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;
1325 pgoff_t __basepage_index(struct page *page)
1327 struct page *page_head = compound_head(page);
1328 pgoff_t index = page_index(page_head);
1329 unsigned long compound_idx;
1331 if (!PageHuge(page_head))
1332 return page_index(page);
1334 if (compound_order(page_head) >= MAX_ORDER)
1335 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1337 compound_idx = page - page_head;
1339 return (index << compound_order(page_head)) + compound_idx;
1342 static struct page *alloc_buddy_huge_page(struct hstate *h,
1343 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1344 nodemask_t *node_alloc_noretry)
1346 int order = huge_page_order(h);
1348 bool alloc_try_hard = true;
1351 * By default we always try hard to allocate the page with
1352 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1353 * a loop (to adjust global huge page counts) and previous allocation
1354 * failed, do not continue to try hard on the same node. Use the
1355 * node_alloc_noretry bitmap to manage this state information.
1357 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1358 alloc_try_hard = false;
1359 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1361 gfp_mask |= __GFP_RETRY_MAYFAIL;
1362 if (nid == NUMA_NO_NODE)
1363 nid = numa_mem_id();
1364 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1366 __count_vm_event(HTLB_BUDDY_PGALLOC);
1368 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1371 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1372 * indicates an overall state change. Clear bit so that we resume
1373 * normal 'try hard' allocations.
1375 if (node_alloc_noretry && page && !alloc_try_hard)
1376 node_clear(nid, *node_alloc_noretry);
1379 * If we tried hard to get a page but failed, set bit so that
1380 * subsequent attempts will not try as hard until there is an
1381 * overall state change.
1383 if (node_alloc_noretry && !page && alloc_try_hard)
1384 node_set(nid, *node_alloc_noretry);
1390 * Common helper to allocate a fresh hugetlb page. All specific allocators
1391 * should use this function to get new hugetlb pages
1393 static struct page *alloc_fresh_huge_page(struct hstate *h,
1394 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1395 nodemask_t *node_alloc_noretry)
1399 if (hstate_is_gigantic(h))
1400 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1402 page = alloc_buddy_huge_page(h, gfp_mask,
1403 nid, nmask, node_alloc_noretry);
1407 if (hstate_is_gigantic(h))
1408 prep_compound_gigantic_page(page, huge_page_order(h));
1409 prep_new_huge_page(h, page, page_to_nid(page));
1415 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1418 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1419 nodemask_t *node_alloc_noretry)
1423 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1425 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1426 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1427 node_alloc_noretry);
1435 put_page(page); /* free it into the hugepage allocator */
1441 * Free huge page from pool from next node to free.
1442 * Attempt to keep persistent huge pages more or less
1443 * balanced over allowed nodes.
1444 * Called with hugetlb_lock locked.
1446 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1452 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1454 * If we're returning unused surplus pages, only examine
1455 * nodes with surplus pages.
1457 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1458 !list_empty(&h->hugepage_freelists[node])) {
1460 list_entry(h->hugepage_freelists[node].next,
1462 list_del(&page->lru);
1463 h->free_huge_pages--;
1464 h->free_huge_pages_node[node]--;
1466 h->surplus_huge_pages--;
1467 h->surplus_huge_pages_node[node]--;
1469 update_and_free_page(h, page);
1479 * Dissolve a given free hugepage into free buddy pages. This function does
1480 * nothing for in-use hugepages and non-hugepages.
1481 * This function returns values like below:
1483 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1484 * (allocated or reserved.)
1485 * 0: successfully dissolved free hugepages or the page is not a
1486 * hugepage (considered as already dissolved)
1488 int dissolve_free_huge_page(struct page *page)
1492 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1493 if (!PageHuge(page))
1496 spin_lock(&hugetlb_lock);
1497 if (!PageHuge(page)) {
1502 if (!page_count(page)) {
1503 struct page *head = compound_head(page);
1504 struct hstate *h = page_hstate(head);
1505 int nid = page_to_nid(head);
1506 if (h->free_huge_pages - h->resv_huge_pages == 0)
1509 * Move PageHWPoison flag from head page to the raw error page,
1510 * which makes any subpages rather than the error page reusable.
1512 if (PageHWPoison(head) && page != head) {
1513 SetPageHWPoison(page);
1514 ClearPageHWPoison(head);
1516 list_del(&head->lru);
1517 h->free_huge_pages--;
1518 h->free_huge_pages_node[nid]--;
1519 h->max_huge_pages--;
1520 update_and_free_page(h, head);
1524 spin_unlock(&hugetlb_lock);
1529 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1530 * make specified memory blocks removable from the system.
1531 * Note that this will dissolve a free gigantic hugepage completely, if any
1532 * part of it lies within the given range.
1533 * Also note that if dissolve_free_huge_page() returns with an error, all
1534 * free hugepages that were dissolved before that error are lost.
1536 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1542 if (!hugepages_supported())
1545 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1546 page = pfn_to_page(pfn);
1547 rc = dissolve_free_huge_page(page);
1556 * Allocates a fresh surplus page from the page allocator.
1558 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1559 int nid, nodemask_t *nmask)
1561 struct page *page = NULL;
1563 if (hstate_is_gigantic(h))
1566 spin_lock(&hugetlb_lock);
1567 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1569 spin_unlock(&hugetlb_lock);
1571 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1575 spin_lock(&hugetlb_lock);
1577 * We could have raced with the pool size change.
1578 * Double check that and simply deallocate the new page
1579 * if we would end up overcommiting the surpluses. Abuse
1580 * temporary page to workaround the nasty free_huge_page
1583 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1584 SetPageHugeTemporary(page);
1585 spin_unlock(&hugetlb_lock);
1589 h->surplus_huge_pages++;
1590 h->surplus_huge_pages_node[page_to_nid(page)]++;
1594 spin_unlock(&hugetlb_lock);
1599 struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1600 int nid, nodemask_t *nmask)
1604 if (hstate_is_gigantic(h))
1607 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1612 * We do not account these pages as surplus because they are only
1613 * temporary and will be released properly on the last reference
1615 SetPageHugeTemporary(page);
1621 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1624 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1625 struct vm_area_struct *vma, unsigned long addr)
1628 struct mempolicy *mpol;
1629 gfp_t gfp_mask = htlb_alloc_mask(h);
1631 nodemask_t *nodemask;
1633 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1634 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1635 mpol_cond_put(mpol);
1640 /* page migration callback function */
1641 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1643 gfp_t gfp_mask = htlb_alloc_mask(h);
1644 struct page *page = NULL;
1646 if (nid != NUMA_NO_NODE)
1647 gfp_mask |= __GFP_THISNODE;
1649 spin_lock(&hugetlb_lock);
1650 if (h->free_huge_pages - h->resv_huge_pages > 0)
1651 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1652 spin_unlock(&hugetlb_lock);
1655 page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1660 /* page migration callback function */
1661 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1664 gfp_t gfp_mask = htlb_alloc_mask(h);
1666 spin_lock(&hugetlb_lock);
1667 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1670 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1672 spin_unlock(&hugetlb_lock);
1676 spin_unlock(&hugetlb_lock);
1678 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1681 /* mempolicy aware migration callback */
1682 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1683 unsigned long address)
1685 struct mempolicy *mpol;
1686 nodemask_t *nodemask;
1691 gfp_mask = htlb_alloc_mask(h);
1692 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1693 page = alloc_huge_page_nodemask(h, node, nodemask);
1694 mpol_cond_put(mpol);
1700 * Increase the hugetlb pool such that it can accommodate a reservation
1703 static int gather_surplus_pages(struct hstate *h, int delta)
1705 struct list_head surplus_list;
1706 struct page *page, *tmp;
1708 int needed, allocated;
1709 bool alloc_ok = true;
1711 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1713 h->resv_huge_pages += delta;
1718 INIT_LIST_HEAD(&surplus_list);
1722 spin_unlock(&hugetlb_lock);
1723 for (i = 0; i < needed; i++) {
1724 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1725 NUMA_NO_NODE, NULL);
1730 list_add(&page->lru, &surplus_list);
1736 * After retaking hugetlb_lock, we need to recalculate 'needed'
1737 * because either resv_huge_pages or free_huge_pages may have changed.
1739 spin_lock(&hugetlb_lock);
1740 needed = (h->resv_huge_pages + delta) -
1741 (h->free_huge_pages + allocated);
1746 * We were not able to allocate enough pages to
1747 * satisfy the entire reservation so we free what
1748 * we've allocated so far.
1753 * The surplus_list now contains _at_least_ the number of extra pages
1754 * needed to accommodate the reservation. Add the appropriate number
1755 * of pages to the hugetlb pool and free the extras back to the buddy
1756 * allocator. Commit the entire reservation here to prevent another
1757 * process from stealing the pages as they are added to the pool but
1758 * before they are reserved.
1760 needed += allocated;
1761 h->resv_huge_pages += delta;
1764 /* Free the needed pages to the hugetlb pool */
1765 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1769 * This page is now managed by the hugetlb allocator and has
1770 * no users -- drop the buddy allocator's reference.
1772 put_page_testzero(page);
1773 VM_BUG_ON_PAGE(page_count(page), page);
1774 enqueue_huge_page(h, page);
1777 spin_unlock(&hugetlb_lock);
1779 /* Free unnecessary surplus pages to the buddy allocator */
1780 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1782 spin_lock(&hugetlb_lock);
1788 * This routine has two main purposes:
1789 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1790 * in unused_resv_pages. This corresponds to the prior adjustments made
1791 * to the associated reservation map.
1792 * 2) Free any unused surplus pages that may have been allocated to satisfy
1793 * the reservation. As many as unused_resv_pages may be freed.
1795 * Called with hugetlb_lock held. However, the lock could be dropped (and
1796 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1797 * we must make sure nobody else can claim pages we are in the process of
1798 * freeing. Do this by ensuring resv_huge_page always is greater than the
1799 * number of huge pages we plan to free when dropping the lock.
1801 static void return_unused_surplus_pages(struct hstate *h,
1802 unsigned long unused_resv_pages)
1804 unsigned long nr_pages;
1806 /* Cannot return gigantic pages currently */
1807 if (hstate_is_gigantic(h))
1811 * Part (or even all) of the reservation could have been backed
1812 * by pre-allocated pages. Only free surplus pages.
1814 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1817 * We want to release as many surplus pages as possible, spread
1818 * evenly across all nodes with memory. Iterate across these nodes
1819 * until we can no longer free unreserved surplus pages. This occurs
1820 * when the nodes with surplus pages have no free pages.
1821 * free_pool_huge_page() will balance the the freed pages across the
1822 * on-line nodes with memory and will handle the hstate accounting.
1824 * Note that we decrement resv_huge_pages as we free the pages. If
1825 * we drop the lock, resv_huge_pages will still be sufficiently large
1826 * to cover subsequent pages we may free.
1828 while (nr_pages--) {
1829 h->resv_huge_pages--;
1830 unused_resv_pages--;
1831 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1833 cond_resched_lock(&hugetlb_lock);
1837 /* Fully uncommit the reservation */
1838 h->resv_huge_pages -= unused_resv_pages;
1843 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1844 * are used by the huge page allocation routines to manage reservations.
1846 * vma_needs_reservation is called to determine if the huge page at addr
1847 * within the vma has an associated reservation. If a reservation is
1848 * needed, the value 1 is returned. The caller is then responsible for
1849 * managing the global reservation and subpool usage counts. After
1850 * the huge page has been allocated, vma_commit_reservation is called
1851 * to add the page to the reservation map. If the page allocation fails,
1852 * the reservation must be ended instead of committed. vma_end_reservation
1853 * is called in such cases.
1855 * In the normal case, vma_commit_reservation returns the same value
1856 * as the preceding vma_needs_reservation call. The only time this
1857 * is not the case is if a reserve map was changed between calls. It
1858 * is the responsibility of the caller to notice the difference and
1859 * take appropriate action.
1861 * vma_add_reservation is used in error paths where a reservation must
1862 * be restored when a newly allocated huge page must be freed. It is
1863 * to be called after calling vma_needs_reservation to determine if a
1864 * reservation exists.
1866 enum vma_resv_mode {
1872 static long __vma_reservation_common(struct hstate *h,
1873 struct vm_area_struct *vma, unsigned long addr,
1874 enum vma_resv_mode mode)
1876 struct resv_map *resv;
1880 resv = vma_resv_map(vma);
1884 idx = vma_hugecache_offset(h, vma, addr);
1886 case VMA_NEEDS_RESV:
1887 ret = region_chg(resv, idx, idx + 1);
1889 case VMA_COMMIT_RESV:
1890 ret = region_add(resv, idx, idx + 1);
1893 region_abort(resv, idx, idx + 1);
1897 if (vma->vm_flags & VM_MAYSHARE)
1898 ret = region_add(resv, idx, idx + 1);
1900 region_abort(resv, idx, idx + 1);
1901 ret = region_del(resv, idx, idx + 1);
1908 if (vma->vm_flags & VM_MAYSHARE)
1910 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1912 * In most cases, reserves always exist for private mappings.
1913 * However, a file associated with mapping could have been
1914 * hole punched or truncated after reserves were consumed.
1915 * As subsequent fault on such a range will not use reserves.
1916 * Subtle - The reserve map for private mappings has the
1917 * opposite meaning than that of shared mappings. If NO
1918 * entry is in the reserve map, it means a reservation exists.
1919 * If an entry exists in the reserve map, it means the
1920 * reservation has already been consumed. As a result, the
1921 * return value of this routine is the opposite of the
1922 * value returned from reserve map manipulation routines above.
1930 return ret < 0 ? ret : 0;
1933 static long vma_needs_reservation(struct hstate *h,
1934 struct vm_area_struct *vma, unsigned long addr)
1936 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1939 static long vma_commit_reservation(struct hstate *h,
1940 struct vm_area_struct *vma, unsigned long addr)
1942 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1945 static void vma_end_reservation(struct hstate *h,
1946 struct vm_area_struct *vma, unsigned long addr)
1948 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1951 static long vma_add_reservation(struct hstate *h,
1952 struct vm_area_struct *vma, unsigned long addr)
1954 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1958 * This routine is called to restore a reservation on error paths. In the
1959 * specific error paths, a huge page was allocated (via alloc_huge_page)
1960 * and is about to be freed. If a reservation for the page existed,
1961 * alloc_huge_page would have consumed the reservation and set PagePrivate
1962 * in the newly allocated page. When the page is freed via free_huge_page,
1963 * the global reservation count will be incremented if PagePrivate is set.
1964 * However, free_huge_page can not adjust the reserve map. Adjust the
1965 * reserve map here to be consistent with global reserve count adjustments
1966 * to be made by free_huge_page.
1968 static void restore_reserve_on_error(struct hstate *h,
1969 struct vm_area_struct *vma, unsigned long address,
1972 if (unlikely(PagePrivate(page))) {
1973 long rc = vma_needs_reservation(h, vma, address);
1975 if (unlikely(rc < 0)) {
1977 * Rare out of memory condition in reserve map
1978 * manipulation. Clear PagePrivate so that
1979 * global reserve count will not be incremented
1980 * by free_huge_page. This will make it appear
1981 * as though the reservation for this page was
1982 * consumed. This may prevent the task from
1983 * faulting in the page at a later time. This
1984 * is better than inconsistent global huge page
1985 * accounting of reserve counts.
1987 ClearPagePrivate(page);
1989 rc = vma_add_reservation(h, vma, address);
1990 if (unlikely(rc < 0))
1992 * See above comment about rare out of
1995 ClearPagePrivate(page);
1997 vma_end_reservation(h, vma, address);
2001 struct page *alloc_huge_page(struct vm_area_struct *vma,
2002 unsigned long addr, int avoid_reserve)
2004 struct hugepage_subpool *spool = subpool_vma(vma);
2005 struct hstate *h = hstate_vma(vma);
2007 long map_chg, map_commit;
2010 struct hugetlb_cgroup *h_cg;
2012 idx = hstate_index(h);
2014 * Examine the region/reserve map to determine if the process
2015 * has a reservation for the page to be allocated. A return
2016 * code of zero indicates a reservation exists (no change).
2018 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2020 return ERR_PTR(-ENOMEM);
2023 * Processes that did not create the mapping will have no
2024 * reserves as indicated by the region/reserve map. Check
2025 * that the allocation will not exceed the subpool limit.
2026 * Allocations for MAP_NORESERVE mappings also need to be
2027 * checked against any subpool limit.
2029 if (map_chg || avoid_reserve) {
2030 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2032 vma_end_reservation(h, vma, addr);
2033 return ERR_PTR(-ENOSPC);
2037 * Even though there was no reservation in the region/reserve
2038 * map, there could be reservations associated with the
2039 * subpool that can be used. This would be indicated if the
2040 * return value of hugepage_subpool_get_pages() is zero.
2041 * However, if avoid_reserve is specified we still avoid even
2042 * the subpool reservations.
2048 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2050 goto out_subpool_put;
2052 spin_lock(&hugetlb_lock);
2054 * glb_chg is passed to indicate whether or not a page must be taken
2055 * from the global free pool (global change). gbl_chg == 0 indicates
2056 * a reservation exists for the allocation.
2058 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2060 spin_unlock(&hugetlb_lock);
2061 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2063 goto out_uncharge_cgroup;
2064 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2065 SetPagePrivate(page);
2066 h->resv_huge_pages--;
2068 spin_lock(&hugetlb_lock);
2069 list_move(&page->lru, &h->hugepage_activelist);
2072 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2073 spin_unlock(&hugetlb_lock);
2075 set_page_private(page, (unsigned long)spool);
2077 map_commit = vma_commit_reservation(h, vma, addr);
2078 if (unlikely(map_chg > map_commit)) {
2080 * The page was added to the reservation map between
2081 * vma_needs_reservation and vma_commit_reservation.
2082 * This indicates a race with hugetlb_reserve_pages.
2083 * Adjust for the subpool count incremented above AND
2084 * in hugetlb_reserve_pages for the same page. Also,
2085 * the reservation count added in hugetlb_reserve_pages
2086 * no longer applies.
2090 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2091 hugetlb_acct_memory(h, -rsv_adjust);
2095 out_uncharge_cgroup:
2096 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2098 if (map_chg || avoid_reserve)
2099 hugepage_subpool_put_pages(spool, 1);
2100 vma_end_reservation(h, vma, addr);
2101 return ERR_PTR(-ENOSPC);
2104 int alloc_bootmem_huge_page(struct hstate *h)
2105 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2106 int __alloc_bootmem_huge_page(struct hstate *h)
2108 struct huge_bootmem_page *m;
2111 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2114 addr = memblock_alloc_try_nid_raw(
2115 huge_page_size(h), huge_page_size(h),
2116 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2119 * Use the beginning of the huge page to store the
2120 * huge_bootmem_page struct (until gather_bootmem
2121 * puts them into the mem_map).
2130 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2131 /* Put them into a private list first because mem_map is not up yet */
2132 INIT_LIST_HEAD(&m->list);
2133 list_add(&m->list, &huge_boot_pages);
2138 static void __init prep_compound_huge_page(struct page *page,
2141 if (unlikely(order > (MAX_ORDER - 1)))
2142 prep_compound_gigantic_page(page, order);
2144 prep_compound_page(page, order);
2147 /* Put bootmem huge pages into the standard lists after mem_map is up */
2148 static void __init gather_bootmem_prealloc(void)
2150 struct huge_bootmem_page *m;
2152 list_for_each_entry(m, &huge_boot_pages, list) {
2153 struct page *page = virt_to_page(m);
2154 struct hstate *h = m->hstate;
2156 WARN_ON(page_count(page) != 1);
2157 prep_compound_huge_page(page, h->order);
2158 WARN_ON(PageReserved(page));
2159 prep_new_huge_page(h, page, page_to_nid(page));
2160 put_page(page); /* free it into the hugepage allocator */
2163 * If we had gigantic hugepages allocated at boot time, we need
2164 * to restore the 'stolen' pages to totalram_pages in order to
2165 * fix confusing memory reports from free(1) and another
2166 * side-effects, like CommitLimit going negative.
2168 if (hstate_is_gigantic(h))
2169 adjust_managed_page_count(page, 1 << h->order);
2174 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2177 nodemask_t *node_alloc_noretry;
2179 if (!hstate_is_gigantic(h)) {
2181 * Bit mask controlling how hard we retry per-node allocations.
2182 * Ignore errors as lower level routines can deal with
2183 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2184 * time, we are likely in bigger trouble.
2186 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2189 /* allocations done at boot time */
2190 node_alloc_noretry = NULL;
2193 /* bit mask controlling how hard we retry per-node allocations */
2194 if (node_alloc_noretry)
2195 nodes_clear(*node_alloc_noretry);
2197 for (i = 0; i < h->max_huge_pages; ++i) {
2198 if (hstate_is_gigantic(h)) {
2199 if (!alloc_bootmem_huge_page(h))
2201 } else if (!alloc_pool_huge_page(h,
2202 &node_states[N_MEMORY],
2203 node_alloc_noretry))
2207 if (i < h->max_huge_pages) {
2210 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2211 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2212 h->max_huge_pages, buf, i);
2213 h->max_huge_pages = i;
2216 kfree(node_alloc_noretry);
2219 static void __init hugetlb_init_hstates(void)
2223 for_each_hstate(h) {
2224 if (minimum_order > huge_page_order(h))
2225 minimum_order = huge_page_order(h);
2227 /* oversize hugepages were init'ed in early boot */
2228 if (!hstate_is_gigantic(h))
2229 hugetlb_hstate_alloc_pages(h);
2231 VM_BUG_ON(minimum_order == UINT_MAX);
2234 static void __init report_hugepages(void)
2238 for_each_hstate(h) {
2241 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2242 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2243 buf, h->free_huge_pages);
2247 #ifdef CONFIG_HIGHMEM
2248 static void try_to_free_low(struct hstate *h, unsigned long count,
2249 nodemask_t *nodes_allowed)
2253 if (hstate_is_gigantic(h))
2256 for_each_node_mask(i, *nodes_allowed) {
2257 struct page *page, *next;
2258 struct list_head *freel = &h->hugepage_freelists[i];
2259 list_for_each_entry_safe(page, next, freel, lru) {
2260 if (count >= h->nr_huge_pages)
2262 if (PageHighMem(page))
2264 list_del(&page->lru);
2265 update_and_free_page(h, page);
2266 h->free_huge_pages--;
2267 h->free_huge_pages_node[page_to_nid(page)]--;
2272 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2273 nodemask_t *nodes_allowed)
2279 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2280 * balanced by operating on them in a round-robin fashion.
2281 * Returns 1 if an adjustment was made.
2283 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2288 VM_BUG_ON(delta != -1 && delta != 1);
2291 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2292 if (h->surplus_huge_pages_node[node])
2296 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2297 if (h->surplus_huge_pages_node[node] <
2298 h->nr_huge_pages_node[node])
2305 h->surplus_huge_pages += delta;
2306 h->surplus_huge_pages_node[node] += delta;
2310 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2311 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2312 nodemask_t *nodes_allowed)
2314 unsigned long min_count, ret;
2315 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2318 * Bit mask controlling how hard we retry per-node allocations.
2319 * If we can not allocate the bit mask, do not attempt to allocate
2320 * the requested huge pages.
2322 if (node_alloc_noretry)
2323 nodes_clear(*node_alloc_noretry);
2327 spin_lock(&hugetlb_lock);
2330 * Check for a node specific request.
2331 * Changing node specific huge page count may require a corresponding
2332 * change to the global count. In any case, the passed node mask
2333 * (nodes_allowed) will restrict alloc/free to the specified node.
2335 if (nid != NUMA_NO_NODE) {
2336 unsigned long old_count = count;
2338 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2340 * User may have specified a large count value which caused the
2341 * above calculation to overflow. In this case, they wanted
2342 * to allocate as many huge pages as possible. Set count to
2343 * largest possible value to align with their intention.
2345 if (count < old_count)
2350 * Gigantic pages runtime allocation depend on the capability for large
2351 * page range allocation.
2352 * If the system does not provide this feature, return an error when
2353 * the user tries to allocate gigantic pages but let the user free the
2354 * boottime allocated gigantic pages.
2356 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2357 if (count > persistent_huge_pages(h)) {
2358 spin_unlock(&hugetlb_lock);
2359 NODEMASK_FREE(node_alloc_noretry);
2362 /* Fall through to decrease pool */
2366 * Increase the pool size
2367 * First take pages out of surplus state. Then make up the
2368 * remaining difference by allocating fresh huge pages.
2370 * We might race with alloc_surplus_huge_page() here and be unable
2371 * to convert a surplus huge page to a normal huge page. That is
2372 * not critical, though, it just means the overall size of the
2373 * pool might be one hugepage larger than it needs to be, but
2374 * within all the constraints specified by the sysctls.
2376 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2377 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2381 while (count > persistent_huge_pages(h)) {
2383 * If this allocation races such that we no longer need the
2384 * page, free_huge_page will handle it by freeing the page
2385 * and reducing the surplus.
2387 spin_unlock(&hugetlb_lock);
2389 /* yield cpu to avoid soft lockup */
2392 ret = alloc_pool_huge_page(h, nodes_allowed,
2393 node_alloc_noretry);
2394 spin_lock(&hugetlb_lock);
2398 /* Bail for signals. Probably ctrl-c from user */
2399 if (signal_pending(current))
2404 * Decrease the pool size
2405 * First return free pages to the buddy allocator (being careful
2406 * to keep enough around to satisfy reservations). Then place
2407 * pages into surplus state as needed so the pool will shrink
2408 * to the desired size as pages become free.
2410 * By placing pages into the surplus state independent of the
2411 * overcommit value, we are allowing the surplus pool size to
2412 * exceed overcommit. There are few sane options here. Since
2413 * alloc_surplus_huge_page() is checking the global counter,
2414 * though, we'll note that we're not allowed to exceed surplus
2415 * and won't grow the pool anywhere else. Not until one of the
2416 * sysctls are changed, or the surplus pages go out of use.
2418 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2419 min_count = max(count, min_count);
2420 try_to_free_low(h, min_count, nodes_allowed);
2421 while (min_count < persistent_huge_pages(h)) {
2422 if (!free_pool_huge_page(h, nodes_allowed, 0))
2424 cond_resched_lock(&hugetlb_lock);
2426 while (count < persistent_huge_pages(h)) {
2427 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2431 h->max_huge_pages = persistent_huge_pages(h);
2432 spin_unlock(&hugetlb_lock);
2434 NODEMASK_FREE(node_alloc_noretry);
2439 #define HSTATE_ATTR_RO(_name) \
2440 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2442 #define HSTATE_ATTR(_name) \
2443 static struct kobj_attribute _name##_attr = \
2444 __ATTR(_name, 0644, _name##_show, _name##_store)
2446 static struct kobject *hugepages_kobj;
2447 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2449 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2451 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2455 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2456 if (hstate_kobjs[i] == kobj) {
2458 *nidp = NUMA_NO_NODE;
2462 return kobj_to_node_hstate(kobj, nidp);
2465 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2466 struct kobj_attribute *attr, char *buf)
2469 unsigned long nr_huge_pages;
2472 h = kobj_to_hstate(kobj, &nid);
2473 if (nid == NUMA_NO_NODE)
2474 nr_huge_pages = h->nr_huge_pages;
2476 nr_huge_pages = h->nr_huge_pages_node[nid];
2478 return sprintf(buf, "%lu\n", nr_huge_pages);
2481 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2482 struct hstate *h, int nid,
2483 unsigned long count, size_t len)
2486 nodemask_t nodes_allowed, *n_mask;
2488 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2491 if (nid == NUMA_NO_NODE) {
2493 * global hstate attribute
2495 if (!(obey_mempolicy &&
2496 init_nodemask_of_mempolicy(&nodes_allowed)))
2497 n_mask = &node_states[N_MEMORY];
2499 n_mask = &nodes_allowed;
2502 * Node specific request. count adjustment happens in
2503 * set_max_huge_pages() after acquiring hugetlb_lock.
2505 init_nodemask_of_node(&nodes_allowed, nid);
2506 n_mask = &nodes_allowed;
2509 err = set_max_huge_pages(h, count, nid, n_mask);
2511 return err ? err : len;
2514 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2515 struct kobject *kobj, const char *buf,
2519 unsigned long count;
2523 err = kstrtoul(buf, 10, &count);
2527 h = kobj_to_hstate(kobj, &nid);
2528 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2531 static ssize_t nr_hugepages_show(struct kobject *kobj,
2532 struct kobj_attribute *attr, char *buf)
2534 return nr_hugepages_show_common(kobj, attr, buf);
2537 static ssize_t nr_hugepages_store(struct kobject *kobj,
2538 struct kobj_attribute *attr, const char *buf, size_t len)
2540 return nr_hugepages_store_common(false, kobj, buf, len);
2542 HSTATE_ATTR(nr_hugepages);
2547 * hstate attribute for optionally mempolicy-based constraint on persistent
2548 * huge page alloc/free.
2550 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2551 struct kobj_attribute *attr, char *buf)
2553 return nr_hugepages_show_common(kobj, attr, buf);
2556 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2557 struct kobj_attribute *attr, const char *buf, size_t len)
2559 return nr_hugepages_store_common(true, kobj, buf, len);
2561 HSTATE_ATTR(nr_hugepages_mempolicy);
2565 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2566 struct kobj_attribute *attr, char *buf)
2568 struct hstate *h = kobj_to_hstate(kobj, NULL);
2569 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2572 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2573 struct kobj_attribute *attr, const char *buf, size_t count)
2576 unsigned long input;
2577 struct hstate *h = kobj_to_hstate(kobj, NULL);
2579 if (hstate_is_gigantic(h))
2582 err = kstrtoul(buf, 10, &input);
2586 spin_lock(&hugetlb_lock);
2587 h->nr_overcommit_huge_pages = input;
2588 spin_unlock(&hugetlb_lock);
2592 HSTATE_ATTR(nr_overcommit_hugepages);
2594 static ssize_t free_hugepages_show(struct kobject *kobj,
2595 struct kobj_attribute *attr, char *buf)
2598 unsigned long free_huge_pages;
2601 h = kobj_to_hstate(kobj, &nid);
2602 if (nid == NUMA_NO_NODE)
2603 free_huge_pages = h->free_huge_pages;
2605 free_huge_pages = h->free_huge_pages_node[nid];
2607 return sprintf(buf, "%lu\n", free_huge_pages);
2609 HSTATE_ATTR_RO(free_hugepages);
2611 static ssize_t resv_hugepages_show(struct kobject *kobj,
2612 struct kobj_attribute *attr, char *buf)
2614 struct hstate *h = kobj_to_hstate(kobj, NULL);
2615 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2617 HSTATE_ATTR_RO(resv_hugepages);
2619 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2620 struct kobj_attribute *attr, char *buf)
2623 unsigned long surplus_huge_pages;
2626 h = kobj_to_hstate(kobj, &nid);
2627 if (nid == NUMA_NO_NODE)
2628 surplus_huge_pages = h->surplus_huge_pages;
2630 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2632 return sprintf(buf, "%lu\n", surplus_huge_pages);
2634 HSTATE_ATTR_RO(surplus_hugepages);
2636 static struct attribute *hstate_attrs[] = {
2637 &nr_hugepages_attr.attr,
2638 &nr_overcommit_hugepages_attr.attr,
2639 &free_hugepages_attr.attr,
2640 &resv_hugepages_attr.attr,
2641 &surplus_hugepages_attr.attr,
2643 &nr_hugepages_mempolicy_attr.attr,
2648 static const struct attribute_group hstate_attr_group = {
2649 .attrs = hstate_attrs,
2652 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2653 struct kobject **hstate_kobjs,
2654 const struct attribute_group *hstate_attr_group)
2657 int hi = hstate_index(h);
2659 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2660 if (!hstate_kobjs[hi])
2663 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2665 kobject_put(hstate_kobjs[hi]);
2670 static void __init hugetlb_sysfs_init(void)
2675 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2676 if (!hugepages_kobj)
2679 for_each_hstate(h) {
2680 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2681 hstate_kobjs, &hstate_attr_group);
2683 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2690 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2691 * with node devices in node_devices[] using a parallel array. The array
2692 * index of a node device or _hstate == node id.
2693 * This is here to avoid any static dependency of the node device driver, in
2694 * the base kernel, on the hugetlb module.
2696 struct node_hstate {
2697 struct kobject *hugepages_kobj;
2698 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2700 static struct node_hstate node_hstates[MAX_NUMNODES];
2703 * A subset of global hstate attributes for node devices
2705 static struct attribute *per_node_hstate_attrs[] = {
2706 &nr_hugepages_attr.attr,
2707 &free_hugepages_attr.attr,
2708 &surplus_hugepages_attr.attr,
2712 static const struct attribute_group per_node_hstate_attr_group = {
2713 .attrs = per_node_hstate_attrs,
2717 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2718 * Returns node id via non-NULL nidp.
2720 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2724 for (nid = 0; nid < nr_node_ids; nid++) {
2725 struct node_hstate *nhs = &node_hstates[nid];
2727 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2728 if (nhs->hstate_kobjs[i] == kobj) {
2740 * Unregister hstate attributes from a single node device.
2741 * No-op if no hstate attributes attached.
2743 static void hugetlb_unregister_node(struct node *node)
2746 struct node_hstate *nhs = &node_hstates[node->dev.id];
2748 if (!nhs->hugepages_kobj)
2749 return; /* no hstate attributes */
2751 for_each_hstate(h) {
2752 int idx = hstate_index(h);
2753 if (nhs->hstate_kobjs[idx]) {
2754 kobject_put(nhs->hstate_kobjs[idx]);
2755 nhs->hstate_kobjs[idx] = NULL;
2759 kobject_put(nhs->hugepages_kobj);
2760 nhs->hugepages_kobj = NULL;
2765 * Register hstate attributes for a single node device.
2766 * No-op if attributes already registered.
2768 static void hugetlb_register_node(struct node *node)
2771 struct node_hstate *nhs = &node_hstates[node->dev.id];
2774 if (nhs->hugepages_kobj)
2775 return; /* already allocated */
2777 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2779 if (!nhs->hugepages_kobj)
2782 for_each_hstate(h) {
2783 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2785 &per_node_hstate_attr_group);
2787 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2788 h->name, node->dev.id);
2789 hugetlb_unregister_node(node);
2796 * hugetlb init time: register hstate attributes for all registered node
2797 * devices of nodes that have memory. All on-line nodes should have
2798 * registered their associated device by this time.
2800 static void __init hugetlb_register_all_nodes(void)
2804 for_each_node_state(nid, N_MEMORY) {
2805 struct node *node = node_devices[nid];
2806 if (node->dev.id == nid)
2807 hugetlb_register_node(node);
2811 * Let the node device driver know we're here so it can
2812 * [un]register hstate attributes on node hotplug.
2814 register_hugetlbfs_with_node(hugetlb_register_node,
2815 hugetlb_unregister_node);
2817 #else /* !CONFIG_NUMA */
2819 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2827 static void hugetlb_register_all_nodes(void) { }
2831 static int __init hugetlb_init(void)
2835 if (!hugepages_supported())
2838 if (!size_to_hstate(default_hstate_size)) {
2839 if (default_hstate_size != 0) {
2840 pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2841 default_hstate_size, HPAGE_SIZE);
2844 default_hstate_size = HPAGE_SIZE;
2845 if (!size_to_hstate(default_hstate_size))
2846 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2848 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2849 if (default_hstate_max_huge_pages) {
2850 if (!default_hstate.max_huge_pages)
2851 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2854 hugetlb_init_hstates();
2855 gather_bootmem_prealloc();
2858 hugetlb_sysfs_init();
2859 hugetlb_register_all_nodes();
2860 hugetlb_cgroup_file_init();
2863 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2865 num_fault_mutexes = 1;
2867 hugetlb_fault_mutex_table =
2868 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
2870 BUG_ON(!hugetlb_fault_mutex_table);
2872 for (i = 0; i < num_fault_mutexes; i++)
2873 mutex_init(&hugetlb_fault_mutex_table[i]);
2876 subsys_initcall(hugetlb_init);
2878 /* Should be called on processing a hugepagesz=... option */
2879 void __init hugetlb_bad_size(void)
2881 parsed_valid_hugepagesz = false;
2884 void __init hugetlb_add_hstate(unsigned int order)
2889 if (size_to_hstate(PAGE_SIZE << order)) {
2890 pr_warn("hugepagesz= specified twice, ignoring\n");
2893 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2895 h = &hstates[hugetlb_max_hstate++];
2897 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2898 h->nr_huge_pages = 0;
2899 h->free_huge_pages = 0;
2900 for (i = 0; i < MAX_NUMNODES; ++i)
2901 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2902 INIT_LIST_HEAD(&h->hugepage_activelist);
2903 h->next_nid_to_alloc = first_memory_node;
2904 h->next_nid_to_free = first_memory_node;
2905 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2906 huge_page_size(h)/1024);
2911 static int __init hugetlb_nrpages_setup(char *s)
2914 static unsigned long *last_mhp;
2916 if (!parsed_valid_hugepagesz) {
2917 pr_warn("hugepages = %s preceded by "
2918 "an unsupported hugepagesz, ignoring\n", s);
2919 parsed_valid_hugepagesz = true;
2923 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2924 * so this hugepages= parameter goes to the "default hstate".
2926 else if (!hugetlb_max_hstate)
2927 mhp = &default_hstate_max_huge_pages;
2929 mhp = &parsed_hstate->max_huge_pages;
2931 if (mhp == last_mhp) {
2932 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2936 if (sscanf(s, "%lu", mhp) <= 0)
2940 * Global state is always initialized later in hugetlb_init.
2941 * But we need to allocate >= MAX_ORDER hstates here early to still
2942 * use the bootmem allocator.
2944 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2945 hugetlb_hstate_alloc_pages(parsed_hstate);
2951 __setup("hugepages=", hugetlb_nrpages_setup);
2953 static int __init hugetlb_default_setup(char *s)
2955 default_hstate_size = memparse(s, &s);
2958 __setup("default_hugepagesz=", hugetlb_default_setup);
2960 static unsigned int cpuset_mems_nr(unsigned int *array)
2963 unsigned int nr = 0;
2965 for_each_node_mask(node, cpuset_current_mems_allowed)
2971 #ifdef CONFIG_SYSCTL
2972 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2973 struct ctl_table *table, int write,
2974 void __user *buffer, size_t *length, loff_t *ppos)
2976 struct hstate *h = &default_hstate;
2977 unsigned long tmp = h->max_huge_pages;
2980 if (!hugepages_supported())
2984 table->maxlen = sizeof(unsigned long);
2985 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2990 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2991 NUMA_NO_NODE, tmp, *length);
2996 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2997 void __user *buffer, size_t *length, loff_t *ppos)
3000 return hugetlb_sysctl_handler_common(false, table, write,
3001 buffer, length, ppos);
3005 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3006 void __user *buffer, size_t *length, loff_t *ppos)
3008 return hugetlb_sysctl_handler_common(true, table, write,
3009 buffer, length, ppos);
3011 #endif /* CONFIG_NUMA */
3013 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3014 void __user *buffer,
3015 size_t *length, loff_t *ppos)
3017 struct hstate *h = &default_hstate;
3021 if (!hugepages_supported())
3024 tmp = h->nr_overcommit_huge_pages;
3026 if (write && hstate_is_gigantic(h))
3030 table->maxlen = sizeof(unsigned long);
3031 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3036 spin_lock(&hugetlb_lock);
3037 h->nr_overcommit_huge_pages = tmp;
3038 spin_unlock(&hugetlb_lock);
3044 #endif /* CONFIG_SYSCTL */
3046 void hugetlb_report_meminfo(struct seq_file *m)
3049 unsigned long total = 0;
3051 if (!hugepages_supported())
3054 for_each_hstate(h) {
3055 unsigned long count = h->nr_huge_pages;
3057 total += (PAGE_SIZE << huge_page_order(h)) * count;
3059 if (h == &default_hstate)
3061 "HugePages_Total: %5lu\n"
3062 "HugePages_Free: %5lu\n"
3063 "HugePages_Rsvd: %5lu\n"
3064 "HugePages_Surp: %5lu\n"
3065 "Hugepagesize: %8lu kB\n",
3069 h->surplus_huge_pages,
3070 (PAGE_SIZE << huge_page_order(h)) / 1024);
3073 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3076 int hugetlb_report_node_meminfo(int nid, char *buf)
3078 struct hstate *h = &default_hstate;
3079 if (!hugepages_supported())
3082 "Node %d HugePages_Total: %5u\n"
3083 "Node %d HugePages_Free: %5u\n"
3084 "Node %d HugePages_Surp: %5u\n",
3085 nid, h->nr_huge_pages_node[nid],
3086 nid, h->free_huge_pages_node[nid],
3087 nid, h->surplus_huge_pages_node[nid]);
3090 void hugetlb_show_meminfo(void)
3095 if (!hugepages_supported())
3098 for_each_node_state(nid, N_MEMORY)
3100 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3102 h->nr_huge_pages_node[nid],
3103 h->free_huge_pages_node[nid],
3104 h->surplus_huge_pages_node[nid],
3105 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3108 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3110 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3111 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3114 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3115 unsigned long hugetlb_total_pages(void)
3118 unsigned long nr_total_pages = 0;
3121 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3122 return nr_total_pages;
3125 static int hugetlb_acct_memory(struct hstate *h, long delta)
3129 spin_lock(&hugetlb_lock);
3131 * When cpuset is configured, it breaks the strict hugetlb page
3132 * reservation as the accounting is done on a global variable. Such
3133 * reservation is completely rubbish in the presence of cpuset because
3134 * the reservation is not checked against page availability for the
3135 * current cpuset. Application can still potentially OOM'ed by kernel
3136 * with lack of free htlb page in cpuset that the task is in.
3137 * Attempt to enforce strict accounting with cpuset is almost
3138 * impossible (or too ugly) because cpuset is too fluid that
3139 * task or memory node can be dynamically moved between cpusets.
3141 * The change of semantics for shared hugetlb mapping with cpuset is
3142 * undesirable. However, in order to preserve some of the semantics,
3143 * we fall back to check against current free page availability as
3144 * a best attempt and hopefully to minimize the impact of changing
3145 * semantics that cpuset has.
3148 if (gather_surplus_pages(h, delta) < 0)
3151 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3152 return_unused_surplus_pages(h, delta);
3159 return_unused_surplus_pages(h, (unsigned long) -delta);
3162 spin_unlock(&hugetlb_lock);
3166 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3168 struct resv_map *resv = vma_resv_map(vma);
3171 * This new VMA should share its siblings reservation map if present.
3172 * The VMA will only ever have a valid reservation map pointer where
3173 * it is being copied for another still existing VMA. As that VMA
3174 * has a reference to the reservation map it cannot disappear until
3175 * after this open call completes. It is therefore safe to take a
3176 * new reference here without additional locking.
3178 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3179 kref_get(&resv->refs);
3182 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3184 struct hstate *h = hstate_vma(vma);
3185 struct resv_map *resv = vma_resv_map(vma);
3186 struct hugepage_subpool *spool = subpool_vma(vma);
3187 unsigned long reserve, start, end;
3190 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3193 start = vma_hugecache_offset(h, vma, vma->vm_start);
3194 end = vma_hugecache_offset(h, vma, vma->vm_end);
3196 reserve = (end - start) - region_count(resv, start, end);
3198 kref_put(&resv->refs, resv_map_release);
3202 * Decrement reserve counts. The global reserve count may be
3203 * adjusted if the subpool has a minimum size.
3205 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3206 hugetlb_acct_memory(h, -gbl_reserve);
3210 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3212 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3217 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3219 struct hstate *hstate = hstate_vma(vma);
3221 return 1UL << huge_page_shift(hstate);
3225 * We cannot handle pagefaults against hugetlb pages at all. They cause
3226 * handle_mm_fault() to try to instantiate regular-sized pages in the
3227 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3230 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3237 * When a new function is introduced to vm_operations_struct and added
3238 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3239 * This is because under System V memory model, mappings created via
3240 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3241 * their original vm_ops are overwritten with shm_vm_ops.
3243 const struct vm_operations_struct hugetlb_vm_ops = {
3244 .fault = hugetlb_vm_op_fault,
3245 .open = hugetlb_vm_op_open,
3246 .close = hugetlb_vm_op_close,
3247 .split = hugetlb_vm_op_split,
3248 .pagesize = hugetlb_vm_op_pagesize,
3251 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3257 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3258 vma->vm_page_prot)));
3260 entry = huge_pte_wrprotect(mk_huge_pte(page,
3261 vma->vm_page_prot));
3263 entry = pte_mkyoung(entry);
3264 entry = pte_mkhuge(entry);
3265 entry = arch_make_huge_pte(entry, vma, page, writable);
3270 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3271 unsigned long address, pte_t *ptep)
3275 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3276 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3277 update_mmu_cache(vma, address, ptep);
3280 bool is_hugetlb_entry_migration(pte_t pte)
3284 if (huge_pte_none(pte) || pte_present(pte))
3286 swp = pte_to_swp_entry(pte);
3287 if (non_swap_entry(swp) && is_migration_entry(swp))
3293 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3297 if (huge_pte_none(pte) || pte_present(pte))
3299 swp = pte_to_swp_entry(pte);
3300 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3306 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3307 struct vm_area_struct *vma)
3309 pte_t *src_pte, *dst_pte, entry, dst_entry;
3310 struct page *ptepage;
3313 struct hstate *h = hstate_vma(vma);
3314 unsigned long sz = huge_page_size(h);
3315 struct mmu_notifier_range range;
3318 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3321 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3324 mmu_notifier_invalidate_range_start(&range);
3327 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3328 spinlock_t *src_ptl, *dst_ptl;
3329 src_pte = huge_pte_offset(src, addr, sz);
3332 dst_pte = huge_pte_alloc(dst, addr, sz);
3339 * If the pagetables are shared don't copy or take references.
3340 * dst_pte == src_pte is the common case of src/dest sharing.
3342 * However, src could have 'unshared' and dst shares with
3343 * another vma. If dst_pte !none, this implies sharing.
3344 * Check here before taking page table lock, and once again
3345 * after taking the lock below.
3347 dst_entry = huge_ptep_get(dst_pte);
3348 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3351 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3352 src_ptl = huge_pte_lockptr(h, src, src_pte);
3353 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3354 entry = huge_ptep_get(src_pte);
3355 dst_entry = huge_ptep_get(dst_pte);
3356 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3358 * Skip if src entry none. Also, skip in the
3359 * unlikely case dst entry !none as this implies
3360 * sharing with another vma.
3363 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3364 is_hugetlb_entry_hwpoisoned(entry))) {
3365 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3367 if (is_write_migration_entry(swp_entry) && cow) {
3369 * COW mappings require pages in both
3370 * parent and child to be set to read.
3372 make_migration_entry_read(&swp_entry);
3373 entry = swp_entry_to_pte(swp_entry);
3374 set_huge_swap_pte_at(src, addr, src_pte,
3377 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3381 * No need to notify as we are downgrading page
3382 * table protection not changing it to point
3385 * See Documentation/vm/mmu_notifier.rst
3387 huge_ptep_set_wrprotect(src, addr, src_pte);
3389 entry = huge_ptep_get(src_pte);
3390 ptepage = pte_page(entry);
3392 page_dup_rmap(ptepage, true);
3393 set_huge_pte_at(dst, addr, dst_pte, entry);
3394 hugetlb_count_add(pages_per_huge_page(h), dst);
3396 spin_unlock(src_ptl);
3397 spin_unlock(dst_ptl);
3401 mmu_notifier_invalidate_range_end(&range);
3406 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3407 unsigned long start, unsigned long end,
3408 struct page *ref_page)
3410 struct mm_struct *mm = vma->vm_mm;
3411 unsigned long address;
3416 struct hstate *h = hstate_vma(vma);
3417 unsigned long sz = huge_page_size(h);
3418 struct mmu_notifier_range range;
3420 WARN_ON(!is_vm_hugetlb_page(vma));
3421 BUG_ON(start & ~huge_page_mask(h));
3422 BUG_ON(end & ~huge_page_mask(h));
3425 * This is a hugetlb vma, all the pte entries should point
3428 tlb_change_page_size(tlb, sz);
3429 tlb_start_vma(tlb, vma);
3432 * If sharing possible, alert mmu notifiers of worst case.
3434 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3436 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3437 mmu_notifier_invalidate_range_start(&range);
3439 for (; address < end; address += sz) {
3440 ptep = huge_pte_offset(mm, address, sz);
3444 ptl = huge_pte_lock(h, mm, ptep);
3445 if (huge_pmd_unshare(mm, &address, ptep)) {
3448 * We just unmapped a page of PMDs by clearing a PUD.
3449 * The caller's TLB flush range should cover this area.
3454 pte = huge_ptep_get(ptep);
3455 if (huge_pte_none(pte)) {
3461 * Migrating hugepage or HWPoisoned hugepage is already
3462 * unmapped and its refcount is dropped, so just clear pte here.
3464 if (unlikely(!pte_present(pte))) {
3465 huge_pte_clear(mm, address, ptep, sz);
3470 page = pte_page(pte);
3472 * If a reference page is supplied, it is because a specific
3473 * page is being unmapped, not a range. Ensure the page we
3474 * are about to unmap is the actual page of interest.
3477 if (page != ref_page) {
3482 * Mark the VMA as having unmapped its page so that
3483 * future faults in this VMA will fail rather than
3484 * looking like data was lost
3486 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3489 pte = huge_ptep_get_and_clear(mm, address, ptep);
3490 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3491 if (huge_pte_dirty(pte))
3492 set_page_dirty(page);
3494 hugetlb_count_sub(pages_per_huge_page(h), mm);
3495 page_remove_rmap(page, true);
3498 tlb_remove_page_size(tlb, page, huge_page_size(h));
3500 * Bail out after unmapping reference page if supplied
3505 mmu_notifier_invalidate_range_end(&range);
3506 tlb_end_vma(tlb, vma);
3509 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3510 struct vm_area_struct *vma, unsigned long start,
3511 unsigned long end, struct page *ref_page)
3513 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3516 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3517 * test will fail on a vma being torn down, and not grab a page table
3518 * on its way out. We're lucky that the flag has such an appropriate
3519 * name, and can in fact be safely cleared here. We could clear it
3520 * before the __unmap_hugepage_range above, but all that's necessary
3521 * is to clear it before releasing the i_mmap_rwsem. This works
3522 * because in the context this is called, the VMA is about to be
3523 * destroyed and the i_mmap_rwsem is held.
3525 vma->vm_flags &= ~VM_MAYSHARE;
3528 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3529 unsigned long end, struct page *ref_page)
3531 struct mm_struct *mm;
3532 struct mmu_gather tlb;
3533 unsigned long tlb_start = start;
3534 unsigned long tlb_end = end;
3537 * If shared PMDs were possibly used within this vma range, adjust
3538 * start/end for worst case tlb flushing.
3539 * Note that we can not be sure if PMDs are shared until we try to
3540 * unmap pages. However, we want to make sure TLB flushing covers
3541 * the largest possible range.
3543 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
3547 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
3548 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3549 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
3553 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3554 * mappping it owns the reserve page for. The intention is to unmap the page
3555 * from other VMAs and let the children be SIGKILLed if they are faulting the
3558 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3559 struct page *page, unsigned long address)
3561 struct hstate *h = hstate_vma(vma);
3562 struct vm_area_struct *iter_vma;
3563 struct address_space *mapping;
3567 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3568 * from page cache lookup which is in HPAGE_SIZE units.
3570 address = address & huge_page_mask(h);
3571 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3573 mapping = vma->vm_file->f_mapping;
3576 * Take the mapping lock for the duration of the table walk. As
3577 * this mapping should be shared between all the VMAs,
3578 * __unmap_hugepage_range() is called as the lock is already held
3580 i_mmap_lock_write(mapping);
3581 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3582 /* Do not unmap the current VMA */
3583 if (iter_vma == vma)
3587 * Shared VMAs have their own reserves and do not affect
3588 * MAP_PRIVATE accounting but it is possible that a shared
3589 * VMA is using the same page so check and skip such VMAs.
3591 if (iter_vma->vm_flags & VM_MAYSHARE)
3595 * Unmap the page from other VMAs without their own reserves.
3596 * They get marked to be SIGKILLed if they fault in these
3597 * areas. This is because a future no-page fault on this VMA
3598 * could insert a zeroed page instead of the data existing
3599 * from the time of fork. This would look like data corruption
3601 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3602 unmap_hugepage_range(iter_vma, address,
3603 address + huge_page_size(h), page);
3605 i_mmap_unlock_write(mapping);
3609 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3610 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3611 * cannot race with other handlers or page migration.
3612 * Keep the pte_same checks anyway to make transition from the mutex easier.
3614 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3615 unsigned long address, pte_t *ptep,
3616 struct page *pagecache_page, spinlock_t *ptl)
3619 struct hstate *h = hstate_vma(vma);
3620 struct page *old_page, *new_page;
3621 int outside_reserve = 0;
3623 unsigned long haddr = address & huge_page_mask(h);
3624 struct mmu_notifier_range range;
3626 pte = huge_ptep_get(ptep);
3627 old_page = pte_page(pte);
3630 /* If no-one else is actually using this page, avoid the copy
3631 * and just make the page writable */
3632 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3633 page_move_anon_rmap(old_page, vma);
3634 set_huge_ptep_writable(vma, haddr, ptep);
3639 * If the process that created a MAP_PRIVATE mapping is about to
3640 * perform a COW due to a shared page count, attempt to satisfy
3641 * the allocation without using the existing reserves. The pagecache
3642 * page is used to determine if the reserve at this address was
3643 * consumed or not. If reserves were used, a partial faulted mapping
3644 * at the time of fork() could consume its reserves on COW instead
3645 * of the full address range.
3647 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3648 old_page != pagecache_page)
3649 outside_reserve = 1;
3654 * Drop page table lock as buddy allocator may be called. It will
3655 * be acquired again before returning to the caller, as expected.
3658 new_page = alloc_huge_page(vma, haddr, outside_reserve);
3660 if (IS_ERR(new_page)) {
3662 * If a process owning a MAP_PRIVATE mapping fails to COW,
3663 * it is due to references held by a child and an insufficient
3664 * huge page pool. To guarantee the original mappers
3665 * reliability, unmap the page from child processes. The child
3666 * may get SIGKILLed if it later faults.
3668 if (outside_reserve) {
3670 BUG_ON(huge_pte_none(pte));
3671 unmap_ref_private(mm, vma, old_page, haddr);
3672 BUG_ON(huge_pte_none(pte));
3674 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3676 pte_same(huge_ptep_get(ptep), pte)))
3677 goto retry_avoidcopy;
3679 * race occurs while re-acquiring page table
3680 * lock, and our job is done.
3685 ret = vmf_error(PTR_ERR(new_page));
3686 goto out_release_old;
3690 * When the original hugepage is shared one, it does not have
3691 * anon_vma prepared.
3693 if (unlikely(anon_vma_prepare(vma))) {
3695 goto out_release_all;
3698 copy_user_huge_page(new_page, old_page, address, vma,
3699 pages_per_huge_page(h));
3700 __SetPageUptodate(new_page);
3702 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
3703 haddr + huge_page_size(h));
3704 mmu_notifier_invalidate_range_start(&range);
3707 * Retake the page table lock to check for racing updates
3708 * before the page tables are altered
3711 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3712 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3713 ClearPagePrivate(new_page);
3716 huge_ptep_clear_flush(vma, haddr, ptep);
3717 mmu_notifier_invalidate_range(mm, range.start, range.end);
3718 set_huge_pte_at(mm, haddr, ptep,
3719 make_huge_pte(vma, new_page, 1));
3720 page_remove_rmap(old_page, true);
3721 hugepage_add_new_anon_rmap(new_page, vma, haddr);
3722 set_page_huge_active(new_page);
3723 /* Make the old page be freed below */
3724 new_page = old_page;
3727 mmu_notifier_invalidate_range_end(&range);
3729 restore_reserve_on_error(h, vma, haddr, new_page);
3734 spin_lock(ptl); /* Caller expects lock to be held */
3738 /* Return the pagecache page at a given address within a VMA */
3739 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3740 struct vm_area_struct *vma, unsigned long address)
3742 struct address_space *mapping;
3745 mapping = vma->vm_file->f_mapping;
3746 idx = vma_hugecache_offset(h, vma, address);
3748 return find_lock_page(mapping, idx);
3752 * Return whether there is a pagecache page to back given address within VMA.
3753 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3755 static bool hugetlbfs_pagecache_present(struct hstate *h,
3756 struct vm_area_struct *vma, unsigned long address)
3758 struct address_space *mapping;
3762 mapping = vma->vm_file->f_mapping;
3763 idx = vma_hugecache_offset(h, vma, address);
3765 page = find_get_page(mapping, idx);
3768 return page != NULL;
3771 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3774 struct inode *inode = mapping->host;
3775 struct hstate *h = hstate_inode(inode);
3776 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3780 ClearPagePrivate(page);
3783 * set page dirty so that it will not be removed from cache/file
3784 * by non-hugetlbfs specific code paths.
3786 set_page_dirty(page);
3788 spin_lock(&inode->i_lock);
3789 inode->i_blocks += blocks_per_huge_page(h);
3790 spin_unlock(&inode->i_lock);
3794 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
3795 struct vm_area_struct *vma,
3796 struct address_space *mapping, pgoff_t idx,
3797 unsigned long address, pte_t *ptep, unsigned int flags)
3799 struct hstate *h = hstate_vma(vma);
3800 vm_fault_t ret = VM_FAULT_SIGBUS;
3806 unsigned long haddr = address & huge_page_mask(h);
3807 bool new_page = false;
3810 * Currently, we are forced to kill the process in the event the
3811 * original mapper has unmapped pages from the child due to a failed
3812 * COW. Warn that such a situation has occurred as it may not be obvious
3814 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3815 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3821 * Use page lock to guard against racing truncation
3822 * before we get page_table_lock.
3825 page = find_lock_page(mapping, idx);
3827 size = i_size_read(mapping->host) >> huge_page_shift(h);
3832 * Check for page in userfault range
3834 if (userfaultfd_missing(vma)) {
3836 struct vm_fault vmf = {
3841 * Hard to debug if it ends up being
3842 * used by a callee that assumes
3843 * something about the other
3844 * uninitialized fields... same as in
3850 * hugetlb_fault_mutex must be dropped before
3851 * handling userfault. Reacquire after handling
3852 * fault to make calling code simpler.
3854 hash = hugetlb_fault_mutex_hash(mapping, idx);
3855 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3856 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3857 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3861 page = alloc_huge_page(vma, haddr, 0);
3864 * Returning error will result in faulting task being
3865 * sent SIGBUS. The hugetlb fault mutex prevents two
3866 * tasks from racing to fault in the same page which
3867 * could result in false unable to allocate errors.
3868 * Page migration does not take the fault mutex, but
3869 * does a clear then write of pte's under page table
3870 * lock. Page fault code could race with migration,
3871 * notice the clear pte and try to allocate a page
3872 * here. Before returning error, get ptl and make
3873 * sure there really is no pte entry.
3875 ptl = huge_pte_lock(h, mm, ptep);
3876 if (!huge_pte_none(huge_ptep_get(ptep))) {
3882 ret = vmf_error(PTR_ERR(page));
3885 clear_huge_page(page, address, pages_per_huge_page(h));
3886 __SetPageUptodate(page);
3889 if (vma->vm_flags & VM_MAYSHARE) {
3890 int err = huge_add_to_page_cache(page, mapping, idx);
3899 if (unlikely(anon_vma_prepare(vma))) {
3901 goto backout_unlocked;
3907 * If memory error occurs between mmap() and fault, some process
3908 * don't have hwpoisoned swap entry for errored virtual address.
3909 * So we need to block hugepage fault by PG_hwpoison bit check.
3911 if (unlikely(PageHWPoison(page))) {
3912 ret = VM_FAULT_HWPOISON |
3913 VM_FAULT_SET_HINDEX(hstate_index(h));
3914 goto backout_unlocked;
3919 * If we are going to COW a private mapping later, we examine the
3920 * pending reservations for this page now. This will ensure that
3921 * any allocations necessary to record that reservation occur outside
3924 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3925 if (vma_needs_reservation(h, vma, haddr) < 0) {
3927 goto backout_unlocked;
3929 /* Just decrements count, does not deallocate */
3930 vma_end_reservation(h, vma, haddr);
3933 ptl = huge_pte_lock(h, mm, ptep);
3934 size = i_size_read(mapping->host) >> huge_page_shift(h);
3939 if (!huge_pte_none(huge_ptep_get(ptep)))
3943 ClearPagePrivate(page);
3944 hugepage_add_new_anon_rmap(page, vma, haddr);
3946 page_dup_rmap(page, true);
3947 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3948 && (vma->vm_flags & VM_SHARED)));
3949 set_huge_pte_at(mm, haddr, ptep, new_pte);
3951 hugetlb_count_add(pages_per_huge_page(h), mm);
3952 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3953 /* Optimization, do the COW without a second fault */
3954 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
3960 * Only make newly allocated pages active. Existing pages found
3961 * in the pagecache could be !page_huge_active() if they have been
3962 * isolated for migration.
3965 set_page_huge_active(page);
3975 restore_reserve_on_error(h, vma, haddr, page);
3981 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
3983 unsigned long key[2];
3986 key[0] = (unsigned long) mapping;
3989 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
3991 return hash & (num_fault_mutexes - 1);
3995 * For uniprocesor systems we always use a single mutex, so just
3996 * return 0 and avoid the hashing overhead.
3998 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4004 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4005 unsigned long address, unsigned int flags)
4012 struct page *page = NULL;
4013 struct page *pagecache_page = NULL;
4014 struct hstate *h = hstate_vma(vma);
4015 struct address_space *mapping;
4016 int need_wait_lock = 0;
4017 unsigned long haddr = address & huge_page_mask(h);
4019 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4021 entry = huge_ptep_get(ptep);
4022 if (unlikely(is_hugetlb_entry_migration(entry))) {
4023 migration_entry_wait_huge(vma, mm, ptep);
4025 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4026 return VM_FAULT_HWPOISON_LARGE |
4027 VM_FAULT_SET_HINDEX(hstate_index(h));
4029 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4031 return VM_FAULT_OOM;
4034 mapping = vma->vm_file->f_mapping;
4035 idx = vma_hugecache_offset(h, vma, haddr);
4038 * Serialize hugepage allocation and instantiation, so that we don't
4039 * get spurious allocation failures if two CPUs race to instantiate
4040 * the same page in the page cache.
4042 hash = hugetlb_fault_mutex_hash(mapping, idx);
4043 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4045 entry = huge_ptep_get(ptep);
4046 if (huge_pte_none(entry)) {
4047 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4054 * entry could be a migration/hwpoison entry at this point, so this
4055 * check prevents the kernel from going below assuming that we have
4056 * a active hugepage in pagecache. This goto expects the 2nd page fault,
4057 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
4060 if (!pte_present(entry))
4064 * If we are going to COW the mapping later, we examine the pending
4065 * reservations for this page now. This will ensure that any
4066 * allocations necessary to record that reservation occur outside the
4067 * spinlock. For private mappings, we also lookup the pagecache
4068 * page now as it is used to determine if a reservation has been
4071 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4072 if (vma_needs_reservation(h, vma, haddr) < 0) {
4076 /* Just decrements count, does not deallocate */
4077 vma_end_reservation(h, vma, haddr);
4079 if (!(vma->vm_flags & VM_MAYSHARE))
4080 pagecache_page = hugetlbfs_pagecache_page(h,
4084 ptl = huge_pte_lock(h, mm, ptep);
4086 /* Check for a racing update before calling hugetlb_cow */
4087 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4091 * hugetlb_cow() requires page locks of pte_page(entry) and
4092 * pagecache_page, so here we need take the former one
4093 * when page != pagecache_page or !pagecache_page.
4095 page = pte_page(entry);
4096 if (page != pagecache_page)
4097 if (!trylock_page(page)) {
4104 if (flags & FAULT_FLAG_WRITE) {
4105 if (!huge_pte_write(entry)) {
4106 ret = hugetlb_cow(mm, vma, address, ptep,
4107 pagecache_page, ptl);
4110 entry = huge_pte_mkdirty(entry);
4112 entry = pte_mkyoung(entry);
4113 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4114 flags & FAULT_FLAG_WRITE))
4115 update_mmu_cache(vma, haddr, ptep);
4117 if (page != pagecache_page)
4123 if (pagecache_page) {
4124 unlock_page(pagecache_page);
4125 put_page(pagecache_page);
4128 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4130 * Generally it's safe to hold refcount during waiting page lock. But
4131 * here we just wait to defer the next page fault to avoid busy loop and
4132 * the page is not used after unlocked before returning from the current
4133 * page fault. So we are safe from accessing freed page, even if we wait
4134 * here without taking refcount.
4137 wait_on_page_locked(page);
4142 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4143 * modifications for huge pages.
4145 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4147 struct vm_area_struct *dst_vma,
4148 unsigned long dst_addr,
4149 unsigned long src_addr,
4150 struct page **pagep)
4152 struct address_space *mapping;
4155 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4156 struct hstate *h = hstate_vma(dst_vma);
4164 page = alloc_huge_page(dst_vma, dst_addr, 0);
4168 ret = copy_huge_page_from_user(page,
4169 (const void __user *) src_addr,
4170 pages_per_huge_page(h), false);
4172 /* fallback to copy_from_user outside mmap_sem */
4173 if (unlikely(ret)) {
4176 /* don't free the page */
4185 * The memory barrier inside __SetPageUptodate makes sure that
4186 * preceding stores to the page contents become visible before
4187 * the set_pte_at() write.
4189 __SetPageUptodate(page);
4191 mapping = dst_vma->vm_file->f_mapping;
4192 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4195 * If shared, add to page cache
4198 size = i_size_read(mapping->host) >> huge_page_shift(h);
4201 goto out_release_nounlock;
4204 * Serialization between remove_inode_hugepages() and
4205 * huge_add_to_page_cache() below happens through the
4206 * hugetlb_fault_mutex_table that here must be hold by
4209 ret = huge_add_to_page_cache(page, mapping, idx);
4211 goto out_release_nounlock;
4214 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4218 * Recheck the i_size after holding PT lock to make sure not
4219 * to leave any page mapped (as page_mapped()) beyond the end
4220 * of the i_size (remove_inode_hugepages() is strict about
4221 * enforcing that). If we bail out here, we'll also leave a
4222 * page in the radix tree in the vm_shared case beyond the end
4223 * of the i_size, but remove_inode_hugepages() will take care
4224 * of it as soon as we drop the hugetlb_fault_mutex_table.
4226 size = i_size_read(mapping->host) >> huge_page_shift(h);
4229 goto out_release_unlock;
4232 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4233 goto out_release_unlock;
4236 page_dup_rmap(page, true);
4238 ClearPagePrivate(page);
4239 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4242 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4243 if (dst_vma->vm_flags & VM_WRITE)
4244 _dst_pte = huge_pte_mkdirty(_dst_pte);
4245 _dst_pte = pte_mkyoung(_dst_pte);
4247 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4249 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4250 dst_vma->vm_flags & VM_WRITE);
4251 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4253 /* No need to invalidate - it was non-present before */
4254 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4257 set_page_huge_active(page);
4267 out_release_nounlock:
4272 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4273 struct page **pages, struct vm_area_struct **vmas,
4274 unsigned long *position, unsigned long *nr_pages,
4275 long i, unsigned int flags, int *locked)
4277 unsigned long pfn_offset;
4278 unsigned long vaddr = *position;
4279 unsigned long remainder = *nr_pages;
4280 struct hstate *h = hstate_vma(vma);
4283 while (vaddr < vma->vm_end && remainder) {
4285 spinlock_t *ptl = NULL;
4290 * If we have a pending SIGKILL, don't keep faulting pages and
4291 * potentially allocating memory.
4293 if (fatal_signal_pending(current)) {
4299 * Some archs (sparc64, sh*) have multiple pte_ts to
4300 * each hugepage. We have to make sure we get the
4301 * first, for the page indexing below to work.
4303 * Note that page table lock is not held when pte is null.
4305 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4308 ptl = huge_pte_lock(h, mm, pte);
4309 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4312 * When coredumping, it suits get_dump_page if we just return
4313 * an error where there's an empty slot with no huge pagecache
4314 * to back it. This way, we avoid allocating a hugepage, and
4315 * the sparse dumpfile avoids allocating disk blocks, but its
4316 * huge holes still show up with zeroes where they need to be.
4318 if (absent && (flags & FOLL_DUMP) &&
4319 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4327 * We need call hugetlb_fault for both hugepages under migration
4328 * (in which case hugetlb_fault waits for the migration,) and
4329 * hwpoisoned hugepages (in which case we need to prevent the
4330 * caller from accessing to them.) In order to do this, we use
4331 * here is_swap_pte instead of is_hugetlb_entry_migration and
4332 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4333 * both cases, and because we can't follow correct pages
4334 * directly from any kind of swap entries.
4336 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4337 ((flags & FOLL_WRITE) &&
4338 !huge_pte_write(huge_ptep_get(pte)))) {
4340 unsigned int fault_flags = 0;
4344 if (flags & FOLL_WRITE)
4345 fault_flags |= FAULT_FLAG_WRITE;
4347 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4348 FAULT_FLAG_KILLABLE;
4349 if (flags & FOLL_NOWAIT)
4350 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4351 FAULT_FLAG_RETRY_NOWAIT;
4352 if (flags & FOLL_TRIED) {
4354 * Note: FAULT_FLAG_ALLOW_RETRY and
4355 * FAULT_FLAG_TRIED can co-exist
4357 fault_flags |= FAULT_FLAG_TRIED;
4359 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4360 if (ret & VM_FAULT_ERROR) {
4361 err = vm_fault_to_errno(ret, flags);
4365 if (ret & VM_FAULT_RETRY) {
4367 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4371 * VM_FAULT_RETRY must not return an
4372 * error, it will return zero
4375 * No need to update "position" as the
4376 * caller will not check it after
4377 * *nr_pages is set to 0.
4384 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4385 page = pte_page(huge_ptep_get(pte));
4388 * If subpage information not requested, update counters
4389 * and skip the same_page loop below.
4391 if (!pages && !vmas && !pfn_offset &&
4392 (vaddr + huge_page_size(h) < vma->vm_end) &&
4393 (remainder >= pages_per_huge_page(h))) {
4394 vaddr += huge_page_size(h);
4395 remainder -= pages_per_huge_page(h);
4396 i += pages_per_huge_page(h);
4403 pages[i] = mem_map_offset(page, pfn_offset);
4405 * try_grab_page() should always succeed here, because:
4406 * a) we hold the ptl lock, and b) we've just checked
4407 * that the huge page is present in the page tables. If
4408 * the huge page is present, then the tail pages must
4409 * also be present. The ptl prevents the head page and
4410 * tail pages from being rearranged in any way. So this
4411 * page must be available at this point, unless the page
4412 * refcount overflowed:
4414 if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) {
4429 if (vaddr < vma->vm_end && remainder &&
4430 pfn_offset < pages_per_huge_page(h)) {
4432 * We use pfn_offset to avoid touching the pageframes
4433 * of this compound page.
4439 *nr_pages = remainder;
4441 * setting position is actually required only if remainder is
4442 * not zero but it's faster not to add a "if (remainder)"
4450 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4452 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4455 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4458 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4459 unsigned long address, unsigned long end, pgprot_t newprot)
4461 struct mm_struct *mm = vma->vm_mm;
4462 unsigned long start = address;
4465 struct hstate *h = hstate_vma(vma);
4466 unsigned long pages = 0;
4467 bool shared_pmd = false;
4468 struct mmu_notifier_range range;
4471 * In the case of shared PMDs, the area to flush could be beyond
4472 * start/end. Set range.start/range.end to cover the maximum possible
4473 * range if PMD sharing is possible.
4475 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
4476 0, vma, mm, start, end);
4477 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4479 BUG_ON(address >= end);
4480 flush_cache_range(vma, range.start, range.end);
4482 mmu_notifier_invalidate_range_start(&range);
4483 i_mmap_lock_write(vma->vm_file->f_mapping);
4484 for (; address < end; address += huge_page_size(h)) {
4486 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4489 ptl = huge_pte_lock(h, mm, ptep);
4490 if (huge_pmd_unshare(mm, &address, ptep)) {
4496 pte = huge_ptep_get(ptep);
4497 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4501 if (unlikely(is_hugetlb_entry_migration(pte))) {
4502 swp_entry_t entry = pte_to_swp_entry(pte);
4504 if (is_write_migration_entry(entry)) {
4507 make_migration_entry_read(&entry);
4508 newpte = swp_entry_to_pte(entry);
4509 set_huge_swap_pte_at(mm, address, ptep,
4510 newpte, huge_page_size(h));
4516 if (!huge_pte_none(pte)) {
4519 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
4520 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
4521 pte = arch_make_huge_pte(pte, vma, NULL, 0);
4522 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
4528 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4529 * may have cleared our pud entry and done put_page on the page table:
4530 * once we release i_mmap_rwsem, another task can do the final put_page
4531 * and that page table be reused and filled with junk. If we actually
4532 * did unshare a page of pmds, flush the range corresponding to the pud.
4535 flush_hugetlb_tlb_range(vma, range.start, range.end);
4537 flush_hugetlb_tlb_range(vma, start, end);
4539 * No need to call mmu_notifier_invalidate_range() we are downgrading
4540 * page table protection not changing it to point to a new page.
4542 * See Documentation/vm/mmu_notifier.rst
4544 i_mmap_unlock_write(vma->vm_file->f_mapping);
4545 mmu_notifier_invalidate_range_end(&range);
4547 return pages << h->order;
4550 int hugetlb_reserve_pages(struct inode *inode,
4552 struct vm_area_struct *vma,
4553 vm_flags_t vm_flags)
4556 struct hstate *h = hstate_inode(inode);
4557 struct hugepage_subpool *spool = subpool_inode(inode);
4558 struct resv_map *resv_map;
4561 /* This should never happen */
4563 VM_WARN(1, "%s called with a negative range\n", __func__);
4568 * Only apply hugepage reservation if asked. At fault time, an
4569 * attempt will be made for VM_NORESERVE to allocate a page
4570 * without using reserves
4572 if (vm_flags & VM_NORESERVE)
4576 * Shared mappings base their reservation on the number of pages that
4577 * are already allocated on behalf of the file. Private mappings need
4578 * to reserve the full area even if read-only as mprotect() may be
4579 * called to make the mapping read-write. Assume !vma is a shm mapping
4581 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4583 * resv_map can not be NULL as hugetlb_reserve_pages is only
4584 * called for inodes for which resv_maps were created (see
4585 * hugetlbfs_get_inode).
4587 resv_map = inode_resv_map(inode);
4589 chg = region_chg(resv_map, from, to);
4592 resv_map = resv_map_alloc();
4598 set_vma_resv_map(vma, resv_map);
4599 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4608 * There must be enough pages in the subpool for the mapping. If
4609 * the subpool has a minimum size, there may be some global
4610 * reservations already in place (gbl_reserve).
4612 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4613 if (gbl_reserve < 0) {
4619 * Check enough hugepages are available for the reservation.
4620 * Hand the pages back to the subpool if there are not
4622 ret = hugetlb_acct_memory(h, gbl_reserve);
4624 /* put back original number of pages, chg */
4625 (void)hugepage_subpool_put_pages(spool, chg);
4630 * Account for the reservations made. Shared mappings record regions
4631 * that have reservations as they are shared by multiple VMAs.
4632 * When the last VMA disappears, the region map says how much
4633 * the reservation was and the page cache tells how much of
4634 * the reservation was consumed. Private mappings are per-VMA and
4635 * only the consumed reservations are tracked. When the VMA
4636 * disappears, the original reservation is the VMA size and the
4637 * consumed reservations are stored in the map. Hence, nothing
4638 * else has to be done for private mappings here
4640 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4641 long add = region_add(resv_map, from, to);
4643 if (unlikely(chg > add)) {
4645 * pages in this range were added to the reserve
4646 * map between region_chg and region_add. This
4647 * indicates a race with alloc_huge_page. Adjust
4648 * the subpool and reserve counts modified above
4649 * based on the difference.
4653 rsv_adjust = hugepage_subpool_put_pages(spool,
4655 hugetlb_acct_memory(h, -rsv_adjust);
4660 if (!vma || vma->vm_flags & VM_MAYSHARE)
4661 /* Don't call region_abort if region_chg failed */
4663 region_abort(resv_map, from, to);
4664 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4665 kref_put(&resv_map->refs, resv_map_release);
4669 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4672 struct hstate *h = hstate_inode(inode);
4673 struct resv_map *resv_map = inode_resv_map(inode);
4675 struct hugepage_subpool *spool = subpool_inode(inode);
4679 * Since this routine can be called in the evict inode path for all
4680 * hugetlbfs inodes, resv_map could be NULL.
4683 chg = region_del(resv_map, start, end);
4685 * region_del() can fail in the rare case where a region
4686 * must be split and another region descriptor can not be
4687 * allocated. If end == LONG_MAX, it will not fail.
4693 spin_lock(&inode->i_lock);
4694 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4695 spin_unlock(&inode->i_lock);
4698 * If the subpool has a minimum size, the number of global
4699 * reservations to be released may be adjusted.
4701 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4702 hugetlb_acct_memory(h, -gbl_reserve);
4707 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4708 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4709 struct vm_area_struct *vma,
4710 unsigned long addr, pgoff_t idx)
4712 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4714 unsigned long sbase = saddr & PUD_MASK;
4715 unsigned long s_end = sbase + PUD_SIZE;
4717 /* Allow segments to share if only one is marked locked */
4718 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4719 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4722 * match the virtual addresses, permission and the alignment of the
4725 if (pmd_index(addr) != pmd_index(saddr) ||
4726 vm_flags != svm_flags ||
4727 sbase < svma->vm_start || svma->vm_end < s_end)
4733 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4735 unsigned long base = addr & PUD_MASK;
4736 unsigned long end = base + PUD_SIZE;
4739 * check on proper vm_flags and page table alignment
4741 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
4747 * Determine if start,end range within vma could be mapped by shared pmd.
4748 * If yes, adjust start and end to cover range associated with possible
4749 * shared pmd mappings.
4751 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4752 unsigned long *start, unsigned long *end)
4754 unsigned long check_addr = *start;
4756 if (!(vma->vm_flags & VM_MAYSHARE))
4759 for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
4760 unsigned long a_start = check_addr & PUD_MASK;
4761 unsigned long a_end = a_start + PUD_SIZE;
4764 * If sharing is possible, adjust start/end if necessary.
4766 if (range_in_vma(vma, a_start, a_end)) {
4767 if (a_start < *start)
4776 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4777 * and returns the corresponding pte. While this is not necessary for the
4778 * !shared pmd case because we can allocate the pmd later as well, it makes the
4779 * code much cleaner. pmd allocation is essential for the shared case because
4780 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4781 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4782 * bad pmd for sharing.
4784 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4786 struct vm_area_struct *vma = find_vma(mm, addr);
4787 struct address_space *mapping = vma->vm_file->f_mapping;
4788 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4790 struct vm_area_struct *svma;
4791 unsigned long saddr;
4796 if (!vma_shareable(vma, addr))
4797 return (pte_t *)pmd_alloc(mm, pud, addr);
4799 i_mmap_lock_read(mapping);
4800 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4804 saddr = page_table_shareable(svma, vma, addr, idx);
4806 spte = huge_pte_offset(svma->vm_mm, saddr,
4807 vma_mmu_pagesize(svma));
4809 get_page(virt_to_page(spte));
4818 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4819 if (pud_none(*pud)) {
4820 pud_populate(mm, pud,
4821 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4824 put_page(virt_to_page(spte));
4828 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4829 i_mmap_unlock_read(mapping);
4834 * unmap huge page backed by shared pte.
4836 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4837 * indicated by page_count > 1, unmap is achieved by clearing pud and
4838 * decrementing the ref count. If count == 1, the pte page is not shared.
4840 * called with page table lock held.
4842 * returns: 1 successfully unmapped a shared pte page
4843 * 0 the underlying pte page is not shared, or it is the last user
4845 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4847 pgd_t *pgd = pgd_offset(mm, *addr);
4848 p4d_t *p4d = p4d_offset(pgd, *addr);
4849 pud_t *pud = pud_offset(p4d, *addr);
4851 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4852 if (page_count(virt_to_page(ptep)) == 1)
4856 put_page(virt_to_page(ptep));
4858 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4861 #define want_pmd_share() (1)
4862 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4863 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4868 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4873 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4874 unsigned long *start, unsigned long *end)
4877 #define want_pmd_share() (0)
4878 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4880 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4881 pte_t *huge_pte_alloc(struct mm_struct *mm,
4882 unsigned long addr, unsigned long sz)
4889 pgd = pgd_offset(mm, addr);
4890 p4d = p4d_alloc(mm, pgd, addr);
4893 pud = pud_alloc(mm, p4d, addr);
4895 if (sz == PUD_SIZE) {
4898 BUG_ON(sz != PMD_SIZE);
4899 if (want_pmd_share() && pud_none(*pud))
4900 pte = huge_pmd_share(mm, addr, pud);
4902 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4905 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4911 * huge_pte_offset() - Walk the page table to resolve the hugepage
4912 * entry at address @addr
4914 * Return: Pointer to page table or swap entry (PUD or PMD) for
4915 * address @addr, or NULL if a p*d_none() entry is encountered and the
4916 * size @sz doesn't match the hugepage size at this level of the page
4919 pte_t *huge_pte_offset(struct mm_struct *mm,
4920 unsigned long addr, unsigned long sz)
4927 pgd = pgd_offset(mm, addr);
4928 if (!pgd_present(*pgd))
4930 p4d = p4d_offset(pgd, addr);
4931 if (!p4d_present(*p4d))
4934 pud = pud_offset(p4d, addr);
4935 if (sz != PUD_SIZE && pud_none(*pud))
4937 /* hugepage or swap? */
4938 if (pud_huge(*pud) || !pud_present(*pud))
4939 return (pte_t *)pud;
4941 pmd = pmd_offset(pud, addr);
4942 if (sz != PMD_SIZE && pmd_none(*pmd))
4944 /* hugepage or swap? */
4945 if (pmd_huge(*pmd) || !pmd_present(*pmd))
4946 return (pte_t *)pmd;
4951 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4954 * These functions are overwritable if your architecture needs its own
4957 struct page * __weak
4958 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4961 return ERR_PTR(-EINVAL);
4964 struct page * __weak
4965 follow_huge_pd(struct vm_area_struct *vma,
4966 unsigned long address, hugepd_t hpd, int flags, int pdshift)
4968 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4972 struct page * __weak
4973 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4974 pmd_t *pmd, int flags)
4976 struct page *page = NULL;
4980 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
4981 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
4982 (FOLL_PIN | FOLL_GET)))
4986 ptl = pmd_lockptr(mm, pmd);
4989 * make sure that the address range covered by this pmd is not
4990 * unmapped from other threads.
4992 if (!pmd_huge(*pmd))
4994 pte = huge_ptep_get((pte_t *)pmd);
4995 if (pte_present(pte)) {
4996 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4998 * try_grab_page() should always succeed here, because: a) we
4999 * hold the pmd (ptl) lock, and b) we've just checked that the
5000 * huge pmd (head) page is present in the page tables. The ptl
5001 * prevents the head page and tail pages from being rearranged
5002 * in any way. So this page must be available at this point,
5003 * unless the page refcount overflowed:
5005 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5010 if (is_hugetlb_entry_migration(pte)) {
5012 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5016 * hwpoisoned entry is treated as no_page_table in
5017 * follow_page_mask().
5025 struct page * __weak
5026 follow_huge_pud(struct mm_struct *mm, unsigned long address,
5027 pud_t *pud, int flags)
5029 if (flags & (FOLL_GET | FOLL_PIN))
5032 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5035 struct page * __weak
5036 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5038 if (flags & (FOLL_GET | FOLL_PIN))
5041 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5044 bool isolate_huge_page(struct page *page, struct list_head *list)
5048 VM_BUG_ON_PAGE(!PageHead(page), page);
5049 spin_lock(&hugetlb_lock);
5050 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
5054 clear_page_huge_active(page);
5055 list_move_tail(&page->lru, list);
5057 spin_unlock(&hugetlb_lock);
5061 void putback_active_hugepage(struct page *page)
5063 VM_BUG_ON_PAGE(!PageHead(page), page);
5064 spin_lock(&hugetlb_lock);
5065 set_page_huge_active(page);
5066 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5067 spin_unlock(&hugetlb_lock);
5071 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5073 struct hstate *h = page_hstate(oldpage);
5075 hugetlb_cgroup_migrate(oldpage, newpage);
5076 set_page_owner_migrate_reason(newpage, reason);
5079 * transfer temporary state of the new huge page. This is
5080 * reverse to other transitions because the newpage is going to
5081 * be final while the old one will be freed so it takes over
5082 * the temporary status.
5084 * Also note that we have to transfer the per-node surplus state
5085 * here as well otherwise the global surplus count will not match
5088 if (PageHugeTemporary(newpage)) {
5089 int old_nid = page_to_nid(oldpage);
5090 int new_nid = page_to_nid(newpage);
5092 SetPageHugeTemporary(oldpage);
5093 ClearPageHugeTemporary(newpage);
5095 spin_lock(&hugetlb_lock);
5096 if (h->surplus_huge_pages_node[old_nid]) {
5097 h->surplus_huge_pages_node[old_nid]--;
5098 h->surplus_huge_pages_node[new_nid]++;
5100 spin_unlock(&hugetlb_lock);