1 // SPDX-License-Identifier: GPL-2.0-only
3 * Generic hugetlb support.
4 * (C) Nadia Yvette Chambers, April 2004
6 #include <linux/list.h>
7 #include <linux/init.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/mmdebug.h>
23 #include <linux/sched/signal.h>
24 #include <linux/rmap.h>
25 #include <linux/string_helpers.h>
26 #include <linux/swap.h>
27 #include <linux/swapops.h>
28 #include <linux/jhash.h>
29 #include <linux/numa.h>
30 #include <linux/llist.h>
33 #include <asm/pgtable.h>
37 #include <linux/hugetlb.h>
38 #include <linux/hugetlb_cgroup.h>
39 #include <linux/node.h>
40 #include <linux/userfaultfd_k.h>
41 #include <linux/page_owner.h>
44 int hugetlb_max_hstate __read_mostly;
45 unsigned int default_hstate_idx;
46 struct hstate hstates[HUGE_MAX_HSTATE];
48 * Minimum page order among possible hugepage sizes, set to a proper value
51 static unsigned int minimum_order __read_mostly = UINT_MAX;
53 __initdata LIST_HEAD(huge_boot_pages);
55 /* for command line parsing */
56 static struct hstate * __initdata parsed_hstate;
57 static unsigned long __initdata default_hstate_max_huge_pages;
58 static unsigned long __initdata default_hstate_size;
59 static bool __initdata parsed_valid_hugepagesz = true;
62 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
63 * free_huge_pages, and surplus_huge_pages.
65 DEFINE_SPINLOCK(hugetlb_lock);
68 * Serializes faults on the same logical page. This is used to
69 * prevent spurious OOMs when the hugepage pool is fully utilized.
71 static int num_fault_mutexes;
72 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
74 /* Forward declaration */
75 static int hugetlb_acct_memory(struct hstate *h, long delta);
77 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
79 bool free = (spool->count == 0) && (spool->used_hpages == 0);
81 spin_unlock(&spool->lock);
83 /* If no pages are used, and no other handles to the subpool
84 * remain, give up any reservations mased on minimum size and
87 if (spool->min_hpages != -1)
88 hugetlb_acct_memory(spool->hstate,
94 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
97 struct hugepage_subpool *spool;
99 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
103 spin_lock_init(&spool->lock);
105 spool->max_hpages = max_hpages;
107 spool->min_hpages = min_hpages;
109 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
113 spool->rsv_hpages = min_hpages;
118 void hugepage_put_subpool(struct hugepage_subpool *spool)
120 spin_lock(&spool->lock);
121 BUG_ON(!spool->count);
123 unlock_or_release_subpool(spool);
127 * Subpool accounting for allocating and reserving pages.
128 * Return -ENOMEM if there are not enough resources to satisfy the
129 * the request. Otherwise, return the number of pages by which the
130 * global pools must be adjusted (upward). The returned value may
131 * only be different than the passed value (delta) in the case where
132 * a subpool minimum size must be manitained.
134 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
142 spin_lock(&spool->lock);
144 if (spool->max_hpages != -1) { /* maximum size accounting */
145 if ((spool->used_hpages + delta) <= spool->max_hpages)
146 spool->used_hpages += delta;
153 /* minimum size accounting */
154 if (spool->min_hpages != -1 && spool->rsv_hpages) {
155 if (delta > spool->rsv_hpages) {
157 * Asking for more reserves than those already taken on
158 * behalf of subpool. Return difference.
160 ret = delta - spool->rsv_hpages;
161 spool->rsv_hpages = 0;
163 ret = 0; /* reserves already accounted for */
164 spool->rsv_hpages -= delta;
169 spin_unlock(&spool->lock);
174 * Subpool accounting for freeing and unreserving pages.
175 * Return the number of global page reservations that must be dropped.
176 * The return value may only be different than the passed value (delta)
177 * in the case where a subpool minimum size must be maintained.
179 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
187 spin_lock(&spool->lock);
189 if (spool->max_hpages != -1) /* maximum size accounting */
190 spool->used_hpages -= delta;
192 /* minimum size accounting */
193 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
194 if (spool->rsv_hpages + delta <= spool->min_hpages)
197 ret = spool->rsv_hpages + delta - spool->min_hpages;
199 spool->rsv_hpages += delta;
200 if (spool->rsv_hpages > spool->min_hpages)
201 spool->rsv_hpages = spool->min_hpages;
205 * If hugetlbfs_put_super couldn't free spool due to an outstanding
206 * quota reference, free it now.
208 unlock_or_release_subpool(spool);
213 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
215 return HUGETLBFS_SB(inode->i_sb)->spool;
218 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
220 return subpool_inode(file_inode(vma->vm_file));
224 * Region tracking -- allows tracking of reservations and instantiated pages
225 * across the pages in a mapping.
227 * The region data structures are embedded into a resv_map and protected
228 * by a resv_map's lock. The set of regions within the resv_map represent
229 * reservations for huge pages, or huge pages that have already been
230 * instantiated within the map. The from and to elements are huge page
231 * indicies into the associated mapping. from indicates the starting index
232 * of the region. to represents the first index past the end of the region.
234 * For example, a file region structure with from == 0 and to == 4 represents
235 * four huge pages in a mapping. It is important to note that the to element
236 * represents the first element past the end of the region. This is used in
237 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
239 * Interval notation of the form [from, to) will be used to indicate that
240 * the endpoint from is inclusive and to is exclusive.
243 struct list_head link;
248 /* Must be called with resv->lock held. Calling this with count_only == true
249 * will count the number of pages to be added but will not modify the linked
252 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
256 struct list_head *head = &resv->regions;
257 struct file_region *rg = NULL, *trg = NULL, *nrg = NULL;
259 /* Locate the region we are before or in. */
260 list_for_each_entry(rg, head, link)
264 /* Round our left edge to the current segment if it encloses us. */
270 /* Check for and consume any regions we now overlap with. */
272 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
273 if (&rg->link == head)
278 /* We overlap with this area, if it extends further than
279 * us then we must extend ourselves. Account for its
280 * existing reservation.
286 chg -= rg->to - rg->from;
288 if (!count_only && rg != nrg) {
303 * Add the huge page range represented by [f, t) to the reserve
304 * map. Existing regions will be expanded to accommodate the specified
305 * range, or a region will be taken from the cache. Sufficient regions
306 * must exist in the cache due to the previous call to region_chg with
309 * Return the number of new huge pages added to the map. This
310 * number is greater than or equal to zero.
312 static long region_add(struct resv_map *resv, long f, long t)
314 struct list_head *head = &resv->regions;
315 struct file_region *rg, *nrg;
318 spin_lock(&resv->lock);
319 /* Locate the region we are either in or before. */
320 list_for_each_entry(rg, head, link)
325 * If no region exists which can be expanded to include the
326 * specified range, pull a region descriptor from the cache
327 * and use it for this range.
329 if (&rg->link == head || t < rg->from) {
330 VM_BUG_ON(resv->region_cache_count <= 0);
332 resv->region_cache_count--;
333 nrg = list_first_entry(&resv->region_cache, struct file_region,
335 list_del(&nrg->link);
339 list_add(&nrg->link, rg->link.prev);
345 add = add_reservation_in_range(resv, f, t, false);
348 resv->adds_in_progress--;
349 spin_unlock(&resv->lock);
355 * Examine the existing reserve map and determine how many
356 * huge pages in the specified range [f, t) are NOT currently
357 * represented. This routine is called before a subsequent
358 * call to region_add that will actually modify the reserve
359 * map to add the specified range [f, t). region_chg does
360 * not change the number of huge pages represented by the
361 * map. A new file_region structure is added to the cache
362 * as a placeholder, so that the subsequent region_add
363 * call will have all the regions it needs and will not fail.
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)
374 spin_lock(&resv->lock);
376 resv->adds_in_progress++;
379 * Check for sufficient descriptors in the cache to accommodate
380 * the number of in progress add operations.
382 if (resv->adds_in_progress > resv->region_cache_count) {
383 struct file_region *trg;
385 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
386 /* Must drop lock to allocate a new descriptor. */
387 resv->adds_in_progress--;
388 spin_unlock(&resv->lock);
390 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
394 spin_lock(&resv->lock);
395 list_add(&trg->link, &resv->region_cache);
396 resv->region_cache_count++;
400 chg = add_reservation_in_range(resv, f, t, true);
402 spin_unlock(&resv->lock);
407 * Abort the in progress add operation. The adds_in_progress field
408 * of the resv_map keeps track of the operations in progress between
409 * calls to region_chg and region_add. Operations are sometimes
410 * aborted after the call to region_chg. In such cases, region_abort
411 * is called to decrement the adds_in_progress counter.
413 * NOTE: The range arguments [f, t) are not needed or used in this
414 * routine. They are kept to make reading the calling code easier as
415 * arguments will match the associated region_chg call.
417 static void region_abort(struct resv_map *resv, long f, long t)
419 spin_lock(&resv->lock);
420 VM_BUG_ON(!resv->region_cache_count);
421 resv->adds_in_progress--;
422 spin_unlock(&resv->lock);
426 * Delete the specified range [f, t) from the reserve map. If the
427 * t parameter is LONG_MAX, this indicates that ALL regions after f
428 * should be deleted. Locate the regions which intersect [f, t)
429 * and either trim, delete or split the existing regions.
431 * Returns the number of huge pages deleted from the reserve map.
432 * In the normal case, the return value is zero or more. In the
433 * case where a region must be split, a new region descriptor must
434 * be allocated. If the allocation fails, -ENOMEM will be returned.
435 * NOTE: If the parameter t == LONG_MAX, then we will never split
436 * a region and possibly return -ENOMEM. Callers specifying
437 * t == LONG_MAX do not need to check for -ENOMEM error.
439 static long region_del(struct resv_map *resv, long f, long t)
441 struct list_head *head = &resv->regions;
442 struct file_region *rg, *trg;
443 struct file_region *nrg = NULL;
447 spin_lock(&resv->lock);
448 list_for_each_entry_safe(rg, trg, head, link) {
450 * Skip regions before the range to be deleted. file_region
451 * ranges are normally of the form [from, to). However, there
452 * may be a "placeholder" entry in the map which is of the form
453 * (from, to) with from == to. Check for placeholder entries
454 * at the beginning of the range to be deleted.
456 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
462 if (f > rg->from && t < rg->to) { /* Must split region */
464 * Check for an entry in the cache before dropping
465 * lock and attempting allocation.
468 resv->region_cache_count > resv->adds_in_progress) {
469 nrg = list_first_entry(&resv->region_cache,
472 list_del(&nrg->link);
473 resv->region_cache_count--;
477 spin_unlock(&resv->lock);
478 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
486 /* New entry for end of split region */
489 INIT_LIST_HEAD(&nrg->link);
491 /* Original entry is trimmed */
494 list_add(&nrg->link, &rg->link);
499 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
500 del += rg->to - rg->from;
506 if (f <= rg->from) { /* Trim beginning of region */
509 } else { /* Trim end of region */
515 spin_unlock(&resv->lock);
521 * A rare out of memory error was encountered which prevented removal of
522 * the reserve map region for a page. The huge page itself was free'ed
523 * and removed from the page cache. This routine will adjust the subpool
524 * usage count, and the global reserve count if needed. By incrementing
525 * these counts, the reserve map entry which could not be deleted will
526 * appear as a "reserved" entry instead of simply dangling with incorrect
529 void hugetlb_fix_reserve_counts(struct inode *inode)
531 struct hugepage_subpool *spool = subpool_inode(inode);
534 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
536 struct hstate *h = hstate_inode(inode);
538 hugetlb_acct_memory(h, 1);
543 * Count and return the number of huge pages in the reserve map
544 * that intersect with the range [f, t).
546 static long region_count(struct resv_map *resv, long f, long t)
548 struct list_head *head = &resv->regions;
549 struct file_region *rg;
552 spin_lock(&resv->lock);
553 /* Locate each segment we overlap with, and count that overlap. */
554 list_for_each_entry(rg, head, link) {
563 seg_from = max(rg->from, f);
564 seg_to = min(rg->to, t);
566 chg += seg_to - seg_from;
568 spin_unlock(&resv->lock);
574 * Convert the address within this vma to the page offset within
575 * the mapping, in pagecache page units; huge pages here.
577 static pgoff_t vma_hugecache_offset(struct hstate *h,
578 struct vm_area_struct *vma, unsigned long address)
580 return ((address - vma->vm_start) >> huge_page_shift(h)) +
581 (vma->vm_pgoff >> huge_page_order(h));
584 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
585 unsigned long address)
587 return vma_hugecache_offset(hstate_vma(vma), vma, address);
589 EXPORT_SYMBOL_GPL(linear_hugepage_index);
592 * Return the size of the pages allocated when backing a VMA. In the majority
593 * cases this will be same size as used by the page table entries.
595 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
597 if (vma->vm_ops && vma->vm_ops->pagesize)
598 return vma->vm_ops->pagesize(vma);
601 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
604 * Return the page size being used by the MMU to back a VMA. In the majority
605 * of cases, the page size used by the kernel matches the MMU size. On
606 * architectures where it differs, an architecture-specific 'strong'
607 * version of this symbol is required.
609 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
611 return vma_kernel_pagesize(vma);
615 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
616 * bits of the reservation map pointer, which are always clear due to
619 #define HPAGE_RESV_OWNER (1UL << 0)
620 #define HPAGE_RESV_UNMAPPED (1UL << 1)
621 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
624 * These helpers are used to track how many pages are reserved for
625 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
626 * is guaranteed to have their future faults succeed.
628 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
629 * the reserve counters are updated with the hugetlb_lock held. It is safe
630 * to reset the VMA at fork() time as it is not in use yet and there is no
631 * chance of the global counters getting corrupted as a result of the values.
633 * The private mapping reservation is represented in a subtly different
634 * manner to a shared mapping. A shared mapping has a region map associated
635 * with the underlying file, this region map represents the backing file
636 * pages which have ever had a reservation assigned which this persists even
637 * after the page is instantiated. A private mapping has a region map
638 * associated with the original mmap which is attached to all VMAs which
639 * reference it, this region map represents those offsets which have consumed
640 * reservation ie. where pages have been instantiated.
642 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
644 return (unsigned long)vma->vm_private_data;
647 static void set_vma_private_data(struct vm_area_struct *vma,
650 vma->vm_private_data = (void *)value;
653 struct resv_map *resv_map_alloc(void)
655 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
656 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
658 if (!resv_map || !rg) {
664 kref_init(&resv_map->refs);
665 spin_lock_init(&resv_map->lock);
666 INIT_LIST_HEAD(&resv_map->regions);
668 resv_map->adds_in_progress = 0;
670 INIT_LIST_HEAD(&resv_map->region_cache);
671 list_add(&rg->link, &resv_map->region_cache);
672 resv_map->region_cache_count = 1;
677 void resv_map_release(struct kref *ref)
679 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
680 struct list_head *head = &resv_map->region_cache;
681 struct file_region *rg, *trg;
683 /* Clear out any active regions before we release the map. */
684 region_del(resv_map, 0, LONG_MAX);
686 /* ... and any entries left in the cache */
687 list_for_each_entry_safe(rg, trg, head, link) {
692 VM_BUG_ON(resv_map->adds_in_progress);
697 static inline struct resv_map *inode_resv_map(struct inode *inode)
700 * At inode evict time, i_mapping may not point to the original
701 * address space within the inode. This original address space
702 * contains the pointer to the resv_map. So, always use the
703 * address space embedded within the inode.
704 * The VERY common case is inode->mapping == &inode->i_data but,
705 * this may not be true for device special inodes.
707 return (struct resv_map *)(&inode->i_data)->private_data;
710 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
712 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
713 if (vma->vm_flags & VM_MAYSHARE) {
714 struct address_space *mapping = vma->vm_file->f_mapping;
715 struct inode *inode = mapping->host;
717 return inode_resv_map(inode);
720 return (struct resv_map *)(get_vma_private_data(vma) &
725 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
727 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
728 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
730 set_vma_private_data(vma, (get_vma_private_data(vma) &
731 HPAGE_RESV_MASK) | (unsigned long)map);
734 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
736 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
737 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
739 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
742 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
744 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
746 return (get_vma_private_data(vma) & flag) != 0;
749 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
750 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
752 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
753 if (!(vma->vm_flags & VM_MAYSHARE))
754 vma->vm_private_data = (void *)0;
757 /* Returns true if the VMA has associated reserve pages */
758 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
760 if (vma->vm_flags & VM_NORESERVE) {
762 * This address is already reserved by other process(chg == 0),
763 * so, we should decrement reserved count. Without decrementing,
764 * reserve count remains after releasing inode, because this
765 * allocated page will go into page cache and is regarded as
766 * coming from reserved pool in releasing step. Currently, we
767 * don't have any other solution to deal with this situation
768 * properly, so add work-around here.
770 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
776 /* Shared mappings always use reserves */
777 if (vma->vm_flags & VM_MAYSHARE) {
779 * We know VM_NORESERVE is not set. Therefore, there SHOULD
780 * be a region map for all pages. The only situation where
781 * there is no region map is if a hole was punched via
782 * fallocate. In this case, there really are no reverves to
783 * use. This situation is indicated if chg != 0.
792 * Only the process that called mmap() has reserves for
795 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
797 * Like the shared case above, a hole punch or truncate
798 * could have been performed on the private mapping.
799 * Examine the value of chg to determine if reserves
800 * actually exist or were previously consumed.
801 * Very Subtle - The value of chg comes from a previous
802 * call to vma_needs_reserves(). The reserve map for
803 * private mappings has different (opposite) semantics
804 * than that of shared mappings. vma_needs_reserves()
805 * has already taken this difference in semantics into
806 * account. Therefore, the meaning of chg is the same
807 * as in the shared case above. Code could easily be
808 * combined, but keeping it separate draws attention to
809 * subtle differences.
820 static void enqueue_huge_page(struct hstate *h, struct page *page)
822 int nid = page_to_nid(page);
823 list_move(&page->lru, &h->hugepage_freelists[nid]);
824 h->free_huge_pages++;
825 h->free_huge_pages_node[nid]++;
828 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
832 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
833 if (!PageHWPoison(page))
836 * if 'non-isolated free hugepage' not found on the list,
837 * the allocation fails.
839 if (&h->hugepage_freelists[nid] == &page->lru)
841 list_move(&page->lru, &h->hugepage_activelist);
842 set_page_refcounted(page);
843 h->free_huge_pages--;
844 h->free_huge_pages_node[nid]--;
848 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
851 unsigned int cpuset_mems_cookie;
852 struct zonelist *zonelist;
855 int node = NUMA_NO_NODE;
857 zonelist = node_zonelist(nid, gfp_mask);
860 cpuset_mems_cookie = read_mems_allowed_begin();
861 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
864 if (!cpuset_zone_allowed(zone, gfp_mask))
867 * no need to ask again on the same node. Pool is node rather than
870 if (zone_to_nid(zone) == node)
872 node = zone_to_nid(zone);
874 page = dequeue_huge_page_node_exact(h, node);
878 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
884 /* Movability of hugepages depends on migration support. */
885 static inline gfp_t htlb_alloc_mask(struct hstate *h)
887 if (hugepage_movable_supported(h))
888 return GFP_HIGHUSER_MOVABLE;
893 static struct page *dequeue_huge_page_vma(struct hstate *h,
894 struct vm_area_struct *vma,
895 unsigned long address, int avoid_reserve,
899 struct mempolicy *mpol;
901 nodemask_t *nodemask;
905 * A child process with MAP_PRIVATE mappings created by their parent
906 * have no page reserves. This check ensures that reservations are
907 * not "stolen". The child may still get SIGKILLed
909 if (!vma_has_reserves(vma, chg) &&
910 h->free_huge_pages - h->resv_huge_pages == 0)
913 /* If reserves cannot be used, ensure enough pages are in the pool */
914 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
917 gfp_mask = htlb_alloc_mask(h);
918 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
919 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
920 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
921 SetPagePrivate(page);
922 h->resv_huge_pages--;
933 * common helper functions for hstate_next_node_to_{alloc|free}.
934 * We may have allocated or freed a huge page based on a different
935 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
936 * be outside of *nodes_allowed. Ensure that we use an allowed
937 * node for alloc or free.
939 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
941 nid = next_node_in(nid, *nodes_allowed);
942 VM_BUG_ON(nid >= MAX_NUMNODES);
947 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
949 if (!node_isset(nid, *nodes_allowed))
950 nid = next_node_allowed(nid, nodes_allowed);
955 * returns the previously saved node ["this node"] from which to
956 * allocate a persistent huge page for the pool and advance the
957 * next node from which to allocate, handling wrap at end of node
960 static int hstate_next_node_to_alloc(struct hstate *h,
961 nodemask_t *nodes_allowed)
965 VM_BUG_ON(!nodes_allowed);
967 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
968 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
974 * helper for free_pool_huge_page() - return the previously saved
975 * node ["this node"] from which to free a huge page. Advance the
976 * next node id whether or not we find a free huge page to free so
977 * that the next attempt to free addresses the next node.
979 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
983 VM_BUG_ON(!nodes_allowed);
985 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
986 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
991 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
992 for (nr_nodes = nodes_weight(*mask); \
994 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
997 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
998 for (nr_nodes = nodes_weight(*mask); \
1000 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1003 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1004 static void destroy_compound_gigantic_page(struct page *page,
1008 int nr_pages = 1 << order;
1009 struct page *p = page + 1;
1011 atomic_set(compound_mapcount_ptr(page), 0);
1012 if (hpage_pincount_available(page))
1013 atomic_set(compound_pincount_ptr(page), 0);
1015 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1016 clear_compound_head(p);
1017 set_page_refcounted(p);
1020 set_compound_order(page, 0);
1021 __ClearPageHead(page);
1024 static void free_gigantic_page(struct page *page, unsigned int order)
1026 free_contig_range(page_to_pfn(page), 1 << order);
1029 #ifdef CONFIG_CONTIG_ALLOC
1030 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1031 int nid, nodemask_t *nodemask)
1033 unsigned long nr_pages = 1UL << huge_page_order(h);
1035 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1038 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1039 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1040 #else /* !CONFIG_CONTIG_ALLOC */
1041 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1042 int nid, nodemask_t *nodemask)
1046 #endif /* CONFIG_CONTIG_ALLOC */
1048 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1049 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1050 int nid, nodemask_t *nodemask)
1054 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1055 static inline void destroy_compound_gigantic_page(struct page *page,
1056 unsigned int order) { }
1059 static void update_and_free_page(struct hstate *h, struct page *page)
1063 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1067 h->nr_huge_pages_node[page_to_nid(page)]--;
1068 for (i = 0; i < pages_per_huge_page(h); i++) {
1069 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1070 1 << PG_referenced | 1 << PG_dirty |
1071 1 << PG_active | 1 << PG_private |
1074 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1075 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1076 set_page_refcounted(page);
1077 if (hstate_is_gigantic(h)) {
1078 destroy_compound_gigantic_page(page, huge_page_order(h));
1079 free_gigantic_page(page, huge_page_order(h));
1081 __free_pages(page, huge_page_order(h));
1085 struct hstate *size_to_hstate(unsigned long size)
1089 for_each_hstate(h) {
1090 if (huge_page_size(h) == size)
1097 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1098 * to hstate->hugepage_activelist.)
1100 * This function can be called for tail pages, but never returns true for them.
1102 bool page_huge_active(struct page *page)
1104 VM_BUG_ON_PAGE(!PageHuge(page), page);
1105 return PageHead(page) && PagePrivate(&page[1]);
1108 /* never called for tail page */
1109 static void set_page_huge_active(struct page *page)
1111 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1112 SetPagePrivate(&page[1]);
1115 static void clear_page_huge_active(struct page *page)
1117 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1118 ClearPagePrivate(&page[1]);
1122 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1125 static inline bool PageHugeTemporary(struct page *page)
1127 if (!PageHuge(page))
1130 return (unsigned long)page[2].mapping == -1U;
1133 static inline void SetPageHugeTemporary(struct page *page)
1135 page[2].mapping = (void *)-1U;
1138 static inline void ClearPageHugeTemporary(struct page *page)
1140 page[2].mapping = NULL;
1143 static void __free_huge_page(struct page *page)
1146 * Can't pass hstate in here because it is called from the
1147 * compound page destructor.
1149 struct hstate *h = page_hstate(page);
1150 int nid = page_to_nid(page);
1151 struct hugepage_subpool *spool =
1152 (struct hugepage_subpool *)page_private(page);
1153 bool restore_reserve;
1155 VM_BUG_ON_PAGE(page_count(page), page);
1156 VM_BUG_ON_PAGE(page_mapcount(page), page);
1158 set_page_private(page, 0);
1159 page->mapping = NULL;
1160 restore_reserve = PagePrivate(page);
1161 ClearPagePrivate(page);
1164 * If PagePrivate() was set on page, page allocation consumed a
1165 * reservation. If the page was associated with a subpool, there
1166 * would have been a page reserved in the subpool before allocation
1167 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1168 * reservtion, do not call hugepage_subpool_put_pages() as this will
1169 * remove the reserved page from the subpool.
1171 if (!restore_reserve) {
1173 * A return code of zero implies that the subpool will be
1174 * under its minimum size if the reservation is not restored
1175 * after page is free. Therefore, force restore_reserve
1178 if (hugepage_subpool_put_pages(spool, 1) == 0)
1179 restore_reserve = true;
1182 spin_lock(&hugetlb_lock);
1183 clear_page_huge_active(page);
1184 hugetlb_cgroup_uncharge_page(hstate_index(h),
1185 pages_per_huge_page(h), page);
1186 if (restore_reserve)
1187 h->resv_huge_pages++;
1189 if (PageHugeTemporary(page)) {
1190 list_del(&page->lru);
1191 ClearPageHugeTemporary(page);
1192 update_and_free_page(h, page);
1193 } else if (h->surplus_huge_pages_node[nid]) {
1194 /* remove the page from active list */
1195 list_del(&page->lru);
1196 update_and_free_page(h, page);
1197 h->surplus_huge_pages--;
1198 h->surplus_huge_pages_node[nid]--;
1200 arch_clear_hugepage_flags(page);
1201 enqueue_huge_page(h, page);
1203 spin_unlock(&hugetlb_lock);
1207 * As free_huge_page() can be called from a non-task context, we have
1208 * to defer the actual freeing in a workqueue to prevent potential
1209 * hugetlb_lock deadlock.
1211 * free_hpage_workfn() locklessly retrieves the linked list of pages to
1212 * be freed and frees them one-by-one. As the page->mapping pointer is
1213 * going to be cleared in __free_huge_page() anyway, it is reused as the
1214 * llist_node structure of a lockless linked list of huge pages to be freed.
1216 static LLIST_HEAD(hpage_freelist);
1218 static void free_hpage_workfn(struct work_struct *work)
1220 struct llist_node *node;
1223 node = llist_del_all(&hpage_freelist);
1226 page = container_of((struct address_space **)node,
1227 struct page, mapping);
1229 __free_huge_page(page);
1232 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1234 void free_huge_page(struct page *page)
1237 * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1241 * Only call schedule_work() if hpage_freelist is previously
1242 * empty. Otherwise, schedule_work() had been called but the
1243 * workfn hasn't retrieved the list yet.
1245 if (llist_add((struct llist_node *)&page->mapping,
1247 schedule_work(&free_hpage_work);
1251 __free_huge_page(page);
1254 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1256 INIT_LIST_HEAD(&page->lru);
1257 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1258 spin_lock(&hugetlb_lock);
1259 set_hugetlb_cgroup(page, NULL);
1261 h->nr_huge_pages_node[nid]++;
1262 spin_unlock(&hugetlb_lock);
1265 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1268 int nr_pages = 1 << order;
1269 struct page *p = page + 1;
1271 /* we rely on prep_new_huge_page to set the destructor */
1272 set_compound_order(page, order);
1273 __ClearPageReserved(page);
1274 __SetPageHead(page);
1275 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1277 * For gigantic hugepages allocated through bootmem at
1278 * boot, it's safer to be consistent with the not-gigantic
1279 * hugepages and clear the PG_reserved bit from all tail pages
1280 * too. Otherwse drivers using get_user_pages() to access tail
1281 * pages may get the reference counting wrong if they see
1282 * PG_reserved set on a tail page (despite the head page not
1283 * having PG_reserved set). Enforcing this consistency between
1284 * head and tail pages allows drivers to optimize away a check
1285 * on the head page when they need know if put_page() is needed
1286 * after get_user_pages().
1288 __ClearPageReserved(p);
1289 set_page_count(p, 0);
1290 set_compound_head(p, page);
1292 atomic_set(compound_mapcount_ptr(page), -1);
1294 if (hpage_pincount_available(page))
1295 atomic_set(compound_pincount_ptr(page), 0);
1299 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1300 * transparent huge pages. See the PageTransHuge() documentation for more
1303 int PageHuge(struct page *page)
1305 if (!PageCompound(page))
1308 page = compound_head(page);
1309 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1311 EXPORT_SYMBOL_GPL(PageHuge);
1314 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1315 * normal or transparent huge pages.
1317 int PageHeadHuge(struct page *page_head)
1319 if (!PageHead(page_head))
1322 return get_compound_page_dtor(page_head) == free_huge_page;
1326 * Find address_space associated with hugetlbfs page.
1327 * Upon entry page is locked and page 'was' mapped although mapped state
1328 * could change. If necessary, use anon_vma to find vma and associated
1329 * address space. The returned mapping may be stale, but it can not be
1330 * invalid as page lock (which is held) is required to destroy mapping.
1332 static struct address_space *_get_hugetlb_page_mapping(struct page *hpage)
1334 struct anon_vma *anon_vma;
1335 pgoff_t pgoff_start, pgoff_end;
1336 struct anon_vma_chain *avc;
1337 struct address_space *mapping = page_mapping(hpage);
1339 /* Simple file based mapping */
1344 * Even anonymous hugetlbfs mappings are associated with an
1345 * underlying hugetlbfs file (see hugetlb_file_setup in mmap
1346 * code). Find a vma associated with the anonymous vma, and
1347 * use the file pointer to get address_space.
1349 anon_vma = page_lock_anon_vma_read(hpage);
1351 return mapping; /* NULL */
1353 /* Use first found vma */
1354 pgoff_start = page_to_pgoff(hpage);
1355 pgoff_end = pgoff_start + hpage_nr_pages(hpage) - 1;
1356 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1357 pgoff_start, pgoff_end) {
1358 struct vm_area_struct *vma = avc->vma;
1360 mapping = vma->vm_file->f_mapping;
1364 anon_vma_unlock_read(anon_vma);
1369 * Find and lock address space (mapping) in write mode.
1371 * Upon entry, the page is locked which allows us to find the mapping
1372 * even in the case of an anon page. However, locking order dictates
1373 * the i_mmap_rwsem be acquired BEFORE the page lock. This is hugetlbfs
1374 * specific. So, we first try to lock the sema while still holding the
1375 * page lock. If this works, great! If not, then we need to drop the
1376 * page lock and then acquire i_mmap_rwsem and reacquire page lock. Of
1377 * course, need to revalidate state along the way.
1379 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1381 struct address_space *mapping, *mapping2;
1383 mapping = _get_hugetlb_page_mapping(hpage);
1389 * If no contention, take lock and return
1391 if (i_mmap_trylock_write(mapping))
1395 * Must drop page lock and wait on mapping sema.
1396 * Note: Once page lock is dropped, mapping could become invalid.
1397 * As a hack, increase map count until we lock page again.
1399 atomic_inc(&hpage->_mapcount);
1401 i_mmap_lock_write(mapping);
1403 atomic_add_negative(-1, &hpage->_mapcount);
1405 /* verify page is still mapped */
1406 if (!page_mapped(hpage)) {
1407 i_mmap_unlock_write(mapping);
1412 * Get address space again and verify it is the same one
1413 * we locked. If not, drop lock and retry.
1415 mapping2 = _get_hugetlb_page_mapping(hpage);
1416 if (mapping2 != mapping) {
1417 i_mmap_unlock_write(mapping);
1425 pgoff_t __basepage_index(struct page *page)
1427 struct page *page_head = compound_head(page);
1428 pgoff_t index = page_index(page_head);
1429 unsigned long compound_idx;
1431 if (!PageHuge(page_head))
1432 return page_index(page);
1434 if (compound_order(page_head) >= MAX_ORDER)
1435 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1437 compound_idx = page - page_head;
1439 return (index << compound_order(page_head)) + compound_idx;
1442 static struct page *alloc_buddy_huge_page(struct hstate *h,
1443 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1444 nodemask_t *node_alloc_noretry)
1446 int order = huge_page_order(h);
1448 bool alloc_try_hard = true;
1451 * By default we always try hard to allocate the page with
1452 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1453 * a loop (to adjust global huge page counts) and previous allocation
1454 * failed, do not continue to try hard on the same node. Use the
1455 * node_alloc_noretry bitmap to manage this state information.
1457 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1458 alloc_try_hard = false;
1459 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1461 gfp_mask |= __GFP_RETRY_MAYFAIL;
1462 if (nid == NUMA_NO_NODE)
1463 nid = numa_mem_id();
1464 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1466 __count_vm_event(HTLB_BUDDY_PGALLOC);
1468 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1471 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1472 * indicates an overall state change. Clear bit so that we resume
1473 * normal 'try hard' allocations.
1475 if (node_alloc_noretry && page && !alloc_try_hard)
1476 node_clear(nid, *node_alloc_noretry);
1479 * If we tried hard to get a page but failed, set bit so that
1480 * subsequent attempts will not try as hard until there is an
1481 * overall state change.
1483 if (node_alloc_noretry && !page && alloc_try_hard)
1484 node_set(nid, *node_alloc_noretry);
1490 * Common helper to allocate a fresh hugetlb page. All specific allocators
1491 * should use this function to get new hugetlb pages
1493 static struct page *alloc_fresh_huge_page(struct hstate *h,
1494 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1495 nodemask_t *node_alloc_noretry)
1499 if (hstate_is_gigantic(h))
1500 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1502 page = alloc_buddy_huge_page(h, gfp_mask,
1503 nid, nmask, node_alloc_noretry);
1507 if (hstate_is_gigantic(h))
1508 prep_compound_gigantic_page(page, huge_page_order(h));
1509 prep_new_huge_page(h, page, page_to_nid(page));
1515 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1518 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1519 nodemask_t *node_alloc_noretry)
1523 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1525 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1526 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1527 node_alloc_noretry);
1535 put_page(page); /* free it into the hugepage allocator */
1541 * Free huge page from pool from next node to free.
1542 * Attempt to keep persistent huge pages more or less
1543 * balanced over allowed nodes.
1544 * Called with hugetlb_lock locked.
1546 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1552 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1554 * If we're returning unused surplus pages, only examine
1555 * nodes with surplus pages.
1557 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1558 !list_empty(&h->hugepage_freelists[node])) {
1560 list_entry(h->hugepage_freelists[node].next,
1562 list_del(&page->lru);
1563 h->free_huge_pages--;
1564 h->free_huge_pages_node[node]--;
1566 h->surplus_huge_pages--;
1567 h->surplus_huge_pages_node[node]--;
1569 update_and_free_page(h, page);
1579 * Dissolve a given free hugepage into free buddy pages. This function does
1580 * nothing for in-use hugepages and non-hugepages.
1581 * This function returns values like below:
1583 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1584 * (allocated or reserved.)
1585 * 0: successfully dissolved free hugepages or the page is not a
1586 * hugepage (considered as already dissolved)
1588 int dissolve_free_huge_page(struct page *page)
1592 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1593 if (!PageHuge(page))
1596 spin_lock(&hugetlb_lock);
1597 if (!PageHuge(page)) {
1602 if (!page_count(page)) {
1603 struct page *head = compound_head(page);
1604 struct hstate *h = page_hstate(head);
1605 int nid = page_to_nid(head);
1606 if (h->free_huge_pages - h->resv_huge_pages == 0)
1609 * Move PageHWPoison flag from head page to the raw error page,
1610 * which makes any subpages rather than the error page reusable.
1612 if (PageHWPoison(head) && page != head) {
1613 SetPageHWPoison(page);
1614 ClearPageHWPoison(head);
1616 list_del(&head->lru);
1617 h->free_huge_pages--;
1618 h->free_huge_pages_node[nid]--;
1619 h->max_huge_pages--;
1620 update_and_free_page(h, head);
1624 spin_unlock(&hugetlb_lock);
1629 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1630 * make specified memory blocks removable from the system.
1631 * Note that this will dissolve a free gigantic hugepage completely, if any
1632 * part of it lies within the given range.
1633 * Also note that if dissolve_free_huge_page() returns with an error, all
1634 * free hugepages that were dissolved before that error are lost.
1636 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1642 if (!hugepages_supported())
1645 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1646 page = pfn_to_page(pfn);
1647 rc = dissolve_free_huge_page(page);
1656 * Allocates a fresh surplus page from the page allocator.
1658 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1659 int nid, nodemask_t *nmask)
1661 struct page *page = NULL;
1663 if (hstate_is_gigantic(h))
1666 spin_lock(&hugetlb_lock);
1667 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1669 spin_unlock(&hugetlb_lock);
1671 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1675 spin_lock(&hugetlb_lock);
1677 * We could have raced with the pool size change.
1678 * Double check that and simply deallocate the new page
1679 * if we would end up overcommiting the surpluses. Abuse
1680 * temporary page to workaround the nasty free_huge_page
1683 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1684 SetPageHugeTemporary(page);
1685 spin_unlock(&hugetlb_lock);
1689 h->surplus_huge_pages++;
1690 h->surplus_huge_pages_node[page_to_nid(page)]++;
1694 spin_unlock(&hugetlb_lock);
1699 struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1700 int nid, nodemask_t *nmask)
1704 if (hstate_is_gigantic(h))
1707 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1712 * We do not account these pages as surplus because they are only
1713 * temporary and will be released properly on the last reference
1715 SetPageHugeTemporary(page);
1721 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1724 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1725 struct vm_area_struct *vma, unsigned long addr)
1728 struct mempolicy *mpol;
1729 gfp_t gfp_mask = htlb_alloc_mask(h);
1731 nodemask_t *nodemask;
1733 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1734 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1735 mpol_cond_put(mpol);
1740 /* page migration callback function */
1741 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1743 gfp_t gfp_mask = htlb_alloc_mask(h);
1744 struct page *page = NULL;
1746 if (nid != NUMA_NO_NODE)
1747 gfp_mask |= __GFP_THISNODE;
1749 spin_lock(&hugetlb_lock);
1750 if (h->free_huge_pages - h->resv_huge_pages > 0)
1751 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1752 spin_unlock(&hugetlb_lock);
1755 page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1760 /* page migration callback function */
1761 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1764 gfp_t gfp_mask = htlb_alloc_mask(h);
1766 spin_lock(&hugetlb_lock);
1767 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1770 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1772 spin_unlock(&hugetlb_lock);
1776 spin_unlock(&hugetlb_lock);
1778 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1781 /* mempolicy aware migration callback */
1782 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1783 unsigned long address)
1785 struct mempolicy *mpol;
1786 nodemask_t *nodemask;
1791 gfp_mask = htlb_alloc_mask(h);
1792 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1793 page = alloc_huge_page_nodemask(h, node, nodemask);
1794 mpol_cond_put(mpol);
1800 * Increase the hugetlb pool such that it can accommodate a reservation
1803 static int gather_surplus_pages(struct hstate *h, int delta)
1805 struct list_head surplus_list;
1806 struct page *page, *tmp;
1808 int needed, allocated;
1809 bool alloc_ok = true;
1811 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1813 h->resv_huge_pages += delta;
1818 INIT_LIST_HEAD(&surplus_list);
1822 spin_unlock(&hugetlb_lock);
1823 for (i = 0; i < needed; i++) {
1824 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1825 NUMA_NO_NODE, NULL);
1830 list_add(&page->lru, &surplus_list);
1836 * After retaking hugetlb_lock, we need to recalculate 'needed'
1837 * because either resv_huge_pages or free_huge_pages may have changed.
1839 spin_lock(&hugetlb_lock);
1840 needed = (h->resv_huge_pages + delta) -
1841 (h->free_huge_pages + allocated);
1846 * We were not able to allocate enough pages to
1847 * satisfy the entire reservation so we free what
1848 * we've allocated so far.
1853 * The surplus_list now contains _at_least_ the number of extra pages
1854 * needed to accommodate the reservation. Add the appropriate number
1855 * of pages to the hugetlb pool and free the extras back to the buddy
1856 * allocator. Commit the entire reservation here to prevent another
1857 * process from stealing the pages as they are added to the pool but
1858 * before they are reserved.
1860 needed += allocated;
1861 h->resv_huge_pages += delta;
1864 /* Free the needed pages to the hugetlb pool */
1865 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1869 * This page is now managed by the hugetlb allocator and has
1870 * no users -- drop the buddy allocator's reference.
1872 put_page_testzero(page);
1873 VM_BUG_ON_PAGE(page_count(page), page);
1874 enqueue_huge_page(h, page);
1877 spin_unlock(&hugetlb_lock);
1879 /* Free unnecessary surplus pages to the buddy allocator */
1880 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1882 spin_lock(&hugetlb_lock);
1888 * This routine has two main purposes:
1889 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1890 * in unused_resv_pages. This corresponds to the prior adjustments made
1891 * to the associated reservation map.
1892 * 2) Free any unused surplus pages that may have been allocated to satisfy
1893 * the reservation. As many as unused_resv_pages may be freed.
1895 * Called with hugetlb_lock held. However, the lock could be dropped (and
1896 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1897 * we must make sure nobody else can claim pages we are in the process of
1898 * freeing. Do this by ensuring resv_huge_page always is greater than the
1899 * number of huge pages we plan to free when dropping the lock.
1901 static void return_unused_surplus_pages(struct hstate *h,
1902 unsigned long unused_resv_pages)
1904 unsigned long nr_pages;
1906 /* Cannot return gigantic pages currently */
1907 if (hstate_is_gigantic(h))
1911 * Part (or even all) of the reservation could have been backed
1912 * by pre-allocated pages. Only free surplus pages.
1914 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1917 * We want to release as many surplus pages as possible, spread
1918 * evenly across all nodes with memory. Iterate across these nodes
1919 * until we can no longer free unreserved surplus pages. This occurs
1920 * when the nodes with surplus pages have no free pages.
1921 * free_pool_huge_page() will balance the the freed pages across the
1922 * on-line nodes with memory and will handle the hstate accounting.
1924 * Note that we decrement resv_huge_pages as we free the pages. If
1925 * we drop the lock, resv_huge_pages will still be sufficiently large
1926 * to cover subsequent pages we may free.
1928 while (nr_pages--) {
1929 h->resv_huge_pages--;
1930 unused_resv_pages--;
1931 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1933 cond_resched_lock(&hugetlb_lock);
1937 /* Fully uncommit the reservation */
1938 h->resv_huge_pages -= unused_resv_pages;
1943 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1944 * are used by the huge page allocation routines to manage reservations.
1946 * vma_needs_reservation is called to determine if the huge page at addr
1947 * within the vma has an associated reservation. If a reservation is
1948 * needed, the value 1 is returned. The caller is then responsible for
1949 * managing the global reservation and subpool usage counts. After
1950 * the huge page has been allocated, vma_commit_reservation is called
1951 * to add the page to the reservation map. If the page allocation fails,
1952 * the reservation must be ended instead of committed. vma_end_reservation
1953 * is called in such cases.
1955 * In the normal case, vma_commit_reservation returns the same value
1956 * as the preceding vma_needs_reservation call. The only time this
1957 * is not the case is if a reserve map was changed between calls. It
1958 * is the responsibility of the caller to notice the difference and
1959 * take appropriate action.
1961 * vma_add_reservation is used in error paths where a reservation must
1962 * be restored when a newly allocated huge page must be freed. It is
1963 * to be called after calling vma_needs_reservation to determine if a
1964 * reservation exists.
1966 enum vma_resv_mode {
1972 static long __vma_reservation_common(struct hstate *h,
1973 struct vm_area_struct *vma, unsigned long addr,
1974 enum vma_resv_mode mode)
1976 struct resv_map *resv;
1980 resv = vma_resv_map(vma);
1984 idx = vma_hugecache_offset(h, vma, addr);
1986 case VMA_NEEDS_RESV:
1987 ret = region_chg(resv, idx, idx + 1);
1989 case VMA_COMMIT_RESV:
1990 ret = region_add(resv, idx, idx + 1);
1993 region_abort(resv, idx, idx + 1);
1997 if (vma->vm_flags & VM_MAYSHARE)
1998 ret = region_add(resv, idx, idx + 1);
2000 region_abort(resv, idx, idx + 1);
2001 ret = region_del(resv, idx, idx + 1);
2008 if (vma->vm_flags & VM_MAYSHARE)
2010 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
2012 * In most cases, reserves always exist for private mappings.
2013 * However, a file associated with mapping could have been
2014 * hole punched or truncated after reserves were consumed.
2015 * As subsequent fault on such a range will not use reserves.
2016 * Subtle - The reserve map for private mappings has the
2017 * opposite meaning than that of shared mappings. If NO
2018 * entry is in the reserve map, it means a reservation exists.
2019 * If an entry exists in the reserve map, it means the
2020 * reservation has already been consumed. As a result, the
2021 * return value of this routine is the opposite of the
2022 * value returned from reserve map manipulation routines above.
2030 return ret < 0 ? ret : 0;
2033 static long vma_needs_reservation(struct hstate *h,
2034 struct vm_area_struct *vma, unsigned long addr)
2036 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2039 static long vma_commit_reservation(struct hstate *h,
2040 struct vm_area_struct *vma, unsigned long addr)
2042 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2045 static void vma_end_reservation(struct hstate *h,
2046 struct vm_area_struct *vma, unsigned long addr)
2048 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2051 static long vma_add_reservation(struct hstate *h,
2052 struct vm_area_struct *vma, unsigned long addr)
2054 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2058 * This routine is called to restore a reservation on error paths. In the
2059 * specific error paths, a huge page was allocated (via alloc_huge_page)
2060 * and is about to be freed. If a reservation for the page existed,
2061 * alloc_huge_page would have consumed the reservation and set PagePrivate
2062 * in the newly allocated page. When the page is freed via free_huge_page,
2063 * the global reservation count will be incremented if PagePrivate is set.
2064 * However, free_huge_page can not adjust the reserve map. Adjust the
2065 * reserve map here to be consistent with global reserve count adjustments
2066 * to be made by free_huge_page.
2068 static void restore_reserve_on_error(struct hstate *h,
2069 struct vm_area_struct *vma, unsigned long address,
2072 if (unlikely(PagePrivate(page))) {
2073 long rc = vma_needs_reservation(h, vma, address);
2075 if (unlikely(rc < 0)) {
2077 * Rare out of memory condition in reserve map
2078 * manipulation. Clear PagePrivate so that
2079 * global reserve count will not be incremented
2080 * by free_huge_page. This will make it appear
2081 * as though the reservation for this page was
2082 * consumed. This may prevent the task from
2083 * faulting in the page at a later time. This
2084 * is better than inconsistent global huge page
2085 * accounting of reserve counts.
2087 ClearPagePrivate(page);
2089 rc = vma_add_reservation(h, vma, address);
2090 if (unlikely(rc < 0))
2092 * See above comment about rare out of
2095 ClearPagePrivate(page);
2097 vma_end_reservation(h, vma, address);
2101 struct page *alloc_huge_page(struct vm_area_struct *vma,
2102 unsigned long addr, int avoid_reserve)
2104 struct hugepage_subpool *spool = subpool_vma(vma);
2105 struct hstate *h = hstate_vma(vma);
2107 long map_chg, map_commit;
2110 struct hugetlb_cgroup *h_cg;
2112 idx = hstate_index(h);
2114 * Examine the region/reserve map to determine if the process
2115 * has a reservation for the page to be allocated. A return
2116 * code of zero indicates a reservation exists (no change).
2118 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2120 return ERR_PTR(-ENOMEM);
2123 * Processes that did not create the mapping will have no
2124 * reserves as indicated by the region/reserve map. Check
2125 * that the allocation will not exceed the subpool limit.
2126 * Allocations for MAP_NORESERVE mappings also need to be
2127 * checked against any subpool limit.
2129 if (map_chg || avoid_reserve) {
2130 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2132 vma_end_reservation(h, vma, addr);
2133 return ERR_PTR(-ENOSPC);
2137 * Even though there was no reservation in the region/reserve
2138 * map, there could be reservations associated with the
2139 * subpool that can be used. This would be indicated if the
2140 * return value of hugepage_subpool_get_pages() is zero.
2141 * However, if avoid_reserve is specified we still avoid even
2142 * the subpool reservations.
2148 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2150 goto out_subpool_put;
2152 spin_lock(&hugetlb_lock);
2154 * glb_chg is passed to indicate whether or not a page must be taken
2155 * from the global free pool (global change). gbl_chg == 0 indicates
2156 * a reservation exists for the allocation.
2158 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2160 spin_unlock(&hugetlb_lock);
2161 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2163 goto out_uncharge_cgroup;
2164 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2165 SetPagePrivate(page);
2166 h->resv_huge_pages--;
2168 spin_lock(&hugetlb_lock);
2169 list_move(&page->lru, &h->hugepage_activelist);
2172 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2173 spin_unlock(&hugetlb_lock);
2175 set_page_private(page, (unsigned long)spool);
2177 map_commit = vma_commit_reservation(h, vma, addr);
2178 if (unlikely(map_chg > map_commit)) {
2180 * The page was added to the reservation map between
2181 * vma_needs_reservation and vma_commit_reservation.
2182 * This indicates a race with hugetlb_reserve_pages.
2183 * Adjust for the subpool count incremented above AND
2184 * in hugetlb_reserve_pages for the same page. Also,
2185 * the reservation count added in hugetlb_reserve_pages
2186 * no longer applies.
2190 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2191 hugetlb_acct_memory(h, -rsv_adjust);
2195 out_uncharge_cgroup:
2196 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2198 if (map_chg || avoid_reserve)
2199 hugepage_subpool_put_pages(spool, 1);
2200 vma_end_reservation(h, vma, addr);
2201 return ERR_PTR(-ENOSPC);
2204 int alloc_bootmem_huge_page(struct hstate *h)
2205 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2206 int __alloc_bootmem_huge_page(struct hstate *h)
2208 struct huge_bootmem_page *m;
2211 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2214 addr = memblock_alloc_try_nid_raw(
2215 huge_page_size(h), huge_page_size(h),
2216 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2219 * Use the beginning of the huge page to store the
2220 * huge_bootmem_page struct (until gather_bootmem
2221 * puts them into the mem_map).
2230 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2231 /* Put them into a private list first because mem_map is not up yet */
2232 INIT_LIST_HEAD(&m->list);
2233 list_add(&m->list, &huge_boot_pages);
2238 static void __init prep_compound_huge_page(struct page *page,
2241 if (unlikely(order > (MAX_ORDER - 1)))
2242 prep_compound_gigantic_page(page, order);
2244 prep_compound_page(page, order);
2247 /* Put bootmem huge pages into the standard lists after mem_map is up */
2248 static void __init gather_bootmem_prealloc(void)
2250 struct huge_bootmem_page *m;
2252 list_for_each_entry(m, &huge_boot_pages, list) {
2253 struct page *page = virt_to_page(m);
2254 struct hstate *h = m->hstate;
2256 WARN_ON(page_count(page) != 1);
2257 prep_compound_huge_page(page, h->order);
2258 WARN_ON(PageReserved(page));
2259 prep_new_huge_page(h, page, page_to_nid(page));
2260 put_page(page); /* free it into the hugepage allocator */
2263 * If we had gigantic hugepages allocated at boot time, we need
2264 * to restore the 'stolen' pages to totalram_pages in order to
2265 * fix confusing memory reports from free(1) and another
2266 * side-effects, like CommitLimit going negative.
2268 if (hstate_is_gigantic(h))
2269 adjust_managed_page_count(page, 1 << h->order);
2274 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2277 nodemask_t *node_alloc_noretry;
2279 if (!hstate_is_gigantic(h)) {
2281 * Bit mask controlling how hard we retry per-node allocations.
2282 * Ignore errors as lower level routines can deal with
2283 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2284 * time, we are likely in bigger trouble.
2286 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2289 /* allocations done at boot time */
2290 node_alloc_noretry = NULL;
2293 /* bit mask controlling how hard we retry per-node allocations */
2294 if (node_alloc_noretry)
2295 nodes_clear(*node_alloc_noretry);
2297 for (i = 0; i < h->max_huge_pages; ++i) {
2298 if (hstate_is_gigantic(h)) {
2299 if (!alloc_bootmem_huge_page(h))
2301 } else if (!alloc_pool_huge_page(h,
2302 &node_states[N_MEMORY],
2303 node_alloc_noretry))
2307 if (i < h->max_huge_pages) {
2310 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2311 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2312 h->max_huge_pages, buf, i);
2313 h->max_huge_pages = i;
2316 kfree(node_alloc_noretry);
2319 static void __init hugetlb_init_hstates(void)
2323 for_each_hstate(h) {
2324 if (minimum_order > huge_page_order(h))
2325 minimum_order = huge_page_order(h);
2327 /* oversize hugepages were init'ed in early boot */
2328 if (!hstate_is_gigantic(h))
2329 hugetlb_hstate_alloc_pages(h);
2331 VM_BUG_ON(minimum_order == UINT_MAX);
2334 static void __init report_hugepages(void)
2338 for_each_hstate(h) {
2341 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2342 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2343 buf, h->free_huge_pages);
2347 #ifdef CONFIG_HIGHMEM
2348 static void try_to_free_low(struct hstate *h, unsigned long count,
2349 nodemask_t *nodes_allowed)
2353 if (hstate_is_gigantic(h))
2356 for_each_node_mask(i, *nodes_allowed) {
2357 struct page *page, *next;
2358 struct list_head *freel = &h->hugepage_freelists[i];
2359 list_for_each_entry_safe(page, next, freel, lru) {
2360 if (count >= h->nr_huge_pages)
2362 if (PageHighMem(page))
2364 list_del(&page->lru);
2365 update_and_free_page(h, page);
2366 h->free_huge_pages--;
2367 h->free_huge_pages_node[page_to_nid(page)]--;
2372 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2373 nodemask_t *nodes_allowed)
2379 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2380 * balanced by operating on them in a round-robin fashion.
2381 * Returns 1 if an adjustment was made.
2383 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2388 VM_BUG_ON(delta != -1 && delta != 1);
2391 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2392 if (h->surplus_huge_pages_node[node])
2396 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2397 if (h->surplus_huge_pages_node[node] <
2398 h->nr_huge_pages_node[node])
2405 h->surplus_huge_pages += delta;
2406 h->surplus_huge_pages_node[node] += delta;
2410 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2411 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2412 nodemask_t *nodes_allowed)
2414 unsigned long min_count, ret;
2415 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2418 * Bit mask controlling how hard we retry per-node allocations.
2419 * If we can not allocate the bit mask, do not attempt to allocate
2420 * the requested huge pages.
2422 if (node_alloc_noretry)
2423 nodes_clear(*node_alloc_noretry);
2427 spin_lock(&hugetlb_lock);
2430 * Check for a node specific request.
2431 * Changing node specific huge page count may require a corresponding
2432 * change to the global count. In any case, the passed node mask
2433 * (nodes_allowed) will restrict alloc/free to the specified node.
2435 if (nid != NUMA_NO_NODE) {
2436 unsigned long old_count = count;
2438 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2440 * User may have specified a large count value which caused the
2441 * above calculation to overflow. In this case, they wanted
2442 * to allocate as many huge pages as possible. Set count to
2443 * largest possible value to align with their intention.
2445 if (count < old_count)
2450 * Gigantic pages runtime allocation depend on the capability for large
2451 * page range allocation.
2452 * If the system does not provide this feature, return an error when
2453 * the user tries to allocate gigantic pages but let the user free the
2454 * boottime allocated gigantic pages.
2456 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2457 if (count > persistent_huge_pages(h)) {
2458 spin_unlock(&hugetlb_lock);
2459 NODEMASK_FREE(node_alloc_noretry);
2462 /* Fall through to decrease pool */
2466 * Increase the pool size
2467 * First take pages out of surplus state. Then make up the
2468 * remaining difference by allocating fresh huge pages.
2470 * We might race with alloc_surplus_huge_page() here and be unable
2471 * to convert a surplus huge page to a normal huge page. That is
2472 * not critical, though, it just means the overall size of the
2473 * pool might be one hugepage larger than it needs to be, but
2474 * within all the constraints specified by the sysctls.
2476 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2477 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2481 while (count > persistent_huge_pages(h)) {
2483 * If this allocation races such that we no longer need the
2484 * page, free_huge_page will handle it by freeing the page
2485 * and reducing the surplus.
2487 spin_unlock(&hugetlb_lock);
2489 /* yield cpu to avoid soft lockup */
2492 ret = alloc_pool_huge_page(h, nodes_allowed,
2493 node_alloc_noretry);
2494 spin_lock(&hugetlb_lock);
2498 /* Bail for signals. Probably ctrl-c from user */
2499 if (signal_pending(current))
2504 * Decrease the pool size
2505 * First return free pages to the buddy allocator (being careful
2506 * to keep enough around to satisfy reservations). Then place
2507 * pages into surplus state as needed so the pool will shrink
2508 * to the desired size as pages become free.
2510 * By placing pages into the surplus state independent of the
2511 * overcommit value, we are allowing the surplus pool size to
2512 * exceed overcommit. There are few sane options here. Since
2513 * alloc_surplus_huge_page() is checking the global counter,
2514 * though, we'll note that we're not allowed to exceed surplus
2515 * and won't grow the pool anywhere else. Not until one of the
2516 * sysctls are changed, or the surplus pages go out of use.
2518 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2519 min_count = max(count, min_count);
2520 try_to_free_low(h, min_count, nodes_allowed);
2521 while (min_count < persistent_huge_pages(h)) {
2522 if (!free_pool_huge_page(h, nodes_allowed, 0))
2524 cond_resched_lock(&hugetlb_lock);
2526 while (count < persistent_huge_pages(h)) {
2527 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2531 h->max_huge_pages = persistent_huge_pages(h);
2532 spin_unlock(&hugetlb_lock);
2534 NODEMASK_FREE(node_alloc_noretry);
2539 #define HSTATE_ATTR_RO(_name) \
2540 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2542 #define HSTATE_ATTR(_name) \
2543 static struct kobj_attribute _name##_attr = \
2544 __ATTR(_name, 0644, _name##_show, _name##_store)
2546 static struct kobject *hugepages_kobj;
2547 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2549 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2551 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2555 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2556 if (hstate_kobjs[i] == kobj) {
2558 *nidp = NUMA_NO_NODE;
2562 return kobj_to_node_hstate(kobj, nidp);
2565 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2566 struct kobj_attribute *attr, char *buf)
2569 unsigned long nr_huge_pages;
2572 h = kobj_to_hstate(kobj, &nid);
2573 if (nid == NUMA_NO_NODE)
2574 nr_huge_pages = h->nr_huge_pages;
2576 nr_huge_pages = h->nr_huge_pages_node[nid];
2578 return sprintf(buf, "%lu\n", nr_huge_pages);
2581 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2582 struct hstate *h, int nid,
2583 unsigned long count, size_t len)
2586 nodemask_t nodes_allowed, *n_mask;
2588 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2591 if (nid == NUMA_NO_NODE) {
2593 * global hstate attribute
2595 if (!(obey_mempolicy &&
2596 init_nodemask_of_mempolicy(&nodes_allowed)))
2597 n_mask = &node_states[N_MEMORY];
2599 n_mask = &nodes_allowed;
2602 * Node specific request. count adjustment happens in
2603 * set_max_huge_pages() after acquiring hugetlb_lock.
2605 init_nodemask_of_node(&nodes_allowed, nid);
2606 n_mask = &nodes_allowed;
2609 err = set_max_huge_pages(h, count, nid, n_mask);
2611 return err ? err : len;
2614 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2615 struct kobject *kobj, const char *buf,
2619 unsigned long count;
2623 err = kstrtoul(buf, 10, &count);
2627 h = kobj_to_hstate(kobj, &nid);
2628 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2631 static ssize_t nr_hugepages_show(struct kobject *kobj,
2632 struct kobj_attribute *attr, char *buf)
2634 return nr_hugepages_show_common(kobj, attr, buf);
2637 static ssize_t nr_hugepages_store(struct kobject *kobj,
2638 struct kobj_attribute *attr, const char *buf, size_t len)
2640 return nr_hugepages_store_common(false, kobj, buf, len);
2642 HSTATE_ATTR(nr_hugepages);
2647 * hstate attribute for optionally mempolicy-based constraint on persistent
2648 * huge page alloc/free.
2650 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2651 struct kobj_attribute *attr, char *buf)
2653 return nr_hugepages_show_common(kobj, attr, buf);
2656 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2657 struct kobj_attribute *attr, const char *buf, size_t len)
2659 return nr_hugepages_store_common(true, kobj, buf, len);
2661 HSTATE_ATTR(nr_hugepages_mempolicy);
2665 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2666 struct kobj_attribute *attr, char *buf)
2668 struct hstate *h = kobj_to_hstate(kobj, NULL);
2669 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2672 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2673 struct kobj_attribute *attr, const char *buf, size_t count)
2676 unsigned long input;
2677 struct hstate *h = kobj_to_hstate(kobj, NULL);
2679 if (hstate_is_gigantic(h))
2682 err = kstrtoul(buf, 10, &input);
2686 spin_lock(&hugetlb_lock);
2687 h->nr_overcommit_huge_pages = input;
2688 spin_unlock(&hugetlb_lock);
2692 HSTATE_ATTR(nr_overcommit_hugepages);
2694 static ssize_t free_hugepages_show(struct kobject *kobj,
2695 struct kobj_attribute *attr, char *buf)
2698 unsigned long free_huge_pages;
2701 h = kobj_to_hstate(kobj, &nid);
2702 if (nid == NUMA_NO_NODE)
2703 free_huge_pages = h->free_huge_pages;
2705 free_huge_pages = h->free_huge_pages_node[nid];
2707 return sprintf(buf, "%lu\n", free_huge_pages);
2709 HSTATE_ATTR_RO(free_hugepages);
2711 static ssize_t resv_hugepages_show(struct kobject *kobj,
2712 struct kobj_attribute *attr, char *buf)
2714 struct hstate *h = kobj_to_hstate(kobj, NULL);
2715 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2717 HSTATE_ATTR_RO(resv_hugepages);
2719 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2720 struct kobj_attribute *attr, char *buf)
2723 unsigned long surplus_huge_pages;
2726 h = kobj_to_hstate(kobj, &nid);
2727 if (nid == NUMA_NO_NODE)
2728 surplus_huge_pages = h->surplus_huge_pages;
2730 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2732 return sprintf(buf, "%lu\n", surplus_huge_pages);
2734 HSTATE_ATTR_RO(surplus_hugepages);
2736 static struct attribute *hstate_attrs[] = {
2737 &nr_hugepages_attr.attr,
2738 &nr_overcommit_hugepages_attr.attr,
2739 &free_hugepages_attr.attr,
2740 &resv_hugepages_attr.attr,
2741 &surplus_hugepages_attr.attr,
2743 &nr_hugepages_mempolicy_attr.attr,
2748 static const struct attribute_group hstate_attr_group = {
2749 .attrs = hstate_attrs,
2752 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2753 struct kobject **hstate_kobjs,
2754 const struct attribute_group *hstate_attr_group)
2757 int hi = hstate_index(h);
2759 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2760 if (!hstate_kobjs[hi])
2763 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2765 kobject_put(hstate_kobjs[hi]);
2770 static void __init hugetlb_sysfs_init(void)
2775 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2776 if (!hugepages_kobj)
2779 for_each_hstate(h) {
2780 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2781 hstate_kobjs, &hstate_attr_group);
2783 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2790 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2791 * with node devices in node_devices[] using a parallel array. The array
2792 * index of a node device or _hstate == node id.
2793 * This is here to avoid any static dependency of the node device driver, in
2794 * the base kernel, on the hugetlb module.
2796 struct node_hstate {
2797 struct kobject *hugepages_kobj;
2798 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2800 static struct node_hstate node_hstates[MAX_NUMNODES];
2803 * A subset of global hstate attributes for node devices
2805 static struct attribute *per_node_hstate_attrs[] = {
2806 &nr_hugepages_attr.attr,
2807 &free_hugepages_attr.attr,
2808 &surplus_hugepages_attr.attr,
2812 static const struct attribute_group per_node_hstate_attr_group = {
2813 .attrs = per_node_hstate_attrs,
2817 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2818 * Returns node id via non-NULL nidp.
2820 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2824 for (nid = 0; nid < nr_node_ids; nid++) {
2825 struct node_hstate *nhs = &node_hstates[nid];
2827 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2828 if (nhs->hstate_kobjs[i] == kobj) {
2840 * Unregister hstate attributes from a single node device.
2841 * No-op if no hstate attributes attached.
2843 static void hugetlb_unregister_node(struct node *node)
2846 struct node_hstate *nhs = &node_hstates[node->dev.id];
2848 if (!nhs->hugepages_kobj)
2849 return; /* no hstate attributes */
2851 for_each_hstate(h) {
2852 int idx = hstate_index(h);
2853 if (nhs->hstate_kobjs[idx]) {
2854 kobject_put(nhs->hstate_kobjs[idx]);
2855 nhs->hstate_kobjs[idx] = NULL;
2859 kobject_put(nhs->hugepages_kobj);
2860 nhs->hugepages_kobj = NULL;
2865 * Register hstate attributes for a single node device.
2866 * No-op if attributes already registered.
2868 static void hugetlb_register_node(struct node *node)
2871 struct node_hstate *nhs = &node_hstates[node->dev.id];
2874 if (nhs->hugepages_kobj)
2875 return; /* already allocated */
2877 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2879 if (!nhs->hugepages_kobj)
2882 for_each_hstate(h) {
2883 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2885 &per_node_hstate_attr_group);
2887 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2888 h->name, node->dev.id);
2889 hugetlb_unregister_node(node);
2896 * hugetlb init time: register hstate attributes for all registered node
2897 * devices of nodes that have memory. All on-line nodes should have
2898 * registered their associated device by this time.
2900 static void __init hugetlb_register_all_nodes(void)
2904 for_each_node_state(nid, N_MEMORY) {
2905 struct node *node = node_devices[nid];
2906 if (node->dev.id == nid)
2907 hugetlb_register_node(node);
2911 * Let the node device driver know we're here so it can
2912 * [un]register hstate attributes on node hotplug.
2914 register_hugetlbfs_with_node(hugetlb_register_node,
2915 hugetlb_unregister_node);
2917 #else /* !CONFIG_NUMA */
2919 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2927 static void hugetlb_register_all_nodes(void) { }
2931 static int __init hugetlb_init(void)
2935 if (!hugepages_supported())
2938 if (!size_to_hstate(default_hstate_size)) {
2939 if (default_hstate_size != 0) {
2940 pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2941 default_hstate_size, HPAGE_SIZE);
2944 default_hstate_size = HPAGE_SIZE;
2945 if (!size_to_hstate(default_hstate_size))
2946 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2948 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2949 if (default_hstate_max_huge_pages) {
2950 if (!default_hstate.max_huge_pages)
2951 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2954 hugetlb_init_hstates();
2955 gather_bootmem_prealloc();
2958 hugetlb_sysfs_init();
2959 hugetlb_register_all_nodes();
2960 hugetlb_cgroup_file_init();
2963 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2965 num_fault_mutexes = 1;
2967 hugetlb_fault_mutex_table =
2968 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
2970 BUG_ON(!hugetlb_fault_mutex_table);
2972 for (i = 0; i < num_fault_mutexes; i++)
2973 mutex_init(&hugetlb_fault_mutex_table[i]);
2976 subsys_initcall(hugetlb_init);
2978 /* Should be called on processing a hugepagesz=... option */
2979 void __init hugetlb_bad_size(void)
2981 parsed_valid_hugepagesz = false;
2984 void __init hugetlb_add_hstate(unsigned int order)
2989 if (size_to_hstate(PAGE_SIZE << order)) {
2990 pr_warn("hugepagesz= specified twice, ignoring\n");
2993 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2995 h = &hstates[hugetlb_max_hstate++];
2997 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2998 h->nr_huge_pages = 0;
2999 h->free_huge_pages = 0;
3000 for (i = 0; i < MAX_NUMNODES; ++i)
3001 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3002 INIT_LIST_HEAD(&h->hugepage_activelist);
3003 h->next_nid_to_alloc = first_memory_node;
3004 h->next_nid_to_free = first_memory_node;
3005 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3006 huge_page_size(h)/1024);
3011 static int __init hugetlb_nrpages_setup(char *s)
3014 static unsigned long *last_mhp;
3016 if (!parsed_valid_hugepagesz) {
3017 pr_warn("hugepages = %s preceded by "
3018 "an unsupported hugepagesz, ignoring\n", s);
3019 parsed_valid_hugepagesz = true;
3023 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
3024 * so this hugepages= parameter goes to the "default hstate".
3026 else if (!hugetlb_max_hstate)
3027 mhp = &default_hstate_max_huge_pages;
3029 mhp = &parsed_hstate->max_huge_pages;
3031 if (mhp == last_mhp) {
3032 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
3036 if (sscanf(s, "%lu", mhp) <= 0)
3040 * Global state is always initialized later in hugetlb_init.
3041 * But we need to allocate >= MAX_ORDER hstates here early to still
3042 * use the bootmem allocator.
3044 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
3045 hugetlb_hstate_alloc_pages(parsed_hstate);
3051 __setup("hugepages=", hugetlb_nrpages_setup);
3053 static int __init hugetlb_default_setup(char *s)
3055 default_hstate_size = memparse(s, &s);
3058 __setup("default_hugepagesz=", hugetlb_default_setup);
3060 static unsigned int cpuset_mems_nr(unsigned int *array)
3063 unsigned int nr = 0;
3065 for_each_node_mask(node, cpuset_current_mems_allowed)
3071 #ifdef CONFIG_SYSCTL
3072 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3073 struct ctl_table *table, int write,
3074 void __user *buffer, size_t *length, loff_t *ppos)
3076 struct hstate *h = &default_hstate;
3077 unsigned long tmp = h->max_huge_pages;
3080 if (!hugepages_supported())
3084 table->maxlen = sizeof(unsigned long);
3085 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3090 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3091 NUMA_NO_NODE, tmp, *length);
3096 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3097 void __user *buffer, size_t *length, loff_t *ppos)
3100 return hugetlb_sysctl_handler_common(false, table, write,
3101 buffer, length, ppos);
3105 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3106 void __user *buffer, size_t *length, loff_t *ppos)
3108 return hugetlb_sysctl_handler_common(true, table, write,
3109 buffer, length, ppos);
3111 #endif /* CONFIG_NUMA */
3113 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3114 void __user *buffer,
3115 size_t *length, loff_t *ppos)
3117 struct hstate *h = &default_hstate;
3121 if (!hugepages_supported())
3124 tmp = h->nr_overcommit_huge_pages;
3126 if (write && hstate_is_gigantic(h))
3130 table->maxlen = sizeof(unsigned long);
3131 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3136 spin_lock(&hugetlb_lock);
3137 h->nr_overcommit_huge_pages = tmp;
3138 spin_unlock(&hugetlb_lock);
3144 #endif /* CONFIG_SYSCTL */
3146 void hugetlb_report_meminfo(struct seq_file *m)
3149 unsigned long total = 0;
3151 if (!hugepages_supported())
3154 for_each_hstate(h) {
3155 unsigned long count = h->nr_huge_pages;
3157 total += (PAGE_SIZE << huge_page_order(h)) * count;
3159 if (h == &default_hstate)
3161 "HugePages_Total: %5lu\n"
3162 "HugePages_Free: %5lu\n"
3163 "HugePages_Rsvd: %5lu\n"
3164 "HugePages_Surp: %5lu\n"
3165 "Hugepagesize: %8lu kB\n",
3169 h->surplus_huge_pages,
3170 (PAGE_SIZE << huge_page_order(h)) / 1024);
3173 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3176 int hugetlb_report_node_meminfo(int nid, char *buf)
3178 struct hstate *h = &default_hstate;
3179 if (!hugepages_supported())
3182 "Node %d HugePages_Total: %5u\n"
3183 "Node %d HugePages_Free: %5u\n"
3184 "Node %d HugePages_Surp: %5u\n",
3185 nid, h->nr_huge_pages_node[nid],
3186 nid, h->free_huge_pages_node[nid],
3187 nid, h->surplus_huge_pages_node[nid]);
3190 void hugetlb_show_meminfo(void)
3195 if (!hugepages_supported())
3198 for_each_node_state(nid, N_MEMORY)
3200 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3202 h->nr_huge_pages_node[nid],
3203 h->free_huge_pages_node[nid],
3204 h->surplus_huge_pages_node[nid],
3205 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3208 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3210 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3211 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3214 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3215 unsigned long hugetlb_total_pages(void)
3218 unsigned long nr_total_pages = 0;
3221 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3222 return nr_total_pages;
3225 static int hugetlb_acct_memory(struct hstate *h, long delta)
3229 spin_lock(&hugetlb_lock);
3231 * When cpuset is configured, it breaks the strict hugetlb page
3232 * reservation as the accounting is done on a global variable. Such
3233 * reservation is completely rubbish in the presence of cpuset because
3234 * the reservation is not checked against page availability for the
3235 * current cpuset. Application can still potentially OOM'ed by kernel
3236 * with lack of free htlb page in cpuset that the task is in.
3237 * Attempt to enforce strict accounting with cpuset is almost
3238 * impossible (or too ugly) because cpuset is too fluid that
3239 * task or memory node can be dynamically moved between cpusets.
3241 * The change of semantics for shared hugetlb mapping with cpuset is
3242 * undesirable. However, in order to preserve some of the semantics,
3243 * we fall back to check against current free page availability as
3244 * a best attempt and hopefully to minimize the impact of changing
3245 * semantics that cpuset has.
3248 if (gather_surplus_pages(h, delta) < 0)
3251 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3252 return_unused_surplus_pages(h, delta);
3259 return_unused_surplus_pages(h, (unsigned long) -delta);
3262 spin_unlock(&hugetlb_lock);
3266 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3268 struct resv_map *resv = vma_resv_map(vma);
3271 * This new VMA should share its siblings reservation map if present.
3272 * The VMA will only ever have a valid reservation map pointer where
3273 * it is being copied for another still existing VMA. As that VMA
3274 * has a reference to the reservation map it cannot disappear until
3275 * after this open call completes. It is therefore safe to take a
3276 * new reference here without additional locking.
3278 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3279 kref_get(&resv->refs);
3282 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3284 struct hstate *h = hstate_vma(vma);
3285 struct resv_map *resv = vma_resv_map(vma);
3286 struct hugepage_subpool *spool = subpool_vma(vma);
3287 unsigned long reserve, start, end;
3290 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3293 start = vma_hugecache_offset(h, vma, vma->vm_start);
3294 end = vma_hugecache_offset(h, vma, vma->vm_end);
3296 reserve = (end - start) - region_count(resv, start, end);
3298 kref_put(&resv->refs, resv_map_release);
3302 * Decrement reserve counts. The global reserve count may be
3303 * adjusted if the subpool has a minimum size.
3305 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3306 hugetlb_acct_memory(h, -gbl_reserve);
3310 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3312 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3317 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3319 struct hstate *hstate = hstate_vma(vma);
3321 return 1UL << huge_page_shift(hstate);
3325 * We cannot handle pagefaults against hugetlb pages at all. They cause
3326 * handle_mm_fault() to try to instantiate regular-sized pages in the
3327 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3330 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3337 * When a new function is introduced to vm_operations_struct and added
3338 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3339 * This is because under System V memory model, mappings created via
3340 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3341 * their original vm_ops are overwritten with shm_vm_ops.
3343 const struct vm_operations_struct hugetlb_vm_ops = {
3344 .fault = hugetlb_vm_op_fault,
3345 .open = hugetlb_vm_op_open,
3346 .close = hugetlb_vm_op_close,
3347 .split = hugetlb_vm_op_split,
3348 .pagesize = hugetlb_vm_op_pagesize,
3351 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3357 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3358 vma->vm_page_prot)));
3360 entry = huge_pte_wrprotect(mk_huge_pte(page,
3361 vma->vm_page_prot));
3363 entry = pte_mkyoung(entry);
3364 entry = pte_mkhuge(entry);
3365 entry = arch_make_huge_pte(entry, vma, page, writable);
3370 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3371 unsigned long address, pte_t *ptep)
3375 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3376 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3377 update_mmu_cache(vma, address, ptep);
3380 bool is_hugetlb_entry_migration(pte_t pte)
3384 if (huge_pte_none(pte) || pte_present(pte))
3386 swp = pte_to_swp_entry(pte);
3387 if (non_swap_entry(swp) && is_migration_entry(swp))
3393 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3397 if (huge_pte_none(pte) || pte_present(pte))
3399 swp = pte_to_swp_entry(pte);
3400 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3406 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3407 struct vm_area_struct *vma)
3409 pte_t *src_pte, *dst_pte, entry, dst_entry;
3410 struct page *ptepage;
3413 struct hstate *h = hstate_vma(vma);
3414 unsigned long sz = huge_page_size(h);
3415 struct address_space *mapping = vma->vm_file->f_mapping;
3416 struct mmu_notifier_range range;
3419 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3422 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3425 mmu_notifier_invalidate_range_start(&range);
3428 * For shared mappings i_mmap_rwsem must be held to call
3429 * huge_pte_alloc, otherwise the returned ptep could go
3430 * away if part of a shared pmd and another thread calls
3433 i_mmap_lock_read(mapping);
3436 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3437 spinlock_t *src_ptl, *dst_ptl;
3438 src_pte = huge_pte_offset(src, addr, sz);
3441 dst_pte = huge_pte_alloc(dst, addr, sz);
3448 * If the pagetables are shared don't copy or take references.
3449 * dst_pte == src_pte is the common case of src/dest sharing.
3451 * However, src could have 'unshared' and dst shares with
3452 * another vma. If dst_pte !none, this implies sharing.
3453 * Check here before taking page table lock, and once again
3454 * after taking the lock below.
3456 dst_entry = huge_ptep_get(dst_pte);
3457 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3460 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3461 src_ptl = huge_pte_lockptr(h, src, src_pte);
3462 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3463 entry = huge_ptep_get(src_pte);
3464 dst_entry = huge_ptep_get(dst_pte);
3465 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3467 * Skip if src entry none. Also, skip in the
3468 * unlikely case dst entry !none as this implies
3469 * sharing with another vma.
3472 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3473 is_hugetlb_entry_hwpoisoned(entry))) {
3474 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3476 if (is_write_migration_entry(swp_entry) && cow) {
3478 * COW mappings require pages in both
3479 * parent and child to be set to read.
3481 make_migration_entry_read(&swp_entry);
3482 entry = swp_entry_to_pte(swp_entry);
3483 set_huge_swap_pte_at(src, addr, src_pte,
3486 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3490 * No need to notify as we are downgrading page
3491 * table protection not changing it to point
3494 * See Documentation/vm/mmu_notifier.rst
3496 huge_ptep_set_wrprotect(src, addr, src_pte);
3498 entry = huge_ptep_get(src_pte);
3499 ptepage = pte_page(entry);
3501 page_dup_rmap(ptepage, true);
3502 set_huge_pte_at(dst, addr, dst_pte, entry);
3503 hugetlb_count_add(pages_per_huge_page(h), dst);
3505 spin_unlock(src_ptl);
3506 spin_unlock(dst_ptl);
3510 mmu_notifier_invalidate_range_end(&range);
3512 i_mmap_unlock_read(mapping);
3517 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3518 unsigned long start, unsigned long end,
3519 struct page *ref_page)
3521 struct mm_struct *mm = vma->vm_mm;
3522 unsigned long address;
3527 struct hstate *h = hstate_vma(vma);
3528 unsigned long sz = huge_page_size(h);
3529 struct mmu_notifier_range range;
3531 WARN_ON(!is_vm_hugetlb_page(vma));
3532 BUG_ON(start & ~huge_page_mask(h));
3533 BUG_ON(end & ~huge_page_mask(h));
3536 * This is a hugetlb vma, all the pte entries should point
3539 tlb_change_page_size(tlb, sz);
3540 tlb_start_vma(tlb, vma);
3543 * If sharing possible, alert mmu notifiers of worst case.
3545 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3547 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3548 mmu_notifier_invalidate_range_start(&range);
3550 for (; address < end; address += sz) {
3551 ptep = huge_pte_offset(mm, address, sz);
3555 ptl = huge_pte_lock(h, mm, ptep);
3556 if (huge_pmd_unshare(mm, &address, ptep)) {
3559 * We just unmapped a page of PMDs by clearing a PUD.
3560 * The caller's TLB flush range should cover this area.
3565 pte = huge_ptep_get(ptep);
3566 if (huge_pte_none(pte)) {
3572 * Migrating hugepage or HWPoisoned hugepage is already
3573 * unmapped and its refcount is dropped, so just clear pte here.
3575 if (unlikely(!pte_present(pte))) {
3576 huge_pte_clear(mm, address, ptep, sz);
3581 page = pte_page(pte);
3583 * If a reference page is supplied, it is because a specific
3584 * page is being unmapped, not a range. Ensure the page we
3585 * are about to unmap is the actual page of interest.
3588 if (page != ref_page) {
3593 * Mark the VMA as having unmapped its page so that
3594 * future faults in this VMA will fail rather than
3595 * looking like data was lost
3597 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3600 pte = huge_ptep_get_and_clear(mm, address, ptep);
3601 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3602 if (huge_pte_dirty(pte))
3603 set_page_dirty(page);
3605 hugetlb_count_sub(pages_per_huge_page(h), mm);
3606 page_remove_rmap(page, true);
3609 tlb_remove_page_size(tlb, page, huge_page_size(h));
3611 * Bail out after unmapping reference page if supplied
3616 mmu_notifier_invalidate_range_end(&range);
3617 tlb_end_vma(tlb, vma);
3620 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3621 struct vm_area_struct *vma, unsigned long start,
3622 unsigned long end, struct page *ref_page)
3624 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3627 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3628 * test will fail on a vma being torn down, and not grab a page table
3629 * on its way out. We're lucky that the flag has such an appropriate
3630 * name, and can in fact be safely cleared here. We could clear it
3631 * before the __unmap_hugepage_range above, but all that's necessary
3632 * is to clear it before releasing the i_mmap_rwsem. This works
3633 * because in the context this is called, the VMA is about to be
3634 * destroyed and the i_mmap_rwsem is held.
3636 vma->vm_flags &= ~VM_MAYSHARE;
3639 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3640 unsigned long end, struct page *ref_page)
3642 struct mm_struct *mm;
3643 struct mmu_gather tlb;
3644 unsigned long tlb_start = start;
3645 unsigned long tlb_end = end;
3648 * If shared PMDs were possibly used within this vma range, adjust
3649 * start/end for worst case tlb flushing.
3650 * Note that we can not be sure if PMDs are shared until we try to
3651 * unmap pages. However, we want to make sure TLB flushing covers
3652 * the largest possible range.
3654 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
3658 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
3659 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3660 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
3664 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3665 * mappping it owns the reserve page for. The intention is to unmap the page
3666 * from other VMAs and let the children be SIGKILLed if they are faulting the
3669 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3670 struct page *page, unsigned long address)
3672 struct hstate *h = hstate_vma(vma);
3673 struct vm_area_struct *iter_vma;
3674 struct address_space *mapping;
3678 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3679 * from page cache lookup which is in HPAGE_SIZE units.
3681 address = address & huge_page_mask(h);
3682 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3684 mapping = vma->vm_file->f_mapping;
3687 * Take the mapping lock for the duration of the table walk. As
3688 * this mapping should be shared between all the VMAs,
3689 * __unmap_hugepage_range() is called as the lock is already held
3691 i_mmap_lock_write(mapping);
3692 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3693 /* Do not unmap the current VMA */
3694 if (iter_vma == vma)
3698 * Shared VMAs have their own reserves and do not affect
3699 * MAP_PRIVATE accounting but it is possible that a shared
3700 * VMA is using the same page so check and skip such VMAs.
3702 if (iter_vma->vm_flags & VM_MAYSHARE)
3706 * Unmap the page from other VMAs without their own reserves.
3707 * They get marked to be SIGKILLed if they fault in these
3708 * areas. This is because a future no-page fault on this VMA
3709 * could insert a zeroed page instead of the data existing
3710 * from the time of fork. This would look like data corruption
3712 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3713 unmap_hugepage_range(iter_vma, address,
3714 address + huge_page_size(h), page);
3716 i_mmap_unlock_write(mapping);
3720 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3721 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3722 * cannot race with other handlers or page migration.
3723 * Keep the pte_same checks anyway to make transition from the mutex easier.
3725 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3726 unsigned long address, pte_t *ptep,
3727 struct page *pagecache_page, spinlock_t *ptl)
3730 struct hstate *h = hstate_vma(vma);
3731 struct page *old_page, *new_page;
3732 int outside_reserve = 0;
3734 unsigned long haddr = address & huge_page_mask(h);
3735 struct mmu_notifier_range range;
3737 pte = huge_ptep_get(ptep);
3738 old_page = pte_page(pte);
3741 /* If no-one else is actually using this page, avoid the copy
3742 * and just make the page writable */
3743 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3744 page_move_anon_rmap(old_page, vma);
3745 set_huge_ptep_writable(vma, haddr, ptep);
3750 * If the process that created a MAP_PRIVATE mapping is about to
3751 * perform a COW due to a shared page count, attempt to satisfy
3752 * the allocation without using the existing reserves. The pagecache
3753 * page is used to determine if the reserve at this address was
3754 * consumed or not. If reserves were used, a partial faulted mapping
3755 * at the time of fork() could consume its reserves on COW instead
3756 * of the full address range.
3758 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3759 old_page != pagecache_page)
3760 outside_reserve = 1;
3765 * Drop page table lock as buddy allocator may be called. It will
3766 * be acquired again before returning to the caller, as expected.
3769 new_page = alloc_huge_page(vma, haddr, outside_reserve);
3771 if (IS_ERR(new_page)) {
3773 * If a process owning a MAP_PRIVATE mapping fails to COW,
3774 * it is due to references held by a child and an insufficient
3775 * huge page pool. To guarantee the original mappers
3776 * reliability, unmap the page from child processes. The child
3777 * may get SIGKILLed if it later faults.
3779 if (outside_reserve) {
3781 BUG_ON(huge_pte_none(pte));
3782 unmap_ref_private(mm, vma, old_page, haddr);
3783 BUG_ON(huge_pte_none(pte));
3785 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3787 pte_same(huge_ptep_get(ptep), pte)))
3788 goto retry_avoidcopy;
3790 * race occurs while re-acquiring page table
3791 * lock, and our job is done.
3796 ret = vmf_error(PTR_ERR(new_page));
3797 goto out_release_old;
3801 * When the original hugepage is shared one, it does not have
3802 * anon_vma prepared.
3804 if (unlikely(anon_vma_prepare(vma))) {
3806 goto out_release_all;
3809 copy_user_huge_page(new_page, old_page, address, vma,
3810 pages_per_huge_page(h));
3811 __SetPageUptodate(new_page);
3813 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
3814 haddr + huge_page_size(h));
3815 mmu_notifier_invalidate_range_start(&range);
3818 * Retake the page table lock to check for racing updates
3819 * before the page tables are altered
3822 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3823 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3824 ClearPagePrivate(new_page);
3827 huge_ptep_clear_flush(vma, haddr, ptep);
3828 mmu_notifier_invalidate_range(mm, range.start, range.end);
3829 set_huge_pte_at(mm, haddr, ptep,
3830 make_huge_pte(vma, new_page, 1));
3831 page_remove_rmap(old_page, true);
3832 hugepage_add_new_anon_rmap(new_page, vma, haddr);
3833 set_page_huge_active(new_page);
3834 /* Make the old page be freed below */
3835 new_page = old_page;
3838 mmu_notifier_invalidate_range_end(&range);
3840 restore_reserve_on_error(h, vma, haddr, new_page);
3845 spin_lock(ptl); /* Caller expects lock to be held */
3849 /* Return the pagecache page at a given address within a VMA */
3850 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3851 struct vm_area_struct *vma, unsigned long address)
3853 struct address_space *mapping;
3856 mapping = vma->vm_file->f_mapping;
3857 idx = vma_hugecache_offset(h, vma, address);
3859 return find_lock_page(mapping, idx);
3863 * Return whether there is a pagecache page to back given address within VMA.
3864 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3866 static bool hugetlbfs_pagecache_present(struct hstate *h,
3867 struct vm_area_struct *vma, unsigned long address)
3869 struct address_space *mapping;
3873 mapping = vma->vm_file->f_mapping;
3874 idx = vma_hugecache_offset(h, vma, address);
3876 page = find_get_page(mapping, idx);
3879 return page != NULL;
3882 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3885 struct inode *inode = mapping->host;
3886 struct hstate *h = hstate_inode(inode);
3887 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3891 ClearPagePrivate(page);
3894 * set page dirty so that it will not be removed from cache/file
3895 * by non-hugetlbfs specific code paths.
3897 set_page_dirty(page);
3899 spin_lock(&inode->i_lock);
3900 inode->i_blocks += blocks_per_huge_page(h);
3901 spin_unlock(&inode->i_lock);
3905 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
3906 struct vm_area_struct *vma,
3907 struct address_space *mapping, pgoff_t idx,
3908 unsigned long address, pte_t *ptep, unsigned int flags)
3910 struct hstate *h = hstate_vma(vma);
3911 vm_fault_t ret = VM_FAULT_SIGBUS;
3917 unsigned long haddr = address & huge_page_mask(h);
3918 bool new_page = false;
3921 * Currently, we are forced to kill the process in the event the
3922 * original mapper has unmapped pages from the child due to a failed
3923 * COW. Warn that such a situation has occurred as it may not be obvious
3925 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3926 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3932 * Use page lock to guard against racing truncation
3933 * before we get page_table_lock.
3936 page = find_lock_page(mapping, idx);
3938 size = i_size_read(mapping->host) >> huge_page_shift(h);
3943 * Check for page in userfault range
3945 if (userfaultfd_missing(vma)) {
3947 struct vm_fault vmf = {
3952 * Hard to debug if it ends up being
3953 * used by a callee that assumes
3954 * something about the other
3955 * uninitialized fields... same as in
3961 * hugetlb_fault_mutex and i_mmap_rwsem must be
3962 * dropped before handling userfault. Reacquire
3963 * after handling fault to make calling code simpler.
3965 hash = hugetlb_fault_mutex_hash(mapping, idx);
3966 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3967 i_mmap_unlock_read(mapping);
3968 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3969 i_mmap_lock_read(mapping);
3970 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3974 page = alloc_huge_page(vma, haddr, 0);
3977 * Returning error will result in faulting task being
3978 * sent SIGBUS. The hugetlb fault mutex prevents two
3979 * tasks from racing to fault in the same page which
3980 * could result in false unable to allocate errors.
3981 * Page migration does not take the fault mutex, but
3982 * does a clear then write of pte's under page table
3983 * lock. Page fault code could race with migration,
3984 * notice the clear pte and try to allocate a page
3985 * here. Before returning error, get ptl and make
3986 * sure there really is no pte entry.
3988 ptl = huge_pte_lock(h, mm, ptep);
3989 if (!huge_pte_none(huge_ptep_get(ptep))) {
3995 ret = vmf_error(PTR_ERR(page));
3998 clear_huge_page(page, address, pages_per_huge_page(h));
3999 __SetPageUptodate(page);
4002 if (vma->vm_flags & VM_MAYSHARE) {
4003 int err = huge_add_to_page_cache(page, mapping, idx);
4012 if (unlikely(anon_vma_prepare(vma))) {
4014 goto backout_unlocked;
4020 * If memory error occurs between mmap() and fault, some process
4021 * don't have hwpoisoned swap entry for errored virtual address.
4022 * So we need to block hugepage fault by PG_hwpoison bit check.
4024 if (unlikely(PageHWPoison(page))) {
4025 ret = VM_FAULT_HWPOISON |
4026 VM_FAULT_SET_HINDEX(hstate_index(h));
4027 goto backout_unlocked;
4032 * If we are going to COW a private mapping later, we examine the
4033 * pending reservations for this page now. This will ensure that
4034 * any allocations necessary to record that reservation occur outside
4037 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4038 if (vma_needs_reservation(h, vma, haddr) < 0) {
4040 goto backout_unlocked;
4042 /* Just decrements count, does not deallocate */
4043 vma_end_reservation(h, vma, haddr);
4046 ptl = huge_pte_lock(h, mm, ptep);
4047 size = i_size_read(mapping->host) >> huge_page_shift(h);
4052 if (!huge_pte_none(huge_ptep_get(ptep)))
4056 ClearPagePrivate(page);
4057 hugepage_add_new_anon_rmap(page, vma, haddr);
4059 page_dup_rmap(page, true);
4060 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4061 && (vma->vm_flags & VM_SHARED)));
4062 set_huge_pte_at(mm, haddr, ptep, new_pte);
4064 hugetlb_count_add(pages_per_huge_page(h), mm);
4065 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4066 /* Optimization, do the COW without a second fault */
4067 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4073 * Only make newly allocated pages active. Existing pages found
4074 * in the pagecache could be !page_huge_active() if they have been
4075 * isolated for migration.
4078 set_page_huge_active(page);
4088 restore_reserve_on_error(h, vma, haddr, page);
4094 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4096 unsigned long key[2];
4099 key[0] = (unsigned long) mapping;
4102 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4104 return hash & (num_fault_mutexes - 1);
4108 * For uniprocesor systems we always use a single mutex, so just
4109 * return 0 and avoid the hashing overhead.
4111 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4117 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4118 unsigned long address, unsigned int flags)
4125 struct page *page = NULL;
4126 struct page *pagecache_page = NULL;
4127 struct hstate *h = hstate_vma(vma);
4128 struct address_space *mapping;
4129 int need_wait_lock = 0;
4130 unsigned long haddr = address & huge_page_mask(h);
4132 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4135 * Since we hold no locks, ptep could be stale. That is
4136 * OK as we are only making decisions based on content and
4137 * not actually modifying content here.
4139 entry = huge_ptep_get(ptep);
4140 if (unlikely(is_hugetlb_entry_migration(entry))) {
4141 migration_entry_wait_huge(vma, mm, ptep);
4143 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4144 return VM_FAULT_HWPOISON_LARGE |
4145 VM_FAULT_SET_HINDEX(hstate_index(h));
4147 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4149 return VM_FAULT_OOM;
4153 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4154 * until finished with ptep. This prevents huge_pmd_unshare from
4155 * being called elsewhere and making the ptep no longer valid.
4157 * ptep could have already be assigned via huge_pte_offset. That
4158 * is OK, as huge_pte_alloc will return the same value unless
4159 * something has changed.
4161 mapping = vma->vm_file->f_mapping;
4162 i_mmap_lock_read(mapping);
4163 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4165 i_mmap_unlock_read(mapping);
4166 return VM_FAULT_OOM;
4170 * Serialize hugepage allocation and instantiation, so that we don't
4171 * get spurious allocation failures if two CPUs race to instantiate
4172 * the same page in the page cache.
4174 idx = vma_hugecache_offset(h, vma, haddr);
4175 hash = hugetlb_fault_mutex_hash(mapping, idx);
4176 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4178 entry = huge_ptep_get(ptep);
4179 if (huge_pte_none(entry)) {
4180 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4187 * entry could be a migration/hwpoison entry at this point, so this
4188 * check prevents the kernel from going below assuming that we have
4189 * a active hugepage in pagecache. This goto expects the 2nd page fault,
4190 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
4193 if (!pte_present(entry))
4197 * If we are going to COW the mapping later, we examine the pending
4198 * reservations for this page now. This will ensure that any
4199 * allocations necessary to record that reservation occur outside the
4200 * spinlock. For private mappings, we also lookup the pagecache
4201 * page now as it is used to determine if a reservation has been
4204 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4205 if (vma_needs_reservation(h, vma, haddr) < 0) {
4209 /* Just decrements count, does not deallocate */
4210 vma_end_reservation(h, vma, haddr);
4212 if (!(vma->vm_flags & VM_MAYSHARE))
4213 pagecache_page = hugetlbfs_pagecache_page(h,
4217 ptl = huge_pte_lock(h, mm, ptep);
4219 /* Check for a racing update before calling hugetlb_cow */
4220 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4224 * hugetlb_cow() requires page locks of pte_page(entry) and
4225 * pagecache_page, so here we need take the former one
4226 * when page != pagecache_page or !pagecache_page.
4228 page = pte_page(entry);
4229 if (page != pagecache_page)
4230 if (!trylock_page(page)) {
4237 if (flags & FAULT_FLAG_WRITE) {
4238 if (!huge_pte_write(entry)) {
4239 ret = hugetlb_cow(mm, vma, address, ptep,
4240 pagecache_page, ptl);
4243 entry = huge_pte_mkdirty(entry);
4245 entry = pte_mkyoung(entry);
4246 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4247 flags & FAULT_FLAG_WRITE))
4248 update_mmu_cache(vma, haddr, ptep);
4250 if (page != pagecache_page)
4256 if (pagecache_page) {
4257 unlock_page(pagecache_page);
4258 put_page(pagecache_page);
4261 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4262 i_mmap_unlock_read(mapping);
4264 * Generally it's safe to hold refcount during waiting page lock. But
4265 * here we just wait to defer the next page fault to avoid busy loop and
4266 * the page is not used after unlocked before returning from the current
4267 * page fault. So we are safe from accessing freed page, even if we wait
4268 * here without taking refcount.
4271 wait_on_page_locked(page);
4276 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4277 * modifications for huge pages.
4279 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4281 struct vm_area_struct *dst_vma,
4282 unsigned long dst_addr,
4283 unsigned long src_addr,
4284 struct page **pagep)
4286 struct address_space *mapping;
4289 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4290 struct hstate *h = hstate_vma(dst_vma);
4298 page = alloc_huge_page(dst_vma, dst_addr, 0);
4302 ret = copy_huge_page_from_user(page,
4303 (const void __user *) src_addr,
4304 pages_per_huge_page(h), false);
4306 /* fallback to copy_from_user outside mmap_sem */
4307 if (unlikely(ret)) {
4310 /* don't free the page */
4319 * The memory barrier inside __SetPageUptodate makes sure that
4320 * preceding stores to the page contents become visible before
4321 * the set_pte_at() write.
4323 __SetPageUptodate(page);
4325 mapping = dst_vma->vm_file->f_mapping;
4326 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4329 * If shared, add to page cache
4332 size = i_size_read(mapping->host) >> huge_page_shift(h);
4335 goto out_release_nounlock;
4338 * Serialization between remove_inode_hugepages() and
4339 * huge_add_to_page_cache() below happens through the
4340 * hugetlb_fault_mutex_table that here must be hold by
4343 ret = huge_add_to_page_cache(page, mapping, idx);
4345 goto out_release_nounlock;
4348 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4352 * Recheck the i_size after holding PT lock to make sure not
4353 * to leave any page mapped (as page_mapped()) beyond the end
4354 * of the i_size (remove_inode_hugepages() is strict about
4355 * enforcing that). If we bail out here, we'll also leave a
4356 * page in the radix tree in the vm_shared case beyond the end
4357 * of the i_size, but remove_inode_hugepages() will take care
4358 * of it as soon as we drop the hugetlb_fault_mutex_table.
4360 size = i_size_read(mapping->host) >> huge_page_shift(h);
4363 goto out_release_unlock;
4366 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4367 goto out_release_unlock;
4370 page_dup_rmap(page, true);
4372 ClearPagePrivate(page);
4373 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4376 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4377 if (dst_vma->vm_flags & VM_WRITE)
4378 _dst_pte = huge_pte_mkdirty(_dst_pte);
4379 _dst_pte = pte_mkyoung(_dst_pte);
4381 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4383 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4384 dst_vma->vm_flags & VM_WRITE);
4385 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4387 /* No need to invalidate - it was non-present before */
4388 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4391 set_page_huge_active(page);
4401 out_release_nounlock:
4406 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4407 struct page **pages, struct vm_area_struct **vmas,
4408 unsigned long *position, unsigned long *nr_pages,
4409 long i, unsigned int flags, int *locked)
4411 unsigned long pfn_offset;
4412 unsigned long vaddr = *position;
4413 unsigned long remainder = *nr_pages;
4414 struct hstate *h = hstate_vma(vma);
4417 while (vaddr < vma->vm_end && remainder) {
4419 spinlock_t *ptl = NULL;
4424 * If we have a pending SIGKILL, don't keep faulting pages and
4425 * potentially allocating memory.
4427 if (fatal_signal_pending(current)) {
4433 * Some archs (sparc64, sh*) have multiple pte_ts to
4434 * each hugepage. We have to make sure we get the
4435 * first, for the page indexing below to work.
4437 * Note that page table lock is not held when pte is null.
4439 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4442 ptl = huge_pte_lock(h, mm, pte);
4443 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4446 * When coredumping, it suits get_dump_page if we just return
4447 * an error where there's an empty slot with no huge pagecache
4448 * to back it. This way, we avoid allocating a hugepage, and
4449 * the sparse dumpfile avoids allocating disk blocks, but its
4450 * huge holes still show up with zeroes where they need to be.
4452 if (absent && (flags & FOLL_DUMP) &&
4453 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4461 * We need call hugetlb_fault for both hugepages under migration
4462 * (in which case hugetlb_fault waits for the migration,) and
4463 * hwpoisoned hugepages (in which case we need to prevent the
4464 * caller from accessing to them.) In order to do this, we use
4465 * here is_swap_pte instead of is_hugetlb_entry_migration and
4466 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4467 * both cases, and because we can't follow correct pages
4468 * directly from any kind of swap entries.
4470 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4471 ((flags & FOLL_WRITE) &&
4472 !huge_pte_write(huge_ptep_get(pte)))) {
4474 unsigned int fault_flags = 0;
4478 if (flags & FOLL_WRITE)
4479 fault_flags |= FAULT_FLAG_WRITE;
4481 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4482 FAULT_FLAG_KILLABLE;
4483 if (flags & FOLL_NOWAIT)
4484 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4485 FAULT_FLAG_RETRY_NOWAIT;
4486 if (flags & FOLL_TRIED) {
4488 * Note: FAULT_FLAG_ALLOW_RETRY and
4489 * FAULT_FLAG_TRIED can co-exist
4491 fault_flags |= FAULT_FLAG_TRIED;
4493 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4494 if (ret & VM_FAULT_ERROR) {
4495 err = vm_fault_to_errno(ret, flags);
4499 if (ret & VM_FAULT_RETRY) {
4501 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4505 * VM_FAULT_RETRY must not return an
4506 * error, it will return zero
4509 * No need to update "position" as the
4510 * caller will not check it after
4511 * *nr_pages is set to 0.
4518 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4519 page = pte_page(huge_ptep_get(pte));
4522 * If subpage information not requested, update counters
4523 * and skip the same_page loop below.
4525 if (!pages && !vmas && !pfn_offset &&
4526 (vaddr + huge_page_size(h) < vma->vm_end) &&
4527 (remainder >= pages_per_huge_page(h))) {
4528 vaddr += huge_page_size(h);
4529 remainder -= pages_per_huge_page(h);
4530 i += pages_per_huge_page(h);
4537 pages[i] = mem_map_offset(page, pfn_offset);
4539 * try_grab_page() should always succeed here, because:
4540 * a) we hold the ptl lock, and b) we've just checked
4541 * that the huge page is present in the page tables. If
4542 * the huge page is present, then the tail pages must
4543 * also be present. The ptl prevents the head page and
4544 * tail pages from being rearranged in any way. So this
4545 * page must be available at this point, unless the page
4546 * refcount overflowed:
4548 if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) {
4563 if (vaddr < vma->vm_end && remainder &&
4564 pfn_offset < pages_per_huge_page(h)) {
4566 * We use pfn_offset to avoid touching the pageframes
4567 * of this compound page.
4573 *nr_pages = remainder;
4575 * setting position is actually required only if remainder is
4576 * not zero but it's faster not to add a "if (remainder)"
4584 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4586 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4589 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4592 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4593 unsigned long address, unsigned long end, pgprot_t newprot)
4595 struct mm_struct *mm = vma->vm_mm;
4596 unsigned long start = address;
4599 struct hstate *h = hstate_vma(vma);
4600 unsigned long pages = 0;
4601 bool shared_pmd = false;
4602 struct mmu_notifier_range range;
4605 * In the case of shared PMDs, the area to flush could be beyond
4606 * start/end. Set range.start/range.end to cover the maximum possible
4607 * range if PMD sharing is possible.
4609 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
4610 0, vma, mm, start, end);
4611 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4613 BUG_ON(address >= end);
4614 flush_cache_range(vma, range.start, range.end);
4616 mmu_notifier_invalidate_range_start(&range);
4617 i_mmap_lock_write(vma->vm_file->f_mapping);
4618 for (; address < end; address += huge_page_size(h)) {
4620 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4623 ptl = huge_pte_lock(h, mm, ptep);
4624 if (huge_pmd_unshare(mm, &address, ptep)) {
4630 pte = huge_ptep_get(ptep);
4631 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4635 if (unlikely(is_hugetlb_entry_migration(pte))) {
4636 swp_entry_t entry = pte_to_swp_entry(pte);
4638 if (is_write_migration_entry(entry)) {
4641 make_migration_entry_read(&entry);
4642 newpte = swp_entry_to_pte(entry);
4643 set_huge_swap_pte_at(mm, address, ptep,
4644 newpte, huge_page_size(h));
4650 if (!huge_pte_none(pte)) {
4653 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
4654 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
4655 pte = arch_make_huge_pte(pte, vma, NULL, 0);
4656 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
4662 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4663 * may have cleared our pud entry and done put_page on the page table:
4664 * once we release i_mmap_rwsem, another task can do the final put_page
4665 * and that page table be reused and filled with junk. If we actually
4666 * did unshare a page of pmds, flush the range corresponding to the pud.
4669 flush_hugetlb_tlb_range(vma, range.start, range.end);
4671 flush_hugetlb_tlb_range(vma, start, end);
4673 * No need to call mmu_notifier_invalidate_range() we are downgrading
4674 * page table protection not changing it to point to a new page.
4676 * See Documentation/vm/mmu_notifier.rst
4678 i_mmap_unlock_write(vma->vm_file->f_mapping);
4679 mmu_notifier_invalidate_range_end(&range);
4681 return pages << h->order;
4684 int hugetlb_reserve_pages(struct inode *inode,
4686 struct vm_area_struct *vma,
4687 vm_flags_t vm_flags)
4690 struct hstate *h = hstate_inode(inode);
4691 struct hugepage_subpool *spool = subpool_inode(inode);
4692 struct resv_map *resv_map;
4695 /* This should never happen */
4697 VM_WARN(1, "%s called with a negative range\n", __func__);
4702 * Only apply hugepage reservation if asked. At fault time, an
4703 * attempt will be made for VM_NORESERVE to allocate a page
4704 * without using reserves
4706 if (vm_flags & VM_NORESERVE)
4710 * Shared mappings base their reservation on the number of pages that
4711 * are already allocated on behalf of the file. Private mappings need
4712 * to reserve the full area even if read-only as mprotect() may be
4713 * called to make the mapping read-write. Assume !vma is a shm mapping
4715 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4717 * resv_map can not be NULL as hugetlb_reserve_pages is only
4718 * called for inodes for which resv_maps were created (see
4719 * hugetlbfs_get_inode).
4721 resv_map = inode_resv_map(inode);
4723 chg = region_chg(resv_map, from, to);
4726 resv_map = resv_map_alloc();
4732 set_vma_resv_map(vma, resv_map);
4733 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4742 * There must be enough pages in the subpool for the mapping. If
4743 * the subpool has a minimum size, there may be some global
4744 * reservations already in place (gbl_reserve).
4746 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4747 if (gbl_reserve < 0) {
4753 * Check enough hugepages are available for the reservation.
4754 * Hand the pages back to the subpool if there are not
4756 ret = hugetlb_acct_memory(h, gbl_reserve);
4758 /* put back original number of pages, chg */
4759 (void)hugepage_subpool_put_pages(spool, chg);
4764 * Account for the reservations made. Shared mappings record regions
4765 * that have reservations as they are shared by multiple VMAs.
4766 * When the last VMA disappears, the region map says how much
4767 * the reservation was and the page cache tells how much of
4768 * the reservation was consumed. Private mappings are per-VMA and
4769 * only the consumed reservations are tracked. When the VMA
4770 * disappears, the original reservation is the VMA size and the
4771 * consumed reservations are stored in the map. Hence, nothing
4772 * else has to be done for private mappings here
4774 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4775 long add = region_add(resv_map, from, to);
4777 if (unlikely(chg > add)) {
4779 * pages in this range were added to the reserve
4780 * map between region_chg and region_add. This
4781 * indicates a race with alloc_huge_page. Adjust
4782 * the subpool and reserve counts modified above
4783 * based on the difference.
4787 rsv_adjust = hugepage_subpool_put_pages(spool,
4789 hugetlb_acct_memory(h, -rsv_adjust);
4794 if (!vma || vma->vm_flags & VM_MAYSHARE)
4795 /* Don't call region_abort if region_chg failed */
4797 region_abort(resv_map, from, to);
4798 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4799 kref_put(&resv_map->refs, resv_map_release);
4803 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4806 struct hstate *h = hstate_inode(inode);
4807 struct resv_map *resv_map = inode_resv_map(inode);
4809 struct hugepage_subpool *spool = subpool_inode(inode);
4813 * Since this routine can be called in the evict inode path for all
4814 * hugetlbfs inodes, resv_map could be NULL.
4817 chg = region_del(resv_map, start, end);
4819 * region_del() can fail in the rare case where a region
4820 * must be split and another region descriptor can not be
4821 * allocated. If end == LONG_MAX, it will not fail.
4827 spin_lock(&inode->i_lock);
4828 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4829 spin_unlock(&inode->i_lock);
4832 * If the subpool has a minimum size, the number of global
4833 * reservations to be released may be adjusted.
4835 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4836 hugetlb_acct_memory(h, -gbl_reserve);
4841 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4842 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4843 struct vm_area_struct *vma,
4844 unsigned long addr, pgoff_t idx)
4846 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4848 unsigned long sbase = saddr & PUD_MASK;
4849 unsigned long s_end = sbase + PUD_SIZE;
4851 /* Allow segments to share if only one is marked locked */
4852 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4853 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4856 * match the virtual addresses, permission and the alignment of the
4859 if (pmd_index(addr) != pmd_index(saddr) ||
4860 vm_flags != svm_flags ||
4861 sbase < svma->vm_start || svma->vm_end < s_end)
4867 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4869 unsigned long base = addr & PUD_MASK;
4870 unsigned long end = base + PUD_SIZE;
4873 * check on proper vm_flags and page table alignment
4875 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
4881 * Determine if start,end range within vma could be mapped by shared pmd.
4882 * If yes, adjust start and end to cover range associated with possible
4883 * shared pmd mappings.
4885 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4886 unsigned long *start, unsigned long *end)
4888 unsigned long check_addr = *start;
4890 if (!(vma->vm_flags & VM_MAYSHARE))
4893 for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
4894 unsigned long a_start = check_addr & PUD_MASK;
4895 unsigned long a_end = a_start + PUD_SIZE;
4898 * If sharing is possible, adjust start/end if necessary.
4900 if (range_in_vma(vma, a_start, a_end)) {
4901 if (a_start < *start)
4910 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4911 * and returns the corresponding pte. While this is not necessary for the
4912 * !shared pmd case because we can allocate the pmd later as well, it makes the
4913 * code much cleaner.
4915 * This routine must be called with i_mmap_rwsem held in at least read mode.
4916 * For hugetlbfs, this prevents removal of any page table entries associated
4917 * with the address space. This is important as we are setting up sharing
4918 * based on existing page table entries (mappings).
4920 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4922 struct vm_area_struct *vma = find_vma(mm, addr);
4923 struct address_space *mapping = vma->vm_file->f_mapping;
4924 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4926 struct vm_area_struct *svma;
4927 unsigned long saddr;
4932 if (!vma_shareable(vma, addr))
4933 return (pte_t *)pmd_alloc(mm, pud, addr);
4935 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4939 saddr = page_table_shareable(svma, vma, addr, idx);
4941 spte = huge_pte_offset(svma->vm_mm, saddr,
4942 vma_mmu_pagesize(svma));
4944 get_page(virt_to_page(spte));
4953 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4954 if (pud_none(*pud)) {
4955 pud_populate(mm, pud,
4956 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4959 put_page(virt_to_page(spte));
4963 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4968 * unmap huge page backed by shared pte.
4970 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4971 * indicated by page_count > 1, unmap is achieved by clearing pud and
4972 * decrementing the ref count. If count == 1, the pte page is not shared.
4974 * Called with page table lock held and i_mmap_rwsem held in write mode.
4976 * returns: 1 successfully unmapped a shared pte page
4977 * 0 the underlying pte page is not shared, or it is the last user
4979 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4981 pgd_t *pgd = pgd_offset(mm, *addr);
4982 p4d_t *p4d = p4d_offset(pgd, *addr);
4983 pud_t *pud = pud_offset(p4d, *addr);
4985 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4986 if (page_count(virt_to_page(ptep)) == 1)
4990 put_page(virt_to_page(ptep));
4992 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4995 #define want_pmd_share() (1)
4996 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4997 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5002 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
5007 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5008 unsigned long *start, unsigned long *end)
5011 #define want_pmd_share() (0)
5012 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5014 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5015 pte_t *huge_pte_alloc(struct mm_struct *mm,
5016 unsigned long addr, unsigned long sz)
5023 pgd = pgd_offset(mm, addr);
5024 p4d = p4d_alloc(mm, pgd, addr);
5027 pud = pud_alloc(mm, p4d, addr);
5029 if (sz == PUD_SIZE) {
5032 BUG_ON(sz != PMD_SIZE);
5033 if (want_pmd_share() && pud_none(*pud))
5034 pte = huge_pmd_share(mm, addr, pud);
5036 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5039 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
5045 * huge_pte_offset() - Walk the page table to resolve the hugepage
5046 * entry at address @addr
5048 * Return: Pointer to page table or swap entry (PUD or PMD) for
5049 * address @addr, or NULL if a p*d_none() entry is encountered and the
5050 * size @sz doesn't match the hugepage size at this level of the page
5053 pte_t *huge_pte_offset(struct mm_struct *mm,
5054 unsigned long addr, unsigned long sz)
5061 pgd = pgd_offset(mm, addr);
5062 if (!pgd_present(*pgd))
5064 p4d = p4d_offset(pgd, addr);
5065 if (!p4d_present(*p4d))
5068 pud = pud_offset(p4d, addr);
5069 if (sz != PUD_SIZE && pud_none(*pud))
5071 /* hugepage or swap? */
5072 if (pud_huge(*pud) || !pud_present(*pud))
5073 return (pte_t *)pud;
5075 pmd = pmd_offset(pud, addr);
5076 if (sz != PMD_SIZE && pmd_none(*pmd))
5078 /* hugepage or swap? */
5079 if (pmd_huge(*pmd) || !pmd_present(*pmd))
5080 return (pte_t *)pmd;
5085 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5088 * These functions are overwritable if your architecture needs its own
5091 struct page * __weak
5092 follow_huge_addr(struct mm_struct *mm, unsigned long address,
5095 return ERR_PTR(-EINVAL);
5098 struct page * __weak
5099 follow_huge_pd(struct vm_area_struct *vma,
5100 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5102 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5106 struct page * __weak
5107 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
5108 pmd_t *pmd, int flags)
5110 struct page *page = NULL;
5114 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5115 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
5116 (FOLL_PIN | FOLL_GET)))
5120 ptl = pmd_lockptr(mm, pmd);
5123 * make sure that the address range covered by this pmd is not
5124 * unmapped from other threads.
5126 if (!pmd_huge(*pmd))
5128 pte = huge_ptep_get((pte_t *)pmd);
5129 if (pte_present(pte)) {
5130 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
5132 * try_grab_page() should always succeed here, because: a) we
5133 * hold the pmd (ptl) lock, and b) we've just checked that the
5134 * huge pmd (head) page is present in the page tables. The ptl
5135 * prevents the head page and tail pages from being rearranged
5136 * in any way. So this page must be available at this point,
5137 * unless the page refcount overflowed:
5139 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5144 if (is_hugetlb_entry_migration(pte)) {
5146 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5150 * hwpoisoned entry is treated as no_page_table in
5151 * follow_page_mask().
5159 struct page * __weak
5160 follow_huge_pud(struct mm_struct *mm, unsigned long address,
5161 pud_t *pud, int flags)
5163 if (flags & (FOLL_GET | FOLL_PIN))
5166 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5169 struct page * __weak
5170 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5172 if (flags & (FOLL_GET | FOLL_PIN))
5175 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5178 bool isolate_huge_page(struct page *page, struct list_head *list)
5182 VM_BUG_ON_PAGE(!PageHead(page), page);
5183 spin_lock(&hugetlb_lock);
5184 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
5188 clear_page_huge_active(page);
5189 list_move_tail(&page->lru, list);
5191 spin_unlock(&hugetlb_lock);
5195 void putback_active_hugepage(struct page *page)
5197 VM_BUG_ON_PAGE(!PageHead(page), page);
5198 spin_lock(&hugetlb_lock);
5199 set_page_huge_active(page);
5200 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5201 spin_unlock(&hugetlb_lock);
5205 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5207 struct hstate *h = page_hstate(oldpage);
5209 hugetlb_cgroup_migrate(oldpage, newpage);
5210 set_page_owner_migrate_reason(newpage, reason);
5213 * transfer temporary state of the new huge page. This is
5214 * reverse to other transitions because the newpage is going to
5215 * be final while the old one will be freed so it takes over
5216 * the temporary status.
5218 * Also note that we have to transfer the per-node surplus state
5219 * here as well otherwise the global surplus count will not match
5222 if (PageHugeTemporary(newpage)) {
5223 int old_nid = page_to_nid(oldpage);
5224 int new_nid = page_to_nid(newpage);
5226 SetPageHugeTemporary(oldpage);
5227 ClearPageHugeTemporary(newpage);
5229 spin_lock(&hugetlb_lock);
5230 if (h->surplus_huge_pages_node[old_nid]) {
5231 h->surplus_huge_pages_node[old_nid]--;
5232 h->surplus_huge_pages_node[new_nid]++;
5234 spin_unlock(&hugetlb_lock);