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 struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1073 int nid, nodemask_t *nodemask)
1075 unsigned long nr_pages = 1UL << huge_page_order(h);
1077 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1080 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1081 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1082 #else /* !CONFIG_CONTIG_ALLOC */
1083 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1084 int nid, nodemask_t *nodemask)
1088 #endif /* CONFIG_CONTIG_ALLOC */
1090 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1091 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1092 int nid, nodemask_t *nodemask)
1096 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1097 static inline void destroy_compound_gigantic_page(struct page *page,
1098 unsigned int order) { }
1101 static void update_and_free_page(struct hstate *h, struct page *page)
1105 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1109 h->nr_huge_pages_node[page_to_nid(page)]--;
1110 for (i = 0; i < pages_per_huge_page(h); i++) {
1111 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1112 1 << PG_referenced | 1 << PG_dirty |
1113 1 << PG_active | 1 << PG_private |
1116 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1117 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1118 set_page_refcounted(page);
1119 if (hstate_is_gigantic(h)) {
1120 destroy_compound_gigantic_page(page, huge_page_order(h));
1121 free_gigantic_page(page, huge_page_order(h));
1123 __free_pages(page, huge_page_order(h));
1127 struct hstate *size_to_hstate(unsigned long size)
1131 for_each_hstate(h) {
1132 if (huge_page_size(h) == size)
1139 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1140 * to hstate->hugepage_activelist.)
1142 * This function can be called for tail pages, but never returns true for them.
1144 bool page_huge_active(struct page *page)
1146 VM_BUG_ON_PAGE(!PageHuge(page), page);
1147 return PageHead(page) && PagePrivate(&page[1]);
1150 /* never called for tail page */
1151 static void set_page_huge_active(struct page *page)
1153 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1154 SetPagePrivate(&page[1]);
1157 static void clear_page_huge_active(struct page *page)
1159 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1160 ClearPagePrivate(&page[1]);
1164 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1167 static inline bool PageHugeTemporary(struct page *page)
1169 if (!PageHuge(page))
1172 return (unsigned long)page[2].mapping == -1U;
1175 static inline void SetPageHugeTemporary(struct page *page)
1177 page[2].mapping = (void *)-1U;
1180 static inline void ClearPageHugeTemporary(struct page *page)
1182 page[2].mapping = NULL;
1185 void free_huge_page(struct page *page)
1188 * Can't pass hstate in here because it is called from the
1189 * compound page destructor.
1191 struct hstate *h = page_hstate(page);
1192 int nid = page_to_nid(page);
1193 struct hugepage_subpool *spool =
1194 (struct hugepage_subpool *)page_private(page);
1195 bool restore_reserve;
1197 VM_BUG_ON_PAGE(page_count(page), page);
1198 VM_BUG_ON_PAGE(page_mapcount(page), page);
1200 set_page_private(page, 0);
1201 page->mapping = NULL;
1202 restore_reserve = PagePrivate(page);
1203 ClearPagePrivate(page);
1206 * If PagePrivate() was set on page, page allocation consumed a
1207 * reservation. If the page was associated with a subpool, there
1208 * would have been a page reserved in the subpool before allocation
1209 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1210 * reservtion, do not call hugepage_subpool_put_pages() as this will
1211 * remove the reserved page from the subpool.
1213 if (!restore_reserve) {
1215 * A return code of zero implies that the subpool will be
1216 * under its minimum size if the reservation is not restored
1217 * after page is free. Therefore, force restore_reserve
1220 if (hugepage_subpool_put_pages(spool, 1) == 0)
1221 restore_reserve = true;
1224 spin_lock(&hugetlb_lock);
1225 clear_page_huge_active(page);
1226 hugetlb_cgroup_uncharge_page(hstate_index(h),
1227 pages_per_huge_page(h), page);
1228 if (restore_reserve)
1229 h->resv_huge_pages++;
1231 if (PageHugeTemporary(page)) {
1232 list_del(&page->lru);
1233 ClearPageHugeTemporary(page);
1234 update_and_free_page(h, page);
1235 } else if (h->surplus_huge_pages_node[nid]) {
1236 /* remove the page from active list */
1237 list_del(&page->lru);
1238 update_and_free_page(h, page);
1239 h->surplus_huge_pages--;
1240 h->surplus_huge_pages_node[nid]--;
1242 arch_clear_hugepage_flags(page);
1243 enqueue_huge_page(h, page);
1245 spin_unlock(&hugetlb_lock);
1248 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1250 INIT_LIST_HEAD(&page->lru);
1251 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1252 spin_lock(&hugetlb_lock);
1253 set_hugetlb_cgroup(page, NULL);
1255 h->nr_huge_pages_node[nid]++;
1256 spin_unlock(&hugetlb_lock);
1259 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1262 int nr_pages = 1 << order;
1263 struct page *p = page + 1;
1265 /* we rely on prep_new_huge_page to set the destructor */
1266 set_compound_order(page, order);
1267 __ClearPageReserved(page);
1268 __SetPageHead(page);
1269 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1271 * For gigantic hugepages allocated through bootmem at
1272 * boot, it's safer to be consistent with the not-gigantic
1273 * hugepages and clear the PG_reserved bit from all tail pages
1274 * too. Otherwse drivers using get_user_pages() to access tail
1275 * pages may get the reference counting wrong if they see
1276 * PG_reserved set on a tail page (despite the head page not
1277 * having PG_reserved set). Enforcing this consistency between
1278 * head and tail pages allows drivers to optimize away a check
1279 * on the head page when they need know if put_page() is needed
1280 * after get_user_pages().
1282 __ClearPageReserved(p);
1283 set_page_count(p, 0);
1284 set_compound_head(p, page);
1286 atomic_set(compound_mapcount_ptr(page), -1);
1290 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1291 * transparent huge pages. See the PageTransHuge() documentation for more
1294 int PageHuge(struct page *page)
1296 if (!PageCompound(page))
1299 page = compound_head(page);
1300 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1302 EXPORT_SYMBOL_GPL(PageHuge);
1305 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1306 * normal or transparent huge pages.
1308 int PageHeadHuge(struct page *page_head)
1310 if (!PageHead(page_head))
1313 return get_compound_page_dtor(page_head) == free_huge_page;
1316 pgoff_t __basepage_index(struct page *page)
1318 struct page *page_head = compound_head(page);
1319 pgoff_t index = page_index(page_head);
1320 unsigned long compound_idx;
1322 if (!PageHuge(page_head))
1323 return page_index(page);
1325 if (compound_order(page_head) >= MAX_ORDER)
1326 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1328 compound_idx = page - page_head;
1330 return (index << compound_order(page_head)) + compound_idx;
1333 static struct page *alloc_buddy_huge_page(struct hstate *h,
1334 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1335 nodemask_t *node_alloc_noretry)
1337 int order = huge_page_order(h);
1339 bool alloc_try_hard = true;
1342 * By default we always try hard to allocate the page with
1343 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1344 * a loop (to adjust global huge page counts) and previous allocation
1345 * failed, do not continue to try hard on the same node. Use the
1346 * node_alloc_noretry bitmap to manage this state information.
1348 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1349 alloc_try_hard = false;
1350 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1352 gfp_mask |= __GFP_RETRY_MAYFAIL;
1353 if (nid == NUMA_NO_NODE)
1354 nid = numa_mem_id();
1355 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1357 __count_vm_event(HTLB_BUDDY_PGALLOC);
1359 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1362 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1363 * indicates an overall state change. Clear bit so that we resume
1364 * normal 'try hard' allocations.
1366 if (node_alloc_noretry && page && !alloc_try_hard)
1367 node_clear(nid, *node_alloc_noretry);
1370 * If we tried hard to get a page but failed, set bit so that
1371 * subsequent attempts will not try as hard until there is an
1372 * overall state change.
1374 if (node_alloc_noretry && !page && alloc_try_hard)
1375 node_set(nid, *node_alloc_noretry);
1381 * Common helper to allocate a fresh hugetlb page. All specific allocators
1382 * should use this function to get new hugetlb pages
1384 static struct page *alloc_fresh_huge_page(struct hstate *h,
1385 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1386 nodemask_t *node_alloc_noretry)
1390 if (hstate_is_gigantic(h))
1391 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1393 page = alloc_buddy_huge_page(h, gfp_mask,
1394 nid, nmask, node_alloc_noretry);
1398 if (hstate_is_gigantic(h))
1399 prep_compound_gigantic_page(page, huge_page_order(h));
1400 prep_new_huge_page(h, page, page_to_nid(page));
1406 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1409 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1410 nodemask_t *node_alloc_noretry)
1414 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1416 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1417 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1418 node_alloc_noretry);
1426 put_page(page); /* free it into the hugepage allocator */
1432 * Free huge page from pool from next node to free.
1433 * Attempt to keep persistent huge pages more or less
1434 * balanced over allowed nodes.
1435 * Called with hugetlb_lock locked.
1437 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1443 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1445 * If we're returning unused surplus pages, only examine
1446 * nodes with surplus pages.
1448 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1449 !list_empty(&h->hugepage_freelists[node])) {
1451 list_entry(h->hugepage_freelists[node].next,
1453 list_del(&page->lru);
1454 h->free_huge_pages--;
1455 h->free_huge_pages_node[node]--;
1457 h->surplus_huge_pages--;
1458 h->surplus_huge_pages_node[node]--;
1460 update_and_free_page(h, page);
1470 * Dissolve a given free hugepage into free buddy pages. This function does
1471 * nothing for in-use hugepages and non-hugepages.
1472 * This function returns values like below:
1474 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1475 * (allocated or reserved.)
1476 * 0: successfully dissolved free hugepages or the page is not a
1477 * hugepage (considered as already dissolved)
1479 int dissolve_free_huge_page(struct page *page)
1483 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1484 if (!PageHuge(page))
1487 spin_lock(&hugetlb_lock);
1488 if (!PageHuge(page)) {
1493 if (!page_count(page)) {
1494 struct page *head = compound_head(page);
1495 struct hstate *h = page_hstate(head);
1496 int nid = page_to_nid(head);
1497 if (h->free_huge_pages - h->resv_huge_pages == 0)
1500 * Move PageHWPoison flag from head page to the raw error page,
1501 * which makes any subpages rather than the error page reusable.
1503 if (PageHWPoison(head) && page != head) {
1504 SetPageHWPoison(page);
1505 ClearPageHWPoison(head);
1507 list_del(&head->lru);
1508 h->free_huge_pages--;
1509 h->free_huge_pages_node[nid]--;
1510 h->max_huge_pages--;
1511 update_and_free_page(h, head);
1515 spin_unlock(&hugetlb_lock);
1520 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1521 * make specified memory blocks removable from the system.
1522 * Note that this will dissolve a free gigantic hugepage completely, if any
1523 * part of it lies within the given range.
1524 * Also note that if dissolve_free_huge_page() returns with an error, all
1525 * free hugepages that were dissolved before that error are lost.
1527 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1533 if (!hugepages_supported())
1536 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1537 page = pfn_to_page(pfn);
1538 rc = dissolve_free_huge_page(page);
1547 * Allocates a fresh surplus page from the page allocator.
1549 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1550 int nid, nodemask_t *nmask)
1552 struct page *page = NULL;
1554 if (hstate_is_gigantic(h))
1557 spin_lock(&hugetlb_lock);
1558 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1560 spin_unlock(&hugetlb_lock);
1562 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1566 spin_lock(&hugetlb_lock);
1568 * We could have raced with the pool size change.
1569 * Double check that and simply deallocate the new page
1570 * if we would end up overcommiting the surpluses. Abuse
1571 * temporary page to workaround the nasty free_huge_page
1574 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1575 SetPageHugeTemporary(page);
1576 spin_unlock(&hugetlb_lock);
1580 h->surplus_huge_pages++;
1581 h->surplus_huge_pages_node[page_to_nid(page)]++;
1585 spin_unlock(&hugetlb_lock);
1590 struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1591 int nid, nodemask_t *nmask)
1595 if (hstate_is_gigantic(h))
1598 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1603 * We do not account these pages as surplus because they are only
1604 * temporary and will be released properly on the last reference
1606 SetPageHugeTemporary(page);
1612 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1615 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1616 struct vm_area_struct *vma, unsigned long addr)
1619 struct mempolicy *mpol;
1620 gfp_t gfp_mask = htlb_alloc_mask(h);
1622 nodemask_t *nodemask;
1624 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1625 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1626 mpol_cond_put(mpol);
1631 /* page migration callback function */
1632 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1634 gfp_t gfp_mask = htlb_alloc_mask(h);
1635 struct page *page = NULL;
1637 if (nid != NUMA_NO_NODE)
1638 gfp_mask |= __GFP_THISNODE;
1640 spin_lock(&hugetlb_lock);
1641 if (h->free_huge_pages - h->resv_huge_pages > 0)
1642 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1643 spin_unlock(&hugetlb_lock);
1646 page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1651 /* page migration callback function */
1652 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1655 gfp_t gfp_mask = htlb_alloc_mask(h);
1657 spin_lock(&hugetlb_lock);
1658 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1661 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1663 spin_unlock(&hugetlb_lock);
1667 spin_unlock(&hugetlb_lock);
1669 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1672 /* mempolicy aware migration callback */
1673 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1674 unsigned long address)
1676 struct mempolicy *mpol;
1677 nodemask_t *nodemask;
1682 gfp_mask = htlb_alloc_mask(h);
1683 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1684 page = alloc_huge_page_nodemask(h, node, nodemask);
1685 mpol_cond_put(mpol);
1691 * Increase the hugetlb pool such that it can accommodate a reservation
1694 static int gather_surplus_pages(struct hstate *h, int delta)
1696 struct list_head surplus_list;
1697 struct page *page, *tmp;
1699 int needed, allocated;
1700 bool alloc_ok = true;
1702 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1704 h->resv_huge_pages += delta;
1709 INIT_LIST_HEAD(&surplus_list);
1713 spin_unlock(&hugetlb_lock);
1714 for (i = 0; i < needed; i++) {
1715 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1716 NUMA_NO_NODE, NULL);
1721 list_add(&page->lru, &surplus_list);
1727 * After retaking hugetlb_lock, we need to recalculate 'needed'
1728 * because either resv_huge_pages or free_huge_pages may have changed.
1730 spin_lock(&hugetlb_lock);
1731 needed = (h->resv_huge_pages + delta) -
1732 (h->free_huge_pages + allocated);
1737 * We were not able to allocate enough pages to
1738 * satisfy the entire reservation so we free what
1739 * we've allocated so far.
1744 * The surplus_list now contains _at_least_ the number of extra pages
1745 * needed to accommodate the reservation. Add the appropriate number
1746 * of pages to the hugetlb pool and free the extras back to the buddy
1747 * allocator. Commit the entire reservation here to prevent another
1748 * process from stealing the pages as they are added to the pool but
1749 * before they are reserved.
1751 needed += allocated;
1752 h->resv_huge_pages += delta;
1755 /* Free the needed pages to the hugetlb pool */
1756 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1760 * This page is now managed by the hugetlb allocator and has
1761 * no users -- drop the buddy allocator's reference.
1763 put_page_testzero(page);
1764 VM_BUG_ON_PAGE(page_count(page), page);
1765 enqueue_huge_page(h, page);
1768 spin_unlock(&hugetlb_lock);
1770 /* Free unnecessary surplus pages to the buddy allocator */
1771 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1773 spin_lock(&hugetlb_lock);
1779 * This routine has two main purposes:
1780 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1781 * in unused_resv_pages. This corresponds to the prior adjustments made
1782 * to the associated reservation map.
1783 * 2) Free any unused surplus pages that may have been allocated to satisfy
1784 * the reservation. As many as unused_resv_pages may be freed.
1786 * Called with hugetlb_lock held. However, the lock could be dropped (and
1787 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1788 * we must make sure nobody else can claim pages we are in the process of
1789 * freeing. Do this by ensuring resv_huge_page always is greater than the
1790 * number of huge pages we plan to free when dropping the lock.
1792 static void return_unused_surplus_pages(struct hstate *h,
1793 unsigned long unused_resv_pages)
1795 unsigned long nr_pages;
1797 /* Cannot return gigantic pages currently */
1798 if (hstate_is_gigantic(h))
1802 * Part (or even all) of the reservation could have been backed
1803 * by pre-allocated pages. Only free surplus pages.
1805 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1808 * We want to release as many surplus pages as possible, spread
1809 * evenly across all nodes with memory. Iterate across these nodes
1810 * until we can no longer free unreserved surplus pages. This occurs
1811 * when the nodes with surplus pages have no free pages.
1812 * free_pool_huge_page() will balance the the freed pages across the
1813 * on-line nodes with memory and will handle the hstate accounting.
1815 * Note that we decrement resv_huge_pages as we free the pages. If
1816 * we drop the lock, resv_huge_pages will still be sufficiently large
1817 * to cover subsequent pages we may free.
1819 while (nr_pages--) {
1820 h->resv_huge_pages--;
1821 unused_resv_pages--;
1822 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1824 cond_resched_lock(&hugetlb_lock);
1828 /* Fully uncommit the reservation */
1829 h->resv_huge_pages -= unused_resv_pages;
1834 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1835 * are used by the huge page allocation routines to manage reservations.
1837 * vma_needs_reservation is called to determine if the huge page at addr
1838 * within the vma has an associated reservation. If a reservation is
1839 * needed, the value 1 is returned. The caller is then responsible for
1840 * managing the global reservation and subpool usage counts. After
1841 * the huge page has been allocated, vma_commit_reservation is called
1842 * to add the page to the reservation map. If the page allocation fails,
1843 * the reservation must be ended instead of committed. vma_end_reservation
1844 * is called in such cases.
1846 * In the normal case, vma_commit_reservation returns the same value
1847 * as the preceding vma_needs_reservation call. The only time this
1848 * is not the case is if a reserve map was changed between calls. It
1849 * is the responsibility of the caller to notice the difference and
1850 * take appropriate action.
1852 * vma_add_reservation is used in error paths where a reservation must
1853 * be restored when a newly allocated huge page must be freed. It is
1854 * to be called after calling vma_needs_reservation to determine if a
1855 * reservation exists.
1857 enum vma_resv_mode {
1863 static long __vma_reservation_common(struct hstate *h,
1864 struct vm_area_struct *vma, unsigned long addr,
1865 enum vma_resv_mode mode)
1867 struct resv_map *resv;
1871 resv = vma_resv_map(vma);
1875 idx = vma_hugecache_offset(h, vma, addr);
1877 case VMA_NEEDS_RESV:
1878 ret = region_chg(resv, idx, idx + 1);
1880 case VMA_COMMIT_RESV:
1881 ret = region_add(resv, idx, idx + 1);
1884 region_abort(resv, idx, idx + 1);
1888 if (vma->vm_flags & VM_MAYSHARE)
1889 ret = region_add(resv, idx, idx + 1);
1891 region_abort(resv, idx, idx + 1);
1892 ret = region_del(resv, idx, idx + 1);
1899 if (vma->vm_flags & VM_MAYSHARE)
1901 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1903 * In most cases, reserves always exist for private mappings.
1904 * However, a file associated with mapping could have been
1905 * hole punched or truncated after reserves were consumed.
1906 * As subsequent fault on such a range will not use reserves.
1907 * Subtle - The reserve map for private mappings has the
1908 * opposite meaning than that of shared mappings. If NO
1909 * entry is in the reserve map, it means a reservation exists.
1910 * If an entry exists in the reserve map, it means the
1911 * reservation has already been consumed. As a result, the
1912 * return value of this routine is the opposite of the
1913 * value returned from reserve map manipulation routines above.
1921 return ret < 0 ? ret : 0;
1924 static long vma_needs_reservation(struct hstate *h,
1925 struct vm_area_struct *vma, unsigned long addr)
1927 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1930 static long vma_commit_reservation(struct hstate *h,
1931 struct vm_area_struct *vma, unsigned long addr)
1933 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1936 static void vma_end_reservation(struct hstate *h,
1937 struct vm_area_struct *vma, unsigned long addr)
1939 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1942 static long vma_add_reservation(struct hstate *h,
1943 struct vm_area_struct *vma, unsigned long addr)
1945 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1949 * This routine is called to restore a reservation on error paths. In the
1950 * specific error paths, a huge page was allocated (via alloc_huge_page)
1951 * and is about to be freed. If a reservation for the page existed,
1952 * alloc_huge_page would have consumed the reservation and set PagePrivate
1953 * in the newly allocated page. When the page is freed via free_huge_page,
1954 * the global reservation count will be incremented if PagePrivate is set.
1955 * However, free_huge_page can not adjust the reserve map. Adjust the
1956 * reserve map here to be consistent with global reserve count adjustments
1957 * to be made by free_huge_page.
1959 static void restore_reserve_on_error(struct hstate *h,
1960 struct vm_area_struct *vma, unsigned long address,
1963 if (unlikely(PagePrivate(page))) {
1964 long rc = vma_needs_reservation(h, vma, address);
1966 if (unlikely(rc < 0)) {
1968 * Rare out of memory condition in reserve map
1969 * manipulation. Clear PagePrivate so that
1970 * global reserve count will not be incremented
1971 * by free_huge_page. This will make it appear
1972 * as though the reservation for this page was
1973 * consumed. This may prevent the task from
1974 * faulting in the page at a later time. This
1975 * is better than inconsistent global huge page
1976 * accounting of reserve counts.
1978 ClearPagePrivate(page);
1980 rc = vma_add_reservation(h, vma, address);
1981 if (unlikely(rc < 0))
1983 * See above comment about rare out of
1986 ClearPagePrivate(page);
1988 vma_end_reservation(h, vma, address);
1992 struct page *alloc_huge_page(struct vm_area_struct *vma,
1993 unsigned long addr, int avoid_reserve)
1995 struct hugepage_subpool *spool = subpool_vma(vma);
1996 struct hstate *h = hstate_vma(vma);
1998 long map_chg, map_commit;
2001 struct hugetlb_cgroup *h_cg;
2003 idx = hstate_index(h);
2005 * Examine the region/reserve map to determine if the process
2006 * has a reservation for the page to be allocated. A return
2007 * code of zero indicates a reservation exists (no change).
2009 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2011 return ERR_PTR(-ENOMEM);
2014 * Processes that did not create the mapping will have no
2015 * reserves as indicated by the region/reserve map. Check
2016 * that the allocation will not exceed the subpool limit.
2017 * Allocations for MAP_NORESERVE mappings also need to be
2018 * checked against any subpool limit.
2020 if (map_chg || avoid_reserve) {
2021 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2023 vma_end_reservation(h, vma, addr);
2024 return ERR_PTR(-ENOSPC);
2028 * Even though there was no reservation in the region/reserve
2029 * map, there could be reservations associated with the
2030 * subpool that can be used. This would be indicated if the
2031 * return value of hugepage_subpool_get_pages() is zero.
2032 * However, if avoid_reserve is specified we still avoid even
2033 * the subpool reservations.
2039 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2041 goto out_subpool_put;
2043 spin_lock(&hugetlb_lock);
2045 * glb_chg is passed to indicate whether or not a page must be taken
2046 * from the global free pool (global change). gbl_chg == 0 indicates
2047 * a reservation exists for the allocation.
2049 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2051 spin_unlock(&hugetlb_lock);
2052 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2054 goto out_uncharge_cgroup;
2055 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2056 SetPagePrivate(page);
2057 h->resv_huge_pages--;
2059 spin_lock(&hugetlb_lock);
2060 list_move(&page->lru, &h->hugepage_activelist);
2063 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2064 spin_unlock(&hugetlb_lock);
2066 set_page_private(page, (unsigned long)spool);
2068 map_commit = vma_commit_reservation(h, vma, addr);
2069 if (unlikely(map_chg > map_commit)) {
2071 * The page was added to the reservation map between
2072 * vma_needs_reservation and vma_commit_reservation.
2073 * This indicates a race with hugetlb_reserve_pages.
2074 * Adjust for the subpool count incremented above AND
2075 * in hugetlb_reserve_pages for the same page. Also,
2076 * the reservation count added in hugetlb_reserve_pages
2077 * no longer applies.
2081 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2082 hugetlb_acct_memory(h, -rsv_adjust);
2086 out_uncharge_cgroup:
2087 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2089 if (map_chg || avoid_reserve)
2090 hugepage_subpool_put_pages(spool, 1);
2091 vma_end_reservation(h, vma, addr);
2092 return ERR_PTR(-ENOSPC);
2095 int alloc_bootmem_huge_page(struct hstate *h)
2096 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2097 int __alloc_bootmem_huge_page(struct hstate *h)
2099 struct huge_bootmem_page *m;
2102 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2105 addr = memblock_alloc_try_nid_raw(
2106 huge_page_size(h), huge_page_size(h),
2107 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2110 * Use the beginning of the huge page to store the
2111 * huge_bootmem_page struct (until gather_bootmem
2112 * puts them into the mem_map).
2121 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2122 /* Put them into a private list first because mem_map is not up yet */
2123 INIT_LIST_HEAD(&m->list);
2124 list_add(&m->list, &huge_boot_pages);
2129 static void __init prep_compound_huge_page(struct page *page,
2132 if (unlikely(order > (MAX_ORDER - 1)))
2133 prep_compound_gigantic_page(page, order);
2135 prep_compound_page(page, order);
2138 /* Put bootmem huge pages into the standard lists after mem_map is up */
2139 static void __init gather_bootmem_prealloc(void)
2141 struct huge_bootmem_page *m;
2143 list_for_each_entry(m, &huge_boot_pages, list) {
2144 struct page *page = virt_to_page(m);
2145 struct hstate *h = m->hstate;
2147 WARN_ON(page_count(page) != 1);
2148 prep_compound_huge_page(page, h->order);
2149 WARN_ON(PageReserved(page));
2150 prep_new_huge_page(h, page, page_to_nid(page));
2151 put_page(page); /* free it into the hugepage allocator */
2154 * If we had gigantic hugepages allocated at boot time, we need
2155 * to restore the 'stolen' pages to totalram_pages in order to
2156 * fix confusing memory reports from free(1) and another
2157 * side-effects, like CommitLimit going negative.
2159 if (hstate_is_gigantic(h))
2160 adjust_managed_page_count(page, 1 << h->order);
2165 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2168 nodemask_t *node_alloc_noretry;
2170 if (!hstate_is_gigantic(h)) {
2172 * Bit mask controlling how hard we retry per-node allocations.
2173 * Ignore errors as lower level routines can deal with
2174 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2175 * time, we are likely in bigger trouble.
2177 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2180 /* allocations done at boot time */
2181 node_alloc_noretry = NULL;
2184 /* bit mask controlling how hard we retry per-node allocations */
2185 if (node_alloc_noretry)
2186 nodes_clear(*node_alloc_noretry);
2188 for (i = 0; i < h->max_huge_pages; ++i) {
2189 if (hstate_is_gigantic(h)) {
2190 if (!alloc_bootmem_huge_page(h))
2192 } else if (!alloc_pool_huge_page(h,
2193 &node_states[N_MEMORY],
2194 node_alloc_noretry))
2198 if (i < h->max_huge_pages) {
2201 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2202 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2203 h->max_huge_pages, buf, i);
2204 h->max_huge_pages = i;
2207 kfree(node_alloc_noretry);
2210 static void __init hugetlb_init_hstates(void)
2214 for_each_hstate(h) {
2215 if (minimum_order > huge_page_order(h))
2216 minimum_order = huge_page_order(h);
2218 /* oversize hugepages were init'ed in early boot */
2219 if (!hstate_is_gigantic(h))
2220 hugetlb_hstate_alloc_pages(h);
2222 VM_BUG_ON(minimum_order == UINT_MAX);
2225 static void __init report_hugepages(void)
2229 for_each_hstate(h) {
2232 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2233 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2234 buf, h->free_huge_pages);
2238 #ifdef CONFIG_HIGHMEM
2239 static void try_to_free_low(struct hstate *h, unsigned long count,
2240 nodemask_t *nodes_allowed)
2244 if (hstate_is_gigantic(h))
2247 for_each_node_mask(i, *nodes_allowed) {
2248 struct page *page, *next;
2249 struct list_head *freel = &h->hugepage_freelists[i];
2250 list_for_each_entry_safe(page, next, freel, lru) {
2251 if (count >= h->nr_huge_pages)
2253 if (PageHighMem(page))
2255 list_del(&page->lru);
2256 update_and_free_page(h, page);
2257 h->free_huge_pages--;
2258 h->free_huge_pages_node[page_to_nid(page)]--;
2263 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2264 nodemask_t *nodes_allowed)
2270 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2271 * balanced by operating on them in a round-robin fashion.
2272 * Returns 1 if an adjustment was made.
2274 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2279 VM_BUG_ON(delta != -1 && delta != 1);
2282 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2283 if (h->surplus_huge_pages_node[node])
2287 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2288 if (h->surplus_huge_pages_node[node] <
2289 h->nr_huge_pages_node[node])
2296 h->surplus_huge_pages += delta;
2297 h->surplus_huge_pages_node[node] += delta;
2301 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2302 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2303 nodemask_t *nodes_allowed)
2305 unsigned long min_count, ret;
2306 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2309 * Bit mask controlling how hard we retry per-node allocations.
2310 * If we can not allocate the bit mask, do not attempt to allocate
2311 * the requested huge pages.
2313 if (node_alloc_noretry)
2314 nodes_clear(*node_alloc_noretry);
2318 spin_lock(&hugetlb_lock);
2321 * Check for a node specific request.
2322 * Changing node specific huge page count may require a corresponding
2323 * change to the global count. In any case, the passed node mask
2324 * (nodes_allowed) will restrict alloc/free to the specified node.
2326 if (nid != NUMA_NO_NODE) {
2327 unsigned long old_count = count;
2329 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2331 * User may have specified a large count value which caused the
2332 * above calculation to overflow. In this case, they wanted
2333 * to allocate as many huge pages as possible. Set count to
2334 * largest possible value to align with their intention.
2336 if (count < old_count)
2341 * Gigantic pages runtime allocation depend on the capability for large
2342 * page range allocation.
2343 * If the system does not provide this feature, return an error when
2344 * the user tries to allocate gigantic pages but let the user free the
2345 * boottime allocated gigantic pages.
2347 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2348 if (count > persistent_huge_pages(h)) {
2349 spin_unlock(&hugetlb_lock);
2350 NODEMASK_FREE(node_alloc_noretry);
2353 /* Fall through to decrease pool */
2357 * Increase the pool size
2358 * First take pages out of surplus state. Then make up the
2359 * remaining difference by allocating fresh huge pages.
2361 * We might race with alloc_surplus_huge_page() here and be unable
2362 * to convert a surplus huge page to a normal huge page. That is
2363 * not critical, though, it just means the overall size of the
2364 * pool might be one hugepage larger than it needs to be, but
2365 * within all the constraints specified by the sysctls.
2367 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2368 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2372 while (count > persistent_huge_pages(h)) {
2374 * If this allocation races such that we no longer need the
2375 * page, free_huge_page will handle it by freeing the page
2376 * and reducing the surplus.
2378 spin_unlock(&hugetlb_lock);
2380 /* yield cpu to avoid soft lockup */
2383 ret = alloc_pool_huge_page(h, nodes_allowed,
2384 node_alloc_noretry);
2385 spin_lock(&hugetlb_lock);
2389 /* Bail for signals. Probably ctrl-c from user */
2390 if (signal_pending(current))
2395 * Decrease the pool size
2396 * First return free pages to the buddy allocator (being careful
2397 * to keep enough around to satisfy reservations). Then place
2398 * pages into surplus state as needed so the pool will shrink
2399 * to the desired size as pages become free.
2401 * By placing pages into the surplus state independent of the
2402 * overcommit value, we are allowing the surplus pool size to
2403 * exceed overcommit. There are few sane options here. Since
2404 * alloc_surplus_huge_page() is checking the global counter,
2405 * though, we'll note that we're not allowed to exceed surplus
2406 * and won't grow the pool anywhere else. Not until one of the
2407 * sysctls are changed, or the surplus pages go out of use.
2409 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2410 min_count = max(count, min_count);
2411 try_to_free_low(h, min_count, nodes_allowed);
2412 while (min_count < persistent_huge_pages(h)) {
2413 if (!free_pool_huge_page(h, nodes_allowed, 0))
2415 cond_resched_lock(&hugetlb_lock);
2417 while (count < persistent_huge_pages(h)) {
2418 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2422 h->max_huge_pages = persistent_huge_pages(h);
2423 spin_unlock(&hugetlb_lock);
2425 NODEMASK_FREE(node_alloc_noretry);
2430 #define HSTATE_ATTR_RO(_name) \
2431 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2433 #define HSTATE_ATTR(_name) \
2434 static struct kobj_attribute _name##_attr = \
2435 __ATTR(_name, 0644, _name##_show, _name##_store)
2437 static struct kobject *hugepages_kobj;
2438 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2440 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2442 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2446 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2447 if (hstate_kobjs[i] == kobj) {
2449 *nidp = NUMA_NO_NODE;
2453 return kobj_to_node_hstate(kobj, nidp);
2456 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2457 struct kobj_attribute *attr, char *buf)
2460 unsigned long nr_huge_pages;
2463 h = kobj_to_hstate(kobj, &nid);
2464 if (nid == NUMA_NO_NODE)
2465 nr_huge_pages = h->nr_huge_pages;
2467 nr_huge_pages = h->nr_huge_pages_node[nid];
2469 return sprintf(buf, "%lu\n", nr_huge_pages);
2472 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2473 struct hstate *h, int nid,
2474 unsigned long count, size_t len)
2477 nodemask_t nodes_allowed, *n_mask;
2479 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2482 if (nid == NUMA_NO_NODE) {
2484 * global hstate attribute
2486 if (!(obey_mempolicy &&
2487 init_nodemask_of_mempolicy(&nodes_allowed)))
2488 n_mask = &node_states[N_MEMORY];
2490 n_mask = &nodes_allowed;
2493 * Node specific request. count adjustment happens in
2494 * set_max_huge_pages() after acquiring hugetlb_lock.
2496 init_nodemask_of_node(&nodes_allowed, nid);
2497 n_mask = &nodes_allowed;
2500 err = set_max_huge_pages(h, count, nid, n_mask);
2502 return err ? err : len;
2505 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2506 struct kobject *kobj, const char *buf,
2510 unsigned long count;
2514 err = kstrtoul(buf, 10, &count);
2518 h = kobj_to_hstate(kobj, &nid);
2519 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2522 static ssize_t nr_hugepages_show(struct kobject *kobj,
2523 struct kobj_attribute *attr, char *buf)
2525 return nr_hugepages_show_common(kobj, attr, buf);
2528 static ssize_t nr_hugepages_store(struct kobject *kobj,
2529 struct kobj_attribute *attr, const char *buf, size_t len)
2531 return nr_hugepages_store_common(false, kobj, buf, len);
2533 HSTATE_ATTR(nr_hugepages);
2538 * hstate attribute for optionally mempolicy-based constraint on persistent
2539 * huge page alloc/free.
2541 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2542 struct kobj_attribute *attr, char *buf)
2544 return nr_hugepages_show_common(kobj, attr, buf);
2547 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2548 struct kobj_attribute *attr, const char *buf, size_t len)
2550 return nr_hugepages_store_common(true, kobj, buf, len);
2552 HSTATE_ATTR(nr_hugepages_mempolicy);
2556 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2557 struct kobj_attribute *attr, char *buf)
2559 struct hstate *h = kobj_to_hstate(kobj, NULL);
2560 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2563 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2564 struct kobj_attribute *attr, const char *buf, size_t count)
2567 unsigned long input;
2568 struct hstate *h = kobj_to_hstate(kobj, NULL);
2570 if (hstate_is_gigantic(h))
2573 err = kstrtoul(buf, 10, &input);
2577 spin_lock(&hugetlb_lock);
2578 h->nr_overcommit_huge_pages = input;
2579 spin_unlock(&hugetlb_lock);
2583 HSTATE_ATTR(nr_overcommit_hugepages);
2585 static ssize_t free_hugepages_show(struct kobject *kobj,
2586 struct kobj_attribute *attr, char *buf)
2589 unsigned long free_huge_pages;
2592 h = kobj_to_hstate(kobj, &nid);
2593 if (nid == NUMA_NO_NODE)
2594 free_huge_pages = h->free_huge_pages;
2596 free_huge_pages = h->free_huge_pages_node[nid];
2598 return sprintf(buf, "%lu\n", free_huge_pages);
2600 HSTATE_ATTR_RO(free_hugepages);
2602 static ssize_t resv_hugepages_show(struct kobject *kobj,
2603 struct kobj_attribute *attr, char *buf)
2605 struct hstate *h = kobj_to_hstate(kobj, NULL);
2606 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2608 HSTATE_ATTR_RO(resv_hugepages);
2610 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2611 struct kobj_attribute *attr, char *buf)
2614 unsigned long surplus_huge_pages;
2617 h = kobj_to_hstate(kobj, &nid);
2618 if (nid == NUMA_NO_NODE)
2619 surplus_huge_pages = h->surplus_huge_pages;
2621 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2623 return sprintf(buf, "%lu\n", surplus_huge_pages);
2625 HSTATE_ATTR_RO(surplus_hugepages);
2627 static struct attribute *hstate_attrs[] = {
2628 &nr_hugepages_attr.attr,
2629 &nr_overcommit_hugepages_attr.attr,
2630 &free_hugepages_attr.attr,
2631 &resv_hugepages_attr.attr,
2632 &surplus_hugepages_attr.attr,
2634 &nr_hugepages_mempolicy_attr.attr,
2639 static const struct attribute_group hstate_attr_group = {
2640 .attrs = hstate_attrs,
2643 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2644 struct kobject **hstate_kobjs,
2645 const struct attribute_group *hstate_attr_group)
2648 int hi = hstate_index(h);
2650 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2651 if (!hstate_kobjs[hi])
2654 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2656 kobject_put(hstate_kobjs[hi]);
2661 static void __init hugetlb_sysfs_init(void)
2666 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2667 if (!hugepages_kobj)
2670 for_each_hstate(h) {
2671 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2672 hstate_kobjs, &hstate_attr_group);
2674 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2681 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2682 * with node devices in node_devices[] using a parallel array. The array
2683 * index of a node device or _hstate == node id.
2684 * This is here to avoid any static dependency of the node device driver, in
2685 * the base kernel, on the hugetlb module.
2687 struct node_hstate {
2688 struct kobject *hugepages_kobj;
2689 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2691 static struct node_hstate node_hstates[MAX_NUMNODES];
2694 * A subset of global hstate attributes for node devices
2696 static struct attribute *per_node_hstate_attrs[] = {
2697 &nr_hugepages_attr.attr,
2698 &free_hugepages_attr.attr,
2699 &surplus_hugepages_attr.attr,
2703 static const struct attribute_group per_node_hstate_attr_group = {
2704 .attrs = per_node_hstate_attrs,
2708 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2709 * Returns node id via non-NULL nidp.
2711 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2715 for (nid = 0; nid < nr_node_ids; nid++) {
2716 struct node_hstate *nhs = &node_hstates[nid];
2718 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2719 if (nhs->hstate_kobjs[i] == kobj) {
2731 * Unregister hstate attributes from a single node device.
2732 * No-op if no hstate attributes attached.
2734 static void hugetlb_unregister_node(struct node *node)
2737 struct node_hstate *nhs = &node_hstates[node->dev.id];
2739 if (!nhs->hugepages_kobj)
2740 return; /* no hstate attributes */
2742 for_each_hstate(h) {
2743 int idx = hstate_index(h);
2744 if (nhs->hstate_kobjs[idx]) {
2745 kobject_put(nhs->hstate_kobjs[idx]);
2746 nhs->hstate_kobjs[idx] = NULL;
2750 kobject_put(nhs->hugepages_kobj);
2751 nhs->hugepages_kobj = NULL;
2756 * Register hstate attributes for a single node device.
2757 * No-op if attributes already registered.
2759 static void hugetlb_register_node(struct node *node)
2762 struct node_hstate *nhs = &node_hstates[node->dev.id];
2765 if (nhs->hugepages_kobj)
2766 return; /* already allocated */
2768 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2770 if (!nhs->hugepages_kobj)
2773 for_each_hstate(h) {
2774 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2776 &per_node_hstate_attr_group);
2778 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2779 h->name, node->dev.id);
2780 hugetlb_unregister_node(node);
2787 * hugetlb init time: register hstate attributes for all registered node
2788 * devices of nodes that have memory. All on-line nodes should have
2789 * registered their associated device by this time.
2791 static void __init hugetlb_register_all_nodes(void)
2795 for_each_node_state(nid, N_MEMORY) {
2796 struct node *node = node_devices[nid];
2797 if (node->dev.id == nid)
2798 hugetlb_register_node(node);
2802 * Let the node device driver know we're here so it can
2803 * [un]register hstate attributes on node hotplug.
2805 register_hugetlbfs_with_node(hugetlb_register_node,
2806 hugetlb_unregister_node);
2808 #else /* !CONFIG_NUMA */
2810 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2818 static void hugetlb_register_all_nodes(void) { }
2822 static int __init hugetlb_init(void)
2826 if (!hugepages_supported())
2829 if (!size_to_hstate(default_hstate_size)) {
2830 if (default_hstate_size != 0) {
2831 pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2832 default_hstate_size, HPAGE_SIZE);
2835 default_hstate_size = HPAGE_SIZE;
2836 if (!size_to_hstate(default_hstate_size))
2837 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2839 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2840 if (default_hstate_max_huge_pages) {
2841 if (!default_hstate.max_huge_pages)
2842 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2845 hugetlb_init_hstates();
2846 gather_bootmem_prealloc();
2849 hugetlb_sysfs_init();
2850 hugetlb_register_all_nodes();
2851 hugetlb_cgroup_file_init();
2854 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2856 num_fault_mutexes = 1;
2858 hugetlb_fault_mutex_table =
2859 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
2861 BUG_ON(!hugetlb_fault_mutex_table);
2863 for (i = 0; i < num_fault_mutexes; i++)
2864 mutex_init(&hugetlb_fault_mutex_table[i]);
2867 subsys_initcall(hugetlb_init);
2869 /* Should be called on processing a hugepagesz=... option */
2870 void __init hugetlb_bad_size(void)
2872 parsed_valid_hugepagesz = false;
2875 void __init hugetlb_add_hstate(unsigned int order)
2880 if (size_to_hstate(PAGE_SIZE << order)) {
2881 pr_warn("hugepagesz= specified twice, ignoring\n");
2884 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2886 h = &hstates[hugetlb_max_hstate++];
2888 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2889 h->nr_huge_pages = 0;
2890 h->free_huge_pages = 0;
2891 for (i = 0; i < MAX_NUMNODES; ++i)
2892 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2893 INIT_LIST_HEAD(&h->hugepage_activelist);
2894 h->next_nid_to_alloc = first_memory_node;
2895 h->next_nid_to_free = first_memory_node;
2896 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2897 huge_page_size(h)/1024);
2902 static int __init hugetlb_nrpages_setup(char *s)
2905 static unsigned long *last_mhp;
2907 if (!parsed_valid_hugepagesz) {
2908 pr_warn("hugepages = %s preceded by "
2909 "an unsupported hugepagesz, ignoring\n", s);
2910 parsed_valid_hugepagesz = true;
2914 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2915 * so this hugepages= parameter goes to the "default hstate".
2917 else if (!hugetlb_max_hstate)
2918 mhp = &default_hstate_max_huge_pages;
2920 mhp = &parsed_hstate->max_huge_pages;
2922 if (mhp == last_mhp) {
2923 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2927 if (sscanf(s, "%lu", mhp) <= 0)
2931 * Global state is always initialized later in hugetlb_init.
2932 * But we need to allocate >= MAX_ORDER hstates here early to still
2933 * use the bootmem allocator.
2935 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2936 hugetlb_hstate_alloc_pages(parsed_hstate);
2942 __setup("hugepages=", hugetlb_nrpages_setup);
2944 static int __init hugetlb_default_setup(char *s)
2946 default_hstate_size = memparse(s, &s);
2949 __setup("default_hugepagesz=", hugetlb_default_setup);
2951 static unsigned int cpuset_mems_nr(unsigned int *array)
2954 unsigned int nr = 0;
2956 for_each_node_mask(node, cpuset_current_mems_allowed)
2962 #ifdef CONFIG_SYSCTL
2963 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2964 struct ctl_table *table, int write,
2965 void __user *buffer, size_t *length, loff_t *ppos)
2967 struct hstate *h = &default_hstate;
2968 unsigned long tmp = h->max_huge_pages;
2971 if (!hugepages_supported())
2975 table->maxlen = sizeof(unsigned long);
2976 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2981 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2982 NUMA_NO_NODE, tmp, *length);
2987 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2988 void __user *buffer, size_t *length, loff_t *ppos)
2991 return hugetlb_sysctl_handler_common(false, table, write,
2992 buffer, length, ppos);
2996 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2997 void __user *buffer, size_t *length, loff_t *ppos)
2999 return hugetlb_sysctl_handler_common(true, table, write,
3000 buffer, length, ppos);
3002 #endif /* CONFIG_NUMA */
3004 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3005 void __user *buffer,
3006 size_t *length, loff_t *ppos)
3008 struct hstate *h = &default_hstate;
3012 if (!hugepages_supported())
3015 tmp = h->nr_overcommit_huge_pages;
3017 if (write && hstate_is_gigantic(h))
3021 table->maxlen = sizeof(unsigned long);
3022 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3027 spin_lock(&hugetlb_lock);
3028 h->nr_overcommit_huge_pages = tmp;
3029 spin_unlock(&hugetlb_lock);
3035 #endif /* CONFIG_SYSCTL */
3037 void hugetlb_report_meminfo(struct seq_file *m)
3040 unsigned long total = 0;
3042 if (!hugepages_supported())
3045 for_each_hstate(h) {
3046 unsigned long count = h->nr_huge_pages;
3048 total += (PAGE_SIZE << huge_page_order(h)) * count;
3050 if (h == &default_hstate)
3052 "HugePages_Total: %5lu\n"
3053 "HugePages_Free: %5lu\n"
3054 "HugePages_Rsvd: %5lu\n"
3055 "HugePages_Surp: %5lu\n"
3056 "Hugepagesize: %8lu kB\n",
3060 h->surplus_huge_pages,
3061 (PAGE_SIZE << huge_page_order(h)) / 1024);
3064 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3067 int hugetlb_report_node_meminfo(int nid, char *buf)
3069 struct hstate *h = &default_hstate;
3070 if (!hugepages_supported())
3073 "Node %d HugePages_Total: %5u\n"
3074 "Node %d HugePages_Free: %5u\n"
3075 "Node %d HugePages_Surp: %5u\n",
3076 nid, h->nr_huge_pages_node[nid],
3077 nid, h->free_huge_pages_node[nid],
3078 nid, h->surplus_huge_pages_node[nid]);
3081 void hugetlb_show_meminfo(void)
3086 if (!hugepages_supported())
3089 for_each_node_state(nid, N_MEMORY)
3091 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3093 h->nr_huge_pages_node[nid],
3094 h->free_huge_pages_node[nid],
3095 h->surplus_huge_pages_node[nid],
3096 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3099 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3101 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3102 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3105 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3106 unsigned long hugetlb_total_pages(void)
3109 unsigned long nr_total_pages = 0;
3112 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3113 return nr_total_pages;
3116 static int hugetlb_acct_memory(struct hstate *h, long delta)
3120 spin_lock(&hugetlb_lock);
3122 * When cpuset is configured, it breaks the strict hugetlb page
3123 * reservation as the accounting is done on a global variable. Such
3124 * reservation is completely rubbish in the presence of cpuset because
3125 * the reservation is not checked against page availability for the
3126 * current cpuset. Application can still potentially OOM'ed by kernel
3127 * with lack of free htlb page in cpuset that the task is in.
3128 * Attempt to enforce strict accounting with cpuset is almost
3129 * impossible (or too ugly) because cpuset is too fluid that
3130 * task or memory node can be dynamically moved between cpusets.
3132 * The change of semantics for shared hugetlb mapping with cpuset is
3133 * undesirable. However, in order to preserve some of the semantics,
3134 * we fall back to check against current free page availability as
3135 * a best attempt and hopefully to minimize the impact of changing
3136 * semantics that cpuset has.
3139 if (gather_surplus_pages(h, delta) < 0)
3142 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3143 return_unused_surplus_pages(h, delta);
3150 return_unused_surplus_pages(h, (unsigned long) -delta);
3153 spin_unlock(&hugetlb_lock);
3157 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3159 struct resv_map *resv = vma_resv_map(vma);
3162 * This new VMA should share its siblings reservation map if present.
3163 * The VMA will only ever have a valid reservation map pointer where
3164 * it is being copied for another still existing VMA. As that VMA
3165 * has a reference to the reservation map it cannot disappear until
3166 * after this open call completes. It is therefore safe to take a
3167 * new reference here without additional locking.
3169 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3170 kref_get(&resv->refs);
3173 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3175 struct hstate *h = hstate_vma(vma);
3176 struct resv_map *resv = vma_resv_map(vma);
3177 struct hugepage_subpool *spool = subpool_vma(vma);
3178 unsigned long reserve, start, end;
3181 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3184 start = vma_hugecache_offset(h, vma, vma->vm_start);
3185 end = vma_hugecache_offset(h, vma, vma->vm_end);
3187 reserve = (end - start) - region_count(resv, start, end);
3189 kref_put(&resv->refs, resv_map_release);
3193 * Decrement reserve counts. The global reserve count may be
3194 * adjusted if the subpool has a minimum size.
3196 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3197 hugetlb_acct_memory(h, -gbl_reserve);
3201 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3203 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3208 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3210 struct hstate *hstate = hstate_vma(vma);
3212 return 1UL << huge_page_shift(hstate);
3216 * We cannot handle pagefaults against hugetlb pages at all. They cause
3217 * handle_mm_fault() to try to instantiate regular-sized pages in the
3218 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3221 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3228 * When a new function is introduced to vm_operations_struct and added
3229 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3230 * This is because under System V memory model, mappings created via
3231 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3232 * their original vm_ops are overwritten with shm_vm_ops.
3234 const struct vm_operations_struct hugetlb_vm_ops = {
3235 .fault = hugetlb_vm_op_fault,
3236 .open = hugetlb_vm_op_open,
3237 .close = hugetlb_vm_op_close,
3238 .split = hugetlb_vm_op_split,
3239 .pagesize = hugetlb_vm_op_pagesize,
3242 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3248 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3249 vma->vm_page_prot)));
3251 entry = huge_pte_wrprotect(mk_huge_pte(page,
3252 vma->vm_page_prot));
3254 entry = pte_mkyoung(entry);
3255 entry = pte_mkhuge(entry);
3256 entry = arch_make_huge_pte(entry, vma, page, writable);
3261 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3262 unsigned long address, pte_t *ptep)
3266 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3267 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3268 update_mmu_cache(vma, address, ptep);
3271 bool is_hugetlb_entry_migration(pte_t pte)
3275 if (huge_pte_none(pte) || pte_present(pte))
3277 swp = pte_to_swp_entry(pte);
3278 if (non_swap_entry(swp) && is_migration_entry(swp))
3284 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3288 if (huge_pte_none(pte) || pte_present(pte))
3290 swp = pte_to_swp_entry(pte);
3291 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3297 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3298 struct vm_area_struct *vma)
3300 pte_t *src_pte, *dst_pte, entry, dst_entry;
3301 struct page *ptepage;
3304 struct hstate *h = hstate_vma(vma);
3305 unsigned long sz = huge_page_size(h);
3306 struct mmu_notifier_range range;
3309 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3312 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3315 mmu_notifier_invalidate_range_start(&range);
3318 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3319 spinlock_t *src_ptl, *dst_ptl;
3320 src_pte = huge_pte_offset(src, addr, sz);
3323 dst_pte = huge_pte_alloc(dst, addr, sz);
3330 * If the pagetables are shared don't copy or take references.
3331 * dst_pte == src_pte is the common case of src/dest sharing.
3333 * However, src could have 'unshared' and dst shares with
3334 * another vma. If dst_pte !none, this implies sharing.
3335 * Check here before taking page table lock, and once again
3336 * after taking the lock below.
3338 dst_entry = huge_ptep_get(dst_pte);
3339 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3342 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3343 src_ptl = huge_pte_lockptr(h, src, src_pte);
3344 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3345 entry = huge_ptep_get(src_pte);
3346 dst_entry = huge_ptep_get(dst_pte);
3347 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3349 * Skip if src entry none. Also, skip in the
3350 * unlikely case dst entry !none as this implies
3351 * sharing with another vma.
3354 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3355 is_hugetlb_entry_hwpoisoned(entry))) {
3356 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3358 if (is_write_migration_entry(swp_entry) && cow) {
3360 * COW mappings require pages in both
3361 * parent and child to be set to read.
3363 make_migration_entry_read(&swp_entry);
3364 entry = swp_entry_to_pte(swp_entry);
3365 set_huge_swap_pte_at(src, addr, src_pte,
3368 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3372 * No need to notify as we are downgrading page
3373 * table protection not changing it to point
3376 * See Documentation/vm/mmu_notifier.rst
3378 huge_ptep_set_wrprotect(src, addr, src_pte);
3380 entry = huge_ptep_get(src_pte);
3381 ptepage = pte_page(entry);
3383 page_dup_rmap(ptepage, true);
3384 set_huge_pte_at(dst, addr, dst_pte, entry);
3385 hugetlb_count_add(pages_per_huge_page(h), dst);
3387 spin_unlock(src_ptl);
3388 spin_unlock(dst_ptl);
3392 mmu_notifier_invalidate_range_end(&range);
3397 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3398 unsigned long start, unsigned long end,
3399 struct page *ref_page)
3401 struct mm_struct *mm = vma->vm_mm;
3402 unsigned long address;
3407 struct hstate *h = hstate_vma(vma);
3408 unsigned long sz = huge_page_size(h);
3409 struct mmu_notifier_range range;
3411 WARN_ON(!is_vm_hugetlb_page(vma));
3412 BUG_ON(start & ~huge_page_mask(h));
3413 BUG_ON(end & ~huge_page_mask(h));
3416 * This is a hugetlb vma, all the pte entries should point
3419 tlb_change_page_size(tlb, sz);
3420 tlb_start_vma(tlb, vma);
3423 * If sharing possible, alert mmu notifiers of worst case.
3425 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3427 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3428 mmu_notifier_invalidate_range_start(&range);
3430 for (; address < end; address += sz) {
3431 ptep = huge_pte_offset(mm, address, sz);
3435 ptl = huge_pte_lock(h, mm, ptep);
3436 if (huge_pmd_unshare(mm, &address, ptep)) {
3439 * We just unmapped a page of PMDs by clearing a PUD.
3440 * The caller's TLB flush range should cover this area.
3445 pte = huge_ptep_get(ptep);
3446 if (huge_pte_none(pte)) {
3452 * Migrating hugepage or HWPoisoned hugepage is already
3453 * unmapped and its refcount is dropped, so just clear pte here.
3455 if (unlikely(!pte_present(pte))) {
3456 huge_pte_clear(mm, address, ptep, sz);
3461 page = pte_page(pte);
3463 * If a reference page is supplied, it is because a specific
3464 * page is being unmapped, not a range. Ensure the page we
3465 * are about to unmap is the actual page of interest.
3468 if (page != ref_page) {
3473 * Mark the VMA as having unmapped its page so that
3474 * future faults in this VMA will fail rather than
3475 * looking like data was lost
3477 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3480 pte = huge_ptep_get_and_clear(mm, address, ptep);
3481 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3482 if (huge_pte_dirty(pte))
3483 set_page_dirty(page);
3485 hugetlb_count_sub(pages_per_huge_page(h), mm);
3486 page_remove_rmap(page, true);
3489 tlb_remove_page_size(tlb, page, huge_page_size(h));
3491 * Bail out after unmapping reference page if supplied
3496 mmu_notifier_invalidate_range_end(&range);
3497 tlb_end_vma(tlb, vma);
3500 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3501 struct vm_area_struct *vma, unsigned long start,
3502 unsigned long end, struct page *ref_page)
3504 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3507 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3508 * test will fail on a vma being torn down, and not grab a page table
3509 * on its way out. We're lucky that the flag has such an appropriate
3510 * name, and can in fact be safely cleared here. We could clear it
3511 * before the __unmap_hugepage_range above, but all that's necessary
3512 * is to clear it before releasing the i_mmap_rwsem. This works
3513 * because in the context this is called, the VMA is about to be
3514 * destroyed and the i_mmap_rwsem is held.
3516 vma->vm_flags &= ~VM_MAYSHARE;
3519 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3520 unsigned long end, struct page *ref_page)
3522 struct mm_struct *mm;
3523 struct mmu_gather tlb;
3524 unsigned long tlb_start = start;
3525 unsigned long tlb_end = end;
3528 * If shared PMDs were possibly used within this vma range, adjust
3529 * start/end for worst case tlb flushing.
3530 * Note that we can not be sure if PMDs are shared until we try to
3531 * unmap pages. However, we want to make sure TLB flushing covers
3532 * the largest possible range.
3534 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
3538 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
3539 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3540 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
3544 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3545 * mappping it owns the reserve page for. The intention is to unmap the page
3546 * from other VMAs and let the children be SIGKILLed if they are faulting the
3549 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3550 struct page *page, unsigned long address)
3552 struct hstate *h = hstate_vma(vma);
3553 struct vm_area_struct *iter_vma;
3554 struct address_space *mapping;
3558 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3559 * from page cache lookup which is in HPAGE_SIZE units.
3561 address = address & huge_page_mask(h);
3562 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3564 mapping = vma->vm_file->f_mapping;
3567 * Take the mapping lock for the duration of the table walk. As
3568 * this mapping should be shared between all the VMAs,
3569 * __unmap_hugepage_range() is called as the lock is already held
3571 i_mmap_lock_write(mapping);
3572 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3573 /* Do not unmap the current VMA */
3574 if (iter_vma == vma)
3578 * Shared VMAs have their own reserves and do not affect
3579 * MAP_PRIVATE accounting but it is possible that a shared
3580 * VMA is using the same page so check and skip such VMAs.
3582 if (iter_vma->vm_flags & VM_MAYSHARE)
3586 * Unmap the page from other VMAs without their own reserves.
3587 * They get marked to be SIGKILLed if they fault in these
3588 * areas. This is because a future no-page fault on this VMA
3589 * could insert a zeroed page instead of the data existing
3590 * from the time of fork. This would look like data corruption
3592 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3593 unmap_hugepage_range(iter_vma, address,
3594 address + huge_page_size(h), page);
3596 i_mmap_unlock_write(mapping);
3600 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3601 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3602 * cannot race with other handlers or page migration.
3603 * Keep the pte_same checks anyway to make transition from the mutex easier.
3605 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3606 unsigned long address, pte_t *ptep,
3607 struct page *pagecache_page, spinlock_t *ptl)
3610 struct hstate *h = hstate_vma(vma);
3611 struct page *old_page, *new_page;
3612 int outside_reserve = 0;
3614 unsigned long haddr = address & huge_page_mask(h);
3615 struct mmu_notifier_range range;
3617 pte = huge_ptep_get(ptep);
3618 old_page = pte_page(pte);
3621 /* If no-one else is actually using this page, avoid the copy
3622 * and just make the page writable */
3623 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3624 page_move_anon_rmap(old_page, vma);
3625 set_huge_ptep_writable(vma, haddr, ptep);
3630 * If the process that created a MAP_PRIVATE mapping is about to
3631 * perform a COW due to a shared page count, attempt to satisfy
3632 * the allocation without using the existing reserves. The pagecache
3633 * page is used to determine if the reserve at this address was
3634 * consumed or not. If reserves were used, a partial faulted mapping
3635 * at the time of fork() could consume its reserves on COW instead
3636 * of the full address range.
3638 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3639 old_page != pagecache_page)
3640 outside_reserve = 1;
3645 * Drop page table lock as buddy allocator may be called. It will
3646 * be acquired again before returning to the caller, as expected.
3649 new_page = alloc_huge_page(vma, haddr, outside_reserve);
3651 if (IS_ERR(new_page)) {
3653 * If a process owning a MAP_PRIVATE mapping fails to COW,
3654 * it is due to references held by a child and an insufficient
3655 * huge page pool. To guarantee the original mappers
3656 * reliability, unmap the page from child processes. The child
3657 * may get SIGKILLed if it later faults.
3659 if (outside_reserve) {
3661 BUG_ON(huge_pte_none(pte));
3662 unmap_ref_private(mm, vma, old_page, haddr);
3663 BUG_ON(huge_pte_none(pte));
3665 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3667 pte_same(huge_ptep_get(ptep), pte)))
3668 goto retry_avoidcopy;
3670 * race occurs while re-acquiring page table
3671 * lock, and our job is done.
3676 ret = vmf_error(PTR_ERR(new_page));
3677 goto out_release_old;
3681 * When the original hugepage is shared one, it does not have
3682 * anon_vma prepared.
3684 if (unlikely(anon_vma_prepare(vma))) {
3686 goto out_release_all;
3689 copy_user_huge_page(new_page, old_page, address, vma,
3690 pages_per_huge_page(h));
3691 __SetPageUptodate(new_page);
3693 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
3694 haddr + huge_page_size(h));
3695 mmu_notifier_invalidate_range_start(&range);
3698 * Retake the page table lock to check for racing updates
3699 * before the page tables are altered
3702 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3703 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3704 ClearPagePrivate(new_page);
3707 huge_ptep_clear_flush(vma, haddr, ptep);
3708 mmu_notifier_invalidate_range(mm, range.start, range.end);
3709 set_huge_pte_at(mm, haddr, ptep,
3710 make_huge_pte(vma, new_page, 1));
3711 page_remove_rmap(old_page, true);
3712 hugepage_add_new_anon_rmap(new_page, vma, haddr);
3713 set_page_huge_active(new_page);
3714 /* Make the old page be freed below */
3715 new_page = old_page;
3718 mmu_notifier_invalidate_range_end(&range);
3720 restore_reserve_on_error(h, vma, haddr, new_page);
3725 spin_lock(ptl); /* Caller expects lock to be held */
3729 /* Return the pagecache page at a given address within a VMA */
3730 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3731 struct vm_area_struct *vma, unsigned long address)
3733 struct address_space *mapping;
3736 mapping = vma->vm_file->f_mapping;
3737 idx = vma_hugecache_offset(h, vma, address);
3739 return find_lock_page(mapping, idx);
3743 * Return whether there is a pagecache page to back given address within VMA.
3744 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3746 static bool hugetlbfs_pagecache_present(struct hstate *h,
3747 struct vm_area_struct *vma, unsigned long address)
3749 struct address_space *mapping;
3753 mapping = vma->vm_file->f_mapping;
3754 idx = vma_hugecache_offset(h, vma, address);
3756 page = find_get_page(mapping, idx);
3759 return page != NULL;
3762 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3765 struct inode *inode = mapping->host;
3766 struct hstate *h = hstate_inode(inode);
3767 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3771 ClearPagePrivate(page);
3774 * set page dirty so that it will not be removed from cache/file
3775 * by non-hugetlbfs specific code paths.
3777 set_page_dirty(page);
3779 spin_lock(&inode->i_lock);
3780 inode->i_blocks += blocks_per_huge_page(h);
3781 spin_unlock(&inode->i_lock);
3785 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
3786 struct vm_area_struct *vma,
3787 struct address_space *mapping, pgoff_t idx,
3788 unsigned long address, pte_t *ptep, unsigned int flags)
3790 struct hstate *h = hstate_vma(vma);
3791 vm_fault_t ret = VM_FAULT_SIGBUS;
3797 unsigned long haddr = address & huge_page_mask(h);
3798 bool new_page = false;
3801 * Currently, we are forced to kill the process in the event the
3802 * original mapper has unmapped pages from the child due to a failed
3803 * COW. Warn that such a situation has occurred as it may not be obvious
3805 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3806 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3812 * Use page lock to guard against racing truncation
3813 * before we get page_table_lock.
3816 page = find_lock_page(mapping, idx);
3818 size = i_size_read(mapping->host) >> huge_page_shift(h);
3823 * Check for page in userfault range
3825 if (userfaultfd_missing(vma)) {
3827 struct vm_fault vmf = {
3832 * Hard to debug if it ends up being
3833 * used by a callee that assumes
3834 * something about the other
3835 * uninitialized fields... same as in
3841 * hugetlb_fault_mutex must be dropped before
3842 * handling userfault. Reacquire after handling
3843 * fault to make calling code simpler.
3845 hash = hugetlb_fault_mutex_hash(h, mapping, idx);
3846 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3847 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3848 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3852 page = alloc_huge_page(vma, haddr, 0);
3855 * Returning error will result in faulting task being
3856 * sent SIGBUS. The hugetlb fault mutex prevents two
3857 * tasks from racing to fault in the same page which
3858 * could result in false unable to allocate errors.
3859 * Page migration does not take the fault mutex, but
3860 * does a clear then write of pte's under page table
3861 * lock. Page fault code could race with migration,
3862 * notice the clear pte and try to allocate a page
3863 * here. Before returning error, get ptl and make
3864 * sure there really is no pte entry.
3866 ptl = huge_pte_lock(h, mm, ptep);
3867 if (!huge_pte_none(huge_ptep_get(ptep))) {
3873 ret = vmf_error(PTR_ERR(page));
3876 clear_huge_page(page, address, pages_per_huge_page(h));
3877 __SetPageUptodate(page);
3880 if (vma->vm_flags & VM_MAYSHARE) {
3881 int err = huge_add_to_page_cache(page, mapping, idx);
3890 if (unlikely(anon_vma_prepare(vma))) {
3892 goto backout_unlocked;
3898 * If memory error occurs between mmap() and fault, some process
3899 * don't have hwpoisoned swap entry for errored virtual address.
3900 * So we need to block hugepage fault by PG_hwpoison bit check.
3902 if (unlikely(PageHWPoison(page))) {
3903 ret = VM_FAULT_HWPOISON |
3904 VM_FAULT_SET_HINDEX(hstate_index(h));
3905 goto backout_unlocked;
3910 * If we are going to COW a private mapping later, we examine the
3911 * pending reservations for this page now. This will ensure that
3912 * any allocations necessary to record that reservation occur outside
3915 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3916 if (vma_needs_reservation(h, vma, haddr) < 0) {
3918 goto backout_unlocked;
3920 /* Just decrements count, does not deallocate */
3921 vma_end_reservation(h, vma, haddr);
3924 ptl = huge_pte_lock(h, mm, ptep);
3925 size = i_size_read(mapping->host) >> huge_page_shift(h);
3930 if (!huge_pte_none(huge_ptep_get(ptep)))
3934 ClearPagePrivate(page);
3935 hugepage_add_new_anon_rmap(page, vma, haddr);
3937 page_dup_rmap(page, true);
3938 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3939 && (vma->vm_flags & VM_SHARED)));
3940 set_huge_pte_at(mm, haddr, ptep, new_pte);
3942 hugetlb_count_add(pages_per_huge_page(h), mm);
3943 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3944 /* Optimization, do the COW without a second fault */
3945 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
3951 * Only make newly allocated pages active. Existing pages found
3952 * in the pagecache could be !page_huge_active() if they have been
3953 * isolated for migration.
3956 set_page_huge_active(page);
3966 restore_reserve_on_error(h, vma, haddr, page);
3972 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping,
3975 unsigned long key[2];
3978 key[0] = (unsigned long) mapping;
3981 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
3983 return hash & (num_fault_mutexes - 1);
3987 * For uniprocesor systems we always use a single mutex, so just
3988 * return 0 and avoid the hashing overhead.
3990 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping,
3997 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3998 unsigned long address, unsigned int flags)
4005 struct page *page = NULL;
4006 struct page *pagecache_page = NULL;
4007 struct hstate *h = hstate_vma(vma);
4008 struct address_space *mapping;
4009 int need_wait_lock = 0;
4010 unsigned long haddr = address & huge_page_mask(h);
4012 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4014 entry = huge_ptep_get(ptep);
4015 if (unlikely(is_hugetlb_entry_migration(entry))) {
4016 migration_entry_wait_huge(vma, mm, ptep);
4018 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4019 return VM_FAULT_HWPOISON_LARGE |
4020 VM_FAULT_SET_HINDEX(hstate_index(h));
4022 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4024 return VM_FAULT_OOM;
4027 mapping = vma->vm_file->f_mapping;
4028 idx = vma_hugecache_offset(h, vma, haddr);
4031 * Serialize hugepage allocation and instantiation, so that we don't
4032 * get spurious allocation failures if two CPUs race to instantiate
4033 * the same page in the page cache.
4035 hash = hugetlb_fault_mutex_hash(h, mapping, idx);
4036 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4038 entry = huge_ptep_get(ptep);
4039 if (huge_pte_none(entry)) {
4040 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4047 * entry could be a migration/hwpoison entry at this point, so this
4048 * check prevents the kernel from going below assuming that we have
4049 * a active hugepage in pagecache. This goto expects the 2nd page fault,
4050 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
4053 if (!pte_present(entry))
4057 * If we are going to COW the mapping later, we examine the pending
4058 * reservations for this page now. This will ensure that any
4059 * allocations necessary to record that reservation occur outside the
4060 * spinlock. For private mappings, we also lookup the pagecache
4061 * page now as it is used to determine if a reservation has been
4064 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4065 if (vma_needs_reservation(h, vma, haddr) < 0) {
4069 /* Just decrements count, does not deallocate */
4070 vma_end_reservation(h, vma, haddr);
4072 if (!(vma->vm_flags & VM_MAYSHARE))
4073 pagecache_page = hugetlbfs_pagecache_page(h,
4077 ptl = huge_pte_lock(h, mm, ptep);
4079 /* Check for a racing update before calling hugetlb_cow */
4080 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4084 * hugetlb_cow() requires page locks of pte_page(entry) and
4085 * pagecache_page, so here we need take the former one
4086 * when page != pagecache_page or !pagecache_page.
4088 page = pte_page(entry);
4089 if (page != pagecache_page)
4090 if (!trylock_page(page)) {
4097 if (flags & FAULT_FLAG_WRITE) {
4098 if (!huge_pte_write(entry)) {
4099 ret = hugetlb_cow(mm, vma, address, ptep,
4100 pagecache_page, ptl);
4103 entry = huge_pte_mkdirty(entry);
4105 entry = pte_mkyoung(entry);
4106 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4107 flags & FAULT_FLAG_WRITE))
4108 update_mmu_cache(vma, haddr, ptep);
4110 if (page != pagecache_page)
4116 if (pagecache_page) {
4117 unlock_page(pagecache_page);
4118 put_page(pagecache_page);
4121 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4123 * Generally it's safe to hold refcount during waiting page lock. But
4124 * here we just wait to defer the next page fault to avoid busy loop and
4125 * the page is not used after unlocked before returning from the current
4126 * page fault. So we are safe from accessing freed page, even if we wait
4127 * here without taking refcount.
4130 wait_on_page_locked(page);
4135 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4136 * modifications for huge pages.
4138 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4140 struct vm_area_struct *dst_vma,
4141 unsigned long dst_addr,
4142 unsigned long src_addr,
4143 struct page **pagep)
4145 struct address_space *mapping;
4148 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4149 struct hstate *h = hstate_vma(dst_vma);
4157 page = alloc_huge_page(dst_vma, dst_addr, 0);
4161 ret = copy_huge_page_from_user(page,
4162 (const void __user *) src_addr,
4163 pages_per_huge_page(h), false);
4165 /* fallback to copy_from_user outside mmap_sem */
4166 if (unlikely(ret)) {
4169 /* don't free the page */
4178 * The memory barrier inside __SetPageUptodate makes sure that
4179 * preceding stores to the page contents become visible before
4180 * the set_pte_at() write.
4182 __SetPageUptodate(page);
4184 mapping = dst_vma->vm_file->f_mapping;
4185 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4188 * If shared, add to page cache
4191 size = i_size_read(mapping->host) >> huge_page_shift(h);
4194 goto out_release_nounlock;
4197 * Serialization between remove_inode_hugepages() and
4198 * huge_add_to_page_cache() below happens through the
4199 * hugetlb_fault_mutex_table that here must be hold by
4202 ret = huge_add_to_page_cache(page, mapping, idx);
4204 goto out_release_nounlock;
4207 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4211 * Recheck the i_size after holding PT lock to make sure not
4212 * to leave any page mapped (as page_mapped()) beyond the end
4213 * of the i_size (remove_inode_hugepages() is strict about
4214 * enforcing that). If we bail out here, we'll also leave a
4215 * page in the radix tree in the vm_shared case beyond the end
4216 * of the i_size, but remove_inode_hugepages() will take care
4217 * of it as soon as we drop the hugetlb_fault_mutex_table.
4219 size = i_size_read(mapping->host) >> huge_page_shift(h);
4222 goto out_release_unlock;
4225 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4226 goto out_release_unlock;
4229 page_dup_rmap(page, true);
4231 ClearPagePrivate(page);
4232 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4235 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4236 if (dst_vma->vm_flags & VM_WRITE)
4237 _dst_pte = huge_pte_mkdirty(_dst_pte);
4238 _dst_pte = pte_mkyoung(_dst_pte);
4240 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4242 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4243 dst_vma->vm_flags & VM_WRITE);
4244 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4246 /* No need to invalidate - it was non-present before */
4247 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4250 set_page_huge_active(page);
4260 out_release_nounlock:
4265 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4266 struct page **pages, struct vm_area_struct **vmas,
4267 unsigned long *position, unsigned long *nr_pages,
4268 long i, unsigned int flags, int *nonblocking)
4270 unsigned long pfn_offset;
4271 unsigned long vaddr = *position;
4272 unsigned long remainder = *nr_pages;
4273 struct hstate *h = hstate_vma(vma);
4276 while (vaddr < vma->vm_end && remainder) {
4278 spinlock_t *ptl = NULL;
4283 * If we have a pending SIGKILL, don't keep faulting pages and
4284 * potentially allocating memory.
4286 if (fatal_signal_pending(current)) {
4292 * Some archs (sparc64, sh*) have multiple pte_ts to
4293 * each hugepage. We have to make sure we get the
4294 * first, for the page indexing below to work.
4296 * Note that page table lock is not held when pte is null.
4298 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4301 ptl = huge_pte_lock(h, mm, pte);
4302 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4305 * When coredumping, it suits get_dump_page if we just return
4306 * an error where there's an empty slot with no huge pagecache
4307 * to back it. This way, we avoid allocating a hugepage, and
4308 * the sparse dumpfile avoids allocating disk blocks, but its
4309 * huge holes still show up with zeroes where they need to be.
4311 if (absent && (flags & FOLL_DUMP) &&
4312 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4320 * We need call hugetlb_fault for both hugepages under migration
4321 * (in which case hugetlb_fault waits for the migration,) and
4322 * hwpoisoned hugepages (in which case we need to prevent the
4323 * caller from accessing to them.) In order to do this, we use
4324 * here is_swap_pte instead of is_hugetlb_entry_migration and
4325 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4326 * both cases, and because we can't follow correct pages
4327 * directly from any kind of swap entries.
4329 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4330 ((flags & FOLL_WRITE) &&
4331 !huge_pte_write(huge_ptep_get(pte)))) {
4333 unsigned int fault_flags = 0;
4337 if (flags & FOLL_WRITE)
4338 fault_flags |= FAULT_FLAG_WRITE;
4340 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4341 if (flags & FOLL_NOWAIT)
4342 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4343 FAULT_FLAG_RETRY_NOWAIT;
4344 if (flags & FOLL_TRIED) {
4345 VM_WARN_ON_ONCE(fault_flags &
4346 FAULT_FLAG_ALLOW_RETRY);
4347 fault_flags |= FAULT_FLAG_TRIED;
4349 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4350 if (ret & VM_FAULT_ERROR) {
4351 err = vm_fault_to_errno(ret, flags);
4355 if (ret & VM_FAULT_RETRY) {
4357 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4361 * VM_FAULT_RETRY must not return an
4362 * error, it will return zero
4365 * No need to update "position" as the
4366 * caller will not check it after
4367 * *nr_pages is set to 0.
4374 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4375 page = pte_page(huge_ptep_get(pte));
4378 * Instead of doing 'try_get_page()' below in the same_page
4379 * loop, just check the count once here.
4381 if (unlikely(page_count(page) <= 0)) {
4391 pages[i] = mem_map_offset(page, pfn_offset);
4402 if (vaddr < vma->vm_end && remainder &&
4403 pfn_offset < pages_per_huge_page(h)) {
4405 * We use pfn_offset to avoid touching the pageframes
4406 * of this compound page.
4412 *nr_pages = remainder;
4414 * setting position is actually required only if remainder is
4415 * not zero but it's faster not to add a "if (remainder)"
4423 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4425 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4428 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4431 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4432 unsigned long address, unsigned long end, pgprot_t newprot)
4434 struct mm_struct *mm = vma->vm_mm;
4435 unsigned long start = address;
4438 struct hstate *h = hstate_vma(vma);
4439 unsigned long pages = 0;
4440 bool shared_pmd = false;
4441 struct mmu_notifier_range range;
4444 * In the case of shared PMDs, the area to flush could be beyond
4445 * start/end. Set range.start/range.end to cover the maximum possible
4446 * range if PMD sharing is possible.
4448 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
4449 0, vma, mm, start, end);
4450 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4452 BUG_ON(address >= end);
4453 flush_cache_range(vma, range.start, range.end);
4455 mmu_notifier_invalidate_range_start(&range);
4456 i_mmap_lock_write(vma->vm_file->f_mapping);
4457 for (; address < end; address += huge_page_size(h)) {
4459 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4462 ptl = huge_pte_lock(h, mm, ptep);
4463 if (huge_pmd_unshare(mm, &address, ptep)) {
4469 pte = huge_ptep_get(ptep);
4470 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4474 if (unlikely(is_hugetlb_entry_migration(pte))) {
4475 swp_entry_t entry = pte_to_swp_entry(pte);
4477 if (is_write_migration_entry(entry)) {
4480 make_migration_entry_read(&entry);
4481 newpte = swp_entry_to_pte(entry);
4482 set_huge_swap_pte_at(mm, address, ptep,
4483 newpte, huge_page_size(h));
4489 if (!huge_pte_none(pte)) {
4492 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
4493 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
4494 pte = arch_make_huge_pte(pte, vma, NULL, 0);
4495 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
4501 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4502 * may have cleared our pud entry and done put_page on the page table:
4503 * once we release i_mmap_rwsem, another task can do the final put_page
4504 * and that page table be reused and filled with junk. If we actually
4505 * did unshare a page of pmds, flush the range corresponding to the pud.
4508 flush_hugetlb_tlb_range(vma, range.start, range.end);
4510 flush_hugetlb_tlb_range(vma, start, end);
4512 * No need to call mmu_notifier_invalidate_range() we are downgrading
4513 * page table protection not changing it to point to a new page.
4515 * See Documentation/vm/mmu_notifier.rst
4517 i_mmap_unlock_write(vma->vm_file->f_mapping);
4518 mmu_notifier_invalidate_range_end(&range);
4520 return pages << h->order;
4523 int hugetlb_reserve_pages(struct inode *inode,
4525 struct vm_area_struct *vma,
4526 vm_flags_t vm_flags)
4529 struct hstate *h = hstate_inode(inode);
4530 struct hugepage_subpool *spool = subpool_inode(inode);
4531 struct resv_map *resv_map;
4534 /* This should never happen */
4536 VM_WARN(1, "%s called with a negative range\n", __func__);
4541 * Only apply hugepage reservation if asked. At fault time, an
4542 * attempt will be made for VM_NORESERVE to allocate a page
4543 * without using reserves
4545 if (vm_flags & VM_NORESERVE)
4549 * Shared mappings base their reservation on the number of pages that
4550 * are already allocated on behalf of the file. Private mappings need
4551 * to reserve the full area even if read-only as mprotect() may be
4552 * called to make the mapping read-write. Assume !vma is a shm mapping
4554 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4556 * resv_map can not be NULL as hugetlb_reserve_pages is only
4557 * called for inodes for which resv_maps were created (see
4558 * hugetlbfs_get_inode).
4560 resv_map = inode_resv_map(inode);
4562 chg = region_chg(resv_map, from, to);
4565 resv_map = resv_map_alloc();
4571 set_vma_resv_map(vma, resv_map);
4572 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4581 * There must be enough pages in the subpool for the mapping. If
4582 * the subpool has a minimum size, there may be some global
4583 * reservations already in place (gbl_reserve).
4585 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4586 if (gbl_reserve < 0) {
4592 * Check enough hugepages are available for the reservation.
4593 * Hand the pages back to the subpool if there are not
4595 ret = hugetlb_acct_memory(h, gbl_reserve);
4597 /* put back original number of pages, chg */
4598 (void)hugepage_subpool_put_pages(spool, chg);
4603 * Account for the reservations made. Shared mappings record regions
4604 * that have reservations as they are shared by multiple VMAs.
4605 * When the last VMA disappears, the region map says how much
4606 * the reservation was and the page cache tells how much of
4607 * the reservation was consumed. Private mappings are per-VMA and
4608 * only the consumed reservations are tracked. When the VMA
4609 * disappears, the original reservation is the VMA size and the
4610 * consumed reservations are stored in the map. Hence, nothing
4611 * else has to be done for private mappings here
4613 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4614 long add = region_add(resv_map, from, to);
4616 if (unlikely(chg > add)) {
4618 * pages in this range were added to the reserve
4619 * map between region_chg and region_add. This
4620 * indicates a race with alloc_huge_page. Adjust
4621 * the subpool and reserve counts modified above
4622 * based on the difference.
4626 rsv_adjust = hugepage_subpool_put_pages(spool,
4628 hugetlb_acct_memory(h, -rsv_adjust);
4633 if (!vma || vma->vm_flags & VM_MAYSHARE)
4634 /* Don't call region_abort if region_chg failed */
4636 region_abort(resv_map, from, to);
4637 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4638 kref_put(&resv_map->refs, resv_map_release);
4642 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4645 struct hstate *h = hstate_inode(inode);
4646 struct resv_map *resv_map = inode_resv_map(inode);
4648 struct hugepage_subpool *spool = subpool_inode(inode);
4652 * Since this routine can be called in the evict inode path for all
4653 * hugetlbfs inodes, resv_map could be NULL.
4656 chg = region_del(resv_map, start, end);
4658 * region_del() can fail in the rare case where a region
4659 * must be split and another region descriptor can not be
4660 * allocated. If end == LONG_MAX, it will not fail.
4666 spin_lock(&inode->i_lock);
4667 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4668 spin_unlock(&inode->i_lock);
4671 * If the subpool has a minimum size, the number of global
4672 * reservations to be released may be adjusted.
4674 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4675 hugetlb_acct_memory(h, -gbl_reserve);
4680 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4681 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4682 struct vm_area_struct *vma,
4683 unsigned long addr, pgoff_t idx)
4685 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4687 unsigned long sbase = saddr & PUD_MASK;
4688 unsigned long s_end = sbase + PUD_SIZE;
4690 /* Allow segments to share if only one is marked locked */
4691 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4692 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4695 * match the virtual addresses, permission and the alignment of the
4698 if (pmd_index(addr) != pmd_index(saddr) ||
4699 vm_flags != svm_flags ||
4700 sbase < svma->vm_start || svma->vm_end < s_end)
4706 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4708 unsigned long base = addr & PUD_MASK;
4709 unsigned long end = base + PUD_SIZE;
4712 * check on proper vm_flags and page table alignment
4714 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
4720 * Determine if start,end range within vma could be mapped by shared pmd.
4721 * If yes, adjust start and end to cover range associated with possible
4722 * shared pmd mappings.
4724 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4725 unsigned long *start, unsigned long *end)
4727 unsigned long check_addr = *start;
4729 if (!(vma->vm_flags & VM_MAYSHARE))
4732 for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
4733 unsigned long a_start = check_addr & PUD_MASK;
4734 unsigned long a_end = a_start + PUD_SIZE;
4737 * If sharing is possible, adjust start/end if necessary.
4739 if (range_in_vma(vma, a_start, a_end)) {
4740 if (a_start < *start)
4749 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4750 * and returns the corresponding pte. While this is not necessary for the
4751 * !shared pmd case because we can allocate the pmd later as well, it makes the
4752 * code much cleaner. pmd allocation is essential for the shared case because
4753 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4754 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4755 * bad pmd for sharing.
4757 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4759 struct vm_area_struct *vma = find_vma(mm, addr);
4760 struct address_space *mapping = vma->vm_file->f_mapping;
4761 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4763 struct vm_area_struct *svma;
4764 unsigned long saddr;
4769 if (!vma_shareable(vma, addr))
4770 return (pte_t *)pmd_alloc(mm, pud, addr);
4772 i_mmap_lock_read(mapping);
4773 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4777 saddr = page_table_shareable(svma, vma, addr, idx);
4779 spte = huge_pte_offset(svma->vm_mm, saddr,
4780 vma_mmu_pagesize(svma));
4782 get_page(virt_to_page(spte));
4791 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4792 if (pud_none(*pud)) {
4793 pud_populate(mm, pud,
4794 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4797 put_page(virt_to_page(spte));
4801 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4802 i_mmap_unlock_read(mapping);
4807 * unmap huge page backed by shared pte.
4809 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4810 * indicated by page_count > 1, unmap is achieved by clearing pud and
4811 * decrementing the ref count. If count == 1, the pte page is not shared.
4813 * called with page table lock held.
4815 * returns: 1 successfully unmapped a shared pte page
4816 * 0 the underlying pte page is not shared, or it is the last user
4818 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4820 pgd_t *pgd = pgd_offset(mm, *addr);
4821 p4d_t *p4d = p4d_offset(pgd, *addr);
4822 pud_t *pud = pud_offset(p4d, *addr);
4824 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4825 if (page_count(virt_to_page(ptep)) == 1)
4829 put_page(virt_to_page(ptep));
4831 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4834 #define want_pmd_share() (1)
4835 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4836 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4841 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4846 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4847 unsigned long *start, unsigned long *end)
4850 #define want_pmd_share() (0)
4851 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4853 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4854 pte_t *huge_pte_alloc(struct mm_struct *mm,
4855 unsigned long addr, unsigned long sz)
4862 pgd = pgd_offset(mm, addr);
4863 p4d = p4d_alloc(mm, pgd, addr);
4866 pud = pud_alloc(mm, p4d, addr);
4868 if (sz == PUD_SIZE) {
4871 BUG_ON(sz != PMD_SIZE);
4872 if (want_pmd_share() && pud_none(*pud))
4873 pte = huge_pmd_share(mm, addr, pud);
4875 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4878 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4884 * huge_pte_offset() - Walk the page table to resolve the hugepage
4885 * entry at address @addr
4887 * Return: Pointer to page table or swap entry (PUD or PMD) for
4888 * address @addr, or NULL if a p*d_none() entry is encountered and the
4889 * size @sz doesn't match the hugepage size at this level of the page
4892 pte_t *huge_pte_offset(struct mm_struct *mm,
4893 unsigned long addr, unsigned long sz)
4900 pgd = pgd_offset(mm, addr);
4901 if (!pgd_present(*pgd))
4903 p4d = p4d_offset(pgd, addr);
4904 if (!p4d_present(*p4d))
4907 pud = pud_offset(p4d, addr);
4908 if (sz != PUD_SIZE && pud_none(*pud))
4910 /* hugepage or swap? */
4911 if (pud_huge(*pud) || !pud_present(*pud))
4912 return (pte_t *)pud;
4914 pmd = pmd_offset(pud, addr);
4915 if (sz != PMD_SIZE && pmd_none(*pmd))
4917 /* hugepage or swap? */
4918 if (pmd_huge(*pmd) || !pmd_present(*pmd))
4919 return (pte_t *)pmd;
4924 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4927 * These functions are overwritable if your architecture needs its own
4930 struct page * __weak
4931 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4934 return ERR_PTR(-EINVAL);
4937 struct page * __weak
4938 follow_huge_pd(struct vm_area_struct *vma,
4939 unsigned long address, hugepd_t hpd, int flags, int pdshift)
4941 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4945 struct page * __weak
4946 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4947 pmd_t *pmd, int flags)
4949 struct page *page = NULL;
4953 ptl = pmd_lockptr(mm, pmd);
4956 * make sure that the address range covered by this pmd is not
4957 * unmapped from other threads.
4959 if (!pmd_huge(*pmd))
4961 pte = huge_ptep_get((pte_t *)pmd);
4962 if (pte_present(pte)) {
4963 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4964 if (flags & FOLL_GET)
4967 if (is_hugetlb_entry_migration(pte)) {
4969 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4973 * hwpoisoned entry is treated as no_page_table in
4974 * follow_page_mask().
4982 struct page * __weak
4983 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4984 pud_t *pud, int flags)
4986 if (flags & FOLL_GET)
4989 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4992 struct page * __weak
4993 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
4995 if (flags & FOLL_GET)
4998 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5001 bool isolate_huge_page(struct page *page, struct list_head *list)
5005 VM_BUG_ON_PAGE(!PageHead(page), page);
5006 spin_lock(&hugetlb_lock);
5007 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
5011 clear_page_huge_active(page);
5012 list_move_tail(&page->lru, list);
5014 spin_unlock(&hugetlb_lock);
5018 void putback_active_hugepage(struct page *page)
5020 VM_BUG_ON_PAGE(!PageHead(page), page);
5021 spin_lock(&hugetlb_lock);
5022 set_page_huge_active(page);
5023 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5024 spin_unlock(&hugetlb_lock);
5028 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5030 struct hstate *h = page_hstate(oldpage);
5032 hugetlb_cgroup_migrate(oldpage, newpage);
5033 set_page_owner_migrate_reason(newpage, reason);
5036 * transfer temporary state of the new huge page. This is
5037 * reverse to other transitions because the newpage is going to
5038 * be final while the old one will be freed so it takes over
5039 * the temporary status.
5041 * Also note that we have to transfer the per-node surplus state
5042 * here as well otherwise the global surplus count will not match
5045 if (PageHugeTemporary(newpage)) {
5046 int old_nid = page_to_nid(oldpage);
5047 int new_nid = page_to_nid(newpage);
5049 SetPageHugeTemporary(oldpage);
5050 ClearPageHugeTemporary(newpage);
5052 spin_lock(&hugetlb_lock);
5053 if (h->surplus_huge_pages_node[old_nid]) {
5054 h->surplus_huge_pages_node[old_nid]--;
5055 h->surplus_huge_pages_node[new_nid]++;
5057 spin_unlock(&hugetlb_lock);