepage_index);
660 * Return the size of the pages allocated when backing a VMA. In the majority
661 * cases this will be same size as used by the page table entries.
663 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
665 if (vma->vm_ops && vma->vm_ops->pagesize)
666 return vma->vm_ops->pagesize(vma);
669 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
672 * Return the page size being used by the MMU to back a VMA. In the majority
673 * of cases, the page size used by the kernel matches the MMU size. On
674 * architectures where it differs, an architecture-specific 'strong'
675 * version of this symbol is required.
677 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
679 return vma_kernel_pagesize(vma);
683 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
684 * bits of the reservation map pointer, which are always clear due to
687 #define HPAGE_RESV_OWNER (1UL << 0)
688 #define HPAGE_RESV_UNMAPPED (1UL << 1)
689 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
692 * These helpers are used to track how many pages are reserved for
693 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
694 * is guaranteed to have their future faults succeed.
696 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
697 * the reserve counters are updated with the hugetlb_lock held. It is safe
698 * to reset the VMA at fork() time as it is not in use yet and there is no
699 * chance of the global counters getting corrupted as a result of the values.
701 * The private mapping reservation is represented in a subtly different
702 * manner to a shared mapping. A shared mapping has a region map associated
703 * with the underlying file, this region map represents the backing file
704 * pages which have ever had a reservation assigned which this persists even
705 * after the page is instantiated. A private mapping has a region map
706 * associated with the original mmap which is attached to all VMAs which
707 * reference it, this region map represents those offsets which have consumed
708 * reservation ie. where pages have been instantiated.
710 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
712 return (unsigned long)vma->vm_private_data;
715 static void set_vma_private_data(struct vm_area_struct *vma,
718 vma->vm_private_data = (void *)value;
721 struct resv_map *resv_map_alloc(void)
723 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
724 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
726 if (!resv_map || !rg) {
732 kref_init(&resv_map->refs);
733 spin_lock_init(&resv_map->lock);
734 INIT_LIST_HEAD(&resv_map->regions);
736 resv_map->adds_in_progress = 0;
738 INIT_LIST_HEAD(&resv_map->region_cache);
739 list_add(&rg->link, &resv_map->region_cache);
740 resv_map->region_cache_count = 1;
745 void resv_map_release(struct kref *ref)
747 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
748 struct list_head *head = &resv_map->region_cache;
749 struct file_region *rg, *trg;
751 /* Clear out any active regions before we release the map. */
752 region_del(resv_map, 0, LONG_MAX);
754 /* ... and any entries left in the cache */
755 list_for_each_entry_safe(rg, trg, head, link) {
760 VM_BUG_ON(resv_map->adds_in_progress);
765 static inline struct resv_map *inode_resv_map(struct inode *inode)
768 * At inode evict time, i_mapping may not point to the original
769 * address space within the inode. This original address space
770 * contains the pointer to the resv_map. So, always use the
771 * address space embedded within the inode.
772 * The VERY common case is inode->mapping == &inode->i_data but,
773 * this may not be true for device special inodes.
775 return (struct resv_map *)(&inode->i_data)->private_data;
778 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
780 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
781 if (vma->vm_flags & VM_MAYSHARE) {
782 struct address_space *mapping = vma->vm_file->f_mapping;
783 struct inode *inode = mapping->host;
785 return inode_resv_map(inode);
788 return (struct resv_map *)(get_vma_private_data(vma) &
793 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
795 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
796 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
798 set_vma_private_data(vma, (get_vma_private_data(vma) &
799 HPAGE_RESV_MASK) | (unsigned long)map);
802 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
804 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
805 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
807 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
810 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
812 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
814 return (get_vma_private_data(vma) & flag) != 0;
817 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
818 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
820 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
821 if (!(vma->vm_flags & VM_MAYSHARE))
822 vma->vm_private_data = (void *)0;
825 /* Returns true if the VMA has associated reserve pages */
826 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
828 if (vma->vm_flags & VM_NORESERVE) {
830 * This address is already reserved by other process(chg == 0),
831 * so, we should decrement reserved count. Without decrementing,
832 * reserve count remains after releasing inode, because this
833 * allocated page will go into page cache and is regarded as
834 * coming from reserved pool in releasing step. Currently, we
835 * don't have any other solution to deal with this situation
836 * properly, so add work-around here.
838 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
844 /* Shared mappings always use reserves */
845 if (vma->vm_flags & VM_MAYSHARE) {
847 * We know VM_NORESERVE is not set. Therefore, there SHOULD
848 * be a region map for all pages. The only situation where
849 * there is no region map is if a hole was punched via
850 * fallocate. In this case, there really are no reverves to
851 * use. This situation is indicated if chg != 0.
860 * Only the process that called mmap() has reserves for
863 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
865 * Like the shared case above, a hole punch or truncate
866 * could have been performed on the private mapping.
867 * Examine the value of chg to determine if reserves
868 * actually exist or were previously consumed.
869 * Very Subtle - The value of chg comes from a previous
870 * call to vma_needs_reserves(). The reserve map for
871 * private mappings has different (opposite) semantics
872 * than that of shared mappings. vma_needs_reserves()
873 * has already taken this difference in semantics into
874 * account. Therefore, the meaning of chg is the same
875 * as in the shared case above. Code could easily be
876 * combined, but keeping it separate draws attention to
877 * subtle differences.
888 static void enqueue_huge_page(struct hstate *h, struct page *page)
890 int nid = page_to_nid(page);
891 list_move(&page->lru, &h->hugepage_freelists[nid]);
892 h->free_huge_pages++;
893 h->free_huge_pages_node[nid]++;
894 SetPageHugeFreed(page);
897 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
901 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
902 if (!PageHWPoison(page))
905 * if 'non-isolated free hugepage' not found on the list,
906 * the allocation fails.
908 if (&h->hugepage_freelists[nid] == &page->lru)
910 list_move(&page->lru, &h->hugepage_activelist);
911 set_page_refcounted(page);
912 ClearPageHugeFreed(page);
913 h->free_huge_pages--;
914 h->free_huge_pages_node[nid]--;
918 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
921 unsigned int cpuset_mems_cookie;
922 struct zonelist *zonelist;
925 int node = NUMA_NO_NODE;
927 zonelist = node_zonelist(nid, gfp_mask);
930 cpuset_mems_cookie = read_mems_allowed_begin();
931 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
934 if (!cpuset_zone_allowed(zone, gfp_mask))
937 * no need to ask again on the same node. Pool is node rather than
940 if (zone_to_nid(zone) == node)
942 node = zone_to_nid(zone);
944 page = dequeue_huge_page_node_exact(h, node);
948 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
954 /* Movability of hugepages depends on migration support. */
955 static inline gfp_t htlb_alloc_mask(struct hstate *h)
957 if (hugepage_movable_supported(h))
958 return GFP_HIGHUSER_MOVABLE;
963 static struct page *dequeue_huge_page_vma(struct hstate *h,
964 struct vm_area_struct *vma,
965 unsigned long address, int avoid_reserve,
969 struct mempolicy *mpol;
971 nodemask_t *nodemask;
975 * A child process with MAP_PRIVATE mappings created by their parent
976 * have no page reserves. This check ensures that reservations are
977 * not "stolen". The child may still get SIGKILLed
979 if (!vma_has_reserves(vma, chg) &&
980 h->free_huge_pages - h->resv_huge_pages == 0)
983 /* If reserves cannot be used, ensure enough pages are in the pool */
984 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
987 gfp_mask = htlb_alloc_mask(h);
988 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
989 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
990 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
991 SetPagePrivate(page);
992 h->resv_huge_pages--;
1003 * common helper functions for hstate_next_node_to_{alloc|free}.
1004 * We may have allocated or freed a huge page based on a different
1005 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1006 * be outside of *nodes_allowed. Ensure that we use an allowed
1007 * node for alloc or free.
1009 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1011 nid = next_node_in(nid, *nodes_allowed);
1012 VM_BUG_ON(nid >= MAX_NUMNODES);
1017 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1019 if (!node_isset(nid, *nodes_allowed))
1020 nid = next_node_allowed(nid, nodes_allowed);
1025 * returns the previously saved node ["this node"] from which to
1026 * allocate a persistent huge page for the pool and advance the
1027 * next node from which to allocate, handling wrap at end of node
1030 static int hstate_next_node_to_alloc(struct hstate *h,
1031 nodemask_t *nodes_allowed)
1035 VM_BUG_ON(!nodes_allowed);
1037 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1038 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1044 * helper for free_pool_huge_page() - return the previously saved
1045 * node ["this node"] from which to free a huge page. Advance the
1046 * next node id whether or not we find a free huge page to free so
1047 * that the next attempt to free addresses the next node.
1049 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1053 VM_BUG_ON(!nodes_allowed);
1055 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1056 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1061 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1062 for (nr_nodes = nodes_weight(*mask); \
1064 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1067 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1068 for (nr_nodes = nodes_weight(*mask); \
1070 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1073 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1074 static void destroy_compound_gigantic_page(struct page *page,
1078 int nr_pages = 1 << order;
1079 struct page *p = page + 1;
1081 atomic_set(compound_mapcount_ptr(page), 0);
1082 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1083 clear_compound_head(p);
1084 set_page_refcounted(p);
1087 set_compound_order(page, 0);
1088 __ClearPageHead(page);
1091 static void free_gigantic_page(struct page *page, unsigned int order)
1093 free_contig_range(page_to_pfn(page), 1 << order);
1096 #ifdef CONFIG_CONTIG_ALLOC
1097 static int __alloc_gigantic_page(unsigned long start_pfn,
1098 unsigned long nr_pages, gfp_t gfp_mask)
1100 unsigned long end_pfn = start_pfn + nr_pages;
1101 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1105 static bool pfn_range_valid_gigantic(struct zone *z,
1106 unsigned long start_pfn, unsigned long nr_pages)
1108 unsigned long i, end_pfn = start_pfn + nr_pages;
1111 for (i = start_pfn; i < end_pfn; i++) {
1112 page = pfn_to_online_page(i);
1116 if (page_zone(page) != z)
1119 if (PageReserved(page))
1122 if (page_count(page) > 0)
1132 static bool zone_spans_last_pfn(const struct zone *zone,
1133 unsigned long start_pfn, unsigned long nr_pages)
1135 unsigned long last_pfn = start_pfn + nr_pages - 1;
1136 return zone_spans_pfn(zone, last_pfn);
1139 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1140 int nid, nodemask_t *nodemask)
1142 unsigned int order = huge_page_order(h);
1143 unsigned long nr_pages = 1 << order;
1144 unsigned long ret, pfn, flags;
1145 struct zonelist *zonelist;
1149 zonelist = node_zonelist(nid, gfp_mask);
1150 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) {
1151 spin_lock_irqsave(&zone->lock, flags);
1153 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
1154 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
1155 if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) {
1157 * We release the zone lock here because
1158 * alloc_contig_range() will also lock the zone
1159 * at some point. If there's an allocation
1160 * spinning on this lock, it may win the race
1161 * and cause alloc_contig_range() to fail...
1163 spin_unlock_irqrestore(&zone->lock, flags);
1164 ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask);
1166 return pfn_to_page(pfn);
1167 spin_lock_irqsave(&zone->lock, flags);
1172 spin_unlock_irqrestore(&zone->lock, flags);
1178 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1179 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1180 #else /* !CONFIG_CONTIG_ALLOC */
1181 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1182 int nid, nodemask_t *nodemask)
1186 #endif /* CONFIG_CONTIG_ALLOC */
1188 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1189 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1190 int nid, nodemask_t *nodemask)
1194 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1195 static inline void destroy_compound_gigantic_page(struct page *page,
1196 unsigned int order) { }
1199 static void update_and_free_page(struct hstate *h, struct page *page)
1202 struct page *subpage = page;
1204 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1208 h->nr_huge_pages_node[page_to_nid(page)]--;
1209 for (i = 0; i < pages_per_huge_page(h);
1210 i++, subpage = mem_map_next(subpage, page, i)) {
1211 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1212 1 << PG_referenced | 1 << PG_dirty |
1213 1 << PG_active | 1 << PG_private |
1216 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1217 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1218 set_page_refcounted(page);
1219 if (hstate_is_gigantic(h)) {
1220 destroy_compound_gigantic_page(page, huge_page_order(h));
1221 free_gigantic_page(page, huge_page_order(h));
1223 __free_pages(page, huge_page_order(h));
1227 struct hstate *size_to_hstate(unsigned long size)
1231 for_each_hstate(h) {
1232 if (huge_page_size(h) == size)
1239 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1240 * to hstate->hugepage_activelist.)
1242 * This function can be called for tail pages, but never returns true for them.
1244 bool page_huge_active(struct page *page)
1246 return PageHeadHuge(page) && PagePrivate(&page[1]);
1249 /* never called for tail page */
1250 void set_page_huge_active(struct page *page)
1252 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1253 SetPagePrivate(&page[1]);
1256 static void clear_page_huge_active(struct page *page)
1258 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1259 ClearPagePrivate(&page[1]);
1263 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1266 static inline bool PageHugeTemporary(struct page *page)
1268 if (!PageHuge(page))
1271 return (unsigned long)page[2].mapping == -1U;
1274 static inline void SetPageHugeTemporary(struct page *page)
1276 page[2].mapping = (void *)-1U;
1279 static inline void ClearPageHugeTemporary(struct page *page)
1281 page[2].mapping = NULL;
1284 static void __free_huge_page(struct page *page)
1287 * Can't pass hstate in here because it is called from the
1288 * compound page destructor.
1290 struct hstate *h = page_hstate(page);
1291 int nid = page_to_nid(page);
1292 struct hugepage_subpool *spool =
1293 (struct hugepage_subpool *)page_private(page);
1294 bool restore_reserve;
1296 VM_BUG_ON_PAGE(page_count(page), page);
1297 VM_BUG_ON_PAGE(page_mapcount(page), page);
1299 set_page_private(page, 0);
1300 page->mapping = NULL;
1301 restore_reserve = PagePrivate(page);
1302 ClearPagePrivate(page);
1305 * If PagePrivate() was set on page, page allocation consumed a
1306 * reservation. If the page was associated with a subpool, there
1307 * would have been a page reserved in the subpool before allocation
1308 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1309 * reservtion, do not call hugepage_subpool_put_pages() as this will
1310 * remove the reserved page from the subpool.
1312 if (!restore_reserve) {
1314 * A return code of zero implies that the subpool will be
1315 * under its minimum size if the reservation is not restored
1316 * after page is free. Therefore, force restore_reserve
1319 if (hugepage_subpool_put_pages(spool, 1) == 0)
1320 restore_reserve = true;
1323 spin_lock(&hugetlb_lock);
1324 clear_page_huge_active(page);
1325 hugetlb_cgroup_uncharge_page(hstate_index(h),
1326 pages_per_huge_page(h), page);
1327 if (restore_reserve)
1328 h->resv_huge_pages++;
1330 if (PageHugeTemporary(page)) {
1331 list_del(&page->lru);
1332 ClearPageHugeTemporary(page);
1333 update_and_free_page(h, page);
1334 } else if (h->surplus_huge_pages_node[nid]) {
1335 /* remove the page from active list */
1336 list_del(&page->lru);
1337 update_and_free_page(h, page);
1338 h->surplus_huge_pages--;
1339 h->surplus_huge_pages_node[nid]--;
1341 arch_clear_hugepage_flags(page);
1342 enqueue_huge_page(h, page);
1344 spin_unlock(&hugetlb_lock);
1348 * As free_huge_page() can be called from a non-task context, we have
1349 * to defer the actual freeing in a workqueue to prevent potential
1350 * hugetlb_lock deadlock.
1352 * free_hpage_workfn() locklessly retrieves the linked list of pages to
1353 * be freed and frees them one-by-one. As the page->mapping pointer is
1354 * going to be cleared in __free_huge_page() anyway, it is reused as the
1355 * llist_node structure of a lockless linked list of huge pages to be freed.
1357 static LLIST_HEAD(hpage_freelist);
1359 static void free_hpage_workfn(struct work_struct *work)
1361 struct llist_node *node;
1364 node = llist_del_all(&hpage_freelist);
1367 page = container_of((struct address_space **)node,
1368 struct page, mapping);
1370 __free_huge_page(page);
1373 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1375 void free_huge_page(struct page *page)
1378 * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1382 * Only call schedule_work() if hpage_freelist is previously
1383 * empty. Otherwise, schedule_work() had been called but the
1384 * workfn hasn't retrieved the list yet.
1386 if (llist_add((struct llist_node *)&page->mapping,
1388 schedule_work(&free_hpage_work);
1392 __free_huge_page(page);
1395 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1397 INIT_LIST_HEAD(&page->lru);
1398 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1399 spin_lock(&hugetlb_lock);
1400 set_hugetlb_cgroup(page, NULL);
1402 h->nr_huge_pages_node[nid]++;
1403 ClearPageHugeFreed(page);
1404 spin_unlock(&hugetlb_lock);
1407 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1410 int nr_pages = 1 << order;
1411 struct page *p = page + 1;
1413 /* we rely on prep_new_huge_page to set the destructor */
1414 set_compound_order(page, order);
1415 __ClearPageReserved(page);
1416 __SetPageHead(page);
1417 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1419 * For gigantic hugepages allocated through bootmem at
1420 * boot, it's safer to be consistent with the not-gigantic
1421 * hugepages and clear the PG_reserved bit from all tail pages
1422 * too. Otherwse drivers using get_user_pages() to access tail
1423 * pages may get the reference counting wrong if they see
1424 * PG_reserved set on a tail page (despite the head page not
1425 * having PG_reserved set). Enforcing this consistency between
1426 * head and tail pages allows drivers to optimize away a check
1427 * on the head page when they need know if put_page() is needed
1428 * after get_user_pages().
1430 __ClearPageReserved(p);
1431 set_page_count(p, 0);
1432 set_compound_head(p, page);
1434 atomic_set(compound_mapcount_ptr(page), -1);
1438 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1439 * transparent huge pages. See the PageTransHuge() documentation for more
1442 int PageHuge(struct page *page)
1444 if (!PageCompound(page))
1447 page = compound_head(page);
1448 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1450 EXPORT_SYMBOL_GPL(PageHuge);
1453 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1454 * normal or transparent huge pages.
1456 int PageHeadHuge(struct page *page_head)
1458 if (!PageHead(page_head))
1461 return get_compound_page_dtor(page_head) == free_huge_page;
1464 pgoff_t hugetlb_basepage_index(struct page *page)
1466 struct page *page_head = compound_head(page);
1467 pgoff_t index = page_index(page_head);
1468 unsigned long compound_idx;
1470 if (compound_order(page_head) >= MAX_ORDER)
1471 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1473 compound_idx = page - page_head;
1475 return (index << compound_order(page_head)) + compound_idx;
1478 static struct page *alloc_buddy_huge_page(struct hstate *h,
1479 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1480 nodemask_t *node_alloc_noretry)
1482 int order = huge_page_order(h);
1484 bool alloc_try_hard = true;
1487 * By default we always try hard to allocate the page with
1488 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1489 * a loop (to adjust global huge page counts) and previous allocation
1490 * failed, do not continue to try hard on the same node. Use the
1491 * node_alloc_noretry bitmap to manage this state information.
1493 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1494 alloc_try_hard = false;
1495 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1497 gfp_mask |= __GFP_RETRY_MAYFAIL;
1498 if (nid == NUMA_NO_NODE)
1499 nid = numa_mem_id();
1500 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1502 __count_vm_event(HTLB_BUDDY_PGALLOC);
1504 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1507 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1508 * indicates an overall state change. Clear bit so that we resume
1509 * normal 'try hard' allocations.
1511 if (node_alloc_noretry && page && !alloc_try_hard)
1512 node_clear(nid, *node_alloc_noretry);
1515 * If we tried hard to get a page but failed, set bit so that
1516 * subsequent attempts will not try as hard until there is an
1517 * overall state change.
1519 if (node_alloc_noretry && !page && alloc_try_hard)
1520 node_set(nid, *node_alloc_noretry);
1526 * Common helper to allocate a fresh hugetlb page. All specific allocators
1527 * should use this function to get new hugetlb pages
1529 static struct page *alloc_fresh_huge_page(struct hstate *h,
1530 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1531 nodemask_t *node_alloc_noretry)
1535 if (hstate_is_gigantic(h))
1536 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1538 page = alloc_buddy_huge_page(h, gfp_mask,
1539 nid, nmask, node_alloc_noretry);
1543 if (hstate_is_gigantic(h))
1544 prep_compound_gigantic_page(page, huge_page_order(h));
1545 prep_new_huge_page(h, page, page_to_nid(page));
1551 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1554 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1555 nodemask_t *node_alloc_noretry)
1559 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1561 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1562 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1563 node_alloc_noretry);
1571 put_page(page); /* free it into the hugepage allocator */
1577 * Free huge page from pool from next node to free.
1578 * Attempt to keep persistent huge pages more or less
1579 * balanced over allowed nodes.
1580 * Called with hugetlb_lock locked.
1582 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1588 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1590 * If we're returning unused surplus pages, only examine
1591 * nodes with surplus pages.
1593 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1594 !list_empty(&h->hugepage_freelists[node])) {
1596 list_entry(h->hugepage_freelists[node].next,
1598 list_del(&page->lru);
1599 h->free_huge_pages--;
1600 h->free_huge_pages_node[node]--;
1602 h->surplus_huge_pages--;
1603 h->surplus_huge_pages_node[node]--;
1605 update_and_free_page(h, page);
1615 * Dissolve a given free hugepage into free buddy pages. This function does
1616 * nothing for in-use hugepages and non-hugepages.
1617 * This function returns values like below:
1619 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1620 * (allocated or reserved.)
1621 * 0: successfully dissolved free hugepages or the page is not a
1622 * hugepage (considered as already dissolved)
1624 int dissolve_free_huge_page(struct page *page)
1629 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1630 if (!PageHuge(page))
1633 spin_lock(&hugetlb_lock);
1634 if (!PageHuge(page)) {
1639 if (!page_count(page)) {
1640 struct page *head = compound_head(page);
1641 struct hstate *h = page_hstate(head);
1642 int nid = page_to_nid(head);
1643 if (h->free_huge_pages - h->resv_huge_pages == 0)
1647 * We should make sure that the page is already on the free list
1648 * when it is dissolved.
1650 if (unlikely(!PageHugeFreed(head))) {
1651 spin_unlock(&hugetlb_lock);
1655 * Theoretically, we should return -EBUSY when we
1656 * encounter this race. In fact, we have a chance
1657 * to successfully dissolve the page if we do a
1658 * retry. Because the race window is quite small.
1659 * If we seize this opportunity, it is an optimization
1660 * for increasing the success rate of dissolving page.
1666 * Move PageHWPoison flag from head page to the raw error page,
1667 * which makes any subpages rather than the error page reusable.
1669 if (PageHWPoison(head) && page != head) {
1670 SetPageHWPoison(page);
1671 ClearPageHWPoison(head);
1673 list_del(&head->lru);
1674 h->free_huge_pages--;
1675 h->free_huge_pages_node[nid]--;
1676 h->max_huge_pages--;
1677 update_and_free_page(h, head);
1681 spin_unlock(&hugetlb_lock);
1686 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1687 * make specified memory blocks removable from the system.
1688 * Note that this will dissolve a free gigantic hugepage completely, if any
1689 * part of it lies within the given range.
1690 * Also note that if dissolve_free_huge_page() returns with an error, all
1691 * free hugepages that were dissolved before that error are lost.
1693 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1699 if (!hugepages_supported())
1702 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1703 page = pfn_to_page(pfn);
1704 rc = dissolve_free_huge_page(page);
1713 * Allocates a fresh surplus page from the page allocator.
1715 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1716 int nid, nodemask_t *nmask)
1718 struct page *page = NULL;
1720 if (hstate_is_gigantic(h))
1723 spin_lock(&hugetlb_lock);
1724 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1726 spin_unlock(&hugetlb_lock);
1728 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1732 spin_lock(&hugetlb_lock);
1734 * We could have raced with the pool size change.
1735 * Double check that and simply deallocate the new page
1736 * if we would end up overcommiting the surpluses. Abuse
1737 * temporary page to workaround the nasty free_huge_page
1740 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1741 SetPageHugeTemporary(page);
1742 spin_unlock(&hugetlb_lock);
1746 h->surplus_huge_pages++;
1747 h->surplus_huge_pages_node[page_to_nid(page)]++;
1751 spin_unlock(&hugetlb_lock);
1756 struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1757 int nid, nodemask_t *nmask)
1761 if (hstate_is_gigantic(h))
1764 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1769 * We do not account these pages as surplus because they are only
1770 * temporary and will be released properly on the last reference
1772 SetPageHugeTemporary(page);
1778 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1781 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1782 struct vm_area_struct *vma, unsigned long addr)
1785 struct mempolicy *mpol;
1786 gfp_t gfp_mask = htlb_alloc_mask(h);
1788 nodemask_t *nodemask;
1790 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1791 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1792 mpol_cond_put(mpol);
1797 /* page migration callback function */
1798 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1800 gfp_t gfp_mask = htlb_alloc_mask(h);
1801 struct page *page = NULL;
1803 if (nid != NUMA_NO_NODE)
1804 gfp_mask |= __GFP_THISNODE;
1806 spin_lock(&hugetlb_lock);
1807 if (h->free_huge_pages - h->resv_huge_pages > 0)
1808 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1809 spin_unlock(&hugetlb_lock);
1812 page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1817 /* page migration callback function */
1818 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1821 gfp_t gfp_mask = htlb_alloc_mask(h);
1823 spin_lock(&hugetlb_lock);
1824 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1827 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1829 spin_unlock(&hugetlb_lock);
1833 spin_unlock(&hugetlb_lock);
1835 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1838 /* mempolicy aware migration callback */
1839 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1840 unsigned long address)
1842 struct mempolicy *mpol;
1843 nodemask_t *nodemask;
1848 gfp_mask = htlb_alloc_mask(h);
1849 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1850 page = alloc_huge_page_nodemask(h, node, nodemask);
1851 mpol_cond_put(mpol);
1857 * Increase the hugetlb pool such that it can accommodate a reservation
1860 static int gather_surplus_pages(struct hstate *h, int delta)
1862 struct list_head surplus_list;
1863 struct page *page, *tmp;
1865 int needed, allocated;
1866 bool alloc_ok = true;
1868 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1870 h->resv_huge_pages += delta;
1875 INIT_LIST_HEAD(&surplus_list);
1879 spin_unlock(&hugetlb_lock);
1880 for (i = 0; i < needed; i++) {
1881 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1882 NUMA_NO_NODE, NULL);
1887 list_add(&page->lru, &surplus_list);
1893 * After retaking hugetlb_lock, we need to recalculate 'needed'
1894 * because either resv_huge_pages or free_huge_pages may have changed.
1896 spin_lock(&hugetlb_lock);
1897 needed = (h->resv_huge_pages + delta) -
1898 (h->free_huge_pages + allocated);
1903 * We were not able to allocate enough pages to
1904 * satisfy the entire reservation so we free what
1905 * we've allocated so far.
1910 * The surplus_list now contains _at_least_ the number of extra pages
1911 * needed to accommodate the reservation. Add the appropriate number
1912 * of pages to the hugetlb pool and free the extras back to the buddy
1913 * allocator. Commit the entire reservation here to prevent another
1914 * process from stealing the pages as they are added to the pool but
1915 * before they are reserved.
1917 needed += allocated;
1918 h->resv_huge_pages += delta;
1921 /* Free the needed pages to the hugetlb pool */
1922 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1926 * This page is now managed by the hugetlb allocator and has
1927 * no users -- drop the buddy allocator's reference.
1929 put_page_testzero(page);
1930 VM_BUG_ON_PAGE(page_count(page), page);
1931 enqueue_huge_page(h, page);
1934 spin_unlock(&hugetlb_lock);
1936 /* Free unnecessary surplus pages to the buddy allocator */
1937 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1939 spin_lock(&hugetlb_lock);
1945 * This routine has two main purposes:
1946 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1947 * in unused_resv_pages. This corresponds to the prior adjustments made
1948 * to the associated reservation map.
1949 * 2) Free any unused surplus pages that may have been allocated to satisfy
1950 * the reservation. As many as unused_resv_pages may be freed.
1952 * Called with hugetlb_lock held. However, the lock could be dropped (and
1953 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1954 * we must make sure nobody else can claim pages we are in the process of
1955 * freeing. Do this by ensuring resv_huge_page always is greater than the
1956 * number of huge pages we plan to free when dropping the lock.
1958 static void return_unused_surplus_pages(struct hstate *h,
1959 unsigned long unused_resv_pages)
1961 unsigned long nr_pages;
1963 /* Cannot return gigantic pages currently */
1964 if (hstate_is_gigantic(h))
1968 * Part (or even all) of the reservation could have been backed
1969 * by pre-allocated pages. Only free surplus pages.
1971 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1974 * We want to release as many surplus pages as possible, spread
1975 * evenly across all nodes with memory. Iterate across these nodes
1976 * until we can no longer free unreserved surplus pages. This occurs
1977 * when the nodes with surplus pages have no free pages.
1978 * free_pool_huge_page() will balance the the freed pages across the
1979 * on-line nodes with memory and will handle the hstate accounting.
1981 * Note that we decrement resv_huge_pages as we free the pages. If
1982 * we drop the lock, resv_huge_pages will still be sufficiently large
1983 * to cover subsequent pages we may free.
1985 while (nr_pages--) {
1986 h->resv_huge_pages--;
1987 unused_resv_pages--;
1988 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1990 cond_resched_lock(&hugetlb_lock);
1994 /* Fully uncommit the reservation */
1995 h->resv_huge_pages -= unused_resv_pages;
2000 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2001 * are used by the huge page allocation routines to manage reservations.
2003 * vma_needs_reservation is called to determine if the huge page at addr
2004 * within the vma has an associated reservation. If a reservation is
2005 * needed, the value 1 is returned. The caller is then responsible for
2006 * managing the global reservation and subpool usage counts. After
2007 * the huge page has been allocated, vma_commit_reservation is called
2008 * to add the page to the reservation map. If the page allocation fails,
2009 * the reservation must be ended instead of committed. vma_end_reservation
2010 * is called in such cases.
2012 * In the normal case, vma_commit_reservation returns the same value
2013 * as the preceding vma_needs_reservation call. The only time this
2014 * is not the case is if a reserve map was changed between calls. It
2015 * is the responsibility of the caller to notice the difference and
2016 * take appropriate action.
2018 * vma_add_reservation is used in error paths where a reservation must
2019 * be restored when a newly allocated huge page must be freed. It is
2020 * to be called after calling vma_needs_reservation to determine if a
2021 * reservation exists.
2023 enum vma_resv_mode {
2029 static long __vma_reservation_common(struct hstate *h,
2030 struct vm_area_struct *vma, unsigned long addr,
2031 enum vma_resv_mode mode)
2033 struct resv_map *resv;
2037 resv = vma_resv_map(vma);
2041 idx = vma_hugecache_offset(h, vma, addr);
2043 case VMA_NEEDS_RESV:
2044 ret = region_chg(resv, idx, idx + 1);
2046 case VMA_COMMIT_RESV:
2047 ret = region_add(resv, idx, idx + 1);
2050 region_abort(resv, idx, idx + 1);
2054 if (vma->vm_flags & VM_MAYSHARE)
2055 ret = region_add(resv, idx, idx + 1);
2057 region_abort(resv, idx, idx + 1);
2058 ret = region_del(resv, idx, idx + 1);
2065 if (vma->vm_flags & VM_MAYSHARE)
2067 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
2069 * In most cases, reserves always exist for private mappings.
2070 * However, a file associated with mapping could have been
2071 * hole punched or truncated after reserves were consumed.
2072 * As subsequent fault on such a range will not use reserves.
2073 * Subtle - The reserve map for private mappings has the
2074 * opposite meaning than that of shared mappings. If NO
2075 * entry is in the reserve map, it means a reservation exists.
2076 * If an entry exists in the reserve map, it means the
2077 * reservation has already been consumed. As a result, the
2078 * return value of this routine is the opposite of the
2079 * value returned from reserve map manipulation routines above.
2087 return ret < 0 ? ret : 0;
2090 static long vma_needs_reservation(struct hstate *h,
2091 struct vm_area_struct *vma, unsigned long addr)
2093 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2096 static long vma_commit_reservation(struct hstate *h,
2097 struct vm_area_struct *vma, unsigned long addr)
2099 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2102 static void vma_end_reservation(struct hstate *h,
2103 struct vm_area_struct *vma, unsigned long addr)
2105 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2108 static long vma_add_reservation(struct hstate *h,
2109 struct vm_area_struct *vma, unsigned long addr)
2111 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2115 * This routine is called to restore a reservation on error paths. In the
2116 * specific error paths, a huge page was allocated (via alloc_huge_page)
2117 * and is about to be freed. If a reservation for the page existed,
2118 * alloc_huge_page would have consumed the reservation and set PagePrivate
2119 * in the newly allocated page. When the page is freed via free_huge_page,
2120 * the global reservation count will be incremented if PagePrivate is set.
2121 * However, free_huge_page can not adjust the reserve map. Adjust the
2122 * reserve map here to be consistent with global reserve count adjustments
2123 * to be made by free_huge_page.
2125 static void restore_reserve_on_error(struct hstate *h,
2126 struct vm_area_struct *vma, unsigned long address,
2129 if (unlikely(PagePrivate(page))) {
2130 long rc = vma_needs_reservation(h, vma, address);
2132 if (unlikely(rc < 0)) {
2134 * Rare out of memory condition in reserve map
2135 * manipulation. Clear PagePrivate so that
2136 * global reserve count will not be incremented
2137 * by free_huge_page. This will make it appear
2138 * as though the reservation for this page was
2139 * consumed. This may prevent the task from
2140 * faulting in the page at a later time. This
2141 * is better than inconsistent global huge page
2142 * accounting of reserve counts.
2144 ClearPagePrivate(page);
2146 rc = vma_add_reservation(h, vma, address);
2147 if (unlikely(rc < 0))
2149 * See above comment about rare out of
2152 ClearPagePrivate(page);
2154 vma_end_reservation(h, vma, address);
2158 struct page *alloc_huge_page(struct vm_area_struct *vma,
2159 unsigned long addr, int avoid_reserve)
2161 struct hugepage_subpool *spool = subpool_vma(vma);
2162 struct hstate *h = hstate_vma(vma);
2164 long map_chg, map_commit;
2167 struct hugetlb_cgroup *h_cg;
2169 idx = hstate_index(h);
2171 * Examine the region/reserve map to determine if the process
2172 * has a reservation for the page to be allocated. A return
2173 * code of zero indicates a reservation exists (no change).
2175 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2177 return ERR_PTR(-ENOMEM);
2180 * Processes that did not create the mapping will have no
2181 * reserves as indicated by the region/reserve map. Check
2182 * that the allocation will not exceed the subpool limit.
2183 * Allocations for MAP_NORESERVE mappings also need to be
2184 * checked against any subpool limit.
2186 if (map_chg || avoid_reserve) {
2187 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2189 vma_end_reservation(h, vma, addr);
2190 return ERR_PTR(-ENOSPC);
2194 * Even though there was no reservation in the region/reserve
2195 * map, there could be reservations associated with the
2196 * subpool that can be used. This would be indicated if the
2197 * return value of hugepage_subpool_get_pages() is zero.
2198 * However, if avoid_reserve is specified we still avoid even
2199 * the subpool reservations.
2205 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2207 goto out_subpool_put;
2209 spin_lock(&hugetlb_lock);
2211 * glb_chg is passed to indicate whether or not a page must be taken
2212 * from the global free pool (global change). gbl_chg == 0 indicates
2213 * a reservation exists for the allocation.
2215 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2217 spin_unlock(&hugetlb_lock);
2218 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2220 goto out_uncharge_cgroup;
2221 spin_lock(&hugetlb_lock);
2222 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2223 SetPagePrivate(page);
2224 h->resv_huge_pages--;
2226 list_move(&page->lru, &h->hugepage_activelist);
2229 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2230 spin_unlock(&hugetlb_lock);
2232 set_page_private(page, (unsigned long)spool);
2234 map_commit = vma_commit_reservation(h, vma, addr);
2235 if (unlikely(map_chg > map_commit)) {
2237 * The page was added to the reservation map between
2238 * vma_needs_reservation and vma_commit_reservation.
2239 * This indicates a race with hugetlb_reserve_pages.
2240 * Adjust for the subpool count incremented above AND
2241 * in hugetlb_reserve_pages for the same page. Also,
2242 * the reservation count added in hugetlb_reserve_pages
2243 * no longer applies.
2247 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2248 hugetlb_acct_memory(h, -rsv_adjust);
2252 out_uncharge_cgroup:
2253 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2255 if (map_chg || avoid_reserve)
2256 hugepage_subpool_put_pages(spool, 1);
2257 vma_end_reservation(h, vma, addr);
2258 return ERR_PTR(-ENOSPC);
2261 int alloc_bootmem_huge_page(struct hstate *h)
2262 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2263 int __alloc_bootmem_huge_page(struct hstate *h)
2265 struct huge_bootmem_page *m;
2268 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2271 addr = memblock_alloc_try_nid_raw(
2272 huge_page_size(h), huge_page_size(h),
2273 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2276 * Use the beginning of the huge page to store the
2277 * huge_bootmem_page struct (until gather_bootmem
2278 * puts them into the mem_map).
2287 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2288 /* Put them into a private list first because mem_map is not up yet */
2289 INIT_LIST_HEAD(&m->list);
2290 list_add(&m->list, &huge_boot_pages);
2295 static void __init prep_compound_huge_page(struct page *page,
2298 if (unlikely(order > (MAX_ORDER - 1)))
2299 prep_compound_gigantic_page(page, order);
2301 prep_compound_page(page, order);
2304 /* Put bootmem huge pages into the standard lists after mem_map is up */
2305 static void __init gather_bootmem_prealloc(void)
2307 struct huge_bootmem_page *m;
2309 list_for_each_entry(m, &huge_boot_pages, list) {
2310 struct page *page = virt_to_page(m);
2311 struct hstate *h = m->hstate;
2313 WARN_ON(page_count(page) != 1);
2314 prep_compound_huge_page(page, h->order);
2315 WARN_ON(PageReserved(page));
2316 prep_new_huge_page(h, page, page_to_nid(page));
2317 put_page(page); /* free it into the hugepage allocator */
2320 * If we had gigantic hugepages allocated at boot time, we need
2321 * to restore the 'stolen' pages to totalram_pages in order to
2322 * fix confusing memory reports from free(1) and another
2323 * side-effects, like CommitLimit going negative.
2325 if (hstate_is_gigantic(h))
2326 adjust_managed_page_count(page, 1 << h->order);
2331 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2334 nodemask_t *node_alloc_noretry;
2336 if (!hstate_is_gigantic(h)) {
2338 * Bit mask controlling how hard we retry per-node allocations.
2339 * Ignore errors as lower level routines can deal with
2340 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2341 * time, we are likely in bigger trouble.
2343 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2346 /* allocations done at boot time */
2347 node_alloc_noretry = NULL;
2350 /* bit mask controlling how hard we retry per-node allocations */
2351 if (node_alloc_noretry)
2352 nodes_clear(*node_alloc_noretry);
2354 for (i = 0; i < h->max_huge_pages; ++i) {
2355 if (hstate_is_gigantic(h)) {
2356 if (!alloc_bootmem_huge_page(h))
2358 } else if (!alloc_pool_huge_page(h,
2359 &node_states[N_MEMORY],
2360 node_alloc_noretry))
2364 if (i < h->max_huge_pages) {
2367 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2368 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2369 h->max_huge_pages, buf, i);
2370 h->max_huge_pages = i;
2373 kfree(node_alloc_noretry);
2376 static void __init hugetlb_init_hstates(void)
2380 for_each_hstate(h) {
2381 if (minimum_order > huge_page_order(h))
2382 minimum_order = huge_page_order(h);
2384 /* oversize hugepages were init'ed in early boot */
2385 if (!hstate_is_gigantic(h))
2386 hugetlb_hstate_alloc_pages(h);
2388 VM_BUG_ON(minimum_order == UINT_MAX);
2391 static void __init report_hugepages(void)
2395 for_each_hstate(h) {
2398 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2399 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2400 buf, h->free_huge_pages);
2404 #ifdef CONFIG_HIGHMEM
2405 static void try_to_free_low(struct hstate *h, unsigned long count,
2406 nodemask_t *nodes_allowed)
2410 if (hstate_is_gigantic(h))
2413 for_each_node_mask(i, *nodes_allowed) {
2414 struct page *page, *next;
2415 struct list_head *freel = &h->hugepage_freelists[i];
2416 list_for_each_entry_safe(page, next, freel, lru) {
2417 if (count >= h->nr_huge_pages)
2419 if (PageHighMem(page))
2421 list_del(&page->lru);
2422 update_and_free_page(h, page);
2423 h->free_huge_pages--;
2424 h->free_huge_pages_node[page_to_nid(page)]--;
2429 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2430 nodemask_t *nodes_allowed)
2436 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2437 * balanced by operating on them in a round-robin fashion.
2438 * Returns 1 if an adjustment was made.
2440 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2445 VM_BUG_ON(delta != -1 && delta != 1);
2448 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2449 if (h->surplus_huge_pages_node[node])
2453 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2454 if (h->surplus_huge_pages_node[node] <
2455 h->nr_huge_pages_node[node])
2462 h->surplus_huge_pages += delta;
2463 h->surplus_huge_pages_node[node] += delta;
2467 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2468 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2469 nodemask_t *nodes_allowed)
2471 unsigned long min_count, ret;
2472 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2475 * Bit mask controlling how hard we retry per-node allocations.
2476 * If we can not allocate the bit mask, do not attempt to allocate
2477 * the requested huge pages.
2479 if (node_alloc_noretry)
2480 nodes_clear(*node_alloc_noretry);
2484 spin_lock(&hugetlb_lock);
2487 * Check for a node specific request.
2488 * Changing node specific huge page count may require a corresponding
2489 * change to the global count. In any case, the passed node mask
2490 * (nodes_allowed) will restrict alloc/free to the specified node.
2492 if (nid != NUMA_NO_NODE) {
2493 unsigned long old_count = count;
2495 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2497 * User may have specified a large count value which caused the
2498 * above calculation to overflow. In this case, they wanted
2499 * to allocate as many huge pages as possible. Set count to
2500 * largest possible value to align with their intention.
2502 if (count < old_count)
2507 * Gigantic pages runtime allocation depend on the capability for large
2508 * page range allocation.
2509 * If the system does not provide this feature, return an error when
2510 * the user tries to allocate gigantic pages but let the user free the
2511 * boottime allocated gigantic pages.
2513 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2514 if (count > persistent_huge_pages(h)) {
2515 spin_unlock(&hugetlb_lock);
2516 NODEMASK_FREE(node_alloc_noretry);
2519 /* Fall through to decrease pool */
2523 * Increase the pool size
2524 * First take pages out of surplus state. Then make up the
2525 * remaining difference by allocating fresh huge pages.
2527 * We might race with alloc_surplus_huge_page() here and be unable
2528 * to convert a surplus huge page to a normal huge page. That is
2529 * not critical, though, it just means the overall size of the
2530 * pool might be one hugepage larger than it needs to be, but
2531 * within all the constraints specified by the sysctls.
2533 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2534 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2538 while (count > persistent_huge_pages(h)) {
2540 * If this allocation races such that we no longer need the
2541 * page, free_huge_page will handle it by freeing the page
2542 * and reducing the surplus.
2544 spin_unlock(&hugetlb_lock);
2546 /* yield cpu to avoid soft lockup */
2549 ret = alloc_pool_huge_page(h, nodes_allowed,
2550 node_alloc_noretry);
2551 spin_lock(&hugetlb_lock);
2555 /* Bail for signals. Probably ctrl-c from user */
2556 if (signal_pending(current))
2561 * Decrease the pool size
2562 * First return free pages to the buddy allocator (being careful
2563 * to keep enough around to satisfy reservations). Then place
2564 * pages into surplus state as needed so the pool will shrink
2565 * to the desired size as pages become free.
2567 * By placing pages into the surplus state independent of the
2568 * overcommit value, we are allowing the surplus pool size to
2569 * exceed overcommit. There are few sane options here. Since
2570 * alloc_surplus_huge_page() is checking the global counter,
2571 * though, we'll note that we're not allowed to exceed surplus
2572 * and won't grow the pool anywhere else. Not until one of the
2573 * sysctls are changed, or the surplus pages go out of use.
2575 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2576 min_count = max(count, min_count);
2577 try_to_free_low(h, min_count, nodes_allowed);
2578 while (min_count < persistent_huge_pages(h)) {
2579 if (!free_pool_huge_page(h, nodes_allowed, 0))
2581 cond_resched_lock(&hugetlb_lock);
2583 while (count < persistent_huge_pages(h)) {
2584 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2588 h->max_huge_pages = persistent_huge_pages(h);
2589 spin_unlock(&hugetlb_lock);
2591 NODEMASK_FREE(node_alloc_noretry);
2596 #define HSTATE_ATTR_RO(_name) \
2597 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2599 #define HSTATE_ATTR(_name) \
2600 static struct kobj_attribute _name##_attr = \
2601 __ATTR(_name, 0644, _name##_show, _name##_store)
2603 static struct kobject *hugepages_kobj;
2604 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2606 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2608 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2612 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2613 if (hstate_kobjs[i] == kobj) {
2615 *nidp = NUMA_NO_NODE;
2619 return kobj_to_node_hstate(kobj, nidp);
2622 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2623 struct kobj_attribute *attr, char *buf)
2626 unsigned long nr_huge_pages;
2629 h = kobj_to_hstate(kobj, &nid);
2630 if (nid == NUMA_NO_NODE)
2631 nr_huge_pages = h->nr_huge_pages;
2633 nr_huge_pages = h->nr_huge_pages_node[nid];
2635 return sprintf(buf, "%lu\n", nr_huge_pages);
2638 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2639 struct hstate *h, int nid,
2640 unsigned long count, size_t len)
2643 nodemask_t nodes_allowed, *n_mask;
2645 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2648 if (nid == NUMA_NO_NODE) {
2650 * global hstate attribute
2652 if (!(obey_mempolicy &&
2653 init_nodemask_of_mempolicy(&nodes_allowed)))
2654 n_mask = &node_states[N_MEMORY];
2656 n_mask = &nodes_allowed;
2659 * Node specific request. count adjustment happens in
2660 * set_max_huge_pages() after acquiring hugetlb_lock.
2662 init_nodemask_of_node(&nodes_allowed, nid);
2663 n_mask = &nodes_allowed;
2666 err = set_max_huge_pages(h, count, nid, n_mask);
2668 return err ? err : len;
2671 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2672 struct kobject *kobj, const char *buf,
2676 unsigned long count;
2680 err = kstrtoul(buf, 10, &count);
2684 h = kobj_to_hstate(kobj, &nid);
2685 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2688 static ssize_t nr_hugepages_show(struct kobject *kobj,
2689 struct kobj_attribute *attr, char *buf)
2691 return nr_hugepages_show_common(kobj, attr, buf);
2694 static ssize_t nr_hugepages_store(struct kobject *kobj,
2695 struct kobj_attribute *attr, const char *buf, size_t len)
2697 return nr_hugepages_store_common(false, kobj, buf, len);
2699 HSTATE_ATTR(nr_hugepages);
2704 * hstate attribute for optionally mempolicy-based constraint on persistent
2705 * huge page alloc/free.
2707 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2708 struct kobj_attribute *attr, char *buf)
2710 return nr_hugepages_show_common(kobj, attr, buf);
2713 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2714 struct kobj_attribute *attr, const char *buf, size_t len)
2716 return nr_hugepages_store_common(true, kobj, buf, len);
2718 HSTATE_ATTR(nr_hugepages_mempolicy);
2722 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2723 struct kobj_attribute *attr, char *buf)
2725 struct hstate *h = kobj_to_hstate(kobj, NULL);
2726 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2729 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2730 struct kobj_attribute *attr, const char *buf, size_t count)
2733 unsigned long input;
2734 struct hstate *h = kobj_to_hstate(kobj, NULL);
2736 if (hstate_is_gigantic(h))
2739 err = kstrtoul(buf, 10, &input);
2743 spin_lock(&hugetlb_lock);
2744 h->nr_overcommit_huge_pages = input;
2745 spin_unlock(&hugetlb_lock);
2749 HSTATE_ATTR(nr_overcommit_hugepages);
2751 static ssize_t free_hugepages_show(struct kobject *kobj,
2752 struct kobj_attribute *attr, char *buf)
2755 unsigned long free_huge_pages;
2758 h = kobj_to_hstate(kobj, &nid);
2759 if (nid == NUMA_NO_NODE)
2760 free_huge_pages = h->free_huge_pages;
2762 free_huge_pages = h->free_huge_pages_node[nid];
2764 return sprintf(buf, "%lu\n", free_huge_pages);
2766 HSTATE_ATTR_RO(free_hugepages);
2768 static ssize_t resv_hugepages_show(struct kobject *kobj,
2769 struct kobj_attribute *attr, char *buf)
2771 struct hstate *h = kobj_to_hstate(kobj, NULL);
2772 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2774 HSTATE_ATTR_RO(resv_hugepages);
2776 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2777 struct kobj_attribute *attr, char *buf)
2780 unsigned long surplus_huge_pages;
2783 h = kobj_to_hstate(kobj, &nid);
2784 if (nid == NUMA_NO_NODE)
2785 surplus_huge_pages = h->surplus_huge_pages;
2787 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2789 return sprintf(buf, "%lu\n", surplus_huge_pages);
2791 HSTATE_ATTR_RO(surplus_hugepages);
2793 static struct attribute *hstate_attrs[] = {
2794 &nr_hugepages_attr.attr,
2795 &nr_overcommit_hugepages_attr.attr,
2796 &free_hugepages_attr.attr,
2797 &resv_hugepages_attr.attr,
2798 &surplus_hugepages_attr.attr,
2800 &nr_hugepages_mempolicy_attr.attr,
2805 static const struct attribute_group hstate_attr_group = {
2806 .attrs = hstate_attrs,
2809 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2810 struct kobject **hstate_kobjs,
2811 const struct attribute_group *hstate_attr_group)
2814 int hi = hstate_index(h);
2816 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2817 if (!hstate_kobjs[hi])
2820 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2822 kobject_put(hstate_kobjs[hi]);
2823 hstate_kobjs[hi] = NULL;
2829 static void __init hugetlb_sysfs_init(void)
2834 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2835 if (!hugepages_kobj)
2838 for_each_hstate(h) {
2839 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2840 hstate_kobjs, &hstate_attr_group);
2842 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2849 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2850 * with node devices in node_devices[] using a parallel array. The array
2851 * index of a node device or _hstate == node id.
2852 * This is here to avoid any static dependency of the node device driver, in
2853 * the base kernel, on the hugetlb module.
2855 struct node_hstate {
2856 struct kobject *hugepages_kobj;
2857 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2859 static struct node_hstate node_hstates[MAX_NUMNODES];
2862 * A subset of global hstate attributes for node devices
2864 static struct attribute *per_node_hstate_attrs[] = {
2865 &nr_hugepages_attr.attr,
2866 &free_hugepages_attr.attr,
2867 &surplus_hugepages_attr.attr,
2871 static const struct attribute_group per_node_hstate_attr_group = {
2872 .attrs = per_node_hstate_attrs,
2876 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2877 * Returns node id via non-NULL nidp.
2879 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2883 for (nid = 0; nid < nr_node_ids; nid++) {
2884 struct node_hstate *nhs = &node_hstates[nid];
2886 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2887 if (nhs->hstate_kobjs[i] == kobj) {
2899 * Unregister hstate attributes from a single node device.
2900 * No-op if no hstate attributes attached.
2902 static void hugetlb_unregister_node(struct node *node)
2905 struct node_hstate *nhs = &node_hstates[node->dev.id];
2907 if (!nhs->hugepages_kobj)
2908 return; /* no hstate attributes */
2910 for_each_hstate(h) {
2911 int idx = hstate_index(h);
2912 if (nhs->hstate_kobjs[idx]) {
2913 kobject_put(nhs->hstate_kobjs[idx]);
2914 nhs->hstate_kobjs[idx] = NULL;
2918 kobject_put(nhs->hugepages_kobj);
2919 nhs->hugepages_kobj = NULL;
2924 * Register hstate attributes for a single node device.
2925 * No-op if attributes already registered.
2927 static void hugetlb_register_node(struct node *node)
2930 struct node_hstate *nhs = &node_hstates[node->dev.id];
2933 if (nhs->hugepages_kobj)
2934 return; /* already allocated */
2936 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2938 if (!nhs->hugepages_kobj)
2941 for_each_hstate(h) {
2942 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2944 &per_node_hstate_attr_group);
2946 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2947 h->name, node->dev.id);
2948 hugetlb_unregister_node(node);
2955 * hugetlb init time: register hstate attributes for all registered node
2956 * devices of nodes that have memory. All on-line nodes should have
2957 * registered their associated device by this time.
2959 static void __init hugetlb_register_all_nodes(void)
2963 for_each_node_state(nid, N_MEMORY) {
2964 struct node *node = node_devices[nid];
2965 if (node->dev.id == nid)
2966 hugetlb_register_node(node);
2970 * Let the node device driver know we're here so it can
2971 * [un]register hstate attributes on node hotplug.
2973 register_hugetlbfs_with_node(hugetlb_register_node,
2974 hugetlb_unregister_node);
2976 #else /* !CONFIG_NUMA */
2978 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2986 static void hugetlb_register_all_nodes(void) { }
2990 static int __init hugetlb_init(void)
2994 if (!hugepages_supported())
2997 if (!size_to_hstate(default_hstate_size)) {
2998 if (default_hstate_size != 0) {
2999 pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
3000 default_hstate_size, HPAGE_SIZE);
3003 default_hstate_size = HPAGE_SIZE;
3004 if (!size_to_hstate(default_hstate_size))
3005 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3007 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
3008 if (default_hstate_max_huge_pages) {
3009 if (!default_hstate.max_huge_pages)
3010 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3013 hugetlb_init_hstates();
3014 gather_bootmem_prealloc();
3017 hugetlb_sysfs_init();
3018 hugetlb_register_all_nodes();
3019 hugetlb_cgroup_file_init();
3022 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3024 num_fault_mutexes = 1;
3026 hugetlb_fault_mutex_table =
3027 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3029 BUG_ON(!hugetlb_fault_mutex_table);
3031 for (i = 0; i < num_fault_mutexes; i++)
3032 mutex_init(&hugetlb_fault_mutex_table[i]);
3035 subsys_initcall(hugetlb_init);
3037 /* Should be called on processing a hugepagesz=... option */
3038 void __init hugetlb_bad_size(void)
3040 parsed_valid_hugepagesz = false;
3043 void __init hugetlb_add_hstate(unsigned int order)
3048 if (size_to_hstate(PAGE_SIZE << order)) {
3049 pr_warn("hugepagesz= specified twice, ignoring\n");
3052 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3054 h = &hstates[hugetlb_max_hstate++];
3056 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
3057 h->nr_huge_pages = 0;
3058 h->free_huge_pages = 0;
3059 for (i = 0; i < MAX_NUMNODES; ++i)
3060 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3061 INIT_LIST_HEAD(&h->hugepage_activelist);
3062 h->next_nid_to_alloc = first_memory_node;
3063 h->next_nid_to_free = first_memory_node;
3064 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3065 huge_page_size(h)/1024);
3070 static int __init hugetlb_nrpages_setup(char *s)
3073 static unsigned long *last_mhp;
3075 if (!parsed_valid_hugepagesz) {
3076 pr_warn("hugepages = %s preceded by "
3077 "an unsupported hugepagesz, ignoring\n", s);
3078 parsed_valid_hugepagesz = true;
3082 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
3083 * so this hugepages= parameter goes to the "default hstate".
3085 else if (!hugetlb_max_hstate)
3086 mhp = &default_hstate_max_huge_pages;
3088 mhp = &parsed_hstate->max_huge_pages;
3090 if (mhp == last_mhp) {
3091 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
3095 if (sscanf(s, "%lu", mhp) <= 0)
3099 * Global state is always initialized later in hugetlb_init.
3100 * But we need to allocate >= MAX_ORDER hstates here early to still
3101 * use the bootmem allocator.
3103 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
3104 hugetlb_hstate_alloc_pages(parsed_hstate);
3110 __setup("hugepages=", hugetlb_nrpages_setup);
3112 static int __init hugetlb_default_setup(char *s)
3114 default_hstate_size = memparse(s, &s);
3117 __setup("default_hugepagesz=", hugetlb_default_setup);
3119 static unsigned int cpuset_mems_nr(unsigned int *array)
3122 unsigned int nr = 0;
3124 for_each_node_mask(node, cpuset_current_mems_allowed)
3130 #ifdef CONFIG_SYSCTL
3131 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3132 void *buffer, size_t *length,
3133 loff_t *ppos, unsigned long *out)
3135 struct ctl_table dup_table;
3138 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3139 * can duplicate the @table and alter the duplicate of it.
3142 dup_table.data = out;
3144 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3147 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3148 struct ctl_table *table, int write,
3149 void __user *buffer, size_t *length, loff_t *ppos)
3151 struct hstate *h = &default_hstate;
3152 unsigned long tmp = h->max_huge_pages;
3155 if (!hugepages_supported())
3158 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3164 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3165 NUMA_NO_NODE, tmp, *length);
3170 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3171 void __user *buffer, size_t *length, loff_t *ppos)
3174 return hugetlb_sysctl_handler_common(false, table, write,
3175 buffer, length, ppos);
3179 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3180 void __user *buffer, size_t *length, loff_t *ppos)
3182 return hugetlb_sysctl_handler_common(true, table, write,
3183 buffer, length, ppos);
3185 #endif /* CONFIG_NUMA */
3187 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3188 void __user *buffer,
3189 size_t *length, loff_t *ppos)
3191 struct hstate *h = &default_hstate;
3195 if (!hugepages_supported())
3198 tmp = h->nr_overcommit_huge_pages;
3200 if (write && hstate_is_gigantic(h))
3203 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3209 spin_lock(&hugetlb_lock);
3210 h->nr_overcommit_huge_pages = tmp;
3211 spin_unlock(&hugetlb_lock);
3217 #endif /* CONFIG_SYSCTL */
3219 void hugetlb_report_meminfo(struct seq_file *m)
3222 unsigned long total = 0;
3224 if (!hugepages_supported())
3227 for_each_hstate(h) {
3228 unsigned long count = h->nr_huge_pages;
3230 total += (PAGE_SIZE << huge_page_order(h)) * count;
3232 if (h == &default_hstate)
3234 "HugePages_Total: %5lu\n"
3235 "HugePages_Free: %5lu\n"
3236 "HugePages_Rsvd: %5lu\n"
3237 "HugePages_Surp: %5lu\n"
3238 "Hugepagesize: %8lu kB\n",
3242 h->surplus_huge_pages,
3243 (PAGE_SIZE << huge_page_order(h)) / 1024);
3246 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3249 int hugetlb_report_node_meminfo(int nid, char *buf)
3251 struct hstate *h = &default_hstate;
3252 if (!hugepages_supported())
3255 "Node %d HugePages_Total: %5u\n"
3256 "Node %d HugePages_Free: %5u\n"
3257 "Node %d HugePages_Surp: %5u\n",
3258 nid, h->nr_huge_pages_node[nid],
3259 nid, h->free_huge_pages_node[nid],
3260 nid, h->surplus_huge_pages_node[nid]);
3263 void hugetlb_show_meminfo(void)
3268 if (!hugepages_supported())
3271 for_each_node_state(nid, N_MEMORY)
3273 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3275 h->nr_huge_pages_node[nid],
3276 h->free_huge_pages_node[nid],
3277 h->surplus_huge_pages_node[nid],
3278 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3281 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3283 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3284 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3287 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3288 unsigned long hugetlb_total_pages(void)
3291 unsigned long nr_total_pages = 0;
3294 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3295 return nr_total_pages;
3298 static int hugetlb_acct_memory(struct hstate *h, long delta)
3302 spin_lock(&hugetlb_lock);
3304 * When cpuset is configured, it breaks the strict hugetlb page
3305 * reservation as the accounting is done on a global variable. Such
3306 * reservation is completely rubbish in the presence of cpuset because
3307 * the reservation is not checked against page availability for the
3308 * current cpuset. Application can still potentially OOM'ed by kernel
3309 * with lack of free htlb page in cpuset that the task is in.
3310 * Attempt to enforce strict accounting with cpuset is almost
3311 * impossible (or too ugly) because cpuset is too fluid that
3312 * task or memory node can be dynamically moved between cpusets.
3314 * The change of semantics for shared hugetlb mapping with cpuset is
3315 * undesirable. However, in order to preserve some of the semantics,
3316 * we fall back to check against current free page availability as
3317 * a best attempt and hopefully to minimize the impact of changing
3318 * semantics that cpuset has.
3321 if (gather_surplus_pages(h, delta) < 0)
3324 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3325 return_unused_surplus_pages(h, delta);
3332 return_unused_surplus_pages(h, (unsigned long) -delta);
3335 spin_unlock(&hugetlb_lock);
3339 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3341 struct resv_map *resv = vma_resv_map(vma);
3344 * This new VMA should share its siblings reservation map if present.
3345 * The VMA will only ever have a valid reservation map pointer where
3346 * it is being copied for another still existing VMA. As that VMA
3347 * has a reference to the reservation map it cannot disappear until
3348 * after this open call completes. It is therefore safe to take a
3349 * new reference here without additional locking.
3351 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3352 kref_get(&resv->refs);
3355 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3357 struct hstate *h = hstate_vma(vma);
3358 struct resv_map *resv = vma_resv_map(vma);
3359 struct hugepage_subpool *spool = subpool_vma(vma);
3360 unsigned long reserve, start, end;
3363 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3366 start = vma_hugecache_offset(h, vma, vma->vm_start);
3367 end = vma_hugecache_offset(h, vma, vma->vm_end);
3369 reserve = (end - start) - region_count(resv, start, end);
3371 kref_put(&resv->refs, resv_map_release);
3375 * Decrement reserve counts. The global reserve count may be
3376 * adjusted if the subpool has a minimum size.
3378 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3379 hugetlb_acct_memory(h, -gbl_reserve);
3383 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3385 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3390 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3392 struct hstate *hstate = hstate_vma(vma);
3394 return 1UL << huge_page_shift(hstate);
3398 * We cannot handle pagefaults against hugetlb pages at all. They cause
3399 * handle_mm_fault() to try to instantiate regular-sized pages in the
3400 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3403 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3410 * When a new function is introduced to vm_operations_struct and added
3411 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3412 * This is because under System V memory model, mappings created via
3413 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3414 * their original vm_ops are overwritten with shm_vm_ops.
3416 const struct vm_operations_struct hugetlb_vm_ops = {
3417 .fault = hugetlb_vm_op_fault,
3418 .open = hugetlb_vm_op_open,
3419 .close = hugetlb_vm_op_close,
3420 .split = hugetlb_vm_op_split,
3421 .pagesize = hugetlb_vm_op_pagesize,
3424 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3430 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3431 vma->vm_page_prot)));
3433 entry = huge_pte_wrprotect(mk_huge_pte(page,
3434 vma->vm_page_prot));
3436 entry = pte_mkyoung(entry);
3437 entry = pte_mkhuge(entry);
3438 entry = arch_make_huge_pte(entry, vma, page, writable);
3443 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3444 unsigned long address, pte_t *ptep)
3448 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3449 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3450 update_mmu_cache(vma, address, ptep);
3453 bool is_hugetlb_entry_migration(pte_t pte)
3457 if (huge_pte_none(pte) || pte_present(pte))
3459 swp = pte_to_swp_entry(pte);
3460 if (non_swap_entry(swp) && is_migration_entry(swp))
3466 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3470 if (huge_pte_none(pte) || pte_present(pte))
3472 swp = pte_to_swp_entry(pte);
3473 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3479 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3480 struct vm_area_struct *vma)
3482 pte_t *src_pte, *dst_pte, entry, dst_entry;
3483 struct page *ptepage;
3486 struct hstate *h = hstate_vma(vma);
3487 unsigned long sz = huge_page_size(h);
3488 struct mmu_notifier_range range;
3491 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3494 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3497 mmu_notifier_invalidate_range_start(&range);
3500 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3501 spinlock_t *src_ptl, *dst_ptl;
3502 src_pte = huge_pte_offset(src, addr, sz);
3505 dst_pte = huge_pte_alloc(dst, addr, sz);
3512 * If the pagetables are shared don't copy or take references.
3513 * dst_pte == src_pte is the common case of src/dest sharing.
3515 * However, src could have 'unshared' and dst shares with
3516 * another vma. If dst_pte !none, this implies sharing.
3517 * Check here before taking page table lock, and once again
3518 * after taking the lock below.
3520 dst_entry = huge_ptep_get(dst_pte);
3521 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3524 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3525 src_ptl = huge_pte_lockptr(h, src, src_pte);
3526 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3527 entry = huge_ptep_get(src_pte);
3528 dst_entry = huge_ptep_get(dst_pte);
3529 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3531 * Skip if src entry none. Also, skip in the
3532 * unlikely case dst entry !none as this implies
3533 * sharing with another vma.
3536 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3537 is_hugetlb_entry_hwpoisoned(entry))) {
3538 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3540 if (is_write_migration_entry(swp_entry) && cow) {
3542 * COW mappings require pages in both
3543 * parent and child to be set to read.
3545 make_migration_entry_read(&swp_entry);
3546 entry = swp_entry_to_pte(swp_entry);
3547 set_huge_swap_pte_at(src, addr, src_pte,
3550 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3554 * No need to notify as we are downgrading page
3555 * table protection not changing it to point
3558 * See Documentation/vm/mmu_notifier.rst
3560 huge_ptep_set_wrprotect(src, addr, src_pte);
3562 entry = huge_ptep_get(src_pte);
3563 ptepage = pte_page(entry);
3565 page_dup_rmap(ptepage, true);
3566 set_huge_pte_at(dst, addr, dst_pte, entry);
3567 hugetlb_count_add(pages_per_huge_page(h), dst);
3569 spin_unlock(src_ptl);
3570 spin_unlock(dst_ptl);
3574 mmu_notifier_invalidate_range_end(&range);
3579 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3580 unsigned long start, unsigned long end,
3581 struct page *ref_page)
3583 struct mm_struct *mm = vma->vm_mm;
3584 unsigned long address;
3589 struct hstate *h = hstate_vma(vma);
3590 unsigned long sz = huge_page_size(h);
3591 struct mmu_notifier_range range;
3592 bool force_flush = false;
3594 WARN_ON(!is_vm_hugetlb_page(vma));
3595 BUG_ON(start & ~huge_page_mask(h));
3596 BUG_ON(end & ~huge_page_mask(h));
3599 * This is a hugetlb vma, all the pte entries should point
3602 tlb_change_page_size(tlb, sz);
3603 tlb_start_vma(tlb, vma);
3606 * If sharing possible, alert mmu notifiers of worst case.
3608 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3610 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3611 mmu_notifier_invalidate_range_start(&range);
3613 for (; address < end; address += sz) {
3614 ptep = huge_pte_offset(mm, address, sz);
3618 ptl = huge_pte_lock(h, mm, ptep);
3619 if (huge_pmd_unshare(mm, &address, ptep)) {
3621 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
3626 pte = huge_ptep_get(ptep);
3627 if (huge_pte_none(pte)) {
3633 * Migrating hugepage or HWPoisoned hugepage is already
3634 * unmapped and its refcount is dropped, so just clear pte here.
3636 if (unlikely(!pte_present(pte))) {
3637 huge_pte_clear(mm, address, ptep, sz);
3642 page = pte_page(pte);
3644 * If a reference page is supplied, it is because a specific
3645 * page is being unmapped, not a range. Ensure the page we
3646 * are about to unmap is the actual page of interest.
3649 if (page != ref_page) {
3654 * Mark the VMA as having unmapped its page so that
3655 * future faults in this VMA will fail rather than
3656 * looking like data was lost
3658 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3661 pte = huge_ptep_get_and_clear(mm, address, ptep);
3662 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3663 if (huge_pte_dirty(pte))
3664 set_page_dirty(page);
3666 hugetlb_count_sub(pages_per_huge_page(h), mm);
3667 page_remove_rmap(page, true);
3670 tlb_remove_page_size(tlb, page, huge_page_size(h));
3672 * Bail out after unmapping reference page if supplied
3677 mmu_notifier_invalidate_range_end(&range);
3678 tlb_end_vma(tlb, vma);
3681 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
3682 * could defer the flush until now, since by holding i_mmap_rwsem we
3683 * guaranteed that the last refernece would not be dropped. But we must
3684 * do the flushing before we return, as otherwise i_mmap_rwsem will be
3685 * dropped and the last reference to the shared PMDs page might be
3688 * In theory we could defer the freeing of the PMD pages as well, but
3689 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
3690 * detect sharing, so we cannot defer the release of the page either.
3691 * Instead, do flush now.
3694 tlb_flush_mmu_tlbonly(tlb);
3697 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3698 struct vm_area_struct *vma, unsigned long start,
3699 unsigned long end, struct page *ref_page)
3701 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3704 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3705 * test will fail on a vma being torn down, and not grab a page table
3706 * on its way out. We're lucky that the flag has such an appropriate
3707 * name, and can in fact be safely cleared here. We could clear it
3708 * before the __unmap_hugepage_range above, but all that's necessary
3709 * is to clear it before releasing the i_mmap_rwsem. This works
3710 * because in the context this is called, the VMA is about to be
3711 * destroyed and the i_mmap_rwsem is held.
3713 vma->vm_flags &= ~VM_MAYSHARE;
3716 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3717 unsigned long end, struct page *ref_page)
3719 struct mm_struct *mm;
3720 struct mmu_gather tlb;
3721 unsigned long tlb_start = start;
3722 unsigned long tlb_end = end;
3725 * If shared PMDs were possibly used within this vma range, adjust
3726 * start/end for worst case tlb flushing.
3727 * Note that we can not be sure if PMDs are shared until we try to
3728 * unmap pages. However, we want to make sure TLB flushing covers
3729 * the largest possible range.
3731 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
3735 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
3736 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3737 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
3741 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3742 * mappping it owns the reserve page for. The intention is to unmap the page
3743 * from other VMAs and let the children be SIGKILLed if they are faulting the
3746 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3747 struct page *page, unsigned long address)
3749 struct hstate *h = hstate_vma(vma);
3750 struct vm_area_struct *iter_vma;
3751 struct address_space *mapping;
3755 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3756 * from page cache lookup which is in HPAGE_SIZE units.
3758 address = address & huge_page_mask(h);
3759 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3761 mapping = vma->vm_file->f_mapping;
3764 * Take the mapping lock for the duration of the table walk. As
3765 * this mapping should be shared between all the VMAs,
3766 * __unmap_hugepage_range() is called as the lock is already held
3768 i_mmap_lock_write(mapping);
3769 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3770 /* Do not unmap the current VMA */
3771 if (iter_vma == vma)
3775 * Shared VMAs have their own reserves and do not affect
3776 * MAP_PRIVATE accounting but it is possible that a shared
3777 * VMA is using the same page so check and skip such VMAs.
3779 if (iter_vma->vm_flags & VM_MAYSHARE)
3783 * Unmap the page from other VMAs without their own reserves.
3784 * They get marked to be SIGKILLed if they fault in these
3785 * areas. This is because a future no-page fault on this VMA
3786 * could insert a zeroed page instead of the data existing
3787 * from the time of fork. This would look like data corruption
3789 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3790 unmap_hugepage_range(iter_vma, address,
3791 address + huge_page_size(h), page);
3793 i_mmap_unlock_write(mapping);
3797 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3798 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3799 * cannot race with other handlers or page migration.
3800 * Keep the pte_same checks anyway to make transition from the mutex easier.
3802 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3803 unsigned long address, pte_t *ptep,
3804 struct page *pagecache_page, spinlock_t *ptl)
3807 struct hstate *h = hstate_vma(vma);
3808 struct page *old_page, *new_page;
3809 int outside_reserve = 0;
3811 unsigned long haddr = address & huge_page_mask(h);
3812 struct mmu_notifier_range range;
3814 pte = huge_ptep_get(ptep);
3815 old_page = pte_page(pte);
3818 /* If no-one else is actually using this page, avoid the copy
3819 * and just make the page writable */
3820 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3821 page_move_anon_rmap(old_page, vma);
3822 set_huge_ptep_writable(vma, haddr, ptep);
3827 * If the process that created a MAP_PRIVATE mapping is about to
3828 * perform a COW due to a shared page count, attempt to satisfy
3829 * the allocation without using the existing reserves. The pagecache
3830 * page is used to determine if the reserve at this address was
3831 * consumed or not. If reserves were used, a partial faulted mapping
3832 * at the time of fork() could consume its reserves on COW instead
3833 * of the full address range.
3835 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3836 old_page != pagecache_page)
3837 outside_reserve = 1;
3842 * Drop page table lock as buddy allocator may be called. It will
3843 * be acquired again before returning to the caller, as expected.
3846 new_page = alloc_huge_page(vma, haddr, outside_reserve);
3848 if (IS_ERR(new_page)) {
3850 * If a process owning a MAP_PRIVATE mapping fails to COW,
3851 * it is due to references held by a child and an insufficient
3852 * huge page pool. To guarantee the original mappers
3853 * reliability, unmap the page from child processes. The child
3854 * may get SIGKILLed if it later faults.
3856 if (outside_reserve) {
3858 BUG_ON(huge_pte_none(pte));
3859 unmap_ref_private(mm, vma, old_page, haddr);
3860 BUG_ON(huge_pte_none(pte));
3862 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3864 pte_same(huge_ptep_get(ptep), pte)))
3865 goto retry_avoidcopy;
3867 * race occurs while re-acquiring page table
3868 * lock, and our job is done.
3873 ret = vmf_error(PTR_ERR(new_page));
3874 goto out_release_old;
3878 * When the original hugepage is shared one, it does not have
3879 * anon_vma prepared.
3881 if (unlikely(anon_vma_prepare(vma))) {
3883 goto out_release_all;
3886 copy_user_huge_page(new_page, old_page, address, vma,
3887 pages_per_huge_page(h));
3888 __SetPageUptodate(new_page);
3890 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
3891 haddr + huge_page_size(h));
3892 mmu_notifier_invalidate_range_start(&range);
3895 * Retake the page table lock to check for racing updates
3896 * before the page tables are altered
3899 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3900 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3901 ClearPagePrivate(new_page);
3904 huge_ptep_clear_flush(vma, haddr, ptep);
3905 mmu_notifier_invalidate_range(mm, range.start, range.end);
3906 set_huge_pte_at(mm, haddr, ptep,
3907 make_huge_pte(vma, new_page, 1));
3908 page_remove_rmap(old_page, true);
3909 hugepage_add_new_anon_rmap(new_page, vma, haddr);
3910 set_page_huge_active(new_page);
3911 /* Make the old page be freed below */
3912 new_page = old_page;
3915 mmu_notifier_invalidate_range_end(&range);
3917 restore_reserve_on_error(h, vma, haddr, new_page);
3922 spin_lock(ptl); /* Caller expects lock to be held */
3926 /* Return the pagecache page at a given address within a VMA */
3927 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3928 struct vm_area_struct *vma, unsigned long address)
3930 struct address_space *mapping;
3933 mapping = vma->vm_file->f_mapping;
3934 idx = vma_hugecache_offset(h, vma, address);
3936 return find_lock_page(mapping, idx);
3940 * Return whether there is a pagecache page to back given address within VMA.
3941 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3943 static bool hugetlbfs_pagecache_present(struct hstate *h,
3944 struct vm_area_struct *vma, unsigned long address)
3946 struct address_space *mapping;
3950 mapping = vma->vm_file->f_mapping;
3951 idx = vma_hugecache_offset(h, vma, address);
3953 page = find_get_page(mapping, idx);
3956 return page != NULL;
3959 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3962 struct inode *inode = mapping->host;
3963 struct hstate *h = hstate_inode(inode);
3964 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3968 ClearPagePrivate(page);
3971 * set page dirty so that it will not be removed from cache/file
3972 * by non-hugetlbfs specific code paths.
3974 set_page_dirty(page);
3976 spin_lock(&inode->i_lock);
3977 inode->i_blocks += blocks_per_huge_page(h);
3978 spin_unlock(&inode->i_lock);
3982 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
3983 struct vm_area_struct *vma,
3984 struct address_space *mapping, pgoff_t idx,
3985 unsigned long address, pte_t *ptep, unsigned int flags)
3987 struct hstate *h = hstate_vma(vma);
3988 vm_fault_t ret = VM_FAULT_SIGBUS;
3994 unsigned long haddr = address & huge_page_mask(h);
3995 bool new_page = false;
3998 * Currently, we are forced to kill the process in the event the
3999 * original mapper has unmapped pages from the child due to a failed
4000 * COW. Warn that such a situation has occurred as it may not be obvious
4002 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4003 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4009 * Use page lock to guard against racing truncation
4010 * before we get page_table_lock.
4013 page = find_lock_page(mapping, idx);
4015 size = i_size_read(mapping->host) >> huge_page_shift(h);
4020 * Check for page in userfault range
4022 if (userfaultfd_missing(vma)) {
4024 struct vm_fault vmf = {
4029 * Hard to debug if it ends up being
4030 * used by a callee that assumes
4031 * something about the other
4032 * uninitialized fields... same as in
4038 * hugetlb_fault_mutex must be dropped before
4039 * handling userfault. Reacquire after handling
4040 * fault to make calling code simpler.
4042 hash = hugetlb_fault_mutex_hash(h, mapping, idx);
4043 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4044 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
4045 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4049 page = alloc_huge_page(vma, haddr, 0);
4052 * Returning error will result in faulting task being
4053 * sent SIGBUS. The hugetlb fault mutex prevents two
4054 * tasks from racing to fault in the same page which
4055 * could result in false unable to allocate errors.
4056 * Page migration does not take the fault mutex, but
4057 * does a clear then write of pte's under page table
4058 * lock. Page fault code could race with migration,
4059 * notice the clear pte and try to allocate a page
4060 * here. Before returning error, get ptl and make
4061 * sure there really is no pte entry.
4063 ptl = huge_pte_lock(h, mm, ptep);
4064 if (!huge_pte_none(huge_ptep_get(ptep))) {
4070 ret = vmf_error(PTR_ERR(page));
4073 clear_huge_page(page, address, pages_per_huge_page(h));
4074 __SetPageUptodate(page);
4077 if (vma->vm_flags & VM_MAYSHARE) {
4078 int err = huge_add_to_page_cache(page, mapping, idx);
4087 if (unlikely(anon_vma_prepare(vma))) {
4089 goto backout_unlocked;
4095 * If memory error occurs between mmap() and fault, some process
4096 * don't have hwpoisoned swap entry for errored virtual address.
4097 * So we need to block hugepage fault by PG_hwpoison bit check.
4099 if (unlikely(PageHWPoison(page))) {
4100 ret = VM_FAULT_HWPOISON_LARGE |
4101 VM_FAULT_SET_HINDEX(hstate_index(h));
4102 goto backout_unlocked;
4107 * If we are going to COW a private mapping later, we examine the
4108 * pending reservations for this page now. This will ensure that
4109 * any allocations necessary to record that reservation occur outside
4112 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4113 if (vma_needs_reservation(h, vma, haddr) < 0) {
4115 goto backout_unlocked;
4117 /* Just decrements count, does not deallocate */
4118 vma_end_reservation(h, vma, haddr);
4121 ptl = huge_pte_lock(h, mm, ptep);
4122 size = i_size_read(mapping->host) >> huge_page_shift(h);
4127 if (!huge_pte_none(huge_ptep_get(ptep)))
4131 ClearPagePrivate(page);
4132 hugepage_add_new_anon_rmap(page, vma, haddr);
4134 page_dup_rmap(page, true);
4135 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4136 && (vma->vm_flags & VM_SHARED)));
4137 set_huge_pte_at(mm, haddr, ptep, new_pte);
4139 hugetlb_count_add(pages_per_huge_page(h), mm);
4140 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4141 /* Optimization, do the COW without a second fault */
4142 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4148 * Only make newly allocated pages active. Existing pages found
4149 * in the pagecache could be !page_huge_active() if they have been
4150 * isolated for migration.
4153 set_page_huge_active(page);
4163 restore_reserve_on_error(h, vma, haddr, page);
4169 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping,
4172 unsigned long key[2];
4175 key[0] = (unsigned long) mapping;
4178 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4180 return hash & (num_fault_mutexes - 1);
4184 * For uniprocesor systems we always use a single mutex, so just
4185 * return 0 and avoid the hashing overhead.
4187 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping,
4194 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4195 unsigned long address, unsigned int flags)
4202 struct page *page = NULL;
4203 struct page *pagecache_page = NULL;
4204 struct hstate *h = hstate_vma(vma);
4205 struct address_space *mapping;
4206 int need_wait_lock = 0;
4207 unsigned long haddr = address & huge_page_mask(h);
4209 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4211 entry = huge_ptep_get(ptep);
4212 if (unlikely(is_hugetlb_entry_migration(entry))) {
4213 migration_entry_wait_huge(vma, mm, ptep);
4215 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4216 return VM_FAULT_HWPOISON_LARGE |
4217 VM_FAULT_SET_HINDEX(hstate_index(h));
4219 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4221 return VM_FAULT_OOM;
4224 mapping = vma->vm_file->f_mapping;
4225 idx = vma_hugecache_offset(h, vma, haddr);
4228 * Serialize hugepage allocation and instantiation, so that we don't
4229 * get spurious allocation failures if two CPUs race to instantiate
4230 * the same page in the page cache.
4232 hash = hugetlb_fault_mutex_hash(h, mapping, idx);
4233 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4235 entry = huge_ptep_get(ptep);
4236 if (huge_pte_none(entry)) {
4237 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4244 * entry could be a migration/hwpoison entry at this point, so this
4245 * check prevents the kernel from going below assuming that we have
4246 * a active hugepage in pagecache. This goto expects the 2nd page fault,
4247 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
4250 if (!pte_present(entry))
4254 * If we are going to COW the mapping later, we examine the pending
4255 * reservations for this page now. This will ensure that any
4256 * allocations necessary to record that reservation occur outside the
4257 * spinlock. For private mappings, we also lookup the pagecache
4258 * page now as it is used to determine if a reservation has been
4261 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4262 if (vma_needs_reservation(h, vma, haddr) < 0) {
4266 /* Just decrements count, does not deallocate */
4267 vma_end_reservation(h, vma, haddr);
4269 if (!(vma->vm_flags & VM_MAYSHARE))
4270 pagecache_page = hugetlbfs_pagecache_page(h,
4274 ptl = huge_pte_lock(h, mm, ptep);
4276 /* Check for a racing update before calling hugetlb_cow */
4277 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4281 * hugetlb_cow() requires page locks of pte_page(entry) and
4282 * pagecache_page, so here we need take the former one
4283 * when page != pagecache_page or !pagecache_page.
4285 page = pte_page(entry);
4286 if (page != pagecache_page)
4287 if (!trylock_page(page)) {
4294 if (flags & FAULT_FLAG_WRITE) {
4295 if (!huge_pte_write(entry)) {
4296 ret = hugetlb_cow(mm, vma, address, ptep,
4297 pagecache_page, ptl);
4300 entry = huge_pte_mkdirty(entry);
4302 entry = pte_mkyoung(entry);
4303 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4304 flags & FAULT_FLAG_WRITE))
4305 update_mmu_cache(vma, haddr, ptep);
4307 if (page != pagecache_page)
4313 if (pagecache_page) {
4314 unlock_page(pagecache_page);
4315 put_page(pagecache_page);
4318 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4320 * Generally it's safe to hold refcount during waiting page lock. But
4321 * here we just wait to defer the next page fault to avoid busy loop and
4322 * the page is not used after unlocked before returning from the current
4323 * page fault. So we are safe from accessing freed page, even if we wait
4324 * here without taking refcount.
4327 wait_on_page_locked(page);
4332 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4333 * modifications for huge pages.
4335 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4337 struct vm_area_struct *dst_vma,
4338 unsigned long dst_addr,
4339 unsigned long src_addr,
4340 struct page **pagep)
4342 struct address_space *mapping;
4345 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4346 struct hstate *h = hstate_vma(dst_vma);
4353 /* If a page already exists, then it's UFFDIO_COPY for
4354 * a non-missing case. Return -EEXIST.
4357 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
4362 page = alloc_huge_page(dst_vma, dst_addr, 0);
4368 ret = copy_huge_page_from_user(page,
4369 (const void __user *) src_addr,
4370 pages_per_huge_page(h), false);
4372 /* fallback to copy_from_user outside mmap_sem */
4373 if (unlikely(ret)) {
4376 /* don't free the page */
4385 * The memory barrier inside __SetPageUptodate makes sure that
4386 * preceding stores to the page contents become visible before
4387 * the set_pte_at() write.
4389 __SetPageUptodate(page);
4391 mapping = dst_vma->vm_file->f_mapping;
4392 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4395 * If shared, add to page cache
4398 size = i_size_read(mapping->host) >> huge_page_shift(h);
4401 goto out_release_nounlock;
4404 * Serialization between remove_inode_hugepages() and
4405 * huge_add_to_page_cache() below happens through the
4406 * hugetlb_fault_mutex_table that here must be hold by
4409 ret = huge_add_to_page_cache(page, mapping, idx);
4411 goto out_release_nounlock;
4414 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4418 * Recheck the i_size after holding PT lock to make sure not
4419 * to leave any page mapped (as page_mapped()) beyond the end
4420 * of the i_size (remove_inode_hugepages() is strict about
4421 * enforcing that). If we bail out here, we'll also leave a
4422 * page in the radix tree in the vm_shared case beyond the end
4423 * of the i_size, but remove_inode_hugepages() will take care
4424 * of it as soon as we drop the hugetlb_fault_mutex_table.
4426 size = i_size_read(mapping->host) >> huge_page_shift(h);
4429 goto out_release_unlock;
4432 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4433 goto out_release_unlock;
4436 page_dup_rmap(page, true);
4438 ClearPagePrivate(page);
4439 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4442 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4443 if (dst_vma->vm_flags & VM_WRITE)
4444 _dst_pte = huge_pte_mkdirty(_dst_pte);
4445 _dst_pte = pte_mkyoung(_dst_pte);
4447 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4449 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4450 dst_vma->vm_flags & VM_WRITE);
4451 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4453 /* No need to invalidate - it was non-present before */
4454 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4457 set_page_huge_active(page);
4467 out_release_nounlock:
4472 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4473 struct page **pages, struct vm_area_struct **vmas,
4474 unsigned long *position, unsigned long *nr_pages,
4475 long i, unsigned int flags, int *nonblocking)
4477 unsigned long pfn_offset;
4478 unsigned long vaddr = *position;
4479 unsigned long remainder = *nr_pages;
4480 struct hstate *h = hstate_vma(vma);
4483 while (vaddr < vma->vm_end && remainder) {
4485 spinlock_t *ptl = NULL;
4490 * If we have a pending SIGKILL, don't keep faulting pages and
4491 * potentially allocating memory.
4493 if (fatal_signal_pending(current)) {
4499 * Some archs (sparc64, sh*) have multiple pte_ts to
4500 * each hugepage. We have to make sure we get the
4501 * first, for the page indexing below to work.
4503 * Note that page table lock is not held when pte is null.
4505 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4508 ptl = huge_pte_lock(h, mm, pte);
4509 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4512 * When coredumping, it suits get_dump_page if we just return
4513 * an error where there's an empty slot with no huge pagecache
4514 * to back it. This way, we avoid allocating a hugepage, and
4515 * the sparse dumpfile avoids allocating disk blocks, but its
4516 * huge holes still show up with zeroes where they need to be.
4518 if (absent && (flags & FOLL_DUMP) &&
4519 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4527 * We need call hugetlb_fault for both hugepages under migration
4528 * (in which case hugetlb_fault waits for the migration,) and
4529 * hwpoisoned hugepages (in which case we need to prevent the
4530 * caller from accessing to them.) In order to do this, we use
4531 * here is_swap_pte instead of is_hugetlb_entry_migration and
4532 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4533 * both cases, and because we can't follow correct pages
4534 * directly from any kind of swap entries.
4536 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4537 ((flags & FOLL_WRITE) &&
4538 !huge_pte_write(huge_ptep_get(pte)))) {
4540 unsigned int fault_flags = 0;
4544 if (flags & FOLL_WRITE)
4545 fault_flags |= FAULT_FLAG_WRITE;
4547 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4548 if (flags & FOLL_NOWAIT)
4549 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4550 FAULT_FLAG_RETRY_NOWAIT;
4551 if (flags & FOLL_TRIED) {
4552 VM_WARN_ON_ONCE(fault_flags &
4553 FAULT_FLAG_ALLOW_RETRY);
4554 fault_flags |= FAULT_FLAG_TRIED;
4556 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4557 if (ret & VM_FAULT_ERROR) {
4558 err = vm_fault_to_errno(ret, flags);
4562 if (ret & VM_FAULT_RETRY) {
4564 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4568 * VM_FAULT_RETRY must not return an
4569 * error, it will return zero
4572 * No need to update "position" as the
4573 * caller will not check it after
4574 * *nr_pages is set to 0.
4581 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4582 page = pte_page(huge_ptep_get(pte));
4585 * Instead of doing 'try_get_page()' below in the same_page
4586 * loop, just check the count once here.
4588 if (unlikely(page_count(page) <= 0)) {
4598 pages[i] = mem_map_offset(page, pfn_offset);
4609 if (vaddr < vma->vm_end && remainder &&
4610 pfn_offset < pages_per_huge_page(h)) {
4612 * We use pfn_offset to avoid touching the pageframes
4613 * of this compound page.
4619 *nr_pages = remainder;
4621 * setting position is actually required only if remainder is
4622 * not zero but it's faster not to add a "if (remainder)"
4630 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4632 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4635 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4638 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4639 unsigned long address, unsigned long end, pgprot_t newprot)
4641 struct mm_struct *mm = vma->vm_mm;
4642 unsigned long start = address;
4645 struct hstate *h = hstate_vma(vma);
4646 unsigned long pages = 0;
4647 bool shared_pmd = false;
4648 struct mmu_notifier_range range;
4651 * In the case of shared PMDs, the area to flush could be beyond
4652 * start/end. Set range.start/range.end to cover the maximum possible
4653 * range if PMD sharing is possible.
4655 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
4656 0, vma, mm, start, end);
4657 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4659 BUG_ON(address >= end);
4660 flush_cache_range(vma, range.start, range.end);
4662 mmu_notifier_invalidate_range_start(&range);
4663 i_mmap_lock_write(vma->vm_file->f_mapping);
4664 for (; address < end; address += huge_page_size(h)) {
4666 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4669 ptl = huge_pte_lock(h, mm, ptep);
4670 if (huge_pmd_unshare(mm, &address, ptep)) {
4676 pte = huge_ptep_get(ptep);
4677 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4681 if (unlikely(is_hugetlb_entry_migration(pte))) {
4682 swp_entry_t entry = pte_to_swp_entry(pte);
4684 if (is_write_migration_entry(entry)) {
4687 make_migration_entry_read(&entry);
4688 newpte = swp_entry_to_pte(entry);
4689 set_huge_swap_pte_at(mm, address, ptep,
4690 newpte, huge_page_size(h));
4696 if (!huge_pte_none(pte)) {
4699 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
4700 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
4701 pte = arch_make_huge_pte(pte, vma, NULL, 0);
4702 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
4708 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4709 * may have cleared our pud entry and done put_page on the page table:
4710 * once we release i_mmap_rwsem, another task can do the final put_page
4711 * and that page table be reused and filled with junk. If we actually
4712 * did unshare a page of pmds, flush the range corresponding to the pud.
4715 flush_hugetlb_tlb_range(vma, range.start, range.end);
4717 flush_hugetlb_tlb_range(vma, start, end);
4719 * No need to call mmu_notifier_invalidate_range() we are downgrading
4720 * page table protection not changing it to point to a new page.
4722 * See Documentation/vm/mmu_notifier.rst
4724 i_mmap_unlock_write(vma->vm_file->f_mapping);
4725 mmu_notifier_invalidate_range_end(&range);
4727 return pages << h->order;
4730 int hugetlb_reserve_pages(struct inode *inode,
4732 struct vm_area_struct *vma,
4733 vm_flags_t vm_flags)
4736 struct hstate *h = hstate_inode(inode);
4737 struct hugepage_subpool *spool = subpool_inode(inode);
4738 struct resv_map *resv_map;
4741 /* This should never happen */
4743 VM_WARN(1, "%s called with a negative range\n", __func__);
4748 * Only apply hugepage reservation if asked. At fault time, an
4749 * attempt will be made for VM_NORESERVE to allocate a page
4750 * without using reserves
4752 if (vm_flags & VM_NORESERVE)
4756 * Shared mappings base their reservation on the number of pages that
4757 * are already allocated on behalf of the file. Private mappings need
4758 * to reserve the full area even if read-only as mprotect() may be
4759 * called to make the mapping read-write. Assume !vma is a shm mapping
4761 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4763 * resv_map can not be NULL as hugetlb_reserve_pages is only
4764 * called for inodes for which resv_maps were created (see
4765 * hugetlbfs_get_inode).
4767 resv_map = inode_resv_map(inode);
4769 chg = region_chg(resv_map, from, to);
4772 resv_map = resv_map_alloc();
4778 set_vma_resv_map(vma, resv_map);
4779 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4788 * There must be enough pages in the subpool for the mapping. If
4789 * the subpool has a minimum size, there may be some global
4790 * reservations already in place (gbl_reserve).
4792 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4793 if (gbl_reserve < 0) {
4799 * Check enough hugepages are available for the reservation.
4800 * Hand the pages back to the subpool if there are not
4802 ret = hugetlb_acct_memory(h, gbl_reserve);
4804 /* put back original number of pages, chg */
4805 (void)hugepage_subpool_put_pages(spool, chg);
4810 * Account for the reservations made. Shared mappings record regions
4811 * that have reservations as they are shared by multiple VMAs.
4812 * When the last VMA disappears, the region map says how much
4813 * the reservation was and the page cache tells how much of
4814 * the reservation was consumed. Private mappings are per-VMA and
4815 * only the consumed reservations are tracked. When the VMA
4816 * disappears, the original reservation is the VMA size and the
4817 * consumed reservations are stored in the map. Hence, nothing
4818 * else has to be done for private mappings here
4820 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4821 long add = region_add(resv_map, from, to);
4823 if (unlikely(chg > add)) {
4825 * pages in this range were added to the reserve
4826 * map between region_chg and region_add. This
4827 * indicates a race with alloc_huge_page. Adjust
4828 * the subpool and reserve counts modified above
4829 * based on the difference.
4833 rsv_adjust = hugepage_subpool_put_pages(spool,
4835 hugetlb_acct_memory(h, -rsv_adjust);
4840 if (!vma || vma->vm_flags & VM_MAYSHARE)
4841 /* Don't call region_abort if region_chg failed */
4843 region_abort(resv_map, from, to);
4844 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4845 kref_put(&resv_map->refs, resv_map_release);
4849 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4852 struct hstate *h = hstate_inode(inode);
4853 struct resv_map *resv_map = inode_resv_map(inode);
4855 struct hugepage_subpool *spool = subpool_inode(inode);
4859 * Since this routine can be called in the evict inode path for all
4860 * hugetlbfs inodes, resv_map could be NULL.
4863 chg = region_del(resv_map, start, end);
4865 * region_del() can fail in the rare case where a region
4866 * must be split and another region descriptor can not be
4867 * allocated. If end == LONG_MAX, it will not fail.
4873 spin_lock(&inode->i_lock);
4874 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4875 spin_unlock(&inode->i_lock);
4878 * If the subpool has a minimum size, the number of global
4879 * reservations to be released may be adjusted.
4881 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4882 hugetlb_acct_memory(h, -gbl_reserve);
4887 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4888 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4889 struct vm_area_struct *vma,
4890 unsigned long addr, pgoff_t idx)
4892 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4894 unsigned long sbase = saddr & PUD_MASK;
4895 unsigned long s_end = sbase + PUD_SIZE;
4897 /* Allow segments to share if only one is marked locked */
4898 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4899 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4902 * match the virtual addresses, permission and the alignment of the
4905 if (pmd_index(addr) != pmd_index(saddr) ||
4906 vm_flags != svm_flags ||
4907 sbase < svma->vm_start || svma->vm_end < s_end)
4913 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4915 unsigned long base = addr & PUD_MASK;
4916 unsigned long end = base + PUD_SIZE;
4919 * check on proper vm_flags and page table alignment
4921 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
4927 * Determine if start,end range within vma could be mapped by shared pmd.
4928 * If yes, adjust start and end to cover range associated with possible
4929 * shared pmd mappings.
4931 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4932 unsigned long *start, unsigned long *end)
4934 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
4935 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
4938 * vma need span at least one aligned PUD size and the start,end range
4939 * must at least partialy within it.
4941 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
4942 (*end <= v_start) || (*start >= v_end))
4945 /* Extend the range to be PUD aligned for a worst case scenario */
4946 if (*start > v_start)
4947 *start = ALIGN_DOWN(*start, PUD_SIZE);
4950 *end = ALIGN(*end, PUD_SIZE);
4954 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4955 * and returns the corresponding pte. While this is not necessary for the
4956 * !shared pmd case because we can allocate the pmd later as well, it makes the
4957 * code much cleaner. pmd allocation is essential for the shared case because
4958 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4959 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4960 * bad pmd for sharing.
4962 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4964 struct vm_area_struct *vma = find_vma(mm, addr);
4965 struct address_space *mapping = vma->vm_file->f_mapping;
4966 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4968 struct vm_area_struct *svma;
4969 unsigned long saddr;
4974 if (!vma_shareable(vma, addr))
4975 return (pte_t *)pmd_alloc(mm, pud, addr);
4977 i_mmap_lock_write(mapping);
4978 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4982 saddr = page_table_shareable(svma, vma, addr, idx);
4984 spte = huge_pte_offset(svma->vm_mm, saddr,
4985 vma_mmu_pagesize(svma));
4987 get_page(virt_to_page(spte));
4996 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4997 if (pud_none(*pud)) {
4998 pud_populate(mm, pud,
4999 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5002 put_page(virt_to_page(spte));
5006 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5007 i_mmap_unlock_write(mapping);
5012 * unmap huge page backed by shared pte.
5014 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5015 * indicated by page_count > 1, unmap is achieved by clearing pud and
5016 * decrementing the ref count. If count == 1, the pte page is not shared.
5018 * called with page table lock held.
5020 * returns: 1 successfully unmapped a shared pte page
5021 * 0 the underlying pte page is not shared, or it is the last user
5023 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
5025 pgd_t *pgd = pgd_offset(mm, *addr);
5026 p4d_t *p4d = p4d_offset(pgd, *addr);
5027 pud_t *pud = pud_offset(p4d, *addr);
5029 BUG_ON(page_count(virt_to_page(ptep)) == 0);
5030 if (page_count(virt_to_page(ptep)) == 1)
5034 put_page(virt_to_page(ptep));
5037 * This update of passed address optimizes loops sequentially
5038 * processing addresses in increments of huge page size (PMD_SIZE
5039 * in this case). By clearing the pud, a PUD_SIZE area is unmapped.
5040 * Update address to the 'last page' in the cleared area so that
5041 * calling loop can move to first page past this area.
5043 *addr |= PUD_SIZE - PMD_SIZE;
5046 #define want_pmd_share() (1)
5047 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5048 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5053 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
5058 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5059 unsigned long *start, unsigned long *end)
5062 #define want_pmd_share() (0)
5063 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5065 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5066 pte_t *huge_pte_alloc(struct mm_struct *mm,
5067 unsigned long addr, unsigned long sz)
5074 pgd = pgd_offset(mm, addr);
5075 p4d = p4d_alloc(mm, pgd, addr);
5078 pud = pud_alloc(mm, p4d, addr);
5080 if (sz == PUD_SIZE) {
5083 BUG_ON(sz != PMD_SIZE);
5084 if (want_pmd_share() && pud_none(*pud))
5085 pte = huge_pmd_share(mm, addr, pud);
5087 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5090 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
5096 * huge_pte_offset() - Walk the page table to resolve the hugepage
5097 * entry at address @addr
5099 * Return: Pointer to page table or swap entry (PUD or PMD) for
5100 * address @addr, or NULL if a p*d_none() entry is encountered and the
5101 * size @sz doesn't match the hugepage size at this level of the page
5104 pte_t *huge_pte_offset(struct mm_struct *mm,
5105 unsigned long addr, unsigned long sz)
5109 pud_t *pud, pud_entry;
5110 pmd_t *pmd, pmd_entry;
5112 pgd = pgd_offset(mm, addr);
5113 if (!pgd_present(*pgd))
5115 p4d = p4d_offset(pgd, addr);
5116 if (!p4d_present(*p4d))
5119 pud = pud_offset(p4d, addr);
5120 pud_entry = READ_ONCE(*pud);
5121 if (sz != PUD_SIZE && pud_none(pud_entry))
5123 /* hugepage or swap? */
5124 if (pud_huge(pud_entry) || !pud_present(pud_entry))
5125 return (pte_t *)pud;
5127 pmd = pmd_offset(pud, addr);
5128 pmd_entry = READ_ONCE(*pmd);
5129 if (sz != PMD_SIZE && pmd_none(pmd_entry))
5131 /* hugepage or swap? */
5132 if (pmd_huge(pmd_entry) || !pmd_present(pmd_entry))
5133 return (pte_t *)pmd;
5138 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5141 * These functions are overwritable if your architecture needs its own
5144 struct page * __weak
5145 follow_huge_addr(struct mm_struct *mm, unsigned long address,
5148 return ERR_PTR(-EINVAL);
5151 struct page * __weak
5152 follow_huge_pd(struct vm_area_struct *vma,
5153 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5155 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5159 struct page * __weak
5160 follow_huge_pmd_pte(struct vm_area_struct *vma, unsigned long address, int flags)
5162 struct hstate *h = hstate_vma(vma);
5163 struct mm_struct *mm = vma->vm_mm;
5164 struct page *page = NULL;
5169 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5173 ptl = huge_pte_lock(h, mm, ptep);
5174 pte = huge_ptep_get(ptep);
5175 if (pte_present(pte)) {
5176 page = pte_page(pte) +
5177 ((address & ~huge_page_mask(h)) >> PAGE_SHIFT);
5178 if (flags & FOLL_GET)
5181 if (is_hugetlb_entry_migration(pte)) {
5183 __migration_entry_wait(mm, ptep, ptl);
5187 * hwpoisoned entry is treated as no_page_table in
5188 * follow_page_mask().
5196 struct page * __weak
5197 follow_huge_pud(struct mm_struct *mm, unsigned long address,
5198 pud_t *pud, int flags)
5200 if (flags & FOLL_GET)
5203 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5206 struct page * __weak
5207 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5209 if (flags & FOLL_GET)
5212 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5215 bool isolate_huge_page(struct page *page, struct list_head *list)
5219 spin_lock(&hugetlb_lock);
5220 if (!PageHeadHuge(page) || !page_huge_active(page) ||
5221 !get_page_unless_zero(page)) {
5225 clear_page_huge_active(page);
5226 list_move_tail(&page->lru, list);
5228 spin_unlock(&hugetlb_lock);
5232 void putback_active_hugepage(struct page *page)
5234 VM_BUG_ON_PAGE(!PageHead(page), page);
5235 spin_lock(&hugetlb_lock);
5236 set_page_huge_active(page);
5237 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5238 spin_unlock(&hugetlb_lock);
5242 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5244 struct hstate *h = page_hstate(oldpage);
5246 hugetlb_cgroup_migrate(oldpage, newpage);
5247 set_page_owner_migrate_reason(newpage, reason);
5250 * transfer temporary state of the new huge page. This is
5251 * reverse to other transitions because the newpage is going to
5252 * be final while the old one will be freed so it takes over
5253 * the temporary status.
5255 * Also note that we have to transfer the per-node surplus state
5256 * here as well otherwise the global surplus count will not match
5259 if (PageHugeTemporary(newpage)) {
5260 int old_nid = page_to_nid(oldpage);
5261 int new_nid = page_to_nid(newpage);
5263 SetPageHugeTemporary(oldpage);
5264 ClearPageHugeTemporary(newpage);
5266 spin_lock(&hugetlb_lock);
5267 if (h->surplus_huge_pages_node[old_nid]) {
5268 h->surplus_huge_pages_node[old_nid]--;
5269 h->surplus_huge_pages_node[new_nid]++;
5271 spin_unlock(&hugetlb_lock);