2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
8 #include <linux/seq_file.h>
9 #include <linux/sysctl.h>
10 #include <linux/highmem.h>
11 #include <linux/mmu_notifier.h>
12 #include <linux/nodemask.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/compiler.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/mmdebug.h>
22 #include <linux/sched/signal.h>
23 #include <linux/rmap.h>
24 #include <linux/string_helpers.h>
25 #include <linux/swap.h>
26 #include <linux/swapops.h>
27 #include <linux/jhash.h>
30 #include <asm/pgtable.h>
34 #include <linux/hugetlb.h>
35 #include <linux/hugetlb_cgroup.h>
36 #include <linux/node.h>
37 #include <linux/userfaultfd_k.h>
38 #include <linux/page_owner.h>
41 int hugetlb_max_hstate __read_mostly;
42 unsigned int default_hstate_idx;
43 struct hstate hstates[HUGE_MAX_HSTATE];
45 * Minimum page order among possible hugepage sizes, set to a proper value
48 static unsigned int minimum_order __read_mostly = UINT_MAX;
50 __initdata LIST_HEAD(huge_boot_pages);
52 /* for command line parsing */
53 static struct hstate * __initdata parsed_hstate;
54 static unsigned long __initdata default_hstate_max_huge_pages;
55 static unsigned long __initdata default_hstate_size;
56 static bool __initdata parsed_valid_hugepagesz = true;
59 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
60 * free_huge_pages, and surplus_huge_pages.
62 DEFINE_SPINLOCK(hugetlb_lock);
65 * Serializes faults on the same logical page. This is used to
66 * prevent spurious OOMs when the hugepage pool is fully utilized.
68 static int num_fault_mutexes;
69 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
71 static inline bool PageHugeFreed(struct page *head)
73 return page_private(head + 4) == -1UL;
76 static inline void SetPageHugeFreed(struct page *head)
78 set_page_private(head + 4, -1UL);
81 static inline void ClearPageHugeFreed(struct page *head)
83 set_page_private(head + 4, 0);
86 /* Forward declaration */
87 static int hugetlb_acct_memory(struct hstate *h, long delta);
89 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
91 bool free = (spool->count == 0) && (spool->used_hpages == 0);
93 spin_unlock(&spool->lock);
95 /* If no pages are used, and no other handles to the subpool
96 * remain, give up any reservations mased on minimum size and
99 if (spool->min_hpages != -1)
100 hugetlb_acct_memory(spool->hstate,
106 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
109 struct hugepage_subpool *spool;
111 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
115 spin_lock_init(&spool->lock);
117 spool->max_hpages = max_hpages;
119 spool->min_hpages = min_hpages;
121 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
125 spool->rsv_hpages = min_hpages;
130 void hugepage_put_subpool(struct hugepage_subpool *spool)
132 spin_lock(&spool->lock);
133 BUG_ON(!spool->count);
135 unlock_or_release_subpool(spool);
139 * Subpool accounting for allocating and reserving pages.
140 * Return -ENOMEM if there are not enough resources to satisfy the
141 * the request. Otherwise, return the number of pages by which the
142 * global pools must be adjusted (upward). The returned value may
143 * only be different than the passed value (delta) in the case where
144 * a subpool minimum size must be manitained.
146 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
154 spin_lock(&spool->lock);
156 if (spool->max_hpages != -1) { /* maximum size accounting */
157 if ((spool->used_hpages + delta) <= spool->max_hpages)
158 spool->used_hpages += delta;
165 /* minimum size accounting */
166 if (spool->min_hpages != -1 && spool->rsv_hpages) {
167 if (delta > spool->rsv_hpages) {
169 * Asking for more reserves than those already taken on
170 * behalf of subpool. Return difference.
172 ret = delta - spool->rsv_hpages;
173 spool->rsv_hpages = 0;
175 ret = 0; /* reserves already accounted for */
176 spool->rsv_hpages -= delta;
181 spin_unlock(&spool->lock);
186 * Subpool accounting for freeing and unreserving pages.
187 * Return the number of global page reservations that must be dropped.
188 * The return value may only be different than the passed value (delta)
189 * in the case where a subpool minimum size must be maintained.
191 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
199 spin_lock(&spool->lock);
201 if (spool->max_hpages != -1) /* maximum size accounting */
202 spool->used_hpages -= delta;
204 /* minimum size accounting */
205 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
206 if (spool->rsv_hpages + delta <= spool->min_hpages)
209 ret = spool->rsv_hpages + delta - spool->min_hpages;
211 spool->rsv_hpages += delta;
212 if (spool->rsv_hpages > spool->min_hpages)
213 spool->rsv_hpages = spool->min_hpages;
217 * If hugetlbfs_put_super couldn't free spool due to an outstanding
218 * quota reference, free it now.
220 unlock_or_release_subpool(spool);
225 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
227 return HUGETLBFS_SB(inode->i_sb)->spool;
230 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
232 return subpool_inode(file_inode(vma->vm_file));
236 * Region tracking -- allows tracking of reservations and instantiated pages
237 * across the pages in a mapping.
239 * The region data structures are embedded into a resv_map and protected
240 * by a resv_map's lock. The set of regions within the resv_map represent
241 * reservations for huge pages, or huge pages that have already been
242 * instantiated within the map. The from and to elements are huge page
243 * indicies into the associated mapping. from indicates the starting index
244 * of the region. to represents the first index past the end of the region.
246 * For example, a file region structure with from == 0 and to == 4 represents
247 * four huge pages in a mapping. It is important to note that the to element
248 * represents the first element past the end of the region. This is used in
249 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
251 * Interval notation of the form [from, to) will be used to indicate that
252 * the endpoint from is inclusive and to is exclusive.
255 struct list_head link;
261 * Add the huge page range represented by [f, t) to the reserve
262 * map. In the normal case, existing regions will be expanded
263 * to accommodate the specified range. Sufficient regions should
264 * exist for expansion due to the previous call to region_chg
265 * with the same range. However, it is possible that region_del
266 * could have been called after region_chg and modifed the map
267 * in such a way that no region exists to be expanded. In this
268 * case, pull a region descriptor from the cache associated with
269 * the map and use that for the new range.
271 * Return the number of new huge pages added to the map. This
272 * number is greater than or equal to zero.
274 static long region_add(struct resv_map *resv, long f, long t)
276 struct list_head *head = &resv->regions;
277 struct file_region *rg, *nrg, *trg;
280 spin_lock(&resv->lock);
281 /* Locate the region we are either in or before. */
282 list_for_each_entry(rg, head, link)
287 * If no region exists which can be expanded to include the
288 * specified range, the list must have been modified by an
289 * interleving call to region_del(). Pull a region descriptor
290 * from the cache and use it for this range.
292 if (&rg->link == head || t < rg->from) {
293 VM_BUG_ON(resv->region_cache_count <= 0);
295 resv->region_cache_count--;
296 nrg = list_first_entry(&resv->region_cache, struct file_region,
298 list_del(&nrg->link);
302 list_add(&nrg->link, rg->link.prev);
308 /* Round our left edge to the current segment if it encloses us. */
312 /* Check for and consume any regions we now overlap with. */
314 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
315 if (&rg->link == head)
320 /* If this area reaches higher then extend our area to
321 * include it completely. If this is not the first area
322 * which we intend to reuse, free it. */
326 /* Decrement return value by the deleted range.
327 * Another range will span this area so that by
328 * end of routine add will be >= zero
330 add -= (rg->to - rg->from);
336 add += (nrg->from - f); /* Added to beginning of region */
338 add += t - nrg->to; /* Added to end of region */
342 resv->adds_in_progress--;
343 spin_unlock(&resv->lock);
349 * Examine the existing reserve map and determine how many
350 * huge pages in the specified range [f, t) are NOT currently
351 * represented. This routine is called before a subsequent
352 * call to region_add that will actually modify the reserve
353 * map to add the specified range [f, t). region_chg does
354 * not change the number of huge pages represented by the
355 * map. However, if the existing regions in the map can not
356 * be expanded to represent the new range, a new file_region
357 * structure is added to the map as a placeholder. This is
358 * so that the subsequent region_add call will have all the
359 * regions it needs and will not fail.
361 * Upon entry, region_chg will also examine the cache of region descriptors
362 * associated with the map. If there are not enough descriptors cached, one
363 * will be allocated for the in progress add operation.
365 * Returns the number of huge pages that need to be added to the existing
366 * reservation map for the range [f, t). This number is greater or equal to
367 * zero. -ENOMEM is returned if a new file_region structure or cache entry
368 * is needed and can not be allocated.
370 static long region_chg(struct resv_map *resv, long f, long t)
372 struct list_head *head = &resv->regions;
373 struct file_region *rg, *nrg = NULL;
377 spin_lock(&resv->lock);
379 resv->adds_in_progress++;
382 * Check for sufficient descriptors in the cache to accommodate
383 * the number of in progress add operations.
385 if (resv->adds_in_progress > resv->region_cache_count) {
386 struct file_region *trg;
388 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
389 /* Must drop lock to allocate a new descriptor. */
390 resv->adds_in_progress--;
391 spin_unlock(&resv->lock);
393 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
399 spin_lock(&resv->lock);
400 list_add(&trg->link, &resv->region_cache);
401 resv->region_cache_count++;
405 /* Locate the region we are before or in. */
406 list_for_each_entry(rg, head, link)
410 /* If we are below the current region then a new region is required.
411 * Subtle, allocate a new region at the position but make it zero
412 * size such that we can guarantee to record the reservation. */
413 if (&rg->link == head || t < rg->from) {
415 resv->adds_in_progress--;
416 spin_unlock(&resv->lock);
417 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
423 INIT_LIST_HEAD(&nrg->link);
427 list_add(&nrg->link, rg->link.prev);
432 /* Round our left edge to the current segment if it encloses us. */
437 /* Check for and consume any regions we now overlap with. */
438 list_for_each_entry(rg, rg->link.prev, link) {
439 if (&rg->link == head)
444 /* We overlap with this area, if it extends further than
445 * us then we must extend ourselves. Account for its
446 * existing reservation. */
451 chg -= rg->to - rg->from;
455 spin_unlock(&resv->lock);
456 /* We already know we raced and no longer need the new region */
460 spin_unlock(&resv->lock);
465 * Abort the in progress add operation. The adds_in_progress field
466 * of the resv_map keeps track of the operations in progress between
467 * calls to region_chg and region_add. Operations are sometimes
468 * aborted after the call to region_chg. In such cases, region_abort
469 * is called to decrement the adds_in_progress counter.
471 * NOTE: The range arguments [f, t) are not needed or used in this
472 * routine. They are kept to make reading the calling code easier as
473 * arguments will match the associated region_chg call.
475 static void region_abort(struct resv_map *resv, long f, long t)
477 spin_lock(&resv->lock);
478 VM_BUG_ON(!resv->region_cache_count);
479 resv->adds_in_progress--;
480 spin_unlock(&resv->lock);
484 * Delete the specified range [f, t) from the reserve map. If the
485 * t parameter is LONG_MAX, this indicates that ALL regions after f
486 * should be deleted. Locate the regions which intersect [f, t)
487 * and either trim, delete or split the existing regions.
489 * Returns the number of huge pages deleted from the reserve map.
490 * In the normal case, the return value is zero or more. In the
491 * case where a region must be split, a new region descriptor must
492 * be allocated. If the allocation fails, -ENOMEM will be returned.
493 * NOTE: If the parameter t == LONG_MAX, then we will never split
494 * a region and possibly return -ENOMEM. Callers specifying
495 * t == LONG_MAX do not need to check for -ENOMEM error.
497 static long region_del(struct resv_map *resv, long f, long t)
499 struct list_head *head = &resv->regions;
500 struct file_region *rg, *trg;
501 struct file_region *nrg = NULL;
505 spin_lock(&resv->lock);
506 list_for_each_entry_safe(rg, trg, head, link) {
508 * Skip regions before the range to be deleted. file_region
509 * ranges are normally of the form [from, to). However, there
510 * may be a "placeholder" entry in the map which is of the form
511 * (from, to) with from == to. Check for placeholder entries
512 * at the beginning of the range to be deleted.
514 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
520 if (f > rg->from && t < rg->to) { /* Must split region */
522 * Check for an entry in the cache before dropping
523 * lock and attempting allocation.
526 resv->region_cache_count > resv->adds_in_progress) {
527 nrg = list_first_entry(&resv->region_cache,
530 list_del(&nrg->link);
531 resv->region_cache_count--;
535 spin_unlock(&resv->lock);
536 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
544 /* New entry for end of split region */
547 INIT_LIST_HEAD(&nrg->link);
549 /* Original entry is trimmed */
552 list_add(&nrg->link, &rg->link);
557 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
558 del += rg->to - rg->from;
564 if (f <= rg->from) { /* Trim beginning of region */
567 } else { /* Trim end of region */
573 spin_unlock(&resv->lock);
579 * A rare out of memory error was encountered which prevented removal of
580 * the reserve map region for a page. The huge page itself was free'ed
581 * and removed from the page cache. This routine will adjust the subpool
582 * usage count, and the global reserve count if needed. By incrementing
583 * these counts, the reserve map entry which could not be deleted will
584 * appear as a "reserved" entry instead of simply dangling with incorrect
587 void hugetlb_fix_reserve_counts(struct inode *inode)
589 struct hugepage_subpool *spool = subpool_inode(inode);
591 bool reserved = false;
593 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
594 if (rsv_adjust > 0) {
595 struct hstate *h = hstate_inode(inode);
597 if (!hugetlb_acct_memory(h, 1))
599 } else if (!rsv_adjust) {
604 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
608 * Count and return the number of huge pages in the reserve map
609 * that intersect with the range [f, t).
611 static long region_count(struct resv_map *resv, long f, long t)
613 struct list_head *head = &resv->regions;
614 struct file_region *rg;
617 spin_lock(&resv->lock);
618 /* Locate each segment we overlap with, and count that overlap. */
619 list_for_each_entry(rg, head, link) {
628 seg_from = max(rg->from, f);
629 seg_to = min(rg->to, t);
631 chg += seg_to - seg_from;
633 spin_unlock(&resv->lock);
639 * Convert the address within this vma to the page offset within
640 * the mapping, in pagecache page units; huge pages here.
642 static pgoff_t vma_hugecache_offset(struct hstate *h,
643 struct vm_area_struct *vma, unsigned long address)
645 return ((address - vma->vm_start) >> huge_page_shift(h)) +
646 (vma->vm_pgoff >> huge_page_order(h));
649 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
650 unsigned long address)
652 return vma_hugecache_offset(hstate_vma(vma), vma, address);
654 EXPORT_SYMBOL_GPL(linear_hugepage_index);
657 * Return the size of the pages allocated when backing a VMA. In the majority
658 * cases this will be same size as used by the page table entries.
660 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
662 if (vma->vm_ops && vma->vm_ops->pagesize)
663 return vma->vm_ops->pagesize(vma);
666 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
669 * Return the page size being used by the MMU to back a VMA. In the majority
670 * of cases, the page size used by the kernel matches the MMU size. On
671 * architectures where it differs, an architecture-specific 'strong'
672 * version of this symbol is required.
674 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
676 return vma_kernel_pagesize(vma);
680 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
681 * bits of the reservation map pointer, which are always clear due to
684 #define HPAGE_RESV_OWNER (1UL << 0)
685 #define HPAGE_RESV_UNMAPPED (1UL << 1)
686 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
689 * These helpers are used to track how many pages are reserved for
690 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
691 * is guaranteed to have their future faults succeed.
693 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
694 * the reserve counters are updated with the hugetlb_lock held. It is safe
695 * to reset the VMA at fork() time as it is not in use yet and there is no
696 * chance of the global counters getting corrupted as a result of the values.
698 * The private mapping reservation is represented in a subtly different
699 * manner to a shared mapping. A shared mapping has a region map associated
700 * with the underlying file, this region map represents the backing file
701 * pages which have ever had a reservation assigned which this persists even
702 * after the page is instantiated. A private mapping has a region map
703 * associated with the original mmap which is attached to all VMAs which
704 * reference it, this region map represents those offsets which have consumed
705 * reservation ie. where pages have been instantiated.
707 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
709 return (unsigned long)vma->vm_private_data;
712 static void set_vma_private_data(struct vm_area_struct *vma,
715 vma->vm_private_data = (void *)value;
718 struct resv_map *resv_map_alloc(void)
720 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
721 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
723 if (!resv_map || !rg) {
729 kref_init(&resv_map->refs);
730 spin_lock_init(&resv_map->lock);
731 INIT_LIST_HEAD(&resv_map->regions);
733 resv_map->adds_in_progress = 0;
735 INIT_LIST_HEAD(&resv_map->region_cache);
736 list_add(&rg->link, &resv_map->region_cache);
737 resv_map->region_cache_count = 1;
742 void resv_map_release(struct kref *ref)
744 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
745 struct list_head *head = &resv_map->region_cache;
746 struct file_region *rg, *trg;
748 /* Clear out any active regions before we release the map. */
749 region_del(resv_map, 0, LONG_MAX);
751 /* ... and any entries left in the cache */
752 list_for_each_entry_safe(rg, trg, head, link) {
757 VM_BUG_ON(resv_map->adds_in_progress);
762 static inline struct resv_map *inode_resv_map(struct inode *inode)
764 return inode->i_mapping->private_data;
767 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
769 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
770 if (vma->vm_flags & VM_MAYSHARE) {
771 struct address_space *mapping = vma->vm_file->f_mapping;
772 struct inode *inode = mapping->host;
774 return inode_resv_map(inode);
777 return (struct resv_map *)(get_vma_private_data(vma) &
782 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
784 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
785 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
787 set_vma_private_data(vma, (get_vma_private_data(vma) &
788 HPAGE_RESV_MASK) | (unsigned long)map);
791 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
793 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
794 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
796 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
799 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
801 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
803 return (get_vma_private_data(vma) & flag) != 0;
806 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
807 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
809 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
810 if (!(vma->vm_flags & VM_MAYSHARE))
811 vma->vm_private_data = (void *)0;
814 /* Returns true if the VMA has associated reserve pages */
815 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
817 if (vma->vm_flags & VM_NORESERVE) {
819 * This address is already reserved by other process(chg == 0),
820 * so, we should decrement reserved count. Without decrementing,
821 * reserve count remains after releasing inode, because this
822 * allocated page will go into page cache and is regarded as
823 * coming from reserved pool in releasing step. Currently, we
824 * don't have any other solution to deal with this situation
825 * properly, so add work-around here.
827 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
833 /* Shared mappings always use reserves */
834 if (vma->vm_flags & VM_MAYSHARE) {
836 * We know VM_NORESERVE is not set. Therefore, there SHOULD
837 * be a region map for all pages. The only situation where
838 * there is no region map is if a hole was punched via
839 * fallocate. In this case, there really are no reverves to
840 * use. This situation is indicated if chg != 0.
849 * Only the process that called mmap() has reserves for
852 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
854 * Like the shared case above, a hole punch or truncate
855 * could have been performed on the private mapping.
856 * Examine the value of chg to determine if reserves
857 * actually exist or were previously consumed.
858 * Very Subtle - The value of chg comes from a previous
859 * call to vma_needs_reserves(). The reserve map for
860 * private mappings has different (opposite) semantics
861 * than that of shared mappings. vma_needs_reserves()
862 * has already taken this difference in semantics into
863 * account. Therefore, the meaning of chg is the same
864 * as in the shared case above. Code could easily be
865 * combined, but keeping it separate draws attention to
866 * subtle differences.
877 static void enqueue_huge_page(struct hstate *h, struct page *page)
879 int nid = page_to_nid(page);
880 list_move(&page->lru, &h->hugepage_freelists[nid]);
881 h->free_huge_pages++;
882 h->free_huge_pages_node[nid]++;
883 SetPageHugeFreed(page);
886 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
890 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
891 if (!PageHWPoison(page))
894 * if 'non-isolated free hugepage' not found on the list,
895 * the allocation fails.
897 if (&h->hugepage_freelists[nid] == &page->lru)
899 list_move(&page->lru, &h->hugepage_activelist);
900 set_page_refcounted(page);
901 ClearPageHugeFreed(page);
902 h->free_huge_pages--;
903 h->free_huge_pages_node[nid]--;
907 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
910 unsigned int cpuset_mems_cookie;
911 struct zonelist *zonelist;
916 zonelist = node_zonelist(nid, gfp_mask);
919 cpuset_mems_cookie = read_mems_allowed_begin();
920 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
923 if (!cpuset_zone_allowed(zone, gfp_mask))
926 * no need to ask again on the same node. Pool is node rather than
929 if (zone_to_nid(zone) == node)
931 node = zone_to_nid(zone);
933 page = dequeue_huge_page_node_exact(h, node);
937 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
943 /* Movability of hugepages depends on migration support. */
944 static inline gfp_t htlb_alloc_mask(struct hstate *h)
946 if (hugepage_migration_supported(h))
947 return GFP_HIGHUSER_MOVABLE;
952 static struct page *dequeue_huge_page_vma(struct hstate *h,
953 struct vm_area_struct *vma,
954 unsigned long address, int avoid_reserve,
958 struct mempolicy *mpol;
960 nodemask_t *nodemask;
964 * A child process with MAP_PRIVATE mappings created by their parent
965 * have no page reserves. This check ensures that reservations are
966 * not "stolen". The child may still get SIGKILLed
968 if (!vma_has_reserves(vma, chg) &&
969 h->free_huge_pages - h->resv_huge_pages == 0)
972 /* If reserves cannot be used, ensure enough pages are in the pool */
973 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
976 gfp_mask = htlb_alloc_mask(h);
977 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
978 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
979 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
980 SetPagePrivate(page);
981 h->resv_huge_pages--;
992 * common helper functions for hstate_next_node_to_{alloc|free}.
993 * We may have allocated or freed a huge page based on a different
994 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
995 * be outside of *nodes_allowed. Ensure that we use an allowed
996 * node for alloc or free.
998 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1000 nid = next_node_in(nid, *nodes_allowed);
1001 VM_BUG_ON(nid >= MAX_NUMNODES);
1006 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1008 if (!node_isset(nid, *nodes_allowed))
1009 nid = next_node_allowed(nid, nodes_allowed);
1014 * returns the previously saved node ["this node"] from which to
1015 * allocate a persistent huge page for the pool and advance the
1016 * next node from which to allocate, handling wrap at end of node
1019 static int hstate_next_node_to_alloc(struct hstate *h,
1020 nodemask_t *nodes_allowed)
1024 VM_BUG_ON(!nodes_allowed);
1026 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1027 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1033 * helper for free_pool_huge_page() - return the previously saved
1034 * node ["this node"] from which to free a huge page. Advance the
1035 * next node id whether or not we find a free huge page to free so
1036 * that the next attempt to free addresses the next node.
1038 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1042 VM_BUG_ON(!nodes_allowed);
1044 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1045 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1050 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1051 for (nr_nodes = nodes_weight(*mask); \
1053 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1056 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1057 for (nr_nodes = nodes_weight(*mask); \
1059 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1062 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1063 static void destroy_compound_gigantic_page(struct page *page,
1067 int nr_pages = 1 << order;
1068 struct page *p = page + 1;
1070 atomic_set(compound_mapcount_ptr(page), 0);
1071 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1072 clear_compound_head(p);
1073 set_page_refcounted(p);
1076 set_compound_order(page, 0);
1077 __ClearPageHead(page);
1080 static void free_gigantic_page(struct page *page, unsigned int order)
1082 free_contig_range(page_to_pfn(page), 1 << order);
1085 static int __alloc_gigantic_page(unsigned long start_pfn,
1086 unsigned long nr_pages, gfp_t gfp_mask)
1088 unsigned long end_pfn = start_pfn + nr_pages;
1089 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1093 static bool pfn_range_valid_gigantic(struct zone *z,
1094 unsigned long start_pfn, unsigned long nr_pages)
1096 unsigned long i, end_pfn = start_pfn + nr_pages;
1099 for (i = start_pfn; i < end_pfn; i++) {
1100 page = pfn_to_online_page(i);
1104 if (page_zone(page) != z)
1107 if (PageReserved(page))
1110 if (page_count(page) > 0)
1120 static bool zone_spans_last_pfn(const struct zone *zone,
1121 unsigned long start_pfn, unsigned long nr_pages)
1123 unsigned long last_pfn = start_pfn + nr_pages - 1;
1124 return zone_spans_pfn(zone, last_pfn);
1127 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1128 int nid, nodemask_t *nodemask)
1130 unsigned int order = huge_page_order(h);
1131 unsigned long nr_pages = 1 << order;
1132 unsigned long ret, pfn, flags;
1133 struct zonelist *zonelist;
1137 zonelist = node_zonelist(nid, gfp_mask);
1138 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) {
1139 spin_lock_irqsave(&zone->lock, flags);
1141 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
1142 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
1143 if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) {
1145 * We release the zone lock here because
1146 * alloc_contig_range() will also lock the zone
1147 * at some point. If there's an allocation
1148 * spinning on this lock, it may win the race
1149 * and cause alloc_contig_range() to fail...
1151 spin_unlock_irqrestore(&zone->lock, flags);
1152 ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask);
1154 return pfn_to_page(pfn);
1155 spin_lock_irqsave(&zone->lock, flags);
1160 spin_unlock_irqrestore(&zone->lock, flags);
1166 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1167 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1169 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1170 static inline bool gigantic_page_supported(void) { return false; }
1171 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1172 int nid, nodemask_t *nodemask) { return NULL; }
1173 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1174 static inline void destroy_compound_gigantic_page(struct page *page,
1175 unsigned int order) { }
1178 static void update_and_free_page(struct hstate *h, struct page *page)
1181 struct page *subpage = page;
1183 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1187 h->nr_huge_pages_node[page_to_nid(page)]--;
1188 for (i = 0; i < pages_per_huge_page(h);
1189 i++, subpage = mem_map_next(subpage, page, i)) {
1190 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1191 1 << PG_referenced | 1 << PG_dirty |
1192 1 << PG_active | 1 << PG_private |
1195 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1196 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1197 set_page_refcounted(page);
1198 if (hstate_is_gigantic(h)) {
1199 destroy_compound_gigantic_page(page, huge_page_order(h));
1200 free_gigantic_page(page, huge_page_order(h));
1202 __free_pages(page, huge_page_order(h));
1206 struct hstate *size_to_hstate(unsigned long size)
1210 for_each_hstate(h) {
1211 if (huge_page_size(h) == size)
1218 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1219 * to hstate->hugepage_activelist.)
1221 * This function can be called for tail pages, but never returns true for them.
1223 bool page_huge_active(struct page *page)
1225 return PageHeadHuge(page) && PagePrivate(&page[1]);
1228 /* never called for tail page */
1229 void set_page_huge_active(struct page *page)
1231 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1232 SetPagePrivate(&page[1]);
1235 static void clear_page_huge_active(struct page *page)
1237 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1238 ClearPagePrivate(&page[1]);
1242 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1245 static inline bool PageHugeTemporary(struct page *page)
1247 if (!PageHuge(page))
1250 return (unsigned long)page[2].mapping == -1U;
1253 static inline void SetPageHugeTemporary(struct page *page)
1255 page[2].mapping = (void *)-1U;
1258 static inline void ClearPageHugeTemporary(struct page *page)
1260 page[2].mapping = NULL;
1263 void free_huge_page(struct page *page)
1266 * Can't pass hstate in here because it is called from the
1267 * compound page destructor.
1269 struct hstate *h = page_hstate(page);
1270 int nid = page_to_nid(page);
1271 struct hugepage_subpool *spool =
1272 (struct hugepage_subpool *)page_private(page);
1273 bool restore_reserve;
1275 set_page_private(page, 0);
1276 page->mapping = NULL;
1277 VM_BUG_ON_PAGE(page_count(page), page);
1278 VM_BUG_ON_PAGE(page_mapcount(page), page);
1279 restore_reserve = PagePrivate(page);
1280 ClearPagePrivate(page);
1283 * If PagePrivate() was set on page, page allocation consumed a
1284 * reservation. If the page was associated with a subpool, there
1285 * would have been a page reserved in the subpool before allocation
1286 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1287 * reservtion, do not call hugepage_subpool_put_pages() as this will
1288 * remove the reserved page from the subpool.
1290 if (!restore_reserve) {
1292 * A return code of zero implies that the subpool will be
1293 * under its minimum size if the reservation is not restored
1294 * after page is free. Therefore, force restore_reserve
1297 if (hugepage_subpool_put_pages(spool, 1) == 0)
1298 restore_reserve = true;
1301 spin_lock(&hugetlb_lock);
1302 clear_page_huge_active(page);
1303 hugetlb_cgroup_uncharge_page(hstate_index(h),
1304 pages_per_huge_page(h), page);
1305 if (restore_reserve)
1306 h->resv_huge_pages++;
1308 if (PageHugeTemporary(page)) {
1309 list_del(&page->lru);
1310 ClearPageHugeTemporary(page);
1311 update_and_free_page(h, page);
1312 } else if (h->surplus_huge_pages_node[nid]) {
1313 /* remove the page from active list */
1314 list_del(&page->lru);
1315 update_and_free_page(h, page);
1316 h->surplus_huge_pages--;
1317 h->surplus_huge_pages_node[nid]--;
1319 arch_clear_hugepage_flags(page);
1320 enqueue_huge_page(h, page);
1322 spin_unlock(&hugetlb_lock);
1325 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1327 INIT_LIST_HEAD(&page->lru);
1328 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1329 spin_lock(&hugetlb_lock);
1330 set_hugetlb_cgroup(page, NULL);
1332 h->nr_huge_pages_node[nid]++;
1333 ClearPageHugeFreed(page);
1334 spin_unlock(&hugetlb_lock);
1337 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1340 int nr_pages = 1 << order;
1341 struct page *p = page + 1;
1343 /* we rely on prep_new_huge_page to set the destructor */
1344 set_compound_order(page, order);
1345 __ClearPageReserved(page);
1346 __SetPageHead(page);
1347 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1349 * For gigantic hugepages allocated through bootmem at
1350 * boot, it's safer to be consistent with the not-gigantic
1351 * hugepages and clear the PG_reserved bit from all tail pages
1352 * too. Otherwse drivers using get_user_pages() to access tail
1353 * pages may get the reference counting wrong if they see
1354 * PG_reserved set on a tail page (despite the head page not
1355 * having PG_reserved set). Enforcing this consistency between
1356 * head and tail pages allows drivers to optimize away a check
1357 * on the head page when they need know if put_page() is needed
1358 * after get_user_pages().
1360 __ClearPageReserved(p);
1361 set_page_count(p, 0);
1362 set_compound_head(p, page);
1364 atomic_set(compound_mapcount_ptr(page), -1);
1368 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1369 * transparent huge pages. See the PageTransHuge() documentation for more
1372 int PageHuge(struct page *page)
1374 if (!PageCompound(page))
1377 page = compound_head(page);
1378 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1380 EXPORT_SYMBOL_GPL(PageHuge);
1383 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1384 * normal or transparent huge pages.
1386 int PageHeadHuge(struct page *page_head)
1388 if (!PageHead(page_head))
1391 return get_compound_page_dtor(page_head) == free_huge_page;
1394 pgoff_t hugetlb_basepage_index(struct page *page)
1396 struct page *page_head = compound_head(page);
1397 pgoff_t index = page_index(page_head);
1398 unsigned long compound_idx;
1400 if (compound_order(page_head) >= MAX_ORDER)
1401 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1403 compound_idx = page - page_head;
1405 return (index << compound_order(page_head)) + compound_idx;
1408 static struct page *alloc_buddy_huge_page(struct hstate *h,
1409 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1411 int order = huge_page_order(h);
1414 gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
1415 if (nid == NUMA_NO_NODE)
1416 nid = numa_mem_id();
1417 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1419 __count_vm_event(HTLB_BUDDY_PGALLOC);
1421 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1427 * Common helper to allocate a fresh hugetlb page. All specific allocators
1428 * should use this function to get new hugetlb pages
1430 static struct page *alloc_fresh_huge_page(struct hstate *h,
1431 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1435 if (hstate_is_gigantic(h))
1436 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1438 page = alloc_buddy_huge_page(h, gfp_mask,
1443 if (hstate_is_gigantic(h))
1444 prep_compound_gigantic_page(page, huge_page_order(h));
1445 prep_new_huge_page(h, page, page_to_nid(page));
1451 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1454 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1458 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1460 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1461 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed);
1469 put_page(page); /* free it into the hugepage allocator */
1475 * Free huge page from pool from next node to free.
1476 * Attempt to keep persistent huge pages more or less
1477 * balanced over allowed nodes.
1478 * Called with hugetlb_lock locked.
1480 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1486 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1488 * If we're returning unused surplus pages, only examine
1489 * nodes with surplus pages.
1491 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1492 !list_empty(&h->hugepage_freelists[node])) {
1494 list_entry(h->hugepage_freelists[node].next,
1496 list_del(&page->lru);
1497 h->free_huge_pages--;
1498 h->free_huge_pages_node[node]--;
1500 h->surplus_huge_pages--;
1501 h->surplus_huge_pages_node[node]--;
1503 update_and_free_page(h, page);
1513 * Dissolve a given free hugepage into free buddy pages. This function does
1514 * nothing for in-use hugepages and non-hugepages.
1515 * This function returns values like below:
1517 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1518 * (allocated or reserved.)
1519 * 0: successfully dissolved free hugepages or the page is not a
1520 * hugepage (considered as already dissolved)
1522 int dissolve_free_huge_page(struct page *page)
1527 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1528 if (!PageHuge(page))
1531 spin_lock(&hugetlb_lock);
1532 if (!PageHuge(page)) {
1537 if (!page_count(page)) {
1538 struct page *head = compound_head(page);
1539 struct hstate *h = page_hstate(head);
1540 int nid = page_to_nid(head);
1541 if (h->free_huge_pages - h->resv_huge_pages == 0)
1545 * We should make sure that the page is already on the free list
1546 * when it is dissolved.
1548 if (unlikely(!PageHugeFreed(head))) {
1549 spin_unlock(&hugetlb_lock);
1553 * Theoretically, we should return -EBUSY when we
1554 * encounter this race. In fact, we have a chance
1555 * to successfully dissolve the page if we do a
1556 * retry. Because the race window is quite small.
1557 * If we seize this opportunity, it is an optimization
1558 * for increasing the success rate of dissolving page.
1564 * Move PageHWPoison flag from head page to the raw error page,
1565 * which makes any subpages rather than the error page reusable.
1567 if (PageHWPoison(head) && page != head) {
1568 SetPageHWPoison(page);
1569 ClearPageHWPoison(head);
1571 list_del(&head->lru);
1572 h->free_huge_pages--;
1573 h->free_huge_pages_node[nid]--;
1574 h->max_huge_pages--;
1575 update_and_free_page(h, head);
1579 spin_unlock(&hugetlb_lock);
1584 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1585 * make specified memory blocks removable from the system.
1586 * Note that this will dissolve a free gigantic hugepage completely, if any
1587 * part of it lies within the given range.
1588 * Also note that if dissolve_free_huge_page() returns with an error, all
1589 * free hugepages that were dissolved before that error are lost.
1591 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1597 if (!hugepages_supported())
1600 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1601 page = pfn_to_page(pfn);
1602 rc = dissolve_free_huge_page(page);
1611 * Allocates a fresh surplus page from the page allocator.
1613 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1614 int nid, nodemask_t *nmask)
1616 struct page *page = NULL;
1618 if (hstate_is_gigantic(h))
1621 spin_lock(&hugetlb_lock);
1622 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1624 spin_unlock(&hugetlb_lock);
1626 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1630 spin_lock(&hugetlb_lock);
1632 * We could have raced with the pool size change.
1633 * Double check that and simply deallocate the new page
1634 * if we would end up overcommiting the surpluses. Abuse
1635 * temporary page to workaround the nasty free_huge_page
1638 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1639 SetPageHugeTemporary(page);
1640 spin_unlock(&hugetlb_lock);
1644 h->surplus_huge_pages++;
1645 h->surplus_huge_pages_node[page_to_nid(page)]++;
1649 spin_unlock(&hugetlb_lock);
1654 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1655 int nid, nodemask_t *nmask)
1659 if (hstate_is_gigantic(h))
1662 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1667 * We do not account these pages as surplus because they are only
1668 * temporary and will be released properly on the last reference
1670 SetPageHugeTemporary(page);
1676 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1679 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1680 struct vm_area_struct *vma, unsigned long addr)
1683 struct mempolicy *mpol;
1684 gfp_t gfp_mask = htlb_alloc_mask(h);
1686 nodemask_t *nodemask;
1688 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1689 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1690 mpol_cond_put(mpol);
1695 /* page migration callback function */
1696 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1698 gfp_t gfp_mask = htlb_alloc_mask(h);
1699 struct page *page = NULL;
1701 if (nid != NUMA_NO_NODE)
1702 gfp_mask |= __GFP_THISNODE;
1704 spin_lock(&hugetlb_lock);
1705 if (h->free_huge_pages - h->resv_huge_pages > 0)
1706 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1707 spin_unlock(&hugetlb_lock);
1710 page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1715 /* page migration callback function */
1716 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1719 gfp_t gfp_mask = htlb_alloc_mask(h);
1721 spin_lock(&hugetlb_lock);
1722 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1725 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1727 spin_unlock(&hugetlb_lock);
1731 spin_unlock(&hugetlb_lock);
1733 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1736 /* mempolicy aware migration callback */
1737 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1738 unsigned long address)
1740 struct mempolicy *mpol;
1741 nodemask_t *nodemask;
1746 gfp_mask = htlb_alloc_mask(h);
1747 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1748 page = alloc_huge_page_nodemask(h, node, nodemask);
1749 mpol_cond_put(mpol);
1755 * Increase the hugetlb pool such that it can accommodate a reservation
1758 static int gather_surplus_pages(struct hstate *h, int delta)
1760 struct list_head surplus_list;
1761 struct page *page, *tmp;
1763 int needed, allocated;
1764 bool alloc_ok = true;
1766 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1768 h->resv_huge_pages += delta;
1773 INIT_LIST_HEAD(&surplus_list);
1777 spin_unlock(&hugetlb_lock);
1778 for (i = 0; i < needed; i++) {
1779 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1780 NUMA_NO_NODE, NULL);
1785 list_add(&page->lru, &surplus_list);
1791 * After retaking hugetlb_lock, we need to recalculate 'needed'
1792 * because either resv_huge_pages or free_huge_pages may have changed.
1794 spin_lock(&hugetlb_lock);
1795 needed = (h->resv_huge_pages + delta) -
1796 (h->free_huge_pages + allocated);
1801 * We were not able to allocate enough pages to
1802 * satisfy the entire reservation so we free what
1803 * we've allocated so far.
1808 * The surplus_list now contains _at_least_ the number of extra pages
1809 * needed to accommodate the reservation. Add the appropriate number
1810 * of pages to the hugetlb pool and free the extras back to the buddy
1811 * allocator. Commit the entire reservation here to prevent another
1812 * process from stealing the pages as they are added to the pool but
1813 * before they are reserved.
1815 needed += allocated;
1816 h->resv_huge_pages += delta;
1819 /* Free the needed pages to the hugetlb pool */
1820 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1824 * This page is now managed by the hugetlb allocator and has
1825 * no users -- drop the buddy allocator's reference.
1827 put_page_testzero(page);
1828 VM_BUG_ON_PAGE(page_count(page), page);
1829 enqueue_huge_page(h, page);
1832 spin_unlock(&hugetlb_lock);
1834 /* Free unnecessary surplus pages to the buddy allocator */
1835 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1837 spin_lock(&hugetlb_lock);
1843 * This routine has two main purposes:
1844 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1845 * in unused_resv_pages. This corresponds to the prior adjustments made
1846 * to the associated reservation map.
1847 * 2) Free any unused surplus pages that may have been allocated to satisfy
1848 * the reservation. As many as unused_resv_pages may be freed.
1850 * Called with hugetlb_lock held. However, the lock could be dropped (and
1851 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1852 * we must make sure nobody else can claim pages we are in the process of
1853 * freeing. Do this by ensuring resv_huge_page always is greater than the
1854 * number of huge pages we plan to free when dropping the lock.
1856 static void return_unused_surplus_pages(struct hstate *h,
1857 unsigned long unused_resv_pages)
1859 unsigned long nr_pages;
1861 /* Cannot return gigantic pages currently */
1862 if (hstate_is_gigantic(h))
1866 * Part (or even all) of the reservation could have been backed
1867 * by pre-allocated pages. Only free surplus pages.
1869 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1872 * We want to release as many surplus pages as possible, spread
1873 * evenly across all nodes with memory. Iterate across these nodes
1874 * until we can no longer free unreserved surplus pages. This occurs
1875 * when the nodes with surplus pages have no free pages.
1876 * free_pool_huge_page() will balance the the freed pages across the
1877 * on-line nodes with memory and will handle the hstate accounting.
1879 * Note that we decrement resv_huge_pages as we free the pages. If
1880 * we drop the lock, resv_huge_pages will still be sufficiently large
1881 * to cover subsequent pages we may free.
1883 while (nr_pages--) {
1884 h->resv_huge_pages--;
1885 unused_resv_pages--;
1886 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1888 cond_resched_lock(&hugetlb_lock);
1892 /* Fully uncommit the reservation */
1893 h->resv_huge_pages -= unused_resv_pages;
1898 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1899 * are used by the huge page allocation routines to manage reservations.
1901 * vma_needs_reservation is called to determine if the huge page at addr
1902 * within the vma has an associated reservation. If a reservation is
1903 * needed, the value 1 is returned. The caller is then responsible for
1904 * managing the global reservation and subpool usage counts. After
1905 * the huge page has been allocated, vma_commit_reservation is called
1906 * to add the page to the reservation map. If the page allocation fails,
1907 * the reservation must be ended instead of committed. vma_end_reservation
1908 * is called in such cases.
1910 * In the normal case, vma_commit_reservation returns the same value
1911 * as the preceding vma_needs_reservation call. The only time this
1912 * is not the case is if a reserve map was changed between calls. It
1913 * is the responsibility of the caller to notice the difference and
1914 * take appropriate action.
1916 * vma_add_reservation is used in error paths where a reservation must
1917 * be restored when a newly allocated huge page must be freed. It is
1918 * to be called after calling vma_needs_reservation to determine if a
1919 * reservation exists.
1921 enum vma_resv_mode {
1927 static long __vma_reservation_common(struct hstate *h,
1928 struct vm_area_struct *vma, unsigned long addr,
1929 enum vma_resv_mode mode)
1931 struct resv_map *resv;
1935 resv = vma_resv_map(vma);
1939 idx = vma_hugecache_offset(h, vma, addr);
1941 case VMA_NEEDS_RESV:
1942 ret = region_chg(resv, idx, idx + 1);
1944 case VMA_COMMIT_RESV:
1945 ret = region_add(resv, idx, idx + 1);
1948 region_abort(resv, idx, idx + 1);
1952 if (vma->vm_flags & VM_MAYSHARE)
1953 ret = region_add(resv, idx, idx + 1);
1955 region_abort(resv, idx, idx + 1);
1956 ret = region_del(resv, idx, idx + 1);
1963 if (vma->vm_flags & VM_MAYSHARE)
1965 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1967 * In most cases, reserves always exist for private mappings.
1968 * However, a file associated with mapping could have been
1969 * hole punched or truncated after reserves were consumed.
1970 * As subsequent fault on such a range will not use reserves.
1971 * Subtle - The reserve map for private mappings has the
1972 * opposite meaning than that of shared mappings. If NO
1973 * entry is in the reserve map, it means a reservation exists.
1974 * If an entry exists in the reserve map, it means the
1975 * reservation has already been consumed. As a result, the
1976 * return value of this routine is the opposite of the
1977 * value returned from reserve map manipulation routines above.
1985 return ret < 0 ? ret : 0;
1988 static long vma_needs_reservation(struct hstate *h,
1989 struct vm_area_struct *vma, unsigned long addr)
1991 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1994 static long vma_commit_reservation(struct hstate *h,
1995 struct vm_area_struct *vma, unsigned long addr)
1997 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2000 static void vma_end_reservation(struct hstate *h,
2001 struct vm_area_struct *vma, unsigned long addr)
2003 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2006 static long vma_add_reservation(struct hstate *h,
2007 struct vm_area_struct *vma, unsigned long addr)
2009 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2013 * This routine is called to restore a reservation on error paths. In the
2014 * specific error paths, a huge page was allocated (via alloc_huge_page)
2015 * and is about to be freed. If a reservation for the page existed,
2016 * alloc_huge_page would have consumed the reservation and set PagePrivate
2017 * in the newly allocated page. When the page is freed via free_huge_page,
2018 * the global reservation count will be incremented if PagePrivate is set.
2019 * However, free_huge_page can not adjust the reserve map. Adjust the
2020 * reserve map here to be consistent with global reserve count adjustments
2021 * to be made by free_huge_page.
2023 static void restore_reserve_on_error(struct hstate *h,
2024 struct vm_area_struct *vma, unsigned long address,
2027 if (unlikely(PagePrivate(page))) {
2028 long rc = vma_needs_reservation(h, vma, address);
2030 if (unlikely(rc < 0)) {
2032 * Rare out of memory condition in reserve map
2033 * manipulation. Clear PagePrivate so that
2034 * global reserve count will not be incremented
2035 * by free_huge_page. This will make it appear
2036 * as though the reservation for this page was
2037 * consumed. This may prevent the task from
2038 * faulting in the page at a later time. This
2039 * is better than inconsistent global huge page
2040 * accounting of reserve counts.
2042 ClearPagePrivate(page);
2044 rc = vma_add_reservation(h, vma, address);
2045 if (unlikely(rc < 0))
2047 * See above comment about rare out of
2050 ClearPagePrivate(page);
2052 vma_end_reservation(h, vma, address);
2056 struct page *alloc_huge_page(struct vm_area_struct *vma,
2057 unsigned long addr, int avoid_reserve)
2059 struct hugepage_subpool *spool = subpool_vma(vma);
2060 struct hstate *h = hstate_vma(vma);
2062 long map_chg, map_commit;
2065 struct hugetlb_cgroup *h_cg;
2067 idx = hstate_index(h);
2069 * Examine the region/reserve map to determine if the process
2070 * has a reservation for the page to be allocated. A return
2071 * code of zero indicates a reservation exists (no change).
2073 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2075 return ERR_PTR(-ENOMEM);
2078 * Processes that did not create the mapping will have no
2079 * reserves as indicated by the region/reserve map. Check
2080 * that the allocation will not exceed the subpool limit.
2081 * Allocations for MAP_NORESERVE mappings also need to be
2082 * checked against any subpool limit.
2084 if (map_chg || avoid_reserve) {
2085 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2087 vma_end_reservation(h, vma, addr);
2088 return ERR_PTR(-ENOSPC);
2092 * Even though there was no reservation in the region/reserve
2093 * map, there could be reservations associated with the
2094 * subpool that can be used. This would be indicated if the
2095 * return value of hugepage_subpool_get_pages() is zero.
2096 * However, if avoid_reserve is specified we still avoid even
2097 * the subpool reservations.
2103 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2105 goto out_subpool_put;
2107 spin_lock(&hugetlb_lock);
2109 * glb_chg is passed to indicate whether or not a page must be taken
2110 * from the global free pool (global change). gbl_chg == 0 indicates
2111 * a reservation exists for the allocation.
2113 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2115 spin_unlock(&hugetlb_lock);
2116 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2118 goto out_uncharge_cgroup;
2119 spin_lock(&hugetlb_lock);
2120 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2121 SetPagePrivate(page);
2122 h->resv_huge_pages--;
2124 list_move(&page->lru, &h->hugepage_activelist);
2127 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2128 spin_unlock(&hugetlb_lock);
2130 set_page_private(page, (unsigned long)spool);
2132 map_commit = vma_commit_reservation(h, vma, addr);
2133 if (unlikely(map_chg > map_commit)) {
2135 * The page was added to the reservation map between
2136 * vma_needs_reservation and vma_commit_reservation.
2137 * This indicates a race with hugetlb_reserve_pages.
2138 * Adjust for the subpool count incremented above AND
2139 * in hugetlb_reserve_pages for the same page. Also,
2140 * the reservation count added in hugetlb_reserve_pages
2141 * no longer applies.
2145 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2146 hugetlb_acct_memory(h, -rsv_adjust);
2150 out_uncharge_cgroup:
2151 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2153 if (map_chg || avoid_reserve)
2154 hugepage_subpool_put_pages(spool, 1);
2155 vma_end_reservation(h, vma, addr);
2156 return ERR_PTR(-ENOSPC);
2159 int alloc_bootmem_huge_page(struct hstate *h)
2160 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2161 int __alloc_bootmem_huge_page(struct hstate *h)
2163 struct huge_bootmem_page *m;
2166 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2169 addr = memblock_virt_alloc_try_nid_raw(
2170 huge_page_size(h), huge_page_size(h),
2171 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
2174 * Use the beginning of the huge page to store the
2175 * huge_bootmem_page struct (until gather_bootmem
2176 * puts them into the mem_map).
2185 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2186 /* Put them into a private list first because mem_map is not up yet */
2187 INIT_LIST_HEAD(&m->list);
2188 list_add(&m->list, &huge_boot_pages);
2193 static void __init prep_compound_huge_page(struct page *page,
2196 if (unlikely(order > (MAX_ORDER - 1)))
2197 prep_compound_gigantic_page(page, order);
2199 prep_compound_page(page, order);
2202 /* Put bootmem huge pages into the standard lists after mem_map is up */
2203 static void __init gather_bootmem_prealloc(void)
2205 struct huge_bootmem_page *m;
2207 list_for_each_entry(m, &huge_boot_pages, list) {
2208 struct page *page = virt_to_page(m);
2209 struct hstate *h = m->hstate;
2211 WARN_ON(page_count(page) != 1);
2212 prep_compound_huge_page(page, h->order);
2213 WARN_ON(PageReserved(page));
2214 prep_new_huge_page(h, page, page_to_nid(page));
2215 put_page(page); /* free it into the hugepage allocator */
2218 * If we had gigantic hugepages allocated at boot time, we need
2219 * to restore the 'stolen' pages to totalram_pages in order to
2220 * fix confusing memory reports from free(1) and another
2221 * side-effects, like CommitLimit going negative.
2223 if (hstate_is_gigantic(h))
2224 adjust_managed_page_count(page, 1 << h->order);
2229 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2233 for (i = 0; i < h->max_huge_pages; ++i) {
2234 if (hstate_is_gigantic(h)) {
2235 if (!alloc_bootmem_huge_page(h))
2237 } else if (!alloc_pool_huge_page(h,
2238 &node_states[N_MEMORY]))
2242 if (i < h->max_huge_pages) {
2245 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2246 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2247 h->max_huge_pages, buf, i);
2248 h->max_huge_pages = i;
2252 static void __init hugetlb_init_hstates(void)
2256 for_each_hstate(h) {
2257 if (minimum_order > huge_page_order(h))
2258 minimum_order = huge_page_order(h);
2260 /* oversize hugepages were init'ed in early boot */
2261 if (!hstate_is_gigantic(h))
2262 hugetlb_hstate_alloc_pages(h);
2264 VM_BUG_ON(minimum_order == UINT_MAX);
2267 static void __init report_hugepages(void)
2271 for_each_hstate(h) {
2274 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2275 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2276 buf, h->free_huge_pages);
2280 #ifdef CONFIG_HIGHMEM
2281 static void try_to_free_low(struct hstate *h, unsigned long count,
2282 nodemask_t *nodes_allowed)
2286 if (hstate_is_gigantic(h))
2289 for_each_node_mask(i, *nodes_allowed) {
2290 struct page *page, *next;
2291 struct list_head *freel = &h->hugepage_freelists[i];
2292 list_for_each_entry_safe(page, next, freel, lru) {
2293 if (count >= h->nr_huge_pages)
2295 if (PageHighMem(page))
2297 list_del(&page->lru);
2298 update_and_free_page(h, page);
2299 h->free_huge_pages--;
2300 h->free_huge_pages_node[page_to_nid(page)]--;
2305 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2306 nodemask_t *nodes_allowed)
2312 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2313 * balanced by operating on them in a round-robin fashion.
2314 * Returns 1 if an adjustment was made.
2316 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2321 VM_BUG_ON(delta != -1 && delta != 1);
2324 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2325 if (h->surplus_huge_pages_node[node])
2329 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2330 if (h->surplus_huge_pages_node[node] <
2331 h->nr_huge_pages_node[node])
2338 h->surplus_huge_pages += delta;
2339 h->surplus_huge_pages_node[node] += delta;
2343 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2344 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2345 nodemask_t *nodes_allowed)
2347 unsigned long min_count, ret;
2349 if (hstate_is_gigantic(h) && !gigantic_page_supported())
2350 return h->max_huge_pages;
2353 * Increase the pool size
2354 * First take pages out of surplus state. Then make up the
2355 * remaining difference by allocating fresh huge pages.
2357 * We might race with alloc_surplus_huge_page() here and be unable
2358 * to convert a surplus huge page to a normal huge page. That is
2359 * not critical, though, it just means the overall size of the
2360 * pool might be one hugepage larger than it needs to be, but
2361 * within all the constraints specified by the sysctls.
2363 spin_lock(&hugetlb_lock);
2364 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2365 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2369 while (count > persistent_huge_pages(h)) {
2371 * If this allocation races such that we no longer need the
2372 * page, free_huge_page will handle it by freeing the page
2373 * and reducing the surplus.
2375 spin_unlock(&hugetlb_lock);
2377 /* yield cpu to avoid soft lockup */
2380 ret = alloc_pool_huge_page(h, nodes_allowed);
2381 spin_lock(&hugetlb_lock);
2385 /* Bail for signals. Probably ctrl-c from user */
2386 if (signal_pending(current))
2391 * Decrease the pool size
2392 * First return free pages to the buddy allocator (being careful
2393 * to keep enough around to satisfy reservations). Then place
2394 * pages into surplus state as needed so the pool will shrink
2395 * to the desired size as pages become free.
2397 * By placing pages into the surplus state independent of the
2398 * overcommit value, we are allowing the surplus pool size to
2399 * exceed overcommit. There are few sane options here. Since
2400 * alloc_surplus_huge_page() is checking the global counter,
2401 * though, we'll note that we're not allowed to exceed surplus
2402 * and won't grow the pool anywhere else. Not until one of the
2403 * sysctls are changed, or the surplus pages go out of use.
2405 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2406 min_count = max(count, min_count);
2407 try_to_free_low(h, min_count, nodes_allowed);
2408 while (min_count < persistent_huge_pages(h)) {
2409 if (!free_pool_huge_page(h, nodes_allowed, 0))
2411 cond_resched_lock(&hugetlb_lock);
2413 while (count < persistent_huge_pages(h)) {
2414 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2418 ret = persistent_huge_pages(h);
2419 spin_unlock(&hugetlb_lock);
2423 #define HSTATE_ATTR_RO(_name) \
2424 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2426 #define HSTATE_ATTR(_name) \
2427 static struct kobj_attribute _name##_attr = \
2428 __ATTR(_name, 0644, _name##_show, _name##_store)
2430 static struct kobject *hugepages_kobj;
2431 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2433 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2435 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2439 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2440 if (hstate_kobjs[i] == kobj) {
2442 *nidp = NUMA_NO_NODE;
2446 return kobj_to_node_hstate(kobj, nidp);
2449 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2450 struct kobj_attribute *attr, char *buf)
2453 unsigned long nr_huge_pages;
2456 h = kobj_to_hstate(kobj, &nid);
2457 if (nid == NUMA_NO_NODE)
2458 nr_huge_pages = h->nr_huge_pages;
2460 nr_huge_pages = h->nr_huge_pages_node[nid];
2462 return sprintf(buf, "%lu\n", nr_huge_pages);
2465 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2466 struct hstate *h, int nid,
2467 unsigned long count, size_t len)
2470 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2472 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2477 if (nid == NUMA_NO_NODE) {
2479 * global hstate attribute
2481 if (!(obey_mempolicy &&
2482 init_nodemask_of_mempolicy(nodes_allowed))) {
2483 NODEMASK_FREE(nodes_allowed);
2484 nodes_allowed = &node_states[N_MEMORY];
2486 } else if (nodes_allowed) {
2488 * per node hstate attribute: adjust count to global,
2489 * but restrict alloc/free to the specified node.
2491 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2492 init_nodemask_of_node(nodes_allowed, nid);
2494 nodes_allowed = &node_states[N_MEMORY];
2496 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2498 if (nodes_allowed != &node_states[N_MEMORY])
2499 NODEMASK_FREE(nodes_allowed);
2503 NODEMASK_FREE(nodes_allowed);
2507 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2508 struct kobject *kobj, const char *buf,
2512 unsigned long count;
2516 err = kstrtoul(buf, 10, &count);
2520 h = kobj_to_hstate(kobj, &nid);
2521 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2524 static ssize_t nr_hugepages_show(struct kobject *kobj,
2525 struct kobj_attribute *attr, char *buf)
2527 return nr_hugepages_show_common(kobj, attr, buf);
2530 static ssize_t nr_hugepages_store(struct kobject *kobj,
2531 struct kobj_attribute *attr, const char *buf, size_t len)
2533 return nr_hugepages_store_common(false, kobj, buf, len);
2535 HSTATE_ATTR(nr_hugepages);
2540 * hstate attribute for optionally mempolicy-based constraint on persistent
2541 * huge page alloc/free.
2543 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2544 struct kobj_attribute *attr, char *buf)
2546 return nr_hugepages_show_common(kobj, attr, buf);
2549 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2550 struct kobj_attribute *attr, const char *buf, size_t len)
2552 return nr_hugepages_store_common(true, kobj, buf, len);
2554 HSTATE_ATTR(nr_hugepages_mempolicy);
2558 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2559 struct kobj_attribute *attr, char *buf)
2561 struct hstate *h = kobj_to_hstate(kobj, NULL);
2562 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2565 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2566 struct kobj_attribute *attr, const char *buf, size_t count)
2569 unsigned long input;
2570 struct hstate *h = kobj_to_hstate(kobj, NULL);
2572 if (hstate_is_gigantic(h))
2575 err = kstrtoul(buf, 10, &input);
2579 spin_lock(&hugetlb_lock);
2580 h->nr_overcommit_huge_pages = input;
2581 spin_unlock(&hugetlb_lock);
2585 HSTATE_ATTR(nr_overcommit_hugepages);
2587 static ssize_t free_hugepages_show(struct kobject *kobj,
2588 struct kobj_attribute *attr, char *buf)
2591 unsigned long free_huge_pages;
2594 h = kobj_to_hstate(kobj, &nid);
2595 if (nid == NUMA_NO_NODE)
2596 free_huge_pages = h->free_huge_pages;
2598 free_huge_pages = h->free_huge_pages_node[nid];
2600 return sprintf(buf, "%lu\n", free_huge_pages);
2602 HSTATE_ATTR_RO(free_hugepages);
2604 static ssize_t resv_hugepages_show(struct kobject *kobj,
2605 struct kobj_attribute *attr, char *buf)
2607 struct hstate *h = kobj_to_hstate(kobj, NULL);
2608 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2610 HSTATE_ATTR_RO(resv_hugepages);
2612 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2613 struct kobj_attribute *attr, char *buf)
2616 unsigned long surplus_huge_pages;
2619 h = kobj_to_hstate(kobj, &nid);
2620 if (nid == NUMA_NO_NODE)
2621 surplus_huge_pages = h->surplus_huge_pages;
2623 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2625 return sprintf(buf, "%lu\n", surplus_huge_pages);
2627 HSTATE_ATTR_RO(surplus_hugepages);
2629 static struct attribute *hstate_attrs[] = {
2630 &nr_hugepages_attr.attr,
2631 &nr_overcommit_hugepages_attr.attr,
2632 &free_hugepages_attr.attr,
2633 &resv_hugepages_attr.attr,
2634 &surplus_hugepages_attr.attr,
2636 &nr_hugepages_mempolicy_attr.attr,
2641 static const struct attribute_group hstate_attr_group = {
2642 .attrs = hstate_attrs,
2645 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2646 struct kobject **hstate_kobjs,
2647 const struct attribute_group *hstate_attr_group)
2650 int hi = hstate_index(h);
2652 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2653 if (!hstate_kobjs[hi])
2656 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2658 kobject_put(hstate_kobjs[hi]);
2659 hstate_kobjs[hi] = NULL;
2665 static void __init hugetlb_sysfs_init(void)
2670 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2671 if (!hugepages_kobj)
2674 for_each_hstate(h) {
2675 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2676 hstate_kobjs, &hstate_attr_group);
2678 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2685 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2686 * with node devices in node_devices[] using a parallel array. The array
2687 * index of a node device or _hstate == node id.
2688 * This is here to avoid any static dependency of the node device driver, in
2689 * the base kernel, on the hugetlb module.
2691 struct node_hstate {
2692 struct kobject *hugepages_kobj;
2693 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2695 static struct node_hstate node_hstates[MAX_NUMNODES];
2698 * A subset of global hstate attributes for node devices
2700 static struct attribute *per_node_hstate_attrs[] = {
2701 &nr_hugepages_attr.attr,
2702 &free_hugepages_attr.attr,
2703 &surplus_hugepages_attr.attr,
2707 static const struct attribute_group per_node_hstate_attr_group = {
2708 .attrs = per_node_hstate_attrs,
2712 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2713 * Returns node id via non-NULL nidp.
2715 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2719 for (nid = 0; nid < nr_node_ids; nid++) {
2720 struct node_hstate *nhs = &node_hstates[nid];
2722 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2723 if (nhs->hstate_kobjs[i] == kobj) {
2735 * Unregister hstate attributes from a single node device.
2736 * No-op if no hstate attributes attached.
2738 static void hugetlb_unregister_node(struct node *node)
2741 struct node_hstate *nhs = &node_hstates[node->dev.id];
2743 if (!nhs->hugepages_kobj)
2744 return; /* no hstate attributes */
2746 for_each_hstate(h) {
2747 int idx = hstate_index(h);
2748 if (nhs->hstate_kobjs[idx]) {
2749 kobject_put(nhs->hstate_kobjs[idx]);
2750 nhs->hstate_kobjs[idx] = NULL;
2754 kobject_put(nhs->hugepages_kobj);
2755 nhs->hugepages_kobj = NULL;
2760 * Register hstate attributes for a single node device.
2761 * No-op if attributes already registered.
2763 static void hugetlb_register_node(struct node *node)
2766 struct node_hstate *nhs = &node_hstates[node->dev.id];
2769 if (nhs->hugepages_kobj)
2770 return; /* already allocated */
2772 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2774 if (!nhs->hugepages_kobj)
2777 for_each_hstate(h) {
2778 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2780 &per_node_hstate_attr_group);
2782 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2783 h->name, node->dev.id);
2784 hugetlb_unregister_node(node);
2791 * hugetlb init time: register hstate attributes for all registered node
2792 * devices of nodes that have memory. All on-line nodes should have
2793 * registered their associated device by this time.
2795 static void __init hugetlb_register_all_nodes(void)
2799 for_each_node_state(nid, N_MEMORY) {
2800 struct node *node = node_devices[nid];
2801 if (node->dev.id == nid)
2802 hugetlb_register_node(node);
2806 * Let the node device driver know we're here so it can
2807 * [un]register hstate attributes on node hotplug.
2809 register_hugetlbfs_with_node(hugetlb_register_node,
2810 hugetlb_unregister_node);
2812 #else /* !CONFIG_NUMA */
2814 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2822 static void hugetlb_register_all_nodes(void) { }
2826 static int __init hugetlb_init(void)
2830 if (!hugepages_supported())
2833 if (!size_to_hstate(default_hstate_size)) {
2834 if (default_hstate_size != 0) {
2835 pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2836 default_hstate_size, HPAGE_SIZE);
2839 default_hstate_size = HPAGE_SIZE;
2840 if (!size_to_hstate(default_hstate_size))
2841 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2843 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2844 if (default_hstate_max_huge_pages) {
2845 if (!default_hstate.max_huge_pages)
2846 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2849 hugetlb_init_hstates();
2850 gather_bootmem_prealloc();
2853 hugetlb_sysfs_init();
2854 hugetlb_register_all_nodes();
2855 hugetlb_cgroup_file_init();
2858 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2860 num_fault_mutexes = 1;
2862 hugetlb_fault_mutex_table =
2863 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
2865 BUG_ON(!hugetlb_fault_mutex_table);
2867 for (i = 0; i < num_fault_mutexes; i++)
2868 mutex_init(&hugetlb_fault_mutex_table[i]);
2871 subsys_initcall(hugetlb_init);
2873 /* Should be called on processing a hugepagesz=... option */
2874 void __init hugetlb_bad_size(void)
2876 parsed_valid_hugepagesz = false;
2879 void __init hugetlb_add_hstate(unsigned int order)
2884 if (size_to_hstate(PAGE_SIZE << order)) {
2885 pr_warn("hugepagesz= specified twice, ignoring\n");
2888 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2890 h = &hstates[hugetlb_max_hstate++];
2892 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2893 h->nr_huge_pages = 0;
2894 h->free_huge_pages = 0;
2895 for (i = 0; i < MAX_NUMNODES; ++i)
2896 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2897 INIT_LIST_HEAD(&h->hugepage_activelist);
2898 h->next_nid_to_alloc = first_memory_node;
2899 h->next_nid_to_free = first_memory_node;
2900 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2901 huge_page_size(h)/1024);
2906 static int __init hugetlb_nrpages_setup(char *s)
2909 static unsigned long *last_mhp;
2911 if (!parsed_valid_hugepagesz) {
2912 pr_warn("hugepages = %s preceded by "
2913 "an unsupported hugepagesz, ignoring\n", s);
2914 parsed_valid_hugepagesz = true;
2918 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2919 * so this hugepages= parameter goes to the "default hstate".
2921 else if (!hugetlb_max_hstate)
2922 mhp = &default_hstate_max_huge_pages;
2924 mhp = &parsed_hstate->max_huge_pages;
2926 if (mhp == last_mhp) {
2927 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2931 if (sscanf(s, "%lu", mhp) <= 0)
2935 * Global state is always initialized later in hugetlb_init.
2936 * But we need to allocate >= MAX_ORDER hstates here early to still
2937 * use the bootmem allocator.
2939 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2940 hugetlb_hstate_alloc_pages(parsed_hstate);
2946 __setup("hugepages=", hugetlb_nrpages_setup);
2948 static int __init hugetlb_default_setup(char *s)
2950 default_hstate_size = memparse(s, &s);
2953 __setup("default_hugepagesz=", hugetlb_default_setup);
2955 static unsigned int cpuset_mems_nr(unsigned int *array)
2958 unsigned int nr = 0;
2960 for_each_node_mask(node, cpuset_current_mems_allowed)
2966 #ifdef CONFIG_SYSCTL
2967 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
2968 void *buffer, size_t *length,
2969 loff_t *ppos, unsigned long *out)
2971 struct ctl_table dup_table;
2974 * In order to avoid races with __do_proc_doulongvec_minmax(), we
2975 * can duplicate the @table and alter the duplicate of it.
2978 dup_table.data = out;
2980 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
2983 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2984 struct ctl_table *table, int write,
2985 void __user *buffer, size_t *length, loff_t *ppos)
2987 struct hstate *h = &default_hstate;
2988 unsigned long tmp = h->max_huge_pages;
2991 if (!hugepages_supported())
2994 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3000 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3001 NUMA_NO_NODE, tmp, *length);
3006 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3007 void __user *buffer, size_t *length, loff_t *ppos)
3010 return hugetlb_sysctl_handler_common(false, table, write,
3011 buffer, length, ppos);
3015 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3016 void __user *buffer, size_t *length, loff_t *ppos)
3018 return hugetlb_sysctl_handler_common(true, table, write,
3019 buffer, length, ppos);
3021 #endif /* CONFIG_NUMA */
3023 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3024 void __user *buffer,
3025 size_t *length, loff_t *ppos)
3027 struct hstate *h = &default_hstate;
3031 if (!hugepages_supported())
3034 tmp = h->nr_overcommit_huge_pages;
3036 if (write && hstate_is_gigantic(h))
3039 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3045 spin_lock(&hugetlb_lock);
3046 h->nr_overcommit_huge_pages = tmp;
3047 spin_unlock(&hugetlb_lock);
3053 #endif /* CONFIG_SYSCTL */
3055 void hugetlb_report_meminfo(struct seq_file *m)
3058 unsigned long total = 0;
3060 if (!hugepages_supported())
3063 for_each_hstate(h) {
3064 unsigned long count = h->nr_huge_pages;
3066 total += (PAGE_SIZE << huge_page_order(h)) * count;
3068 if (h == &default_hstate)
3070 "HugePages_Total: %5lu\n"
3071 "HugePages_Free: %5lu\n"
3072 "HugePages_Rsvd: %5lu\n"
3073 "HugePages_Surp: %5lu\n"
3074 "Hugepagesize: %8lu kB\n",
3078 h->surplus_huge_pages,
3079 (PAGE_SIZE << huge_page_order(h)) / 1024);
3082 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3085 int hugetlb_report_node_meminfo(int nid, char *buf)
3087 struct hstate *h = &default_hstate;
3088 if (!hugepages_supported())
3091 "Node %d HugePages_Total: %5u\n"
3092 "Node %d HugePages_Free: %5u\n"
3093 "Node %d HugePages_Surp: %5u\n",
3094 nid, h->nr_huge_pages_node[nid],
3095 nid, h->free_huge_pages_node[nid],
3096 nid, h->surplus_huge_pages_node[nid]);
3099 void hugetlb_show_meminfo(void)
3104 if (!hugepages_supported())
3107 for_each_node_state(nid, N_MEMORY)
3109 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3111 h->nr_huge_pages_node[nid],
3112 h->free_huge_pages_node[nid],
3113 h->surplus_huge_pages_node[nid],
3114 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3117 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3119 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3120 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3123 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3124 unsigned long hugetlb_total_pages(void)
3127 unsigned long nr_total_pages = 0;
3130 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3131 return nr_total_pages;
3134 static int hugetlb_acct_memory(struct hstate *h, long delta)
3138 spin_lock(&hugetlb_lock);
3140 * When cpuset is configured, it breaks the strict hugetlb page
3141 * reservation as the accounting is done on a global variable. Such
3142 * reservation is completely rubbish in the presence of cpuset because
3143 * the reservation is not checked against page availability for the
3144 * current cpuset. Application can still potentially OOM'ed by kernel
3145 * with lack of free htlb page in cpuset that the task is in.
3146 * Attempt to enforce strict accounting with cpuset is almost
3147 * impossible (or too ugly) because cpuset is too fluid that
3148 * task or memory node can be dynamically moved between cpusets.
3150 * The change of semantics for shared hugetlb mapping with cpuset is
3151 * undesirable. However, in order to preserve some of the semantics,
3152 * we fall back to check against current free page availability as
3153 * a best attempt and hopefully to minimize the impact of changing
3154 * semantics that cpuset has.
3157 if (gather_surplus_pages(h, delta) < 0)
3160 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3161 return_unused_surplus_pages(h, delta);
3168 return_unused_surplus_pages(h, (unsigned long) -delta);
3171 spin_unlock(&hugetlb_lock);
3175 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3177 struct resv_map *resv = vma_resv_map(vma);
3180 * This new VMA should share its siblings reservation map if present.
3181 * The VMA will only ever have a valid reservation map pointer where
3182 * it is being copied for another still existing VMA. As that VMA
3183 * has a reference to the reservation map it cannot disappear until
3184 * after this open call completes. It is therefore safe to take a
3185 * new reference here without additional locking.
3187 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3188 kref_get(&resv->refs);
3191 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3193 struct hstate *h = hstate_vma(vma);
3194 struct resv_map *resv = vma_resv_map(vma);
3195 struct hugepage_subpool *spool = subpool_vma(vma);
3196 unsigned long reserve, start, end;
3199 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3202 start = vma_hugecache_offset(h, vma, vma->vm_start);
3203 end = vma_hugecache_offset(h, vma, vma->vm_end);
3205 reserve = (end - start) - region_count(resv, start, end);
3207 kref_put(&resv->refs, resv_map_release);
3211 * Decrement reserve counts. The global reserve count may be
3212 * adjusted if the subpool has a minimum size.
3214 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3215 hugetlb_acct_memory(h, -gbl_reserve);
3219 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3221 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3226 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3228 struct hstate *hstate = hstate_vma(vma);
3230 return 1UL << huge_page_shift(hstate);
3234 * We cannot handle pagefaults against hugetlb pages at all. They cause
3235 * handle_mm_fault() to try to instantiate regular-sized pages in the
3236 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3239 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3246 * When a new function is introduced to vm_operations_struct and added
3247 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3248 * This is because under System V memory model, mappings created via
3249 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3250 * their original vm_ops are overwritten with shm_vm_ops.
3252 const struct vm_operations_struct hugetlb_vm_ops = {
3253 .fault = hugetlb_vm_op_fault,
3254 .open = hugetlb_vm_op_open,
3255 .close = hugetlb_vm_op_close,
3256 .split = hugetlb_vm_op_split,
3257 .pagesize = hugetlb_vm_op_pagesize,
3260 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3266 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3267 vma->vm_page_prot)));
3269 entry = huge_pte_wrprotect(mk_huge_pte(page,
3270 vma->vm_page_prot));
3272 entry = pte_mkyoung(entry);
3273 entry = pte_mkhuge(entry);
3274 entry = arch_make_huge_pte(entry, vma, page, writable);
3279 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3280 unsigned long address, pte_t *ptep)
3284 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3285 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3286 update_mmu_cache(vma, address, ptep);
3289 bool is_hugetlb_entry_migration(pte_t pte)
3293 if (huge_pte_none(pte) || pte_present(pte))
3295 swp = pte_to_swp_entry(pte);
3296 if (non_swap_entry(swp) && is_migration_entry(swp))
3302 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3306 if (huge_pte_none(pte) || pte_present(pte))
3308 swp = pte_to_swp_entry(pte);
3309 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3315 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3316 struct vm_area_struct *vma)
3318 pte_t *src_pte, *dst_pte, entry, dst_entry;
3319 struct page *ptepage;
3322 struct hstate *h = hstate_vma(vma);
3323 unsigned long sz = huge_page_size(h);
3324 unsigned long mmun_start; /* For mmu_notifiers */
3325 unsigned long mmun_end; /* For mmu_notifiers */
3328 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3330 mmun_start = vma->vm_start;
3331 mmun_end = vma->vm_end;
3333 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3335 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3336 spinlock_t *src_ptl, *dst_ptl;
3337 src_pte = huge_pte_offset(src, addr, sz);
3340 dst_pte = huge_pte_alloc(dst, addr, sz);
3347 * If the pagetables are shared don't copy or take references.
3348 * dst_pte == src_pte is the common case of src/dest sharing.
3350 * However, src could have 'unshared' and dst shares with
3351 * another vma. If dst_pte !none, this implies sharing.
3352 * Check here before taking page table lock, and once again
3353 * after taking the lock below.
3355 dst_entry = huge_ptep_get(dst_pte);
3356 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3359 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3360 src_ptl = huge_pte_lockptr(h, src, src_pte);
3361 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3362 entry = huge_ptep_get(src_pte);
3363 dst_entry = huge_ptep_get(dst_pte);
3364 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3366 * Skip if src entry none. Also, skip in the
3367 * unlikely case dst entry !none as this implies
3368 * sharing with another vma.
3371 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3372 is_hugetlb_entry_hwpoisoned(entry))) {
3373 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3375 if (is_write_migration_entry(swp_entry) && cow) {
3377 * COW mappings require pages in both
3378 * parent and child to be set to read.
3380 make_migration_entry_read(&swp_entry);
3381 entry = swp_entry_to_pte(swp_entry);
3382 set_huge_swap_pte_at(src, addr, src_pte,
3385 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3389 * No need to notify as we are downgrading page
3390 * table protection not changing it to point
3393 * See Documentation/vm/mmu_notifier.rst
3395 huge_ptep_set_wrprotect(src, addr, src_pte);
3397 entry = huge_ptep_get(src_pte);
3398 ptepage = pte_page(entry);
3400 page_dup_rmap(ptepage, true);
3401 set_huge_pte_at(dst, addr, dst_pte, entry);
3402 hugetlb_count_add(pages_per_huge_page(h), dst);
3404 spin_unlock(src_ptl);
3405 spin_unlock(dst_ptl);
3409 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3414 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3415 unsigned long start, unsigned long end,
3416 struct page *ref_page)
3418 struct mm_struct *mm = vma->vm_mm;
3419 unsigned long address;
3424 struct hstate *h = hstate_vma(vma);
3425 unsigned long sz = huge_page_size(h);
3426 unsigned long mmun_start = start; /* For mmu_notifiers */
3427 unsigned long mmun_end = end; /* For mmu_notifiers */
3428 bool force_flush = false;
3430 WARN_ON(!is_vm_hugetlb_page(vma));
3431 BUG_ON(start & ~huge_page_mask(h));
3432 BUG_ON(end & ~huge_page_mask(h));
3435 * This is a hugetlb vma, all the pte entries should point
3438 tlb_remove_check_page_size_change(tlb, sz);
3439 tlb_start_vma(tlb, vma);
3442 * If sharing possible, alert mmu notifiers of worst case.
3444 adjust_range_if_pmd_sharing_possible(vma, &mmun_start, &mmun_end);
3445 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3447 for (; address < end; address += sz) {
3448 ptep = huge_pte_offset(mm, address, sz);
3452 ptl = huge_pte_lock(h, mm, ptep);
3453 if (huge_pmd_unshare(mm, &address, ptep)) {
3455 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
3460 pte = huge_ptep_get(ptep);
3461 if (huge_pte_none(pte)) {
3467 * Migrating hugepage or HWPoisoned hugepage is already
3468 * unmapped and its refcount is dropped, so just clear pte here.
3470 if (unlikely(!pte_present(pte))) {
3471 huge_pte_clear(mm, address, ptep, sz);
3476 page = pte_page(pte);
3478 * If a reference page is supplied, it is because a specific
3479 * page is being unmapped, not a range. Ensure the page we
3480 * are about to unmap is the actual page of interest.
3483 if (page != ref_page) {
3488 * Mark the VMA as having unmapped its page so that
3489 * future faults in this VMA will fail rather than
3490 * looking like data was lost
3492 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3495 pte = huge_ptep_get_and_clear(mm, address, ptep);
3496 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3497 if (huge_pte_dirty(pte))
3498 set_page_dirty(page);
3500 hugetlb_count_sub(pages_per_huge_page(h), mm);
3501 page_remove_rmap(page, true);
3504 tlb_remove_page_size(tlb, page, huge_page_size(h));
3506 * Bail out after unmapping reference page if supplied
3511 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3512 tlb_end_vma(tlb, vma);
3515 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
3516 * could defer the flush until now, since by holding i_mmap_rwsem we
3517 * guaranteed that the last refernece would not be dropped. But we must
3518 * do the flushing before we return, as otherwise i_mmap_rwsem will be
3519 * dropped and the last reference to the shared PMDs page might be
3522 * In theory we could defer the freeing of the PMD pages as well, but
3523 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
3524 * detect sharing, so we cannot defer the release of the page either.
3525 * Instead, do flush now.
3528 tlb_flush_mmu_tlbonly(tlb);
3531 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3532 struct vm_area_struct *vma, unsigned long start,
3533 unsigned long end, struct page *ref_page)
3535 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3538 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3539 * test will fail on a vma being torn down, and not grab a page table
3540 * on its way out. We're lucky that the flag has such an appropriate
3541 * name, and can in fact be safely cleared here. We could clear it
3542 * before the __unmap_hugepage_range above, but all that's necessary
3543 * is to clear it before releasing the i_mmap_rwsem. This works
3544 * because in the context this is called, the VMA is about to be
3545 * destroyed and the i_mmap_rwsem is held.
3547 vma->vm_flags &= ~VM_MAYSHARE;
3550 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3551 unsigned long end, struct page *ref_page)
3553 struct mm_struct *mm;
3554 struct mmu_gather tlb;
3555 unsigned long tlb_start = start;
3556 unsigned long tlb_end = end;
3559 * If shared PMDs were possibly used within this vma range, adjust
3560 * start/end for worst case tlb flushing.
3561 * Note that we can not be sure if PMDs are shared until we try to
3562 * unmap pages. However, we want to make sure TLB flushing covers
3563 * the largest possible range.
3565 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
3569 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
3570 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3571 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
3575 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3576 * mappping it owns the reserve page for. The intention is to unmap the page
3577 * from other VMAs and let the children be SIGKILLed if they are faulting the
3580 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3581 struct page *page, unsigned long address)
3583 struct hstate *h = hstate_vma(vma);
3584 struct vm_area_struct *iter_vma;
3585 struct address_space *mapping;
3589 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3590 * from page cache lookup which is in HPAGE_SIZE units.
3592 address = address & huge_page_mask(h);
3593 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3595 mapping = vma->vm_file->f_mapping;
3598 * Take the mapping lock for the duration of the table walk. As
3599 * this mapping should be shared between all the VMAs,
3600 * __unmap_hugepage_range() is called as the lock is already held
3602 i_mmap_lock_write(mapping);
3603 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3604 /* Do not unmap the current VMA */
3605 if (iter_vma == vma)
3609 * Shared VMAs have their own reserves and do not affect
3610 * MAP_PRIVATE accounting but it is possible that a shared
3611 * VMA is using the same page so check and skip such VMAs.
3613 if (iter_vma->vm_flags & VM_MAYSHARE)
3617 * Unmap the page from other VMAs without their own reserves.
3618 * They get marked to be SIGKILLed if they fault in these
3619 * areas. This is because a future no-page fault on this VMA
3620 * could insert a zeroed page instead of the data existing
3621 * from the time of fork. This would look like data corruption
3623 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3624 unmap_hugepage_range(iter_vma, address,
3625 address + huge_page_size(h), page);
3627 i_mmap_unlock_write(mapping);
3631 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3632 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3633 * cannot race with other handlers or page migration.
3634 * Keep the pte_same checks anyway to make transition from the mutex easier.
3636 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3637 unsigned long address, pte_t *ptep,
3638 struct page *pagecache_page, spinlock_t *ptl)
3641 struct hstate *h = hstate_vma(vma);
3642 struct page *old_page, *new_page;
3643 int outside_reserve = 0;
3645 unsigned long mmun_start; /* For mmu_notifiers */
3646 unsigned long mmun_end; /* For mmu_notifiers */
3647 unsigned long haddr = address & huge_page_mask(h);
3649 pte = huge_ptep_get(ptep);
3650 old_page = pte_page(pte);
3653 /* If no-one else is actually using this page, avoid the copy
3654 * and just make the page writable */
3655 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3656 page_move_anon_rmap(old_page, vma);
3657 set_huge_ptep_writable(vma, haddr, ptep);
3662 * If the process that created a MAP_PRIVATE mapping is about to
3663 * perform a COW due to a shared page count, attempt to satisfy
3664 * the allocation without using the existing reserves. The pagecache
3665 * page is used to determine if the reserve at this address was
3666 * consumed or not. If reserves were used, a partial faulted mapping
3667 * at the time of fork() could consume its reserves on COW instead
3668 * of the full address range.
3670 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3671 old_page != pagecache_page)
3672 outside_reserve = 1;
3677 * Drop page table lock as buddy allocator may be called. It will
3678 * be acquired again before returning to the caller, as expected.
3681 new_page = alloc_huge_page(vma, haddr, outside_reserve);
3683 if (IS_ERR(new_page)) {
3685 * If a process owning a MAP_PRIVATE mapping fails to COW,
3686 * it is due to references held by a child and an insufficient
3687 * huge page pool. To guarantee the original mappers
3688 * reliability, unmap the page from child processes. The child
3689 * may get SIGKILLed if it later faults.
3691 if (outside_reserve) {
3693 BUG_ON(huge_pte_none(pte));
3694 unmap_ref_private(mm, vma, old_page, haddr);
3695 BUG_ON(huge_pte_none(pte));
3697 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3699 pte_same(huge_ptep_get(ptep), pte)))
3700 goto retry_avoidcopy;
3702 * race occurs while re-acquiring page table
3703 * lock, and our job is done.
3708 ret = vmf_error(PTR_ERR(new_page));
3709 goto out_release_old;
3713 * When the original hugepage is shared one, it does not have
3714 * anon_vma prepared.
3716 if (unlikely(anon_vma_prepare(vma))) {
3718 goto out_release_all;
3721 copy_user_huge_page(new_page, old_page, address, vma,
3722 pages_per_huge_page(h));
3723 __SetPageUptodate(new_page);
3726 mmun_end = mmun_start + huge_page_size(h);
3727 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3730 * Retake the page table lock to check for racing updates
3731 * before the page tables are altered
3734 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3735 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3736 ClearPagePrivate(new_page);
3739 huge_ptep_clear_flush(vma, haddr, ptep);
3740 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3741 set_huge_pte_at(mm, haddr, ptep,
3742 make_huge_pte(vma, new_page, 1));
3743 page_remove_rmap(old_page, true);
3744 hugepage_add_new_anon_rmap(new_page, vma, haddr);
3745 set_page_huge_active(new_page);
3746 /* Make the old page be freed below */
3747 new_page = old_page;
3750 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3752 restore_reserve_on_error(h, vma, haddr, new_page);
3757 spin_lock(ptl); /* Caller expects lock to be held */
3761 /* Return the pagecache page at a given address within a VMA */
3762 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3763 struct vm_area_struct *vma, unsigned long address)
3765 struct address_space *mapping;
3768 mapping = vma->vm_file->f_mapping;
3769 idx = vma_hugecache_offset(h, vma, address);
3771 return find_lock_page(mapping, idx);
3775 * Return whether there is a pagecache page to back given address within VMA.
3776 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3778 static bool hugetlbfs_pagecache_present(struct hstate *h,
3779 struct vm_area_struct *vma, unsigned long address)
3781 struct address_space *mapping;
3785 mapping = vma->vm_file->f_mapping;
3786 idx = vma_hugecache_offset(h, vma, address);
3788 page = find_get_page(mapping, idx);
3791 return page != NULL;
3794 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3797 struct inode *inode = mapping->host;
3798 struct hstate *h = hstate_inode(inode);
3799 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3803 ClearPagePrivate(page);
3806 * set page dirty so that it will not be removed from cache/file
3807 * by non-hugetlbfs specific code paths.
3809 set_page_dirty(page);
3811 spin_lock(&inode->i_lock);
3812 inode->i_blocks += blocks_per_huge_page(h);
3813 spin_unlock(&inode->i_lock);
3817 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
3818 struct vm_area_struct *vma,
3819 struct address_space *mapping, pgoff_t idx,
3820 unsigned long address, pte_t *ptep, unsigned int flags)
3822 struct hstate *h = hstate_vma(vma);
3823 vm_fault_t ret = VM_FAULT_SIGBUS;
3829 unsigned long haddr = address & huge_page_mask(h);
3830 bool new_page = false;
3833 * Currently, we are forced to kill the process in the event the
3834 * original mapper has unmapped pages from the child due to a failed
3835 * COW. Warn that such a situation has occurred as it may not be obvious
3837 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3838 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3844 * Use page lock to guard against racing truncation
3845 * before we get page_table_lock.
3848 page = find_lock_page(mapping, idx);
3850 size = i_size_read(mapping->host) >> huge_page_shift(h);
3855 * Check for page in userfault range
3857 if (userfaultfd_missing(vma)) {
3859 struct vm_fault vmf = {
3864 * Hard to debug if it ends up being
3865 * used by a callee that assumes
3866 * something about the other
3867 * uninitialized fields... same as in
3873 * hugetlb_fault_mutex must be dropped before
3874 * handling userfault. Reacquire after handling
3875 * fault to make calling code simpler.
3877 hash = hugetlb_fault_mutex_hash(h, mapping, idx);
3878 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3879 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3880 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3884 page = alloc_huge_page(vma, haddr, 0);
3886 ret = vmf_error(PTR_ERR(page));
3889 clear_huge_page(page, address, pages_per_huge_page(h));
3890 __SetPageUptodate(page);
3893 if (vma->vm_flags & VM_MAYSHARE) {
3894 int err = huge_add_to_page_cache(page, mapping, idx);
3903 if (unlikely(anon_vma_prepare(vma))) {
3905 goto backout_unlocked;
3911 * If memory error occurs between mmap() and fault, some process
3912 * don't have hwpoisoned swap entry for errored virtual address.
3913 * So we need to block hugepage fault by PG_hwpoison bit check.
3915 if (unlikely(PageHWPoison(page))) {
3916 ret = VM_FAULT_HWPOISON_LARGE |
3917 VM_FAULT_SET_HINDEX(hstate_index(h));
3918 goto backout_unlocked;
3923 * If we are going to COW a private mapping later, we examine the
3924 * pending reservations for this page now. This will ensure that
3925 * any allocations necessary to record that reservation occur outside
3928 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3929 if (vma_needs_reservation(h, vma, haddr) < 0) {
3931 goto backout_unlocked;
3933 /* Just decrements count, does not deallocate */
3934 vma_end_reservation(h, vma, haddr);
3937 ptl = huge_pte_lock(h, mm, ptep);
3938 size = i_size_read(mapping->host) >> huge_page_shift(h);
3943 if (!huge_pte_none(huge_ptep_get(ptep)))
3947 ClearPagePrivate(page);
3948 hugepage_add_new_anon_rmap(page, vma, haddr);
3950 page_dup_rmap(page, true);
3951 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3952 && (vma->vm_flags & VM_SHARED)));
3953 set_huge_pte_at(mm, haddr, ptep, new_pte);
3955 hugetlb_count_add(pages_per_huge_page(h), mm);
3956 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3957 /* Optimization, do the COW without a second fault */
3958 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
3964 * Only make newly allocated pages active. Existing pages found
3965 * in the pagecache could be !page_huge_active() if they have been
3966 * isolated for migration.
3969 set_page_huge_active(page);
3979 restore_reserve_on_error(h, vma, haddr, page);
3985 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping,
3988 unsigned long key[2];
3991 key[0] = (unsigned long) mapping;
3994 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
3996 return hash & (num_fault_mutexes - 1);
4000 * For uniprocesor systems we always use a single mutex, so just
4001 * return 0 and avoid the hashing overhead.
4003 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping,
4010 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4011 unsigned long address, unsigned int flags)
4018 struct page *page = NULL;
4019 struct page *pagecache_page = NULL;
4020 struct hstate *h = hstate_vma(vma);
4021 struct address_space *mapping;
4022 int need_wait_lock = 0;
4023 unsigned long haddr = address & huge_page_mask(h);
4025 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4027 entry = huge_ptep_get(ptep);
4028 if (unlikely(is_hugetlb_entry_migration(entry))) {
4029 migration_entry_wait_huge(vma, mm, ptep);
4031 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4032 return VM_FAULT_HWPOISON_LARGE |
4033 VM_FAULT_SET_HINDEX(hstate_index(h));
4035 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4037 return VM_FAULT_OOM;
4040 mapping = vma->vm_file->f_mapping;
4041 idx = vma_hugecache_offset(h, vma, haddr);
4044 * Serialize hugepage allocation and instantiation, so that we don't
4045 * get spurious allocation failures if two CPUs race to instantiate
4046 * the same page in the page cache.
4048 hash = hugetlb_fault_mutex_hash(h, mapping, idx);
4049 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4051 entry = huge_ptep_get(ptep);
4052 if (huge_pte_none(entry)) {
4053 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4060 * entry could be a migration/hwpoison entry at this point, so this
4061 * check prevents the kernel from going below assuming that we have
4062 * a active hugepage in pagecache. This goto expects the 2nd page fault,
4063 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
4066 if (!pte_present(entry))
4070 * If we are going to COW the mapping later, we examine the pending
4071 * reservations for this page now. This will ensure that any
4072 * allocations necessary to record that reservation occur outside the
4073 * spinlock. For private mappings, we also lookup the pagecache
4074 * page now as it is used to determine if a reservation has been
4077 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4078 if (vma_needs_reservation(h, vma, haddr) < 0) {
4082 /* Just decrements count, does not deallocate */
4083 vma_end_reservation(h, vma, haddr);
4085 if (!(vma->vm_flags & VM_MAYSHARE))
4086 pagecache_page = hugetlbfs_pagecache_page(h,
4090 ptl = huge_pte_lock(h, mm, ptep);
4092 /* Check for a racing update before calling hugetlb_cow */
4093 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4097 * hugetlb_cow() requires page locks of pte_page(entry) and
4098 * pagecache_page, so here we need take the former one
4099 * when page != pagecache_page or !pagecache_page.
4101 page = pte_page(entry);
4102 if (page != pagecache_page)
4103 if (!trylock_page(page)) {
4110 if (flags & FAULT_FLAG_WRITE) {
4111 if (!huge_pte_write(entry)) {
4112 ret = hugetlb_cow(mm, vma, address, ptep,
4113 pagecache_page, ptl);
4116 entry = huge_pte_mkdirty(entry);
4118 entry = pte_mkyoung(entry);
4119 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4120 flags & FAULT_FLAG_WRITE))
4121 update_mmu_cache(vma, haddr, ptep);
4123 if (page != pagecache_page)
4129 if (pagecache_page) {
4130 unlock_page(pagecache_page);
4131 put_page(pagecache_page);
4134 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4136 * Generally it's safe to hold refcount during waiting page lock. But
4137 * here we just wait to defer the next page fault to avoid busy loop and
4138 * the page is not used after unlocked before returning from the current
4139 * page fault. So we are safe from accessing freed page, even if we wait
4140 * here without taking refcount.
4143 wait_on_page_locked(page);
4148 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4149 * modifications for huge pages.
4151 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4153 struct vm_area_struct *dst_vma,
4154 unsigned long dst_addr,
4155 unsigned long src_addr,
4156 struct page **pagep)
4158 struct address_space *mapping;
4161 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4162 struct hstate *h = hstate_vma(dst_vma);
4169 /* If a page already exists, then it's UFFDIO_COPY for
4170 * a non-missing case. Return -EEXIST.
4173 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
4178 page = alloc_huge_page(dst_vma, dst_addr, 0);
4184 ret = copy_huge_page_from_user(page,
4185 (const void __user *) src_addr,
4186 pages_per_huge_page(h), false);
4188 /* fallback to copy_from_user outside mmap_sem */
4189 if (unlikely(ret)) {
4192 /* don't free the page */
4201 * The memory barrier inside __SetPageUptodate makes sure that
4202 * preceding stores to the page contents become visible before
4203 * the set_pte_at() write.
4205 __SetPageUptodate(page);
4207 mapping = dst_vma->vm_file->f_mapping;
4208 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4211 * If shared, add to page cache
4214 size = i_size_read(mapping->host) >> huge_page_shift(h);
4217 goto out_release_nounlock;
4220 * Serialization between remove_inode_hugepages() and
4221 * huge_add_to_page_cache() below happens through the
4222 * hugetlb_fault_mutex_table that here must be hold by
4225 ret = huge_add_to_page_cache(page, mapping, idx);
4227 goto out_release_nounlock;
4230 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4234 * Recheck the i_size after holding PT lock to make sure not
4235 * to leave any page mapped (as page_mapped()) beyond the end
4236 * of the i_size (remove_inode_hugepages() is strict about
4237 * enforcing that). If we bail out here, we'll also leave a
4238 * page in the radix tree in the vm_shared case beyond the end
4239 * of the i_size, but remove_inode_hugepages() will take care
4240 * of it as soon as we drop the hugetlb_fault_mutex_table.
4242 size = i_size_read(mapping->host) >> huge_page_shift(h);
4245 goto out_release_unlock;
4248 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4249 goto out_release_unlock;
4252 page_dup_rmap(page, true);
4254 ClearPagePrivate(page);
4255 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4258 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4259 if (dst_vma->vm_flags & VM_WRITE)
4260 _dst_pte = huge_pte_mkdirty(_dst_pte);
4261 _dst_pte = pte_mkyoung(_dst_pte);
4263 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4265 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4266 dst_vma->vm_flags & VM_WRITE);
4267 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4269 /* No need to invalidate - it was non-present before */
4270 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4273 set_page_huge_active(page);
4283 out_release_nounlock:
4288 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4289 struct page **pages, struct vm_area_struct **vmas,
4290 unsigned long *position, unsigned long *nr_pages,
4291 long i, unsigned int flags, int *nonblocking)
4293 unsigned long pfn_offset;
4294 unsigned long vaddr = *position;
4295 unsigned long remainder = *nr_pages;
4296 struct hstate *h = hstate_vma(vma);
4299 while (vaddr < vma->vm_end && remainder) {
4301 spinlock_t *ptl = NULL;
4306 * If we have a pending SIGKILL, don't keep faulting pages and
4307 * potentially allocating memory.
4309 if (unlikely(fatal_signal_pending(current))) {
4315 * Some archs (sparc64, sh*) have multiple pte_ts to
4316 * each hugepage. We have to make sure we get the
4317 * first, for the page indexing below to work.
4319 * Note that page table lock is not held when pte is null.
4321 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4324 ptl = huge_pte_lock(h, mm, pte);
4325 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4328 * When coredumping, it suits get_dump_page if we just return
4329 * an error where there's an empty slot with no huge pagecache
4330 * to back it. This way, we avoid allocating a hugepage, and
4331 * the sparse dumpfile avoids allocating disk blocks, but its
4332 * huge holes still show up with zeroes where they need to be.
4334 if (absent && (flags & FOLL_DUMP) &&
4335 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4343 * We need call hugetlb_fault for both hugepages under migration
4344 * (in which case hugetlb_fault waits for the migration,) and
4345 * hwpoisoned hugepages (in which case we need to prevent the
4346 * caller from accessing to them.) In order to do this, we use
4347 * here is_swap_pte instead of is_hugetlb_entry_migration and
4348 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4349 * both cases, and because we can't follow correct pages
4350 * directly from any kind of swap entries.
4352 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4353 ((flags & FOLL_WRITE) &&
4354 !huge_pte_write(huge_ptep_get(pte)))) {
4356 unsigned int fault_flags = 0;
4360 if (flags & FOLL_WRITE)
4361 fault_flags |= FAULT_FLAG_WRITE;
4363 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4364 if (flags & FOLL_NOWAIT)
4365 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4366 FAULT_FLAG_RETRY_NOWAIT;
4367 if (flags & FOLL_TRIED) {
4368 VM_WARN_ON_ONCE(fault_flags &
4369 FAULT_FLAG_ALLOW_RETRY);
4370 fault_flags |= FAULT_FLAG_TRIED;
4372 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4373 if (ret & VM_FAULT_ERROR) {
4374 err = vm_fault_to_errno(ret, flags);
4378 if (ret & VM_FAULT_RETRY) {
4380 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4384 * VM_FAULT_RETRY must not return an
4385 * error, it will return zero
4388 * No need to update "position" as the
4389 * caller will not check it after
4390 * *nr_pages is set to 0.
4397 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4398 page = pte_page(huge_ptep_get(pte));
4401 * Instead of doing 'try_get_page()' below in the same_page
4402 * loop, just check the count once here.
4404 if (unlikely(page_count(page) <= 0)) {
4414 pages[i] = mem_map_offset(page, pfn_offset);
4425 if (vaddr < vma->vm_end && remainder &&
4426 pfn_offset < pages_per_huge_page(h)) {
4428 * We use pfn_offset to avoid touching the pageframes
4429 * of this compound page.
4435 *nr_pages = remainder;
4437 * setting position is actually required only if remainder is
4438 * not zero but it's faster not to add a "if (remainder)"
4446 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4448 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4451 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4454 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4455 unsigned long address, unsigned long end, pgprot_t newprot)
4457 struct mm_struct *mm = vma->vm_mm;
4458 unsigned long start = address;
4461 struct hstate *h = hstate_vma(vma);
4462 unsigned long pages = 0;
4463 unsigned long f_start = start;
4464 unsigned long f_end = end;
4465 bool shared_pmd = false;
4468 * In the case of shared PMDs, the area to flush could be beyond
4469 * start/end. Set f_start/f_end to cover the maximum possible
4470 * range if PMD sharing is possible.
4472 adjust_range_if_pmd_sharing_possible(vma, &f_start, &f_end);
4474 BUG_ON(address >= end);
4475 flush_cache_range(vma, f_start, f_end);
4477 mmu_notifier_invalidate_range_start(mm, f_start, f_end);
4478 i_mmap_lock_write(vma->vm_file->f_mapping);
4479 for (; address < end; address += huge_page_size(h)) {
4481 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4484 ptl = huge_pte_lock(h, mm, ptep);
4485 if (huge_pmd_unshare(mm, &address, ptep)) {
4491 pte = huge_ptep_get(ptep);
4492 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4496 if (unlikely(is_hugetlb_entry_migration(pte))) {
4497 swp_entry_t entry = pte_to_swp_entry(pte);
4499 if (is_write_migration_entry(entry)) {
4502 make_migration_entry_read(&entry);
4503 newpte = swp_entry_to_pte(entry);
4504 set_huge_swap_pte_at(mm, address, ptep,
4505 newpte, huge_page_size(h));
4511 if (!huge_pte_none(pte)) {
4512 pte = huge_ptep_get_and_clear(mm, address, ptep);
4513 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
4514 pte = arch_make_huge_pte(pte, vma, NULL, 0);
4515 set_huge_pte_at(mm, address, ptep, pte);
4521 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4522 * may have cleared our pud entry and done put_page on the page table:
4523 * once we release i_mmap_rwsem, another task can do the final put_page
4524 * and that page table be reused and filled with junk. If we actually
4525 * did unshare a page of pmds, flush the range corresponding to the pud.
4528 flush_hugetlb_tlb_range(vma, f_start, f_end);
4530 flush_hugetlb_tlb_range(vma, start, end);
4532 * No need to call mmu_notifier_invalidate_range() we are downgrading
4533 * page table protection not changing it to point to a new page.
4535 * See Documentation/vm/mmu_notifier.rst
4537 i_mmap_unlock_write(vma->vm_file->f_mapping);
4538 mmu_notifier_invalidate_range_end(mm, f_start, f_end);
4540 return pages << h->order;
4543 int hugetlb_reserve_pages(struct inode *inode,
4545 struct vm_area_struct *vma,
4546 vm_flags_t vm_flags)
4549 struct hstate *h = hstate_inode(inode);
4550 struct hugepage_subpool *spool = subpool_inode(inode);
4551 struct resv_map *resv_map;
4554 /* This should never happen */
4556 VM_WARN(1, "%s called with a negative range\n", __func__);
4561 * Only apply hugepage reservation if asked. At fault time, an
4562 * attempt will be made for VM_NORESERVE to allocate a page
4563 * without using reserves
4565 if (vm_flags & VM_NORESERVE)
4569 * Shared mappings base their reservation on the number of pages that
4570 * are already allocated on behalf of the file. Private mappings need
4571 * to reserve the full area even if read-only as mprotect() may be
4572 * called to make the mapping read-write. Assume !vma is a shm mapping
4574 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4575 resv_map = inode_resv_map(inode);
4577 chg = region_chg(resv_map, from, to);
4580 resv_map = resv_map_alloc();
4586 set_vma_resv_map(vma, resv_map);
4587 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4596 * There must be enough pages in the subpool for the mapping. If
4597 * the subpool has a minimum size, there may be some global
4598 * reservations already in place (gbl_reserve).
4600 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4601 if (gbl_reserve < 0) {
4607 * Check enough hugepages are available for the reservation.
4608 * Hand the pages back to the subpool if there are not
4610 ret = hugetlb_acct_memory(h, gbl_reserve);
4612 /* put back original number of pages, chg */
4613 (void)hugepage_subpool_put_pages(spool, chg);
4618 * Account for the reservations made. Shared mappings record regions
4619 * that have reservations as they are shared by multiple VMAs.
4620 * When the last VMA disappears, the region map says how much
4621 * the reservation was and the page cache tells how much of
4622 * the reservation was consumed. Private mappings are per-VMA and
4623 * only the consumed reservations are tracked. When the VMA
4624 * disappears, the original reservation is the VMA size and the
4625 * consumed reservations are stored in the map. Hence, nothing
4626 * else has to be done for private mappings here
4628 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4629 long add = region_add(resv_map, from, to);
4631 if (unlikely(chg > add)) {
4633 * pages in this range were added to the reserve
4634 * map between region_chg and region_add. This
4635 * indicates a race with alloc_huge_page. Adjust
4636 * the subpool and reserve counts modified above
4637 * based on the difference.
4641 rsv_adjust = hugepage_subpool_put_pages(spool,
4643 hugetlb_acct_memory(h, -rsv_adjust);
4648 if (!vma || vma->vm_flags & VM_MAYSHARE)
4649 /* Don't call region_abort if region_chg failed */
4651 region_abort(resv_map, from, to);
4652 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4653 kref_put(&resv_map->refs, resv_map_release);
4657 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4660 struct hstate *h = hstate_inode(inode);
4661 struct resv_map *resv_map = inode_resv_map(inode);
4663 struct hugepage_subpool *spool = subpool_inode(inode);
4667 chg = region_del(resv_map, start, end);
4669 * region_del() can fail in the rare case where a region
4670 * must be split and another region descriptor can not be
4671 * allocated. If end == LONG_MAX, it will not fail.
4677 spin_lock(&inode->i_lock);
4678 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4679 spin_unlock(&inode->i_lock);
4682 * If the subpool has a minimum size, the number of global
4683 * reservations to be released may be adjusted.
4685 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4686 hugetlb_acct_memory(h, -gbl_reserve);
4691 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4692 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4693 struct vm_area_struct *vma,
4694 unsigned long addr, pgoff_t idx)
4696 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4698 unsigned long sbase = saddr & PUD_MASK;
4699 unsigned long s_end = sbase + PUD_SIZE;
4701 /* Allow segments to share if only one is marked locked */
4702 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4703 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4706 * match the virtual addresses, permission and the alignment of the
4709 if (pmd_index(addr) != pmd_index(saddr) ||
4710 vm_flags != svm_flags ||
4711 sbase < svma->vm_start || svma->vm_end < s_end)
4717 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4719 unsigned long base = addr & PUD_MASK;
4720 unsigned long end = base + PUD_SIZE;
4723 * check on proper vm_flags and page table alignment
4725 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
4731 * Determine if start,end range within vma could be mapped by shared pmd.
4732 * If yes, adjust start and end to cover range associated with possible
4733 * shared pmd mappings.
4735 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4736 unsigned long *start, unsigned long *end)
4738 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
4739 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
4742 * vma need span at least one aligned PUD size and the start,end range
4743 * must at least partialy within it.
4745 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
4746 (*end <= v_start) || (*start >= v_end))
4749 /* Extend the range to be PUD aligned for a worst case scenario */
4750 if (*start > v_start)
4751 *start = ALIGN_DOWN(*start, PUD_SIZE);
4754 *end = ALIGN(*end, PUD_SIZE);
4758 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4759 * and returns the corresponding pte. While this is not necessary for the
4760 * !shared pmd case because we can allocate the pmd later as well, it makes the
4761 * code much cleaner. pmd allocation is essential for the shared case because
4762 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4763 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4764 * bad pmd for sharing.
4766 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4768 struct vm_area_struct *vma = find_vma(mm, addr);
4769 struct address_space *mapping = vma->vm_file->f_mapping;
4770 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4772 struct vm_area_struct *svma;
4773 unsigned long saddr;
4778 if (!vma_shareable(vma, addr))
4779 return (pte_t *)pmd_alloc(mm, pud, addr);
4781 i_mmap_lock_write(mapping);
4782 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4786 saddr = page_table_shareable(svma, vma, addr, idx);
4788 spte = huge_pte_offset(svma->vm_mm, saddr,
4789 vma_mmu_pagesize(svma));
4791 get_page(virt_to_page(spte));
4800 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4801 if (pud_none(*pud)) {
4802 pud_populate(mm, pud,
4803 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4806 put_page(virt_to_page(spte));
4810 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4811 i_mmap_unlock_write(mapping);
4816 * unmap huge page backed by shared pte.
4818 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4819 * indicated by page_count > 1, unmap is achieved by clearing pud and
4820 * decrementing the ref count. If count == 1, the pte page is not shared.
4822 * called with page table lock held.
4824 * returns: 1 successfully unmapped a shared pte page
4825 * 0 the underlying pte page is not shared, or it is the last user
4827 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4829 pgd_t *pgd = pgd_offset(mm, *addr);
4830 p4d_t *p4d = p4d_offset(pgd, *addr);
4831 pud_t *pud = pud_offset(p4d, *addr);
4833 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4834 if (page_count(virt_to_page(ptep)) == 1)
4838 put_page(virt_to_page(ptep));
4841 * This update of passed address optimizes loops sequentially
4842 * processing addresses in increments of huge page size (PMD_SIZE
4843 * in this case). By clearing the pud, a PUD_SIZE area is unmapped.
4844 * Update address to the 'last page' in the cleared area so that
4845 * calling loop can move to first page past this area.
4847 *addr |= PUD_SIZE - PMD_SIZE;
4850 #define want_pmd_share() (1)
4851 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4852 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4857 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4862 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4863 unsigned long *start, unsigned long *end)
4866 #define want_pmd_share() (0)
4867 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4869 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4870 pte_t *huge_pte_alloc(struct mm_struct *mm,
4871 unsigned long addr, unsigned long sz)
4878 pgd = pgd_offset(mm, addr);
4879 p4d = p4d_alloc(mm, pgd, addr);
4882 pud = pud_alloc(mm, p4d, addr);
4884 if (sz == PUD_SIZE) {
4887 BUG_ON(sz != PMD_SIZE);
4888 if (want_pmd_share() && pud_none(*pud))
4889 pte = huge_pmd_share(mm, addr, pud);
4891 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4894 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4900 * huge_pte_offset() - Walk the page table to resolve the hugepage
4901 * entry at address @addr
4903 * Return: Pointer to page table or swap entry (PUD or PMD) for
4904 * address @addr, or NULL if a p*d_none() entry is encountered and the
4905 * size @sz doesn't match the hugepage size at this level of the page
4908 pte_t *huge_pte_offset(struct mm_struct *mm,
4909 unsigned long addr, unsigned long sz)
4913 pud_t *pud, pud_entry;
4914 pmd_t *pmd, pmd_entry;
4916 pgd = pgd_offset(mm, addr);
4917 if (!pgd_present(*pgd))
4919 p4d = p4d_offset(pgd, addr);
4920 if (!p4d_present(*p4d))
4923 pud = pud_offset(p4d, addr);
4924 pud_entry = READ_ONCE(*pud);
4925 if (sz != PUD_SIZE && pud_none(pud_entry))
4927 /* hugepage or swap? */
4928 if (pud_huge(pud_entry) || !pud_present(pud_entry))
4929 return (pte_t *)pud;
4931 pmd = pmd_offset(pud, addr);
4932 pmd_entry = READ_ONCE(*pmd);
4933 if (sz != PMD_SIZE && pmd_none(pmd_entry))
4935 /* hugepage or swap? */
4936 if (pmd_huge(pmd_entry) || !pmd_present(pmd_entry))
4937 return (pte_t *)pmd;
4942 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4945 * These functions are overwritable if your architecture needs its own
4948 struct page * __weak
4949 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4952 return ERR_PTR(-EINVAL);
4955 struct page * __weak
4956 follow_huge_pd(struct vm_area_struct *vma,
4957 unsigned long address, hugepd_t hpd, int flags, int pdshift)
4959 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4963 struct page * __weak
4964 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4965 pmd_t *pmd, int flags)
4967 struct page *page = NULL;
4971 ptl = pmd_lockptr(mm, pmd);
4974 * make sure that the address range covered by this pmd is not
4975 * unmapped from other threads.
4977 if (!pmd_huge(*pmd))
4979 pte = huge_ptep_get((pte_t *)pmd);
4980 if (pte_present(pte)) {
4981 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4982 if (flags & FOLL_GET)
4985 if (is_hugetlb_entry_migration(pte)) {
4987 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4991 * hwpoisoned entry is treated as no_page_table in
4992 * follow_page_mask().
5000 struct page * __weak
5001 follow_huge_pud(struct mm_struct *mm, unsigned long address,
5002 pud_t *pud, int flags)
5004 if (flags & FOLL_GET)
5007 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5010 struct page * __weak
5011 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5013 if (flags & FOLL_GET)
5016 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5019 bool isolate_huge_page(struct page *page, struct list_head *list)
5023 spin_lock(&hugetlb_lock);
5024 if (!PageHeadHuge(page) || !page_huge_active(page) ||
5025 !get_page_unless_zero(page)) {
5029 clear_page_huge_active(page);
5030 list_move_tail(&page->lru, list);
5032 spin_unlock(&hugetlb_lock);
5036 void putback_active_hugepage(struct page *page)
5038 VM_BUG_ON_PAGE(!PageHead(page), page);
5039 spin_lock(&hugetlb_lock);
5040 set_page_huge_active(page);
5041 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5042 spin_unlock(&hugetlb_lock);
5046 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5048 struct hstate *h = page_hstate(oldpage);
5050 hugetlb_cgroup_migrate(oldpage, newpage);
5051 set_page_owner_migrate_reason(newpage, reason);
5054 * transfer temporary state of the new huge page. This is
5055 * reverse to other transitions because the newpage is going to
5056 * be final while the old one will be freed so it takes over
5057 * the temporary status.
5059 * Also note that we have to transfer the per-node surplus state
5060 * here as well otherwise the global surplus count will not match
5063 if (PageHugeTemporary(newpage)) {
5064 int old_nid = page_to_nid(oldpage);
5065 int new_nid = page_to_nid(newpage);
5067 SetPageHugeTemporary(oldpage);
5068 ClearPageHugeTemporary(newpage);
5070 spin_lock(&hugetlb_lock);
5071 if (h->surplus_huge_pages_node[old_nid]) {
5072 h->surplus_huge_pages_node[old_nid]--;
5073 h->surplus_huge_pages_node[new_nid]++;
5075 spin_unlock(&hugetlb_lock);