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>
32 #include <asm/pgtable.h>
36 #include <linux/hugetlb.h>
37 #include <linux/hugetlb_cgroup.h>
38 #include <linux/node.h>
39 #include <linux/userfaultfd_k.h>
40 #include <linux/page_owner.h>
43 int hugetlb_max_hstate __read_mostly;
44 unsigned int default_hstate_idx;
45 struct hstate hstates[HUGE_MAX_HSTATE];
47 * Minimum page order among possible hugepage sizes, set to a proper value
50 static unsigned int minimum_order __read_mostly = UINT_MAX;
52 __initdata LIST_HEAD(huge_boot_pages);
54 /* for command line parsing */
55 static struct hstate * __initdata parsed_hstate;
56 static unsigned long __initdata default_hstate_max_huge_pages;
57 static unsigned long __initdata default_hstate_size;
58 static bool __initdata parsed_valid_hugepagesz = true;
61 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
62 * free_huge_pages, and surplus_huge_pages.
64 DEFINE_SPINLOCK(hugetlb_lock);
67 * Serializes faults on the same logical page. This is used to
68 * prevent spurious OOMs when the hugepage pool is fully utilized.
70 static int num_fault_mutexes;
71 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
73 /* Forward declaration */
74 static int hugetlb_acct_memory(struct hstate *h, long delta);
76 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
78 bool free = (spool->count == 0) && (spool->used_hpages == 0);
80 spin_unlock(&spool->lock);
82 /* If no pages are used, and no other handles to the subpool
83 * remain, give up any reservations mased on minimum size and
86 if (spool->min_hpages != -1)
87 hugetlb_acct_memory(spool->hstate,
93 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
96 struct hugepage_subpool *spool;
98 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
102 spin_lock_init(&spool->lock);
104 spool->max_hpages = max_hpages;
106 spool->min_hpages = min_hpages;
108 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
112 spool->rsv_hpages = min_hpages;
117 void hugepage_put_subpool(struct hugepage_subpool *spool)
119 spin_lock(&spool->lock);
120 BUG_ON(!spool->count);
122 unlock_or_release_subpool(spool);
126 * Subpool accounting for allocating and reserving pages.
127 * Return -ENOMEM if there are not enough resources to satisfy the
128 * the request. Otherwise, return the number of pages by which the
129 * global pools must be adjusted (upward). The returned value may
130 * only be different than the passed value (delta) in the case where
131 * a subpool minimum size must be manitained.
133 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
141 spin_lock(&spool->lock);
143 if (spool->max_hpages != -1) { /* maximum size accounting */
144 if ((spool->used_hpages + delta) <= spool->max_hpages)
145 spool->used_hpages += delta;
152 /* minimum size accounting */
153 if (spool->min_hpages != -1 && spool->rsv_hpages) {
154 if (delta > spool->rsv_hpages) {
156 * Asking for more reserves than those already taken on
157 * behalf of subpool. Return difference.
159 ret = delta - spool->rsv_hpages;
160 spool->rsv_hpages = 0;
162 ret = 0; /* reserves already accounted for */
163 spool->rsv_hpages -= delta;
168 spin_unlock(&spool->lock);
173 * Subpool accounting for freeing and unreserving pages.
174 * Return the number of global page reservations that must be dropped.
175 * The return value may only be different than the passed value (delta)
176 * in the case where a subpool minimum size must be maintained.
178 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
186 spin_lock(&spool->lock);
188 if (spool->max_hpages != -1) /* maximum size accounting */
189 spool->used_hpages -= delta;
191 /* minimum size accounting */
192 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
193 if (spool->rsv_hpages + delta <= spool->min_hpages)
196 ret = spool->rsv_hpages + delta - spool->min_hpages;
198 spool->rsv_hpages += delta;
199 if (spool->rsv_hpages > spool->min_hpages)
200 spool->rsv_hpages = spool->min_hpages;
204 * If hugetlbfs_put_super couldn't free spool due to an outstanding
205 * quota reference, free it now.
207 unlock_or_release_subpool(spool);
212 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
214 return HUGETLBFS_SB(inode->i_sb)->spool;
217 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
219 return subpool_inode(file_inode(vma->vm_file));
223 * Region tracking -- allows tracking of reservations and instantiated pages
224 * across the pages in a mapping.
226 * The region data structures are embedded into a resv_map and protected
227 * by a resv_map's lock. The set of regions within the resv_map represent
228 * reservations for huge pages, or huge pages that have already been
229 * instantiated within the map. The from and to elements are huge page
230 * indicies into the associated mapping. from indicates the starting index
231 * of the region. to represents the first index past the end of the region.
233 * For example, a file region structure with from == 0 and to == 4 represents
234 * four huge pages in a mapping. It is important to note that the to element
235 * represents the first element past the end of the region. This is used in
236 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
238 * Interval notation of the form [from, to) will be used to indicate that
239 * the endpoint from is inclusive and to is exclusive.
242 struct list_head link;
248 * Add the huge page range represented by [f, t) to the reserve
249 * map. In the normal case, existing regions will be expanded
250 * to accommodate the specified range. Sufficient regions should
251 * exist for expansion due to the previous call to region_chg
252 * with the same range. However, it is possible that region_del
253 * could have been called after region_chg and modifed the map
254 * in such a way that no region exists to be expanded. In this
255 * case, pull a region descriptor from the cache associated with
256 * the map and use that for the new range.
258 * Return the number of new huge pages added to the map. This
259 * number is greater than or equal to zero.
261 static long region_add(struct resv_map *resv, long f, long t)
263 struct list_head *head = &resv->regions;
264 struct file_region *rg, *nrg, *trg;
267 spin_lock(&resv->lock);
268 /* Locate the region we are either in or before. */
269 list_for_each_entry(rg, head, link)
274 * If no region exists which can be expanded to include the
275 * specified range, the list must have been modified by an
276 * interleving call to region_del(). Pull a region descriptor
277 * from the cache and use it for this range.
279 if (&rg->link == head || t < rg->from) {
280 VM_BUG_ON(resv->region_cache_count <= 0);
282 resv->region_cache_count--;
283 nrg = list_first_entry(&resv->region_cache, struct file_region,
285 list_del(&nrg->link);
289 list_add(&nrg->link, rg->link.prev);
295 /* Round our left edge to the current segment if it encloses us. */
299 /* Check for and consume any regions we now overlap with. */
301 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
302 if (&rg->link == head)
307 /* If this area reaches higher then extend our area to
308 * include it completely. If this is not the first area
309 * which we intend to reuse, free it. */
313 /* Decrement return value by the deleted range.
314 * Another range will span this area so that by
315 * end of routine add will be >= zero
317 add -= (rg->to - rg->from);
323 add += (nrg->from - f); /* Added to beginning of region */
325 add += t - nrg->to; /* Added to end of region */
329 resv->adds_in_progress--;
330 spin_unlock(&resv->lock);
336 * Examine the existing reserve map and determine how many
337 * huge pages in the specified range [f, t) are NOT currently
338 * represented. This routine is called before a subsequent
339 * call to region_add that will actually modify the reserve
340 * map to add the specified range [f, t). region_chg does
341 * not change the number of huge pages represented by the
342 * map. However, if the existing regions in the map can not
343 * be expanded to represent the new range, a new file_region
344 * structure is added to the map as a placeholder. This is
345 * so that the subsequent region_add call will have all the
346 * regions it needs and will not fail.
348 * Upon entry, region_chg will also examine the cache of region descriptors
349 * associated with the map. If there are not enough descriptors cached, one
350 * will be allocated for the in progress add operation.
352 * Returns the number of huge pages that need to be added to the existing
353 * reservation map for the range [f, t). This number is greater or equal to
354 * zero. -ENOMEM is returned if a new file_region structure or cache entry
355 * is needed and can not be allocated.
357 static long region_chg(struct resv_map *resv, long f, long t)
359 struct list_head *head = &resv->regions;
360 struct file_region *rg, *nrg = NULL;
364 spin_lock(&resv->lock);
366 resv->adds_in_progress++;
369 * Check for sufficient descriptors in the cache to accommodate
370 * the number of in progress add operations.
372 if (resv->adds_in_progress > resv->region_cache_count) {
373 struct file_region *trg;
375 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
376 /* Must drop lock to allocate a new descriptor. */
377 resv->adds_in_progress--;
378 spin_unlock(&resv->lock);
380 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
386 spin_lock(&resv->lock);
387 list_add(&trg->link, &resv->region_cache);
388 resv->region_cache_count++;
392 /* Locate the region we are before or in. */
393 list_for_each_entry(rg, head, link)
397 /* If we are below the current region then a new region is required.
398 * Subtle, allocate a new region at the position but make it zero
399 * size such that we can guarantee to record the reservation. */
400 if (&rg->link == head || t < rg->from) {
402 resv->adds_in_progress--;
403 spin_unlock(&resv->lock);
404 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
410 INIT_LIST_HEAD(&nrg->link);
414 list_add(&nrg->link, rg->link.prev);
419 /* Round our left edge to the current segment if it encloses us. */
424 /* Check for and consume any regions we now overlap with. */
425 list_for_each_entry(rg, rg->link.prev, link) {
426 if (&rg->link == head)
431 /* We overlap with this area, if it extends further than
432 * us then we must extend ourselves. Account for its
433 * existing reservation. */
438 chg -= rg->to - rg->from;
442 spin_unlock(&resv->lock);
443 /* We already know we raced and no longer need the new region */
447 spin_unlock(&resv->lock);
452 * Abort the in progress add operation. The adds_in_progress field
453 * of the resv_map keeps track of the operations in progress between
454 * calls to region_chg and region_add. Operations are sometimes
455 * aborted after the call to region_chg. In such cases, region_abort
456 * is called to decrement the adds_in_progress counter.
458 * NOTE: The range arguments [f, t) are not needed or used in this
459 * routine. They are kept to make reading the calling code easier as
460 * arguments will match the associated region_chg call.
462 static void region_abort(struct resv_map *resv, long f, long t)
464 spin_lock(&resv->lock);
465 VM_BUG_ON(!resv->region_cache_count);
466 resv->adds_in_progress--;
467 spin_unlock(&resv->lock);
471 * Delete the specified range [f, t) from the reserve map. If the
472 * t parameter is LONG_MAX, this indicates that ALL regions after f
473 * should be deleted. Locate the regions which intersect [f, t)
474 * and either trim, delete or split the existing regions.
476 * Returns the number of huge pages deleted from the reserve map.
477 * In the normal case, the return value is zero or more. In the
478 * case where a region must be split, a new region descriptor must
479 * be allocated. If the allocation fails, -ENOMEM will be returned.
480 * NOTE: If the parameter t == LONG_MAX, then we will never split
481 * a region and possibly return -ENOMEM. Callers specifying
482 * t == LONG_MAX do not need to check for -ENOMEM error.
484 static long region_del(struct resv_map *resv, long f, long t)
486 struct list_head *head = &resv->regions;
487 struct file_region *rg, *trg;
488 struct file_region *nrg = NULL;
492 spin_lock(&resv->lock);
493 list_for_each_entry_safe(rg, trg, head, link) {
495 * Skip regions before the range to be deleted. file_region
496 * ranges are normally of the form [from, to). However, there
497 * may be a "placeholder" entry in the map which is of the form
498 * (from, to) with from == to. Check for placeholder entries
499 * at the beginning of the range to be deleted.
501 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
507 if (f > rg->from && t < rg->to) { /* Must split region */
509 * Check for an entry in the cache before dropping
510 * lock and attempting allocation.
513 resv->region_cache_count > resv->adds_in_progress) {
514 nrg = list_first_entry(&resv->region_cache,
517 list_del(&nrg->link);
518 resv->region_cache_count--;
522 spin_unlock(&resv->lock);
523 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
531 /* New entry for end of split region */
534 INIT_LIST_HEAD(&nrg->link);
536 /* Original entry is trimmed */
539 list_add(&nrg->link, &rg->link);
544 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
545 del += rg->to - rg->from;
551 if (f <= rg->from) { /* Trim beginning of region */
554 } else { /* Trim end of region */
560 spin_unlock(&resv->lock);
566 * A rare out of memory error was encountered which prevented removal of
567 * the reserve map region for a page. The huge page itself was free'ed
568 * and removed from the page cache. This routine will adjust the subpool
569 * usage count, and the global reserve count if needed. By incrementing
570 * these counts, the reserve map entry which could not be deleted will
571 * appear as a "reserved" entry instead of simply dangling with incorrect
574 void hugetlb_fix_reserve_counts(struct inode *inode)
576 struct hugepage_subpool *spool = subpool_inode(inode);
579 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
581 struct hstate *h = hstate_inode(inode);
583 hugetlb_acct_memory(h, 1);
588 * Count and return the number of huge pages in the reserve map
589 * that intersect with the range [f, t).
591 static long region_count(struct resv_map *resv, long f, long t)
593 struct list_head *head = &resv->regions;
594 struct file_region *rg;
597 spin_lock(&resv->lock);
598 /* Locate each segment we overlap with, and count that overlap. */
599 list_for_each_entry(rg, head, link) {
608 seg_from = max(rg->from, f);
609 seg_to = min(rg->to, t);
611 chg += seg_to - seg_from;
613 spin_unlock(&resv->lock);
619 * Convert the address within this vma to the page offset within
620 * the mapping, in pagecache page units; huge pages here.
622 static pgoff_t vma_hugecache_offset(struct hstate *h,
623 struct vm_area_struct *vma, unsigned long address)
625 return ((address - vma->vm_start) >> huge_page_shift(h)) +
626 (vma->vm_pgoff >> huge_page_order(h));
629 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
630 unsigned long address)
632 return vma_hugecache_offset(hstate_vma(vma), vma, address);
634 EXPORT_SYMBOL_GPL(linear_hugepage_index);
637 * Return the size of the pages allocated when backing a VMA. In the majority
638 * cases this will be same size as used by the page table entries.
640 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
642 if (vma->vm_ops && vma->vm_ops->pagesize)
643 return vma->vm_ops->pagesize(vma);
646 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
649 * Return the page size being used by the MMU to back a VMA. In the majority
650 * of cases, the page size used by the kernel matches the MMU size. On
651 * architectures where it differs, an architecture-specific 'strong'
652 * version of this symbol is required.
654 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
656 return vma_kernel_pagesize(vma);
660 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
661 * bits of the reservation map pointer, which are always clear due to
664 #define HPAGE_RESV_OWNER (1UL << 0)
665 #define HPAGE_RESV_UNMAPPED (1UL << 1)
666 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
669 * These helpers are used to track how many pages are reserved for
670 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
671 * is guaranteed to have their future faults succeed.
673 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
674 * the reserve counters are updated with the hugetlb_lock held. It is safe
675 * to reset the VMA at fork() time as it is not in use yet and there is no
676 * chance of the global counters getting corrupted as a result of the values.
678 * The private mapping reservation is represented in a subtly different
679 * manner to a shared mapping. A shared mapping has a region map associated
680 * with the underlying file, this region map represents the backing file
681 * pages which have ever had a reservation assigned which this persists even
682 * after the page is instantiated. A private mapping has a region map
683 * associated with the original mmap which is attached to all VMAs which
684 * reference it, this region map represents those offsets which have consumed
685 * reservation ie. where pages have been instantiated.
687 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
689 return (unsigned long)vma->vm_private_data;
692 static void set_vma_private_data(struct vm_area_struct *vma,
695 vma->vm_private_data = (void *)value;
698 struct resv_map *resv_map_alloc(void)
700 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
701 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
703 if (!resv_map || !rg) {
709 kref_init(&resv_map->refs);
710 spin_lock_init(&resv_map->lock);
711 INIT_LIST_HEAD(&resv_map->regions);
713 resv_map->adds_in_progress = 0;
715 INIT_LIST_HEAD(&resv_map->region_cache);
716 list_add(&rg->link, &resv_map->region_cache);
717 resv_map->region_cache_count = 1;
722 void resv_map_release(struct kref *ref)
724 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
725 struct list_head *head = &resv_map->region_cache;
726 struct file_region *rg, *trg;
728 /* Clear out any active regions before we release the map. */
729 region_del(resv_map, 0, LONG_MAX);
731 /* ... and any entries left in the cache */
732 list_for_each_entry_safe(rg, trg, head, link) {
737 VM_BUG_ON(resv_map->adds_in_progress);
742 static inline struct resv_map *inode_resv_map(struct inode *inode)
745 * At inode evict time, i_mapping may not point to the original
746 * address space within the inode. This original address space
747 * contains the pointer to the resv_map. So, always use the
748 * address space embedded within the inode.
749 * The VERY common case is inode->mapping == &inode->i_data but,
750 * this may not be true for device special inodes.
752 return (struct resv_map *)(&inode->i_data)->private_data;
755 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
757 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
758 if (vma->vm_flags & VM_MAYSHARE) {
759 struct address_space *mapping = vma->vm_file->f_mapping;
760 struct inode *inode = mapping->host;
762 return inode_resv_map(inode);
765 return (struct resv_map *)(get_vma_private_data(vma) &
770 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
772 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
773 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
775 set_vma_private_data(vma, (get_vma_private_data(vma) &
776 HPAGE_RESV_MASK) | (unsigned long)map);
779 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
781 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
782 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
784 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
787 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
789 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
791 return (get_vma_private_data(vma) & flag) != 0;
794 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
795 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
797 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
798 if (!(vma->vm_flags & VM_MAYSHARE))
799 vma->vm_private_data = (void *)0;
802 /* Returns true if the VMA has associated reserve pages */
803 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
805 if (vma->vm_flags & VM_NORESERVE) {
807 * This address is already reserved by other process(chg == 0),
808 * so, we should decrement reserved count. Without decrementing,
809 * reserve count remains after releasing inode, because this
810 * allocated page will go into page cache and is regarded as
811 * coming from reserved pool in releasing step. Currently, we
812 * don't have any other solution to deal with this situation
813 * properly, so add work-around here.
815 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
821 /* Shared mappings always use reserves */
822 if (vma->vm_flags & VM_MAYSHARE) {
824 * We know VM_NORESERVE is not set. Therefore, there SHOULD
825 * be a region map for all pages. The only situation where
826 * there is no region map is if a hole was punched via
827 * fallocate. In this case, there really are no reverves to
828 * use. This situation is indicated if chg != 0.
837 * Only the process that called mmap() has reserves for
840 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
842 * Like the shared case above, a hole punch or truncate
843 * could have been performed on the private mapping.
844 * Examine the value of chg to determine if reserves
845 * actually exist or were previously consumed.
846 * Very Subtle - The value of chg comes from a previous
847 * call to vma_needs_reserves(). The reserve map for
848 * private mappings has different (opposite) semantics
849 * than that of shared mappings. vma_needs_reserves()
850 * has already taken this difference in semantics into
851 * account. Therefore, the meaning of chg is the same
852 * as in the shared case above. Code could easily be
853 * combined, but keeping it separate draws attention to
854 * subtle differences.
865 static void enqueue_huge_page(struct hstate *h, struct page *page)
867 int nid = page_to_nid(page);
868 list_move(&page->lru, &h->hugepage_freelists[nid]);
869 h->free_huge_pages++;
870 h->free_huge_pages_node[nid]++;
873 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
877 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
878 if (!PageHWPoison(page))
881 * if 'non-isolated free hugepage' not found on the list,
882 * the allocation fails.
884 if (&h->hugepage_freelists[nid] == &page->lru)
886 list_move(&page->lru, &h->hugepage_activelist);
887 set_page_refcounted(page);
888 h->free_huge_pages--;
889 h->free_huge_pages_node[nid]--;
893 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
896 unsigned int cpuset_mems_cookie;
897 struct zonelist *zonelist;
900 int node = NUMA_NO_NODE;
902 zonelist = node_zonelist(nid, gfp_mask);
905 cpuset_mems_cookie = read_mems_allowed_begin();
906 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
909 if (!cpuset_zone_allowed(zone, gfp_mask))
912 * no need to ask again on the same node. Pool is node rather than
915 if (zone_to_nid(zone) == node)
917 node = zone_to_nid(zone);
919 page = dequeue_huge_page_node_exact(h, node);
923 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
929 /* Movability of hugepages depends on migration support. */
930 static inline gfp_t htlb_alloc_mask(struct hstate *h)
932 if (hugepage_movable_supported(h))
933 return GFP_HIGHUSER_MOVABLE;
938 static struct page *dequeue_huge_page_vma(struct hstate *h,
939 struct vm_area_struct *vma,
940 unsigned long address, int avoid_reserve,
944 struct mempolicy *mpol;
946 nodemask_t *nodemask;
950 * A child process with MAP_PRIVATE mappings created by their parent
951 * have no page reserves. This check ensures that reservations are
952 * not "stolen". The child may still get SIGKILLed
954 if (!vma_has_reserves(vma, chg) &&
955 h->free_huge_pages - h->resv_huge_pages == 0)
958 /* If reserves cannot be used, ensure enough pages are in the pool */
959 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
962 gfp_mask = htlb_alloc_mask(h);
963 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
964 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
965 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
966 SetPagePrivate(page);
967 h->resv_huge_pages--;
978 * common helper functions for hstate_next_node_to_{alloc|free}.
979 * We may have allocated or freed a huge page based on a different
980 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
981 * be outside of *nodes_allowed. Ensure that we use an allowed
982 * node for alloc or free.
984 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
986 nid = next_node_in(nid, *nodes_allowed);
987 VM_BUG_ON(nid >= MAX_NUMNODES);
992 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
994 if (!node_isset(nid, *nodes_allowed))
995 nid = next_node_allowed(nid, nodes_allowed);
1000 * returns the previously saved node ["this node"] from which to
1001 * allocate a persistent huge page for the pool and advance the
1002 * next node from which to allocate, handling wrap at end of node
1005 static int hstate_next_node_to_alloc(struct hstate *h,
1006 nodemask_t *nodes_allowed)
1010 VM_BUG_ON(!nodes_allowed);
1012 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1013 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1019 * helper for free_pool_huge_page() - return the previously saved
1020 * node ["this node"] from which to free a huge page. Advance the
1021 * next node id whether or not we find a free huge page to free so
1022 * that the next attempt to free addresses the next node.
1024 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1028 VM_BUG_ON(!nodes_allowed);
1030 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1031 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1036 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1037 for (nr_nodes = nodes_weight(*mask); \
1039 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1042 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1043 for (nr_nodes = nodes_weight(*mask); \
1045 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1048 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1049 static void destroy_compound_gigantic_page(struct page *page,
1053 int nr_pages = 1 << order;
1054 struct page *p = page + 1;
1056 atomic_set(compound_mapcount_ptr(page), 0);
1057 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1058 clear_compound_head(p);
1059 set_page_refcounted(p);
1062 set_compound_order(page, 0);
1063 __ClearPageHead(page);
1066 static void free_gigantic_page(struct page *page, unsigned int order)
1068 free_contig_range(page_to_pfn(page), 1 << order);
1071 #ifdef CONFIG_CONTIG_ALLOC
1072 static int __alloc_gigantic_page(unsigned long start_pfn,
1073 unsigned long nr_pages, gfp_t gfp_mask)
1075 unsigned long end_pfn = start_pfn + nr_pages;
1076 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1080 static bool pfn_range_valid_gigantic(struct zone *z,
1081 unsigned long start_pfn, unsigned long nr_pages)
1083 unsigned long i, end_pfn = start_pfn + nr_pages;
1086 for (i = start_pfn; i < end_pfn; i++) {
1087 page = pfn_to_online_page(i);
1091 if (page_zone(page) != z)
1094 if (PageReserved(page))
1097 if (page_count(page) > 0)
1107 static bool zone_spans_last_pfn(const struct zone *zone,
1108 unsigned long start_pfn, unsigned long nr_pages)
1110 unsigned long last_pfn = start_pfn + nr_pages - 1;
1111 return zone_spans_pfn(zone, last_pfn);
1114 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1115 int nid, nodemask_t *nodemask)
1117 unsigned int order = huge_page_order(h);
1118 unsigned long nr_pages = 1 << order;
1119 unsigned long ret, pfn, flags;
1120 struct zonelist *zonelist;
1124 zonelist = node_zonelist(nid, gfp_mask);
1125 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) {
1126 spin_lock_irqsave(&zone->lock, flags);
1128 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
1129 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
1130 if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) {
1132 * We release the zone lock here because
1133 * alloc_contig_range() will also lock the zone
1134 * at some point. If there's an allocation
1135 * spinning on this lock, it may win the race
1136 * and cause alloc_contig_range() to fail...
1138 spin_unlock_irqrestore(&zone->lock, flags);
1139 ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask);
1141 return pfn_to_page(pfn);
1142 spin_lock_irqsave(&zone->lock, flags);
1147 spin_unlock_irqrestore(&zone->lock, flags);
1153 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1154 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1155 #else /* !CONFIG_CONTIG_ALLOC */
1156 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1157 int nid, nodemask_t *nodemask)
1161 #endif /* CONFIG_CONTIG_ALLOC */
1163 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1164 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1165 int nid, nodemask_t *nodemask)
1169 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1170 static inline void destroy_compound_gigantic_page(struct page *page,
1171 unsigned int order) { }
1174 static void update_and_free_page(struct hstate *h, struct page *page)
1178 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1182 h->nr_huge_pages_node[page_to_nid(page)]--;
1183 for (i = 0; i < pages_per_huge_page(h); i++) {
1184 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1185 1 << PG_referenced | 1 << PG_dirty |
1186 1 << PG_active | 1 << PG_private |
1189 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1190 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1191 set_page_refcounted(page);
1192 if (hstate_is_gigantic(h)) {
1193 destroy_compound_gigantic_page(page, huge_page_order(h));
1194 free_gigantic_page(page, huge_page_order(h));
1196 __free_pages(page, huge_page_order(h));
1200 struct hstate *size_to_hstate(unsigned long size)
1204 for_each_hstate(h) {
1205 if (huge_page_size(h) == size)
1212 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1213 * to hstate->hugepage_activelist.)
1215 * This function can be called for tail pages, but never returns true for them.
1217 bool page_huge_active(struct page *page)
1219 VM_BUG_ON_PAGE(!PageHuge(page), page);
1220 return PageHead(page) && PagePrivate(&page[1]);
1223 /* never called for tail page */
1224 static void set_page_huge_active(struct page *page)
1226 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1227 SetPagePrivate(&page[1]);
1230 static void clear_page_huge_active(struct page *page)
1232 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1233 ClearPagePrivate(&page[1]);
1237 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1240 static inline bool PageHugeTemporary(struct page *page)
1242 if (!PageHuge(page))
1245 return (unsigned long)page[2].mapping == -1U;
1248 static inline void SetPageHugeTemporary(struct page *page)
1250 page[2].mapping = (void *)-1U;
1253 static inline void ClearPageHugeTemporary(struct page *page)
1255 page[2].mapping = NULL;
1258 void free_huge_page(struct page *page)
1261 * Can't pass hstate in here because it is called from the
1262 * compound page destructor.
1264 struct hstate *h = page_hstate(page);
1265 int nid = page_to_nid(page);
1266 struct hugepage_subpool *spool =
1267 (struct hugepage_subpool *)page_private(page);
1268 bool restore_reserve;
1270 VM_BUG_ON_PAGE(page_count(page), page);
1271 VM_BUG_ON_PAGE(page_mapcount(page), page);
1273 set_page_private(page, 0);
1274 page->mapping = NULL;
1275 restore_reserve = PagePrivate(page);
1276 ClearPagePrivate(page);
1279 * If PagePrivate() was set on page, page allocation consumed a
1280 * reservation. If the page was associated with a subpool, there
1281 * would have been a page reserved in the subpool before allocation
1282 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1283 * reservtion, do not call hugepage_subpool_put_pages() as this will
1284 * remove the reserved page from the subpool.
1286 if (!restore_reserve) {
1288 * A return code of zero implies that the subpool will be
1289 * under its minimum size if the reservation is not restored
1290 * after page is free. Therefore, force restore_reserve
1293 if (hugepage_subpool_put_pages(spool, 1) == 0)
1294 restore_reserve = true;
1297 spin_lock(&hugetlb_lock);
1298 clear_page_huge_active(page);
1299 hugetlb_cgroup_uncharge_page(hstate_index(h),
1300 pages_per_huge_page(h), page);
1301 if (restore_reserve)
1302 h->resv_huge_pages++;
1304 if (PageHugeTemporary(page)) {
1305 list_del(&page->lru);
1306 ClearPageHugeTemporary(page);
1307 update_and_free_page(h, page);
1308 } else if (h->surplus_huge_pages_node[nid]) {
1309 /* remove the page from active list */
1310 list_del(&page->lru);
1311 update_and_free_page(h, page);
1312 h->surplus_huge_pages--;
1313 h->surplus_huge_pages_node[nid]--;
1315 arch_clear_hugepage_flags(page);
1316 enqueue_huge_page(h, page);
1318 spin_unlock(&hugetlb_lock);
1321 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1323 INIT_LIST_HEAD(&page->lru);
1324 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1325 spin_lock(&hugetlb_lock);
1326 set_hugetlb_cgroup(page, NULL);
1328 h->nr_huge_pages_node[nid]++;
1329 spin_unlock(&hugetlb_lock);
1332 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1335 int nr_pages = 1 << order;
1336 struct page *p = page + 1;
1338 /* we rely on prep_new_huge_page to set the destructor */
1339 set_compound_order(page, order);
1340 __ClearPageReserved(page);
1341 __SetPageHead(page);
1342 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1344 * For gigantic hugepages allocated through bootmem at
1345 * boot, it's safer to be consistent with the not-gigantic
1346 * hugepages and clear the PG_reserved bit from all tail pages
1347 * too. Otherwse drivers using get_user_pages() to access tail
1348 * pages may get the reference counting wrong if they see
1349 * PG_reserved set on a tail page (despite the head page not
1350 * having PG_reserved set). Enforcing this consistency between
1351 * head and tail pages allows drivers to optimize away a check
1352 * on the head page when they need know if put_page() is needed
1353 * after get_user_pages().
1355 __ClearPageReserved(p);
1356 set_page_count(p, 0);
1357 set_compound_head(p, page);
1359 atomic_set(compound_mapcount_ptr(page), -1);
1363 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1364 * transparent huge pages. See the PageTransHuge() documentation for more
1367 int PageHuge(struct page *page)
1369 if (!PageCompound(page))
1372 page = compound_head(page);
1373 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1375 EXPORT_SYMBOL_GPL(PageHuge);
1378 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1379 * normal or transparent huge pages.
1381 int PageHeadHuge(struct page *page_head)
1383 if (!PageHead(page_head))
1386 return get_compound_page_dtor(page_head) == free_huge_page;
1389 pgoff_t __basepage_index(struct page *page)
1391 struct page *page_head = compound_head(page);
1392 pgoff_t index = page_index(page_head);
1393 unsigned long compound_idx;
1395 if (!PageHuge(page_head))
1396 return page_index(page);
1398 if (compound_order(page_head) >= MAX_ORDER)
1399 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1401 compound_idx = page - page_head;
1403 return (index << compound_order(page_head)) + compound_idx;
1406 static struct page *alloc_buddy_huge_page(struct hstate *h,
1407 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1408 nodemask_t *node_alloc_noretry)
1410 int order = huge_page_order(h);
1412 bool alloc_try_hard = true;
1415 * By default we always try hard to allocate the page with
1416 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1417 * a loop (to adjust global huge page counts) and previous allocation
1418 * failed, do not continue to try hard on the same node. Use the
1419 * node_alloc_noretry bitmap to manage this state information.
1421 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1422 alloc_try_hard = false;
1423 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1425 gfp_mask |= __GFP_RETRY_MAYFAIL;
1426 if (nid == NUMA_NO_NODE)
1427 nid = numa_mem_id();
1428 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1430 __count_vm_event(HTLB_BUDDY_PGALLOC);
1432 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1435 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1436 * indicates an overall state change. Clear bit so that we resume
1437 * normal 'try hard' allocations.
1439 if (node_alloc_noretry && page && !alloc_try_hard)
1440 node_clear(nid, *node_alloc_noretry);
1443 * If we tried hard to get a page but failed, set bit so that
1444 * subsequent attempts will not try as hard until there is an
1445 * overall state change.
1447 if (node_alloc_noretry && !page && alloc_try_hard)
1448 node_set(nid, *node_alloc_noretry);
1454 * Common helper to allocate a fresh hugetlb page. All specific allocators
1455 * should use this function to get new hugetlb pages
1457 static struct page *alloc_fresh_huge_page(struct hstate *h,
1458 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1459 nodemask_t *node_alloc_noretry)
1463 if (hstate_is_gigantic(h))
1464 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1466 page = alloc_buddy_huge_page(h, gfp_mask,
1467 nid, nmask, node_alloc_noretry);
1471 if (hstate_is_gigantic(h))
1472 prep_compound_gigantic_page(page, huge_page_order(h));
1473 prep_new_huge_page(h, page, page_to_nid(page));
1479 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1482 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1483 nodemask_t *node_alloc_noretry)
1487 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1489 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1490 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1491 node_alloc_noretry);
1499 put_page(page); /* free it into the hugepage allocator */
1505 * Free huge page from pool from next node to free.
1506 * Attempt to keep persistent huge pages more or less
1507 * balanced over allowed nodes.
1508 * Called with hugetlb_lock locked.
1510 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1516 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1518 * If we're returning unused surplus pages, only examine
1519 * nodes with surplus pages.
1521 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1522 !list_empty(&h->hugepage_freelists[node])) {
1524 list_entry(h->hugepage_freelists[node].next,
1526 list_del(&page->lru);
1527 h->free_huge_pages--;
1528 h->free_huge_pages_node[node]--;
1530 h->surplus_huge_pages--;
1531 h->surplus_huge_pages_node[node]--;
1533 update_and_free_page(h, page);
1543 * Dissolve a given free hugepage into free buddy pages. This function does
1544 * nothing for in-use hugepages and non-hugepages.
1545 * This function returns values like below:
1547 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1548 * (allocated or reserved.)
1549 * 0: successfully dissolved free hugepages or the page is not a
1550 * hugepage (considered as already dissolved)
1552 int dissolve_free_huge_page(struct page *page)
1556 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1557 if (!PageHuge(page))
1560 spin_lock(&hugetlb_lock);
1561 if (!PageHuge(page)) {
1566 if (!page_count(page)) {
1567 struct page *head = compound_head(page);
1568 struct hstate *h = page_hstate(head);
1569 int nid = page_to_nid(head);
1570 if (h->free_huge_pages - h->resv_huge_pages == 0)
1573 * Move PageHWPoison flag from head page to the raw error page,
1574 * which makes any subpages rather than the error page reusable.
1576 if (PageHWPoison(head) && page != head) {
1577 SetPageHWPoison(page);
1578 ClearPageHWPoison(head);
1580 list_del(&head->lru);
1581 h->free_huge_pages--;
1582 h->free_huge_pages_node[nid]--;
1583 h->max_huge_pages--;
1584 update_and_free_page(h, head);
1588 spin_unlock(&hugetlb_lock);
1593 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1594 * make specified memory blocks removable from the system.
1595 * Note that this will dissolve a free gigantic hugepage completely, if any
1596 * part of it lies within the given range.
1597 * Also note that if dissolve_free_huge_page() returns with an error, all
1598 * free hugepages that were dissolved before that error are lost.
1600 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1606 if (!hugepages_supported())
1609 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1610 page = pfn_to_page(pfn);
1611 rc = dissolve_free_huge_page(page);
1620 * Allocates a fresh surplus page from the page allocator.
1622 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1623 int nid, nodemask_t *nmask)
1625 struct page *page = NULL;
1627 if (hstate_is_gigantic(h))
1630 spin_lock(&hugetlb_lock);
1631 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1633 spin_unlock(&hugetlb_lock);
1635 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1639 spin_lock(&hugetlb_lock);
1641 * We could have raced with the pool size change.
1642 * Double check that and simply deallocate the new page
1643 * if we would end up overcommiting the surpluses. Abuse
1644 * temporary page to workaround the nasty free_huge_page
1647 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1648 SetPageHugeTemporary(page);
1649 spin_unlock(&hugetlb_lock);
1653 h->surplus_huge_pages++;
1654 h->surplus_huge_pages_node[page_to_nid(page)]++;
1658 spin_unlock(&hugetlb_lock);
1663 struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1664 int nid, nodemask_t *nmask)
1668 if (hstate_is_gigantic(h))
1671 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1676 * We do not account these pages as surplus because they are only
1677 * temporary and will be released properly on the last reference
1679 SetPageHugeTemporary(page);
1685 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1688 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1689 struct vm_area_struct *vma, unsigned long addr)
1692 struct mempolicy *mpol;
1693 gfp_t gfp_mask = htlb_alloc_mask(h);
1695 nodemask_t *nodemask;
1697 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1698 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1699 mpol_cond_put(mpol);
1704 /* page migration callback function */
1705 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1707 gfp_t gfp_mask = htlb_alloc_mask(h);
1708 struct page *page = NULL;
1710 if (nid != NUMA_NO_NODE)
1711 gfp_mask |= __GFP_THISNODE;
1713 spin_lock(&hugetlb_lock);
1714 if (h->free_huge_pages - h->resv_huge_pages > 0)
1715 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1716 spin_unlock(&hugetlb_lock);
1719 page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1724 /* page migration callback function */
1725 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1728 gfp_t gfp_mask = htlb_alloc_mask(h);
1730 spin_lock(&hugetlb_lock);
1731 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1734 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1736 spin_unlock(&hugetlb_lock);
1740 spin_unlock(&hugetlb_lock);
1742 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1745 /* mempolicy aware migration callback */
1746 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1747 unsigned long address)
1749 struct mempolicy *mpol;
1750 nodemask_t *nodemask;
1755 gfp_mask = htlb_alloc_mask(h);
1756 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1757 page = alloc_huge_page_nodemask(h, node, nodemask);
1758 mpol_cond_put(mpol);
1764 * Increase the hugetlb pool such that it can accommodate a reservation
1767 static int gather_surplus_pages(struct hstate *h, int delta)
1769 struct list_head surplus_list;
1770 struct page *page, *tmp;
1772 int needed, allocated;
1773 bool alloc_ok = true;
1775 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1777 h->resv_huge_pages += delta;
1782 INIT_LIST_HEAD(&surplus_list);
1786 spin_unlock(&hugetlb_lock);
1787 for (i = 0; i < needed; i++) {
1788 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1789 NUMA_NO_NODE, NULL);
1794 list_add(&page->lru, &surplus_list);
1800 * After retaking hugetlb_lock, we need to recalculate 'needed'
1801 * because either resv_huge_pages or free_huge_pages may have changed.
1803 spin_lock(&hugetlb_lock);
1804 needed = (h->resv_huge_pages + delta) -
1805 (h->free_huge_pages + allocated);
1810 * We were not able to allocate enough pages to
1811 * satisfy the entire reservation so we free what
1812 * we've allocated so far.
1817 * The surplus_list now contains _at_least_ the number of extra pages
1818 * needed to accommodate the reservation. Add the appropriate number
1819 * of pages to the hugetlb pool and free the extras back to the buddy
1820 * allocator. Commit the entire reservation here to prevent another
1821 * process from stealing the pages as they are added to the pool but
1822 * before they are reserved.
1824 needed += allocated;
1825 h->resv_huge_pages += delta;
1828 /* Free the needed pages to the hugetlb pool */
1829 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1833 * This page is now managed by the hugetlb allocator and has
1834 * no users -- drop the buddy allocator's reference.
1836 put_page_testzero(page);
1837 VM_BUG_ON_PAGE(page_count(page), page);
1838 enqueue_huge_page(h, page);
1841 spin_unlock(&hugetlb_lock);
1843 /* Free unnecessary surplus pages to the buddy allocator */
1844 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1846 spin_lock(&hugetlb_lock);
1852 * This routine has two main purposes:
1853 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1854 * in unused_resv_pages. This corresponds to the prior adjustments made
1855 * to the associated reservation map.
1856 * 2) Free any unused surplus pages that may have been allocated to satisfy
1857 * the reservation. As many as unused_resv_pages may be freed.
1859 * Called with hugetlb_lock held. However, the lock could be dropped (and
1860 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1861 * we must make sure nobody else can claim pages we are in the process of
1862 * freeing. Do this by ensuring resv_huge_page always is greater than the
1863 * number of huge pages we plan to free when dropping the lock.
1865 static void return_unused_surplus_pages(struct hstate *h,
1866 unsigned long unused_resv_pages)
1868 unsigned long nr_pages;
1870 /* Cannot return gigantic pages currently */
1871 if (hstate_is_gigantic(h))
1875 * Part (or even all) of the reservation could have been backed
1876 * by pre-allocated pages. Only free surplus pages.
1878 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1881 * We want to release as many surplus pages as possible, spread
1882 * evenly across all nodes with memory. Iterate across these nodes
1883 * until we can no longer free unreserved surplus pages. This occurs
1884 * when the nodes with surplus pages have no free pages.
1885 * free_pool_huge_page() will balance the the freed pages across the
1886 * on-line nodes with memory and will handle the hstate accounting.
1888 * Note that we decrement resv_huge_pages as we free the pages. If
1889 * we drop the lock, resv_huge_pages will still be sufficiently large
1890 * to cover subsequent pages we may free.
1892 while (nr_pages--) {
1893 h->resv_huge_pages--;
1894 unused_resv_pages--;
1895 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1897 cond_resched_lock(&hugetlb_lock);
1901 /* Fully uncommit the reservation */
1902 h->resv_huge_pages -= unused_resv_pages;
1907 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1908 * are used by the huge page allocation routines to manage reservations.
1910 * vma_needs_reservation is called to determine if the huge page at addr
1911 * within the vma has an associated reservation. If a reservation is
1912 * needed, the value 1 is returned. The caller is then responsible for
1913 * managing the global reservation and subpool usage counts. After
1914 * the huge page has been allocated, vma_commit_reservation is called
1915 * to add the page to the reservation map. If the page allocation fails,
1916 * the reservation must be ended instead of committed. vma_end_reservation
1917 * is called in such cases.
1919 * In the normal case, vma_commit_reservation returns the same value
1920 * as the preceding vma_needs_reservation call. The only time this
1921 * is not the case is if a reserve map was changed between calls. It
1922 * is the responsibility of the caller to notice the difference and
1923 * take appropriate action.
1925 * vma_add_reservation is used in error paths where a reservation must
1926 * be restored when a newly allocated huge page must be freed. It is
1927 * to be called after calling vma_needs_reservation to determine if a
1928 * reservation exists.
1930 enum vma_resv_mode {
1936 static long __vma_reservation_common(struct hstate *h,
1937 struct vm_area_struct *vma, unsigned long addr,
1938 enum vma_resv_mode mode)
1940 struct resv_map *resv;
1944 resv = vma_resv_map(vma);
1948 idx = vma_hugecache_offset(h, vma, addr);
1950 case VMA_NEEDS_RESV:
1951 ret = region_chg(resv, idx, idx + 1);
1953 case VMA_COMMIT_RESV:
1954 ret = region_add(resv, idx, idx + 1);
1957 region_abort(resv, idx, idx + 1);
1961 if (vma->vm_flags & VM_MAYSHARE)
1962 ret = region_add(resv, idx, idx + 1);
1964 region_abort(resv, idx, idx + 1);
1965 ret = region_del(resv, idx, idx + 1);
1972 if (vma->vm_flags & VM_MAYSHARE)
1974 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1976 * In most cases, reserves always exist for private mappings.
1977 * However, a file associated with mapping could have been
1978 * hole punched or truncated after reserves were consumed.
1979 * As subsequent fault on such a range will not use reserves.
1980 * Subtle - The reserve map for private mappings has the
1981 * opposite meaning than that of shared mappings. If NO
1982 * entry is in the reserve map, it means a reservation exists.
1983 * If an entry exists in the reserve map, it means the
1984 * reservation has already been consumed. As a result, the
1985 * return value of this routine is the opposite of the
1986 * value returned from reserve map manipulation routines above.
1994 return ret < 0 ? ret : 0;
1997 static long vma_needs_reservation(struct hstate *h,
1998 struct vm_area_struct *vma, unsigned long addr)
2000 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2003 static long vma_commit_reservation(struct hstate *h,
2004 struct vm_area_struct *vma, unsigned long addr)
2006 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2009 static void vma_end_reservation(struct hstate *h,
2010 struct vm_area_struct *vma, unsigned long addr)
2012 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2015 static long vma_add_reservation(struct hstate *h,
2016 struct vm_area_struct *vma, unsigned long addr)
2018 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2022 * This routine is called to restore a reservation on error paths. In the
2023 * specific error paths, a huge page was allocated (via alloc_huge_page)
2024 * and is about to be freed. If a reservation for the page existed,
2025 * alloc_huge_page would have consumed the reservation and set PagePrivate
2026 * in the newly allocated page. When the page is freed via free_huge_page,
2027 * the global reservation count will be incremented if PagePrivate is set.
2028 * However, free_huge_page can not adjust the reserve map. Adjust the
2029 * reserve map here to be consistent with global reserve count adjustments
2030 * to be made by free_huge_page.
2032 static void restore_reserve_on_error(struct hstate *h,
2033 struct vm_area_struct *vma, unsigned long address,
2036 if (unlikely(PagePrivate(page))) {
2037 long rc = vma_needs_reservation(h, vma, address);
2039 if (unlikely(rc < 0)) {
2041 * Rare out of memory condition in reserve map
2042 * manipulation. Clear PagePrivate so that
2043 * global reserve count will not be incremented
2044 * by free_huge_page. This will make it appear
2045 * as though the reservation for this page was
2046 * consumed. This may prevent the task from
2047 * faulting in the page at a later time. This
2048 * is better than inconsistent global huge page
2049 * accounting of reserve counts.
2051 ClearPagePrivate(page);
2053 rc = vma_add_reservation(h, vma, address);
2054 if (unlikely(rc < 0))
2056 * See above comment about rare out of
2059 ClearPagePrivate(page);
2061 vma_end_reservation(h, vma, address);
2065 struct page *alloc_huge_page(struct vm_area_struct *vma,
2066 unsigned long addr, int avoid_reserve)
2068 struct hugepage_subpool *spool = subpool_vma(vma);
2069 struct hstate *h = hstate_vma(vma);
2071 long map_chg, map_commit;
2074 struct hugetlb_cgroup *h_cg;
2076 idx = hstate_index(h);
2078 * Examine the region/reserve map to determine if the process
2079 * has a reservation for the page to be allocated. A return
2080 * code of zero indicates a reservation exists (no change).
2082 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2084 return ERR_PTR(-ENOMEM);
2087 * Processes that did not create the mapping will have no
2088 * reserves as indicated by the region/reserve map. Check
2089 * that the allocation will not exceed the subpool limit.
2090 * Allocations for MAP_NORESERVE mappings also need to be
2091 * checked against any subpool limit.
2093 if (map_chg || avoid_reserve) {
2094 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2096 vma_end_reservation(h, vma, addr);
2097 return ERR_PTR(-ENOSPC);
2101 * Even though there was no reservation in the region/reserve
2102 * map, there could be reservations associated with the
2103 * subpool that can be used. This would be indicated if the
2104 * return value of hugepage_subpool_get_pages() is zero.
2105 * However, if avoid_reserve is specified we still avoid even
2106 * the subpool reservations.
2112 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2114 goto out_subpool_put;
2116 spin_lock(&hugetlb_lock);
2118 * glb_chg is passed to indicate whether or not a page must be taken
2119 * from the global free pool (global change). gbl_chg == 0 indicates
2120 * a reservation exists for the allocation.
2122 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2124 spin_unlock(&hugetlb_lock);
2125 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2127 goto out_uncharge_cgroup;
2128 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2129 SetPagePrivate(page);
2130 h->resv_huge_pages--;
2132 spin_lock(&hugetlb_lock);
2133 list_move(&page->lru, &h->hugepage_activelist);
2136 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2137 spin_unlock(&hugetlb_lock);
2139 set_page_private(page, (unsigned long)spool);
2141 map_commit = vma_commit_reservation(h, vma, addr);
2142 if (unlikely(map_chg > map_commit)) {
2144 * The page was added to the reservation map between
2145 * vma_needs_reservation and vma_commit_reservation.
2146 * This indicates a race with hugetlb_reserve_pages.
2147 * Adjust for the subpool count incremented above AND
2148 * in hugetlb_reserve_pages for the same page. Also,
2149 * the reservation count added in hugetlb_reserve_pages
2150 * no longer applies.
2154 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2155 hugetlb_acct_memory(h, -rsv_adjust);
2159 out_uncharge_cgroup:
2160 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2162 if (map_chg || avoid_reserve)
2163 hugepage_subpool_put_pages(spool, 1);
2164 vma_end_reservation(h, vma, addr);
2165 return ERR_PTR(-ENOSPC);
2168 int alloc_bootmem_huge_page(struct hstate *h)
2169 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2170 int __alloc_bootmem_huge_page(struct hstate *h)
2172 struct huge_bootmem_page *m;
2175 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2178 addr = memblock_alloc_try_nid_raw(
2179 huge_page_size(h), huge_page_size(h),
2180 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2183 * Use the beginning of the huge page to store the
2184 * huge_bootmem_page struct (until gather_bootmem
2185 * puts them into the mem_map).
2194 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2195 /* Put them into a private list first because mem_map is not up yet */
2196 INIT_LIST_HEAD(&m->list);
2197 list_add(&m->list, &huge_boot_pages);
2202 static void __init prep_compound_huge_page(struct page *page,
2205 if (unlikely(order > (MAX_ORDER - 1)))
2206 prep_compound_gigantic_page(page, order);
2208 prep_compound_page(page, order);
2211 /* Put bootmem huge pages into the standard lists after mem_map is up */
2212 static void __init gather_bootmem_prealloc(void)
2214 struct huge_bootmem_page *m;
2216 list_for_each_entry(m, &huge_boot_pages, list) {
2217 struct page *page = virt_to_page(m);
2218 struct hstate *h = m->hstate;
2220 WARN_ON(page_count(page) != 1);
2221 prep_compound_huge_page(page, h->order);
2222 WARN_ON(PageReserved(page));
2223 prep_new_huge_page(h, page, page_to_nid(page));
2224 put_page(page); /* free it into the hugepage allocator */
2227 * If we had gigantic hugepages allocated at boot time, we need
2228 * to restore the 'stolen' pages to totalram_pages in order to
2229 * fix confusing memory reports from free(1) and another
2230 * side-effects, like CommitLimit going negative.
2232 if (hstate_is_gigantic(h))
2233 adjust_managed_page_count(page, 1 << h->order);
2238 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2241 nodemask_t *node_alloc_noretry;
2243 if (!hstate_is_gigantic(h)) {
2245 * Bit mask controlling how hard we retry per-node allocations.
2246 * Ignore errors as lower level routines can deal with
2247 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2248 * time, we are likely in bigger trouble.
2250 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2253 /* allocations done at boot time */
2254 node_alloc_noretry = NULL;
2257 /* bit mask controlling how hard we retry per-node allocations */
2258 if (node_alloc_noretry)
2259 nodes_clear(*node_alloc_noretry);
2261 for (i = 0; i < h->max_huge_pages; ++i) {
2262 if (hstate_is_gigantic(h)) {
2263 if (!alloc_bootmem_huge_page(h))
2265 } else if (!alloc_pool_huge_page(h,
2266 &node_states[N_MEMORY],
2267 node_alloc_noretry))
2271 if (i < h->max_huge_pages) {
2274 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2275 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2276 h->max_huge_pages, buf, i);
2277 h->max_huge_pages = i;
2280 kfree(node_alloc_noretry);
2283 static void __init hugetlb_init_hstates(void)
2287 for_each_hstate(h) {
2288 if (minimum_order > huge_page_order(h))
2289 minimum_order = huge_page_order(h);
2291 /* oversize hugepages were init'ed in early boot */
2292 if (!hstate_is_gigantic(h))
2293 hugetlb_hstate_alloc_pages(h);
2295 VM_BUG_ON(minimum_order == UINT_MAX);
2298 static void __init report_hugepages(void)
2302 for_each_hstate(h) {
2305 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2306 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2307 buf, h->free_huge_pages);
2311 #ifdef CONFIG_HIGHMEM
2312 static void try_to_free_low(struct hstate *h, unsigned long count,
2313 nodemask_t *nodes_allowed)
2317 if (hstate_is_gigantic(h))
2320 for_each_node_mask(i, *nodes_allowed) {
2321 struct page *page, *next;
2322 struct list_head *freel = &h->hugepage_freelists[i];
2323 list_for_each_entry_safe(page, next, freel, lru) {
2324 if (count >= h->nr_huge_pages)
2326 if (PageHighMem(page))
2328 list_del(&page->lru);
2329 update_and_free_page(h, page);
2330 h->free_huge_pages--;
2331 h->free_huge_pages_node[page_to_nid(page)]--;
2336 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2337 nodemask_t *nodes_allowed)
2343 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2344 * balanced by operating on them in a round-robin fashion.
2345 * Returns 1 if an adjustment was made.
2347 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2352 VM_BUG_ON(delta != -1 && delta != 1);
2355 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2356 if (h->surplus_huge_pages_node[node])
2360 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2361 if (h->surplus_huge_pages_node[node] <
2362 h->nr_huge_pages_node[node])
2369 h->surplus_huge_pages += delta;
2370 h->surplus_huge_pages_node[node] += delta;
2374 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2375 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2376 nodemask_t *nodes_allowed)
2378 unsigned long min_count, ret;
2379 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2382 * Bit mask controlling how hard we retry per-node allocations.
2383 * If we can not allocate the bit mask, do not attempt to allocate
2384 * the requested huge pages.
2386 if (node_alloc_noretry)
2387 nodes_clear(*node_alloc_noretry);
2391 spin_lock(&hugetlb_lock);
2394 * Check for a node specific request.
2395 * Changing node specific huge page count may require a corresponding
2396 * change to the global count. In any case, the passed node mask
2397 * (nodes_allowed) will restrict alloc/free to the specified node.
2399 if (nid != NUMA_NO_NODE) {
2400 unsigned long old_count = count;
2402 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2404 * User may have specified a large count value which caused the
2405 * above calculation to overflow. In this case, they wanted
2406 * to allocate as many huge pages as possible. Set count to
2407 * largest possible value to align with their intention.
2409 if (count < old_count)
2414 * Gigantic pages runtime allocation depend on the capability for large
2415 * page range allocation.
2416 * If the system does not provide this feature, return an error when
2417 * the user tries to allocate gigantic pages but let the user free the
2418 * boottime allocated gigantic pages.
2420 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2421 if (count > persistent_huge_pages(h)) {
2422 spin_unlock(&hugetlb_lock);
2423 NODEMASK_FREE(node_alloc_noretry);
2426 /* Fall through to decrease pool */
2430 * Increase the pool size
2431 * First take pages out of surplus state. Then make up the
2432 * remaining difference by allocating fresh huge pages.
2434 * We might race with alloc_surplus_huge_page() here and be unable
2435 * to convert a surplus huge page to a normal huge page. That is
2436 * not critical, though, it just means the overall size of the
2437 * pool might be one hugepage larger than it needs to be, but
2438 * within all the constraints specified by the sysctls.
2440 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2441 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2445 while (count > persistent_huge_pages(h)) {
2447 * If this allocation races such that we no longer need the
2448 * page, free_huge_page will handle it by freeing the page
2449 * and reducing the surplus.
2451 spin_unlock(&hugetlb_lock);
2453 /* yield cpu to avoid soft lockup */
2456 ret = alloc_pool_huge_page(h, nodes_allowed,
2457 node_alloc_noretry);
2458 spin_lock(&hugetlb_lock);
2462 /* Bail for signals. Probably ctrl-c from user */
2463 if (signal_pending(current))
2468 * Decrease the pool size
2469 * First return free pages to the buddy allocator (being careful
2470 * to keep enough around to satisfy reservations). Then place
2471 * pages into surplus state as needed so the pool will shrink
2472 * to the desired size as pages become free.
2474 * By placing pages into the surplus state independent of the
2475 * overcommit value, we are allowing the surplus pool size to
2476 * exceed overcommit. There are few sane options here. Since
2477 * alloc_surplus_huge_page() is checking the global counter,
2478 * though, we'll note that we're not allowed to exceed surplus
2479 * and won't grow the pool anywhere else. Not until one of the
2480 * sysctls are changed, or the surplus pages go out of use.
2482 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2483 min_count = max(count, min_count);
2484 try_to_free_low(h, min_count, nodes_allowed);
2485 while (min_count < persistent_huge_pages(h)) {
2486 if (!free_pool_huge_page(h, nodes_allowed, 0))
2488 cond_resched_lock(&hugetlb_lock);
2490 while (count < persistent_huge_pages(h)) {
2491 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2495 h->max_huge_pages = persistent_huge_pages(h);
2496 spin_unlock(&hugetlb_lock);
2498 NODEMASK_FREE(node_alloc_noretry);
2503 #define HSTATE_ATTR_RO(_name) \
2504 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2506 #define HSTATE_ATTR(_name) \
2507 static struct kobj_attribute _name##_attr = \
2508 __ATTR(_name, 0644, _name##_show, _name##_store)
2510 static struct kobject *hugepages_kobj;
2511 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2513 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2515 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2519 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2520 if (hstate_kobjs[i] == kobj) {
2522 *nidp = NUMA_NO_NODE;
2526 return kobj_to_node_hstate(kobj, nidp);
2529 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2530 struct kobj_attribute *attr, char *buf)
2533 unsigned long nr_huge_pages;
2536 h = kobj_to_hstate(kobj, &nid);
2537 if (nid == NUMA_NO_NODE)
2538 nr_huge_pages = h->nr_huge_pages;
2540 nr_huge_pages = h->nr_huge_pages_node[nid];
2542 return sprintf(buf, "%lu\n", nr_huge_pages);
2545 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2546 struct hstate *h, int nid,
2547 unsigned long count, size_t len)
2550 nodemask_t nodes_allowed, *n_mask;
2552 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2555 if (nid == NUMA_NO_NODE) {
2557 * global hstate attribute
2559 if (!(obey_mempolicy &&
2560 init_nodemask_of_mempolicy(&nodes_allowed)))
2561 n_mask = &node_states[N_MEMORY];
2563 n_mask = &nodes_allowed;
2566 * Node specific request. count adjustment happens in
2567 * set_max_huge_pages() after acquiring hugetlb_lock.
2569 init_nodemask_of_node(&nodes_allowed, nid);
2570 n_mask = &nodes_allowed;
2573 err = set_max_huge_pages(h, count, nid, n_mask);
2575 return err ? err : len;
2578 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2579 struct kobject *kobj, const char *buf,
2583 unsigned long count;
2587 err = kstrtoul(buf, 10, &count);
2591 h = kobj_to_hstate(kobj, &nid);
2592 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2595 static ssize_t nr_hugepages_show(struct kobject *kobj,
2596 struct kobj_attribute *attr, char *buf)
2598 return nr_hugepages_show_common(kobj, attr, buf);
2601 static ssize_t nr_hugepages_store(struct kobject *kobj,
2602 struct kobj_attribute *attr, const char *buf, size_t len)
2604 return nr_hugepages_store_common(false, kobj, buf, len);
2606 HSTATE_ATTR(nr_hugepages);
2611 * hstate attribute for optionally mempolicy-based constraint on persistent
2612 * huge page alloc/free.
2614 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2615 struct kobj_attribute *attr, char *buf)
2617 return nr_hugepages_show_common(kobj, attr, buf);
2620 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2621 struct kobj_attribute *attr, const char *buf, size_t len)
2623 return nr_hugepages_store_common(true, kobj, buf, len);
2625 HSTATE_ATTR(nr_hugepages_mempolicy);
2629 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2630 struct kobj_attribute *attr, char *buf)
2632 struct hstate *h = kobj_to_hstate(kobj, NULL);
2633 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2636 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2637 struct kobj_attribute *attr, const char *buf, size_t count)
2640 unsigned long input;
2641 struct hstate *h = kobj_to_hstate(kobj, NULL);
2643 if (hstate_is_gigantic(h))
2646 err = kstrtoul(buf, 10, &input);
2650 spin_lock(&hugetlb_lock);
2651 h->nr_overcommit_huge_pages = input;
2652 spin_unlock(&hugetlb_lock);
2656 HSTATE_ATTR(nr_overcommit_hugepages);
2658 static ssize_t free_hugepages_show(struct kobject *kobj,
2659 struct kobj_attribute *attr, char *buf)
2662 unsigned long free_huge_pages;
2665 h = kobj_to_hstate(kobj, &nid);
2666 if (nid == NUMA_NO_NODE)
2667 free_huge_pages = h->free_huge_pages;
2669 free_huge_pages = h->free_huge_pages_node[nid];
2671 return sprintf(buf, "%lu\n", free_huge_pages);
2673 HSTATE_ATTR_RO(free_hugepages);
2675 static ssize_t resv_hugepages_show(struct kobject *kobj,
2676 struct kobj_attribute *attr, char *buf)
2678 struct hstate *h = kobj_to_hstate(kobj, NULL);
2679 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2681 HSTATE_ATTR_RO(resv_hugepages);
2683 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2684 struct kobj_attribute *attr, char *buf)
2687 unsigned long surplus_huge_pages;
2690 h = kobj_to_hstate(kobj, &nid);
2691 if (nid == NUMA_NO_NODE)
2692 surplus_huge_pages = h->surplus_huge_pages;
2694 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2696 return sprintf(buf, "%lu\n", surplus_huge_pages);
2698 HSTATE_ATTR_RO(surplus_hugepages);
2700 static struct attribute *hstate_attrs[] = {
2701 &nr_hugepages_attr.attr,
2702 &nr_overcommit_hugepages_attr.attr,
2703 &free_hugepages_attr.attr,
2704 &resv_hugepages_attr.attr,
2705 &surplus_hugepages_attr.attr,
2707 &nr_hugepages_mempolicy_attr.attr,
2712 static const struct attribute_group hstate_attr_group = {
2713 .attrs = hstate_attrs,
2716 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2717 struct kobject **hstate_kobjs,
2718 const struct attribute_group *hstate_attr_group)
2721 int hi = hstate_index(h);
2723 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2724 if (!hstate_kobjs[hi])
2727 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2729 kobject_put(hstate_kobjs[hi]);
2734 static void __init hugetlb_sysfs_init(void)
2739 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2740 if (!hugepages_kobj)
2743 for_each_hstate(h) {
2744 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2745 hstate_kobjs, &hstate_attr_group);
2747 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2754 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2755 * with node devices in node_devices[] using a parallel array. The array
2756 * index of a node device or _hstate == node id.
2757 * This is here to avoid any static dependency of the node device driver, in
2758 * the base kernel, on the hugetlb module.
2760 struct node_hstate {
2761 struct kobject *hugepages_kobj;
2762 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2764 static struct node_hstate node_hstates[MAX_NUMNODES];
2767 * A subset of global hstate attributes for node devices
2769 static struct attribute *per_node_hstate_attrs[] = {
2770 &nr_hugepages_attr.attr,
2771 &free_hugepages_attr.attr,
2772 &surplus_hugepages_attr.attr,
2776 static const struct attribute_group per_node_hstate_attr_group = {
2777 .attrs = per_node_hstate_attrs,
2781 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2782 * Returns node id via non-NULL nidp.
2784 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2788 for (nid = 0; nid < nr_node_ids; nid++) {
2789 struct node_hstate *nhs = &node_hstates[nid];
2791 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2792 if (nhs->hstate_kobjs[i] == kobj) {
2804 * Unregister hstate attributes from a single node device.
2805 * No-op if no hstate attributes attached.
2807 static void hugetlb_unregister_node(struct node *node)
2810 struct node_hstate *nhs = &node_hstates[node->dev.id];
2812 if (!nhs->hugepages_kobj)
2813 return; /* no hstate attributes */
2815 for_each_hstate(h) {
2816 int idx = hstate_index(h);
2817 if (nhs->hstate_kobjs[idx]) {
2818 kobject_put(nhs->hstate_kobjs[idx]);
2819 nhs->hstate_kobjs[idx] = NULL;
2823 kobject_put(nhs->hugepages_kobj);
2824 nhs->hugepages_kobj = NULL;
2829 * Register hstate attributes for a single node device.
2830 * No-op if attributes already registered.
2832 static void hugetlb_register_node(struct node *node)
2835 struct node_hstate *nhs = &node_hstates[node->dev.id];
2838 if (nhs->hugepages_kobj)
2839 return; /* already allocated */
2841 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2843 if (!nhs->hugepages_kobj)
2846 for_each_hstate(h) {
2847 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2849 &per_node_hstate_attr_group);
2851 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2852 h->name, node->dev.id);
2853 hugetlb_unregister_node(node);
2860 * hugetlb init time: register hstate attributes for all registered node
2861 * devices of nodes that have memory. All on-line nodes should have
2862 * registered their associated device by this time.
2864 static void __init hugetlb_register_all_nodes(void)
2868 for_each_node_state(nid, N_MEMORY) {
2869 struct node *node = node_devices[nid];
2870 if (node->dev.id == nid)
2871 hugetlb_register_node(node);
2875 * Let the node device driver know we're here so it can
2876 * [un]register hstate attributes on node hotplug.
2878 register_hugetlbfs_with_node(hugetlb_register_node,
2879 hugetlb_unregister_node);
2881 #else /* !CONFIG_NUMA */
2883 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2891 static void hugetlb_register_all_nodes(void) { }
2895 static int __init hugetlb_init(void)
2899 if (!hugepages_supported())
2902 if (!size_to_hstate(default_hstate_size)) {
2903 if (default_hstate_size != 0) {
2904 pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2905 default_hstate_size, HPAGE_SIZE);
2908 default_hstate_size = HPAGE_SIZE;
2909 if (!size_to_hstate(default_hstate_size))
2910 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2912 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2913 if (default_hstate_max_huge_pages) {
2914 if (!default_hstate.max_huge_pages)
2915 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2918 hugetlb_init_hstates();
2919 gather_bootmem_prealloc();
2922 hugetlb_sysfs_init();
2923 hugetlb_register_all_nodes();
2924 hugetlb_cgroup_file_init();
2927 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2929 num_fault_mutexes = 1;
2931 hugetlb_fault_mutex_table =
2932 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
2934 BUG_ON(!hugetlb_fault_mutex_table);
2936 for (i = 0; i < num_fault_mutexes; i++)
2937 mutex_init(&hugetlb_fault_mutex_table[i]);
2940 subsys_initcall(hugetlb_init);
2942 /* Should be called on processing a hugepagesz=... option */
2943 void __init hugetlb_bad_size(void)
2945 parsed_valid_hugepagesz = false;
2948 void __init hugetlb_add_hstate(unsigned int order)
2953 if (size_to_hstate(PAGE_SIZE << order)) {
2954 pr_warn("hugepagesz= specified twice, ignoring\n");
2957 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2959 h = &hstates[hugetlb_max_hstate++];
2961 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2962 h->nr_huge_pages = 0;
2963 h->free_huge_pages = 0;
2964 for (i = 0; i < MAX_NUMNODES; ++i)
2965 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2966 INIT_LIST_HEAD(&h->hugepage_activelist);
2967 h->next_nid_to_alloc = first_memory_node;
2968 h->next_nid_to_free = first_memory_node;
2969 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2970 huge_page_size(h)/1024);
2975 static int __init hugetlb_nrpages_setup(char *s)
2978 static unsigned long *last_mhp;
2980 if (!parsed_valid_hugepagesz) {
2981 pr_warn("hugepages = %s preceded by "
2982 "an unsupported hugepagesz, ignoring\n", s);
2983 parsed_valid_hugepagesz = true;
2987 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2988 * so this hugepages= parameter goes to the "default hstate".
2990 else if (!hugetlb_max_hstate)
2991 mhp = &default_hstate_max_huge_pages;
2993 mhp = &parsed_hstate->max_huge_pages;
2995 if (mhp == last_mhp) {
2996 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
3000 if (sscanf(s, "%lu", mhp) <= 0)
3004 * Global state is always initialized later in hugetlb_init.
3005 * But we need to allocate >= MAX_ORDER hstates here early to still
3006 * use the bootmem allocator.
3008 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
3009 hugetlb_hstate_alloc_pages(parsed_hstate);
3015 __setup("hugepages=", hugetlb_nrpages_setup);
3017 static int __init hugetlb_default_setup(char *s)
3019 default_hstate_size = memparse(s, &s);
3022 __setup("default_hugepagesz=", hugetlb_default_setup);
3024 static unsigned int cpuset_mems_nr(unsigned int *array)
3027 unsigned int nr = 0;
3029 for_each_node_mask(node, cpuset_current_mems_allowed)
3035 #ifdef CONFIG_SYSCTL
3036 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3037 struct ctl_table *table, int write,
3038 void __user *buffer, size_t *length, loff_t *ppos)
3040 struct hstate *h = &default_hstate;
3041 unsigned long tmp = h->max_huge_pages;
3044 if (!hugepages_supported())
3048 table->maxlen = sizeof(unsigned long);
3049 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3054 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3055 NUMA_NO_NODE, tmp, *length);
3060 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3061 void __user *buffer, size_t *length, loff_t *ppos)
3064 return hugetlb_sysctl_handler_common(false, table, write,
3065 buffer, length, ppos);
3069 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3070 void __user *buffer, size_t *length, loff_t *ppos)
3072 return hugetlb_sysctl_handler_common(true, table, write,
3073 buffer, length, ppos);
3075 #endif /* CONFIG_NUMA */
3077 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3078 void __user *buffer,
3079 size_t *length, loff_t *ppos)
3081 struct hstate *h = &default_hstate;
3085 if (!hugepages_supported())
3088 tmp = h->nr_overcommit_huge_pages;
3090 if (write && hstate_is_gigantic(h))
3094 table->maxlen = sizeof(unsigned long);
3095 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3100 spin_lock(&hugetlb_lock);
3101 h->nr_overcommit_huge_pages = tmp;
3102 spin_unlock(&hugetlb_lock);
3108 #endif /* CONFIG_SYSCTL */
3110 void hugetlb_report_meminfo(struct seq_file *m)
3113 unsigned long total = 0;
3115 if (!hugepages_supported())
3118 for_each_hstate(h) {
3119 unsigned long count = h->nr_huge_pages;
3121 total += (PAGE_SIZE << huge_page_order(h)) * count;
3123 if (h == &default_hstate)
3125 "HugePages_Total: %5lu\n"
3126 "HugePages_Free: %5lu\n"
3127 "HugePages_Rsvd: %5lu\n"
3128 "HugePages_Surp: %5lu\n"
3129 "Hugepagesize: %8lu kB\n",
3133 h->surplus_huge_pages,
3134 (PAGE_SIZE << huge_page_order(h)) / 1024);
3137 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3140 int hugetlb_report_node_meminfo(int nid, char *buf)
3142 struct hstate *h = &default_hstate;
3143 if (!hugepages_supported())
3146 "Node %d HugePages_Total: %5u\n"
3147 "Node %d HugePages_Free: %5u\n"
3148 "Node %d HugePages_Surp: %5u\n",
3149 nid, h->nr_huge_pages_node[nid],
3150 nid, h->free_huge_pages_node[nid],
3151 nid, h->surplus_huge_pages_node[nid]);
3154 void hugetlb_show_meminfo(void)
3159 if (!hugepages_supported())
3162 for_each_node_state(nid, N_MEMORY)
3164 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3166 h->nr_huge_pages_node[nid],
3167 h->free_huge_pages_node[nid],
3168 h->surplus_huge_pages_node[nid],
3169 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3172 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3174 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3175 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3178 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3179 unsigned long hugetlb_total_pages(void)
3182 unsigned long nr_total_pages = 0;
3185 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3186 return nr_total_pages;
3189 static int hugetlb_acct_memory(struct hstate *h, long delta)
3193 spin_lock(&hugetlb_lock);
3195 * When cpuset is configured, it breaks the strict hugetlb page
3196 * reservation as the accounting is done on a global variable. Such
3197 * reservation is completely rubbish in the presence of cpuset because
3198 * the reservation is not checked against page availability for the
3199 * current cpuset. Application can still potentially OOM'ed by kernel
3200 * with lack of free htlb page in cpuset that the task is in.
3201 * Attempt to enforce strict accounting with cpuset is almost
3202 * impossible (or too ugly) because cpuset is too fluid that
3203 * task or memory node can be dynamically moved between cpusets.
3205 * The change of semantics for shared hugetlb mapping with cpuset is
3206 * undesirable. However, in order to preserve some of the semantics,
3207 * we fall back to check against current free page availability as
3208 * a best attempt and hopefully to minimize the impact of changing
3209 * semantics that cpuset has.
3212 if (gather_surplus_pages(h, delta) < 0)
3215 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3216 return_unused_surplus_pages(h, delta);
3223 return_unused_surplus_pages(h, (unsigned long) -delta);
3226 spin_unlock(&hugetlb_lock);
3230 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3232 struct resv_map *resv = vma_resv_map(vma);
3235 * This new VMA should share its siblings reservation map if present.
3236 * The VMA will only ever have a valid reservation map pointer where
3237 * it is being copied for another still existing VMA. As that VMA
3238 * has a reference to the reservation map it cannot disappear until
3239 * after this open call completes. It is therefore safe to take a
3240 * new reference here without additional locking.
3242 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3243 kref_get(&resv->refs);
3246 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3248 struct hstate *h = hstate_vma(vma);
3249 struct resv_map *resv = vma_resv_map(vma);
3250 struct hugepage_subpool *spool = subpool_vma(vma);
3251 unsigned long reserve, start, end;
3254 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3257 start = vma_hugecache_offset(h, vma, vma->vm_start);
3258 end = vma_hugecache_offset(h, vma, vma->vm_end);
3260 reserve = (end - start) - region_count(resv, start, end);
3262 kref_put(&resv->refs, resv_map_release);
3266 * Decrement reserve counts. The global reserve count may be
3267 * adjusted if the subpool has a minimum size.
3269 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3270 hugetlb_acct_memory(h, -gbl_reserve);
3274 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3276 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3281 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3283 struct hstate *hstate = hstate_vma(vma);
3285 return 1UL << huge_page_shift(hstate);
3289 * We cannot handle pagefaults against hugetlb pages at all. They cause
3290 * handle_mm_fault() to try to instantiate regular-sized pages in the
3291 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3294 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3301 * When a new function is introduced to vm_operations_struct and added
3302 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3303 * This is because under System V memory model, mappings created via
3304 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3305 * their original vm_ops are overwritten with shm_vm_ops.
3307 const struct vm_operations_struct hugetlb_vm_ops = {
3308 .fault = hugetlb_vm_op_fault,
3309 .open = hugetlb_vm_op_open,
3310 .close = hugetlb_vm_op_close,
3311 .split = hugetlb_vm_op_split,
3312 .pagesize = hugetlb_vm_op_pagesize,
3315 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3321 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3322 vma->vm_page_prot)));
3324 entry = huge_pte_wrprotect(mk_huge_pte(page,
3325 vma->vm_page_prot));
3327 entry = pte_mkyoung(entry);
3328 entry = pte_mkhuge(entry);
3329 entry = arch_make_huge_pte(entry, vma, page, writable);
3334 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3335 unsigned long address, pte_t *ptep)
3339 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3340 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3341 update_mmu_cache(vma, address, ptep);
3344 bool is_hugetlb_entry_migration(pte_t pte)
3348 if (huge_pte_none(pte) || pte_present(pte))
3350 swp = pte_to_swp_entry(pte);
3351 if (non_swap_entry(swp) && is_migration_entry(swp))
3357 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3361 if (huge_pte_none(pte) || pte_present(pte))
3363 swp = pte_to_swp_entry(pte);
3364 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3370 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3371 struct vm_area_struct *vma)
3373 pte_t *src_pte, *dst_pte, entry, dst_entry;
3374 struct page *ptepage;
3377 struct hstate *h = hstate_vma(vma);
3378 unsigned long sz = huge_page_size(h);
3379 struct mmu_notifier_range range;
3382 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3385 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3388 mmu_notifier_invalidate_range_start(&range);
3391 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3392 spinlock_t *src_ptl, *dst_ptl;
3393 src_pte = huge_pte_offset(src, addr, sz);
3396 dst_pte = huge_pte_alloc(dst, addr, sz);
3403 * If the pagetables are shared don't copy or take references.
3404 * dst_pte == src_pte is the common case of src/dest sharing.
3406 * However, src could have 'unshared' and dst shares with
3407 * another vma. If dst_pte !none, this implies sharing.
3408 * Check here before taking page table lock, and once again
3409 * after taking the lock below.
3411 dst_entry = huge_ptep_get(dst_pte);
3412 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3415 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3416 src_ptl = huge_pte_lockptr(h, src, src_pte);
3417 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3418 entry = huge_ptep_get(src_pte);
3419 dst_entry = huge_ptep_get(dst_pte);
3420 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3422 * Skip if src entry none. Also, skip in the
3423 * unlikely case dst entry !none as this implies
3424 * sharing with another vma.
3427 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3428 is_hugetlb_entry_hwpoisoned(entry))) {
3429 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3431 if (is_write_migration_entry(swp_entry) && cow) {
3433 * COW mappings require pages in both
3434 * parent and child to be set to read.
3436 make_migration_entry_read(&swp_entry);
3437 entry = swp_entry_to_pte(swp_entry);
3438 set_huge_swap_pte_at(src, addr, src_pte,
3441 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3445 * No need to notify as we are downgrading page
3446 * table protection not changing it to point
3449 * See Documentation/vm/mmu_notifier.rst
3451 huge_ptep_set_wrprotect(src, addr, src_pte);
3453 entry = huge_ptep_get(src_pte);
3454 ptepage = pte_page(entry);
3456 page_dup_rmap(ptepage, true);
3457 set_huge_pte_at(dst, addr, dst_pte, entry);
3458 hugetlb_count_add(pages_per_huge_page(h), dst);
3460 spin_unlock(src_ptl);
3461 spin_unlock(dst_ptl);
3465 mmu_notifier_invalidate_range_end(&range);
3470 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3471 unsigned long start, unsigned long end,
3472 struct page *ref_page)
3474 struct mm_struct *mm = vma->vm_mm;
3475 unsigned long address;
3480 struct hstate *h = hstate_vma(vma);
3481 unsigned long sz = huge_page_size(h);
3482 struct mmu_notifier_range range;
3484 WARN_ON(!is_vm_hugetlb_page(vma));
3485 BUG_ON(start & ~huge_page_mask(h));
3486 BUG_ON(end & ~huge_page_mask(h));
3489 * This is a hugetlb vma, all the pte entries should point
3492 tlb_change_page_size(tlb, sz);
3493 tlb_start_vma(tlb, vma);
3496 * If sharing possible, alert mmu notifiers of worst case.
3498 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3500 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3501 mmu_notifier_invalidate_range_start(&range);
3503 for (; address < end; address += sz) {
3504 ptep = huge_pte_offset(mm, address, sz);
3508 ptl = huge_pte_lock(h, mm, ptep);
3509 if (huge_pmd_unshare(mm, &address, ptep)) {
3512 * We just unmapped a page of PMDs by clearing a PUD.
3513 * The caller's TLB flush range should cover this area.
3518 pte = huge_ptep_get(ptep);
3519 if (huge_pte_none(pte)) {
3525 * Migrating hugepage or HWPoisoned hugepage is already
3526 * unmapped and its refcount is dropped, so just clear pte here.
3528 if (unlikely(!pte_present(pte))) {
3529 huge_pte_clear(mm, address, ptep, sz);
3534 page = pte_page(pte);
3536 * If a reference page is supplied, it is because a specific
3537 * page is being unmapped, not a range. Ensure the page we
3538 * are about to unmap is the actual page of interest.
3541 if (page != ref_page) {
3546 * Mark the VMA as having unmapped its page so that
3547 * future faults in this VMA will fail rather than
3548 * looking like data was lost
3550 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3553 pte = huge_ptep_get_and_clear(mm, address, ptep);
3554 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3555 if (huge_pte_dirty(pte))
3556 set_page_dirty(page);
3558 hugetlb_count_sub(pages_per_huge_page(h), mm);
3559 page_remove_rmap(page, true);
3562 tlb_remove_page_size(tlb, page, huge_page_size(h));
3564 * Bail out after unmapping reference page if supplied
3569 mmu_notifier_invalidate_range_end(&range);
3570 tlb_end_vma(tlb, vma);
3573 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3574 struct vm_area_struct *vma, unsigned long start,
3575 unsigned long end, struct page *ref_page)
3577 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3580 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3581 * test will fail on a vma being torn down, and not grab a page table
3582 * on its way out. We're lucky that the flag has such an appropriate
3583 * name, and can in fact be safely cleared here. We could clear it
3584 * before the __unmap_hugepage_range above, but all that's necessary
3585 * is to clear it before releasing the i_mmap_rwsem. This works
3586 * because in the context this is called, the VMA is about to be
3587 * destroyed and the i_mmap_rwsem is held.
3589 vma->vm_flags &= ~VM_MAYSHARE;
3592 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3593 unsigned long end, struct page *ref_page)
3595 struct mm_struct *mm;
3596 struct mmu_gather tlb;
3597 unsigned long tlb_start = start;
3598 unsigned long tlb_end = end;
3601 * If shared PMDs were possibly used within this vma range, adjust
3602 * start/end for worst case tlb flushing.
3603 * Note that we can not be sure if PMDs are shared until we try to
3604 * unmap pages. However, we want to make sure TLB flushing covers
3605 * the largest possible range.
3607 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
3611 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
3612 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3613 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
3617 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3618 * mappping it owns the reserve page for. The intention is to unmap the page
3619 * from other VMAs and let the children be SIGKILLed if they are faulting the
3622 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3623 struct page *page, unsigned long address)
3625 struct hstate *h = hstate_vma(vma);
3626 struct vm_area_struct *iter_vma;
3627 struct address_space *mapping;
3631 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3632 * from page cache lookup which is in HPAGE_SIZE units.
3634 address = address & huge_page_mask(h);
3635 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3637 mapping = vma->vm_file->f_mapping;
3640 * Take the mapping lock for the duration of the table walk. As
3641 * this mapping should be shared between all the VMAs,
3642 * __unmap_hugepage_range() is called as the lock is already held
3644 i_mmap_lock_write(mapping);
3645 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3646 /* Do not unmap the current VMA */
3647 if (iter_vma == vma)
3651 * Shared VMAs have their own reserves and do not affect
3652 * MAP_PRIVATE accounting but it is possible that a shared
3653 * VMA is using the same page so check and skip such VMAs.
3655 if (iter_vma->vm_flags & VM_MAYSHARE)
3659 * Unmap the page from other VMAs without their own reserves.
3660 * They get marked to be SIGKILLed if they fault in these
3661 * areas. This is because a future no-page fault on this VMA
3662 * could insert a zeroed page instead of the data existing
3663 * from the time of fork. This would look like data corruption
3665 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3666 unmap_hugepage_range(iter_vma, address,
3667 address + huge_page_size(h), page);
3669 i_mmap_unlock_write(mapping);
3673 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3674 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3675 * cannot race with other handlers or page migration.
3676 * Keep the pte_same checks anyway to make transition from the mutex easier.
3678 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3679 unsigned long address, pte_t *ptep,
3680 struct page *pagecache_page, spinlock_t *ptl)
3683 struct hstate *h = hstate_vma(vma);
3684 struct page *old_page, *new_page;
3685 int outside_reserve = 0;
3687 unsigned long haddr = address & huge_page_mask(h);
3688 struct mmu_notifier_range range;
3690 pte = huge_ptep_get(ptep);
3691 old_page = pte_page(pte);
3694 /* If no-one else is actually using this page, avoid the copy
3695 * and just make the page writable */
3696 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3697 page_move_anon_rmap(old_page, vma);
3698 set_huge_ptep_writable(vma, haddr, ptep);
3703 * If the process that created a MAP_PRIVATE mapping is about to
3704 * perform a COW due to a shared page count, attempt to satisfy
3705 * the allocation without using the existing reserves. The pagecache
3706 * page is used to determine if the reserve at this address was
3707 * consumed or not. If reserves were used, a partial faulted mapping
3708 * at the time of fork() could consume its reserves on COW instead
3709 * of the full address range.
3711 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3712 old_page != pagecache_page)
3713 outside_reserve = 1;
3718 * Drop page table lock as buddy allocator may be called. It will
3719 * be acquired again before returning to the caller, as expected.
3722 new_page = alloc_huge_page(vma, haddr, outside_reserve);
3724 if (IS_ERR(new_page)) {
3726 * If a process owning a MAP_PRIVATE mapping fails to COW,
3727 * it is due to references held by a child and an insufficient
3728 * huge page pool. To guarantee the original mappers
3729 * reliability, unmap the page from child processes. The child
3730 * may get SIGKILLed if it later faults.
3732 if (outside_reserve) {
3734 BUG_ON(huge_pte_none(pte));
3735 unmap_ref_private(mm, vma, old_page, haddr);
3736 BUG_ON(huge_pte_none(pte));
3738 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3740 pte_same(huge_ptep_get(ptep), pte)))
3741 goto retry_avoidcopy;
3743 * race occurs while re-acquiring page table
3744 * lock, and our job is done.
3749 ret = vmf_error(PTR_ERR(new_page));
3750 goto out_release_old;
3754 * When the original hugepage is shared one, it does not have
3755 * anon_vma prepared.
3757 if (unlikely(anon_vma_prepare(vma))) {
3759 goto out_release_all;
3762 copy_user_huge_page(new_page, old_page, address, vma,
3763 pages_per_huge_page(h));
3764 __SetPageUptodate(new_page);
3766 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
3767 haddr + huge_page_size(h));
3768 mmu_notifier_invalidate_range_start(&range);
3771 * Retake the page table lock to check for racing updates
3772 * before the page tables are altered
3775 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3776 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3777 ClearPagePrivate(new_page);
3780 huge_ptep_clear_flush(vma, haddr, ptep);
3781 mmu_notifier_invalidate_range(mm, range.start, range.end);
3782 set_huge_pte_at(mm, haddr, ptep,
3783 make_huge_pte(vma, new_page, 1));
3784 page_remove_rmap(old_page, true);
3785 hugepage_add_new_anon_rmap(new_page, vma, haddr);
3786 set_page_huge_active(new_page);
3787 /* Make the old page be freed below */
3788 new_page = old_page;
3791 mmu_notifier_invalidate_range_end(&range);
3793 restore_reserve_on_error(h, vma, haddr, new_page);
3798 spin_lock(ptl); /* Caller expects lock to be held */
3802 /* Return the pagecache page at a given address within a VMA */
3803 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3804 struct vm_area_struct *vma, unsigned long address)
3806 struct address_space *mapping;
3809 mapping = vma->vm_file->f_mapping;
3810 idx = vma_hugecache_offset(h, vma, address);
3812 return find_lock_page(mapping, idx);
3816 * Return whether there is a pagecache page to back given address within VMA.
3817 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3819 static bool hugetlbfs_pagecache_present(struct hstate *h,
3820 struct vm_area_struct *vma, unsigned long address)
3822 struct address_space *mapping;
3826 mapping = vma->vm_file->f_mapping;
3827 idx = vma_hugecache_offset(h, vma, address);
3829 page = find_get_page(mapping, idx);
3832 return page != NULL;
3835 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3838 struct inode *inode = mapping->host;
3839 struct hstate *h = hstate_inode(inode);
3840 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3844 ClearPagePrivate(page);
3847 * set page dirty so that it will not be removed from cache/file
3848 * by non-hugetlbfs specific code paths.
3850 set_page_dirty(page);
3852 spin_lock(&inode->i_lock);
3853 inode->i_blocks += blocks_per_huge_page(h);
3854 spin_unlock(&inode->i_lock);
3858 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
3859 struct vm_area_struct *vma,
3860 struct address_space *mapping, pgoff_t idx,
3861 unsigned long address, pte_t *ptep, unsigned int flags)
3863 struct hstate *h = hstate_vma(vma);
3864 vm_fault_t ret = VM_FAULT_SIGBUS;
3870 unsigned long haddr = address & huge_page_mask(h);
3871 bool new_page = false;
3874 * Currently, we are forced to kill the process in the event the
3875 * original mapper has unmapped pages from the child due to a failed
3876 * COW. Warn that such a situation has occurred as it may not be obvious
3878 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3879 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3885 * Use page lock to guard against racing truncation
3886 * before we get page_table_lock.
3889 page = find_lock_page(mapping, idx);
3891 size = i_size_read(mapping->host) >> huge_page_shift(h);
3896 * Check for page in userfault range
3898 if (userfaultfd_missing(vma)) {
3900 struct vm_fault vmf = {
3905 * Hard to debug if it ends up being
3906 * used by a callee that assumes
3907 * something about the other
3908 * uninitialized fields... same as in
3914 * hugetlb_fault_mutex must be dropped before
3915 * handling userfault. Reacquire after handling
3916 * fault to make calling code simpler.
3918 hash = hugetlb_fault_mutex_hash(h, mapping, idx, haddr);
3919 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3920 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3921 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3925 page = alloc_huge_page(vma, haddr, 0);
3928 * Returning error will result in faulting task being
3929 * sent SIGBUS. The hugetlb fault mutex prevents two
3930 * tasks from racing to fault in the same page which
3931 * could result in false unable to allocate errors.
3932 * Page migration does not take the fault mutex, but
3933 * does a clear then write of pte's under page table
3934 * lock. Page fault code could race with migration,
3935 * notice the clear pte and try to allocate a page
3936 * here. Before returning error, get ptl and make
3937 * sure there really is no pte entry.
3939 ptl = huge_pte_lock(h, mm, ptep);
3940 if (!huge_pte_none(huge_ptep_get(ptep))) {
3946 ret = vmf_error(PTR_ERR(page));
3949 clear_huge_page(page, address, pages_per_huge_page(h));
3950 __SetPageUptodate(page);
3953 if (vma->vm_flags & VM_MAYSHARE) {
3954 int err = huge_add_to_page_cache(page, mapping, idx);
3963 if (unlikely(anon_vma_prepare(vma))) {
3965 goto backout_unlocked;
3971 * If memory error occurs between mmap() and fault, some process
3972 * don't have hwpoisoned swap entry for errored virtual address.
3973 * So we need to block hugepage fault by PG_hwpoison bit check.
3975 if (unlikely(PageHWPoison(page))) {
3976 ret = VM_FAULT_HWPOISON |
3977 VM_FAULT_SET_HINDEX(hstate_index(h));
3978 goto backout_unlocked;
3983 * If we are going to COW a private mapping later, we examine the
3984 * pending reservations for this page now. This will ensure that
3985 * any allocations necessary to record that reservation occur outside
3988 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3989 if (vma_needs_reservation(h, vma, haddr) < 0) {
3991 goto backout_unlocked;
3993 /* Just decrements count, does not deallocate */
3994 vma_end_reservation(h, vma, haddr);
3997 ptl = huge_pte_lock(h, mm, ptep);
3998 size = i_size_read(mapping->host) >> huge_page_shift(h);
4003 if (!huge_pte_none(huge_ptep_get(ptep)))
4007 ClearPagePrivate(page);
4008 hugepage_add_new_anon_rmap(page, vma, haddr);
4010 page_dup_rmap(page, true);
4011 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4012 && (vma->vm_flags & VM_SHARED)));
4013 set_huge_pte_at(mm, haddr, ptep, new_pte);
4015 hugetlb_count_add(pages_per_huge_page(h), mm);
4016 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4017 /* Optimization, do the COW without a second fault */
4018 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4024 * Only make newly allocated pages active. Existing pages found
4025 * in the pagecache could be !page_huge_active() if they have been
4026 * isolated for migration.
4029 set_page_huge_active(page);
4039 restore_reserve_on_error(h, vma, haddr, page);
4045 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping,
4046 pgoff_t idx, unsigned long address)
4048 unsigned long key[2];
4051 key[0] = (unsigned long) mapping;
4054 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
4056 return hash & (num_fault_mutexes - 1);
4060 * For uniprocesor systems we always use a single mutex, so just
4061 * return 0 and avoid the hashing overhead.
4063 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping,
4064 pgoff_t idx, unsigned long address)
4070 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4071 unsigned long address, unsigned int flags)
4078 struct page *page = NULL;
4079 struct page *pagecache_page = NULL;
4080 struct hstate *h = hstate_vma(vma);
4081 struct address_space *mapping;
4082 int need_wait_lock = 0;
4083 unsigned long haddr = address & huge_page_mask(h);
4085 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4087 entry = huge_ptep_get(ptep);
4088 if (unlikely(is_hugetlb_entry_migration(entry))) {
4089 migration_entry_wait_huge(vma, mm, ptep);
4091 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4092 return VM_FAULT_HWPOISON_LARGE |
4093 VM_FAULT_SET_HINDEX(hstate_index(h));
4095 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4097 return VM_FAULT_OOM;
4100 mapping = vma->vm_file->f_mapping;
4101 idx = vma_hugecache_offset(h, vma, haddr);
4104 * Serialize hugepage allocation and instantiation, so that we don't
4105 * get spurious allocation failures if two CPUs race to instantiate
4106 * the same page in the page cache.
4108 hash = hugetlb_fault_mutex_hash(h, mapping, idx, haddr);
4109 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4111 entry = huge_ptep_get(ptep);
4112 if (huge_pte_none(entry)) {
4113 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4120 * entry could be a migration/hwpoison entry at this point, so this
4121 * check prevents the kernel from going below assuming that we have
4122 * a active hugepage in pagecache. This goto expects the 2nd page fault,
4123 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
4126 if (!pte_present(entry))
4130 * If we are going to COW the mapping later, we examine the pending
4131 * reservations for this page now. This will ensure that any
4132 * allocations necessary to record that reservation occur outside the
4133 * spinlock. For private mappings, we also lookup the pagecache
4134 * page now as it is used to determine if a reservation has been
4137 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4138 if (vma_needs_reservation(h, vma, haddr) < 0) {
4142 /* Just decrements count, does not deallocate */
4143 vma_end_reservation(h, vma, haddr);
4145 if (!(vma->vm_flags & VM_MAYSHARE))
4146 pagecache_page = hugetlbfs_pagecache_page(h,
4150 ptl = huge_pte_lock(h, mm, ptep);
4152 /* Check for a racing update before calling hugetlb_cow */
4153 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4157 * hugetlb_cow() requires page locks of pte_page(entry) and
4158 * pagecache_page, so here we need take the former one
4159 * when page != pagecache_page or !pagecache_page.
4161 page = pte_page(entry);
4162 if (page != pagecache_page)
4163 if (!trylock_page(page)) {
4170 if (flags & FAULT_FLAG_WRITE) {
4171 if (!huge_pte_write(entry)) {
4172 ret = hugetlb_cow(mm, vma, address, ptep,
4173 pagecache_page, ptl);
4176 entry = huge_pte_mkdirty(entry);
4178 entry = pte_mkyoung(entry);
4179 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4180 flags & FAULT_FLAG_WRITE))
4181 update_mmu_cache(vma, haddr, ptep);
4183 if (page != pagecache_page)
4189 if (pagecache_page) {
4190 unlock_page(pagecache_page);
4191 put_page(pagecache_page);
4194 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4196 * Generally it's safe to hold refcount during waiting page lock. But
4197 * here we just wait to defer the next page fault to avoid busy loop and
4198 * the page is not used after unlocked before returning from the current
4199 * page fault. So we are safe from accessing freed page, even if we wait
4200 * here without taking refcount.
4203 wait_on_page_locked(page);
4208 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4209 * modifications for huge pages.
4211 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4213 struct vm_area_struct *dst_vma,
4214 unsigned long dst_addr,
4215 unsigned long src_addr,
4216 struct page **pagep)
4218 struct address_space *mapping;
4221 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4222 struct hstate *h = hstate_vma(dst_vma);
4230 page = alloc_huge_page(dst_vma, dst_addr, 0);
4234 ret = copy_huge_page_from_user(page,
4235 (const void __user *) src_addr,
4236 pages_per_huge_page(h), false);
4238 /* fallback to copy_from_user outside mmap_sem */
4239 if (unlikely(ret)) {
4242 /* don't free the page */
4251 * The memory barrier inside __SetPageUptodate makes sure that
4252 * preceding stores to the page contents become visible before
4253 * the set_pte_at() write.
4255 __SetPageUptodate(page);
4257 mapping = dst_vma->vm_file->f_mapping;
4258 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4261 * If shared, add to page cache
4264 size = i_size_read(mapping->host) >> huge_page_shift(h);
4267 goto out_release_nounlock;
4270 * Serialization between remove_inode_hugepages() and
4271 * huge_add_to_page_cache() below happens through the
4272 * hugetlb_fault_mutex_table that here must be hold by
4275 ret = huge_add_to_page_cache(page, mapping, idx);
4277 goto out_release_nounlock;
4280 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4284 * Recheck the i_size after holding PT lock to make sure not
4285 * to leave any page mapped (as page_mapped()) beyond the end
4286 * of the i_size (remove_inode_hugepages() is strict about
4287 * enforcing that). If we bail out here, we'll also leave a
4288 * page in the radix tree in the vm_shared case beyond the end
4289 * of the i_size, but remove_inode_hugepages() will take care
4290 * of it as soon as we drop the hugetlb_fault_mutex_table.
4292 size = i_size_read(mapping->host) >> huge_page_shift(h);
4295 goto out_release_unlock;
4298 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4299 goto out_release_unlock;
4302 page_dup_rmap(page, true);
4304 ClearPagePrivate(page);
4305 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4308 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4309 if (dst_vma->vm_flags & VM_WRITE)
4310 _dst_pte = huge_pte_mkdirty(_dst_pte);
4311 _dst_pte = pte_mkyoung(_dst_pte);
4313 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4315 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4316 dst_vma->vm_flags & VM_WRITE);
4317 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4319 /* No need to invalidate - it was non-present before */
4320 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4323 set_page_huge_active(page);
4333 out_release_nounlock:
4338 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4339 struct page **pages, struct vm_area_struct **vmas,
4340 unsigned long *position, unsigned long *nr_pages,
4341 long i, unsigned int flags, int *nonblocking)
4343 unsigned long pfn_offset;
4344 unsigned long vaddr = *position;
4345 unsigned long remainder = *nr_pages;
4346 struct hstate *h = hstate_vma(vma);
4349 while (vaddr < vma->vm_end && remainder) {
4351 spinlock_t *ptl = NULL;
4356 * If we have a pending SIGKILL, don't keep faulting pages and
4357 * potentially allocating memory.
4359 if (fatal_signal_pending(current)) {
4365 * Some archs (sparc64, sh*) have multiple pte_ts to
4366 * each hugepage. We have to make sure we get the
4367 * first, for the page indexing below to work.
4369 * Note that page table lock is not held when pte is null.
4371 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4374 ptl = huge_pte_lock(h, mm, pte);
4375 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4378 * When coredumping, it suits get_dump_page if we just return
4379 * an error where there's an empty slot with no huge pagecache
4380 * to back it. This way, we avoid allocating a hugepage, and
4381 * the sparse dumpfile avoids allocating disk blocks, but its
4382 * huge holes still show up with zeroes where they need to be.
4384 if (absent && (flags & FOLL_DUMP) &&
4385 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4393 * We need call hugetlb_fault for both hugepages under migration
4394 * (in which case hugetlb_fault waits for the migration,) and
4395 * hwpoisoned hugepages (in which case we need to prevent the
4396 * caller from accessing to them.) In order to do this, we use
4397 * here is_swap_pte instead of is_hugetlb_entry_migration and
4398 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4399 * both cases, and because we can't follow correct pages
4400 * directly from any kind of swap entries.
4402 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4403 ((flags & FOLL_WRITE) &&
4404 !huge_pte_write(huge_ptep_get(pte)))) {
4406 unsigned int fault_flags = 0;
4410 if (flags & FOLL_WRITE)
4411 fault_flags |= FAULT_FLAG_WRITE;
4413 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4414 if (flags & FOLL_NOWAIT)
4415 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4416 FAULT_FLAG_RETRY_NOWAIT;
4417 if (flags & FOLL_TRIED) {
4418 VM_WARN_ON_ONCE(fault_flags &
4419 FAULT_FLAG_ALLOW_RETRY);
4420 fault_flags |= FAULT_FLAG_TRIED;
4422 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4423 if (ret & VM_FAULT_ERROR) {
4424 err = vm_fault_to_errno(ret, flags);
4428 if (ret & VM_FAULT_RETRY) {
4430 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4434 * VM_FAULT_RETRY must not return an
4435 * error, it will return zero
4438 * No need to update "position" as the
4439 * caller will not check it after
4440 * *nr_pages is set to 0.
4447 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4448 page = pte_page(huge_ptep_get(pte));
4451 * Instead of doing 'try_get_page()' below in the same_page
4452 * loop, just check the count once here.
4454 if (unlikely(page_count(page) <= 0)) {
4464 pages[i] = mem_map_offset(page, pfn_offset);
4475 if (vaddr < vma->vm_end && remainder &&
4476 pfn_offset < pages_per_huge_page(h)) {
4478 * We use pfn_offset to avoid touching the pageframes
4479 * of this compound page.
4485 *nr_pages = remainder;
4487 * setting position is actually required only if remainder is
4488 * not zero but it's faster not to add a "if (remainder)"
4496 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4498 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4501 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4504 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4505 unsigned long address, unsigned long end, pgprot_t newprot)
4507 struct mm_struct *mm = vma->vm_mm;
4508 unsigned long start = address;
4511 struct hstate *h = hstate_vma(vma);
4512 unsigned long pages = 0;
4513 bool shared_pmd = false;
4514 struct mmu_notifier_range range;
4517 * In the case of shared PMDs, the area to flush could be beyond
4518 * start/end. Set range.start/range.end to cover the maximum possible
4519 * range if PMD sharing is possible.
4521 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
4522 0, vma, mm, start, end);
4523 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4525 BUG_ON(address >= end);
4526 flush_cache_range(vma, range.start, range.end);
4528 mmu_notifier_invalidate_range_start(&range);
4529 i_mmap_lock_write(vma->vm_file->f_mapping);
4530 for (; address < end; address += huge_page_size(h)) {
4532 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4535 ptl = huge_pte_lock(h, mm, ptep);
4536 if (huge_pmd_unshare(mm, &address, ptep)) {
4542 pte = huge_ptep_get(ptep);
4543 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4547 if (unlikely(is_hugetlb_entry_migration(pte))) {
4548 swp_entry_t entry = pte_to_swp_entry(pte);
4550 if (is_write_migration_entry(entry)) {
4553 make_migration_entry_read(&entry);
4554 newpte = swp_entry_to_pte(entry);
4555 set_huge_swap_pte_at(mm, address, ptep,
4556 newpte, huge_page_size(h));
4562 if (!huge_pte_none(pte)) {
4565 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
4566 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
4567 pte = arch_make_huge_pte(pte, vma, NULL, 0);
4568 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
4574 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4575 * may have cleared our pud entry and done put_page on the page table:
4576 * once we release i_mmap_rwsem, another task can do the final put_page
4577 * and that page table be reused and filled with junk. If we actually
4578 * did unshare a page of pmds, flush the range corresponding to the pud.
4581 flush_hugetlb_tlb_range(vma, range.start, range.end);
4583 flush_hugetlb_tlb_range(vma, start, end);
4585 * No need to call mmu_notifier_invalidate_range() we are downgrading
4586 * page table protection not changing it to point to a new page.
4588 * See Documentation/vm/mmu_notifier.rst
4590 i_mmap_unlock_write(vma->vm_file->f_mapping);
4591 mmu_notifier_invalidate_range_end(&range);
4593 return pages << h->order;
4596 int hugetlb_reserve_pages(struct inode *inode,
4598 struct vm_area_struct *vma,
4599 vm_flags_t vm_flags)
4602 struct hstate *h = hstate_inode(inode);
4603 struct hugepage_subpool *spool = subpool_inode(inode);
4604 struct resv_map *resv_map;
4607 /* This should never happen */
4609 VM_WARN(1, "%s called with a negative range\n", __func__);
4614 * Only apply hugepage reservation if asked. At fault time, an
4615 * attempt will be made for VM_NORESERVE to allocate a page
4616 * without using reserves
4618 if (vm_flags & VM_NORESERVE)
4622 * Shared mappings base their reservation on the number of pages that
4623 * are already allocated on behalf of the file. Private mappings need
4624 * to reserve the full area even if read-only as mprotect() may be
4625 * called to make the mapping read-write. Assume !vma is a shm mapping
4627 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4629 * resv_map can not be NULL as hugetlb_reserve_pages is only
4630 * called for inodes for which resv_maps were created (see
4631 * hugetlbfs_get_inode).
4633 resv_map = inode_resv_map(inode);
4635 chg = region_chg(resv_map, from, to);
4638 resv_map = resv_map_alloc();
4644 set_vma_resv_map(vma, resv_map);
4645 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4654 * There must be enough pages in the subpool for the mapping. If
4655 * the subpool has a minimum size, there may be some global
4656 * reservations already in place (gbl_reserve).
4658 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4659 if (gbl_reserve < 0) {
4665 * Check enough hugepages are available for the reservation.
4666 * Hand the pages back to the subpool if there are not
4668 ret = hugetlb_acct_memory(h, gbl_reserve);
4670 /* put back original number of pages, chg */
4671 (void)hugepage_subpool_put_pages(spool, chg);
4676 * Account for the reservations made. Shared mappings record regions
4677 * that have reservations as they are shared by multiple VMAs.
4678 * When the last VMA disappears, the region map says how much
4679 * the reservation was and the page cache tells how much of
4680 * the reservation was consumed. Private mappings are per-VMA and
4681 * only the consumed reservations are tracked. When the VMA
4682 * disappears, the original reservation is the VMA size and the
4683 * consumed reservations are stored in the map. Hence, nothing
4684 * else has to be done for private mappings here
4686 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4687 long add = region_add(resv_map, from, to);
4689 if (unlikely(chg > add)) {
4691 * pages in this range were added to the reserve
4692 * map between region_chg and region_add. This
4693 * indicates a race with alloc_huge_page. Adjust
4694 * the subpool and reserve counts modified above
4695 * based on the difference.
4699 rsv_adjust = hugepage_subpool_put_pages(spool,
4701 hugetlb_acct_memory(h, -rsv_adjust);
4706 if (!vma || vma->vm_flags & VM_MAYSHARE)
4707 /* Don't call region_abort if region_chg failed */
4709 region_abort(resv_map, from, to);
4710 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4711 kref_put(&resv_map->refs, resv_map_release);
4715 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4718 struct hstate *h = hstate_inode(inode);
4719 struct resv_map *resv_map = inode_resv_map(inode);
4721 struct hugepage_subpool *spool = subpool_inode(inode);
4725 * Since this routine can be called in the evict inode path for all
4726 * hugetlbfs inodes, resv_map could be NULL.
4729 chg = region_del(resv_map, start, end);
4731 * region_del() can fail in the rare case where a region
4732 * must be split and another region descriptor can not be
4733 * allocated. If end == LONG_MAX, it will not fail.
4739 spin_lock(&inode->i_lock);
4740 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4741 spin_unlock(&inode->i_lock);
4744 * If the subpool has a minimum size, the number of global
4745 * reservations to be released may be adjusted.
4747 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4748 hugetlb_acct_memory(h, -gbl_reserve);
4753 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4754 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4755 struct vm_area_struct *vma,
4756 unsigned long addr, pgoff_t idx)
4758 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4760 unsigned long sbase = saddr & PUD_MASK;
4761 unsigned long s_end = sbase + PUD_SIZE;
4763 /* Allow segments to share if only one is marked locked */
4764 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4765 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4768 * match the virtual addresses, permission and the alignment of the
4771 if (pmd_index(addr) != pmd_index(saddr) ||
4772 vm_flags != svm_flags ||
4773 sbase < svma->vm_start || svma->vm_end < s_end)
4779 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4781 unsigned long base = addr & PUD_MASK;
4782 unsigned long end = base + PUD_SIZE;
4785 * check on proper vm_flags and page table alignment
4787 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
4793 * Determine if start,end range within vma could be mapped by shared pmd.
4794 * If yes, adjust start and end to cover range associated with possible
4795 * shared pmd mappings.
4797 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4798 unsigned long *start, unsigned long *end)
4800 unsigned long check_addr = *start;
4802 if (!(vma->vm_flags & VM_MAYSHARE))
4805 for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
4806 unsigned long a_start = check_addr & PUD_MASK;
4807 unsigned long a_end = a_start + PUD_SIZE;
4810 * If sharing is possible, adjust start/end if necessary.
4812 if (range_in_vma(vma, a_start, a_end)) {
4813 if (a_start < *start)
4822 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4823 * and returns the corresponding pte. While this is not necessary for the
4824 * !shared pmd case because we can allocate the pmd later as well, it makes the
4825 * code much cleaner. pmd allocation is essential for the shared case because
4826 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4827 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4828 * bad pmd for sharing.
4830 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4832 struct vm_area_struct *vma = find_vma(mm, addr);
4833 struct address_space *mapping = vma->vm_file->f_mapping;
4834 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4836 struct vm_area_struct *svma;
4837 unsigned long saddr;
4842 if (!vma_shareable(vma, addr))
4843 return (pte_t *)pmd_alloc(mm, pud, addr);
4845 i_mmap_lock_write(mapping);
4846 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4850 saddr = page_table_shareable(svma, vma, addr, idx);
4852 spte = huge_pte_offset(svma->vm_mm, saddr,
4853 vma_mmu_pagesize(svma));
4855 get_page(virt_to_page(spte));
4864 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4865 if (pud_none(*pud)) {
4866 pud_populate(mm, pud,
4867 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4870 put_page(virt_to_page(spte));
4874 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4875 i_mmap_unlock_write(mapping);
4880 * unmap huge page backed by shared pte.
4882 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4883 * indicated by page_count > 1, unmap is achieved by clearing pud and
4884 * decrementing the ref count. If count == 1, the pte page is not shared.
4886 * called with page table lock held.
4888 * returns: 1 successfully unmapped a shared pte page
4889 * 0 the underlying pte page is not shared, or it is the last user
4891 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4893 pgd_t *pgd = pgd_offset(mm, *addr);
4894 p4d_t *p4d = p4d_offset(pgd, *addr);
4895 pud_t *pud = pud_offset(p4d, *addr);
4897 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4898 if (page_count(virt_to_page(ptep)) == 1)
4902 put_page(virt_to_page(ptep));
4904 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4907 #define want_pmd_share() (1)
4908 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4909 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4914 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4919 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4920 unsigned long *start, unsigned long *end)
4923 #define want_pmd_share() (0)
4924 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4926 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4927 pte_t *huge_pte_alloc(struct mm_struct *mm,
4928 unsigned long addr, unsigned long sz)
4935 pgd = pgd_offset(mm, addr);
4936 p4d = p4d_alloc(mm, pgd, addr);
4939 pud = pud_alloc(mm, p4d, addr);
4941 if (sz == PUD_SIZE) {
4944 BUG_ON(sz != PMD_SIZE);
4945 if (want_pmd_share() && pud_none(*pud))
4946 pte = huge_pmd_share(mm, addr, pud);
4948 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4951 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4957 * huge_pte_offset() - Walk the page table to resolve the hugepage
4958 * entry at address @addr
4960 * Return: Pointer to page table or swap entry (PUD or PMD) for
4961 * address @addr, or NULL if a p*d_none() entry is encountered and the
4962 * size @sz doesn't match the hugepage size at this level of the page
4965 pte_t *huge_pte_offset(struct mm_struct *mm,
4966 unsigned long addr, unsigned long sz)
4973 pgd = pgd_offset(mm, addr);
4974 if (!pgd_present(*pgd))
4976 p4d = p4d_offset(pgd, addr);
4977 if (!p4d_present(*p4d))
4980 pud = pud_offset(p4d, addr);
4981 if (sz != PUD_SIZE && pud_none(*pud))
4983 /* hugepage or swap? */
4984 if (pud_huge(*pud) || !pud_present(*pud))
4985 return (pte_t *)pud;
4987 pmd = pmd_offset(pud, addr);
4988 if (sz != PMD_SIZE && pmd_none(*pmd))
4990 /* hugepage or swap? */
4991 if (pmd_huge(*pmd) || !pmd_present(*pmd))
4992 return (pte_t *)pmd;
4997 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5000 * These functions are overwritable if your architecture needs its own
5003 struct page * __weak
5004 follow_huge_addr(struct mm_struct *mm, unsigned long address,
5007 return ERR_PTR(-EINVAL);
5010 struct page * __weak
5011 follow_huge_pd(struct vm_area_struct *vma,
5012 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5014 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5018 struct page * __weak
5019 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
5020 pmd_t *pmd, int flags)
5022 struct page *page = NULL;
5026 ptl = pmd_lockptr(mm, pmd);
5029 * make sure that the address range covered by this pmd is not
5030 * unmapped from other threads.
5032 if (!pmd_huge(*pmd))
5034 pte = huge_ptep_get((pte_t *)pmd);
5035 if (pte_present(pte)) {
5036 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
5037 if (flags & FOLL_GET)
5040 if (is_hugetlb_entry_migration(pte)) {
5042 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5046 * hwpoisoned entry is treated as no_page_table in
5047 * follow_page_mask().
5055 struct page * __weak
5056 follow_huge_pud(struct mm_struct *mm, unsigned long address,
5057 pud_t *pud, int flags)
5059 if (flags & FOLL_GET)
5062 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5065 struct page * __weak
5066 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5068 if (flags & FOLL_GET)
5071 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5074 bool isolate_huge_page(struct page *page, struct list_head *list)
5078 VM_BUG_ON_PAGE(!PageHead(page), page);
5079 spin_lock(&hugetlb_lock);
5080 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
5084 clear_page_huge_active(page);
5085 list_move_tail(&page->lru, list);
5087 spin_unlock(&hugetlb_lock);
5091 void putback_active_hugepage(struct page *page)
5093 VM_BUG_ON_PAGE(!PageHead(page), page);
5094 spin_lock(&hugetlb_lock);
5095 set_page_huge_active(page);
5096 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5097 spin_unlock(&hugetlb_lock);
5101 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5103 struct hstate *h = page_hstate(oldpage);
5105 hugetlb_cgroup_migrate(oldpage, newpage);
5106 set_page_owner_migrate_reason(newpage, reason);
5109 * transfer temporary state of the new huge page. This is
5110 * reverse to other transitions because the newpage is going to
5111 * be final while the old one will be freed so it takes over
5112 * the temporary status.
5114 * Also note that we have to transfer the per-node surplus state
5115 * here as well otherwise the global surplus count will not match
5118 if (PageHugeTemporary(newpage)) {
5119 int old_nid = page_to_nid(oldpage);
5120 int new_nid = page_to_nid(newpage);
5122 SetPageHugeTemporary(oldpage);
5123 ClearPageHugeTemporary(newpage);
5125 spin_lock(&hugetlb_lock);
5126 if (h->surplus_huge_pages_node[old_nid]) {
5127 h->surplus_huge_pages_node[old_nid]--;
5128 h->surplus_huge_pages_node[new_nid]++;
5130 spin_unlock(&hugetlb_lock);