2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
8 #include <linux/seq_file.h>
9 #include <linux/sysctl.h>
10 #include <linux/highmem.h>
11 #include <linux/mmu_notifier.h>
12 #include <linux/nodemask.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/compiler.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/memblock.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/mmdebug.h>
22 #include <linux/sched/signal.h>
23 #include <linux/rmap.h>
24 #include <linux/string_helpers.h>
25 #include <linux/swap.h>
26 #include <linux/swapops.h>
27 #include <linux/jhash.h>
28 #include <linux/numa.h>
31 #include <asm/pgtable.h>
35 #include <linux/hugetlb.h>
36 #include <linux/hugetlb_cgroup.h>
37 #include <linux/node.h>
38 #include <linux/userfaultfd_k.h>
39 #include <linux/page_owner.h>
42 int hugetlb_max_hstate __read_mostly;
43 unsigned int default_hstate_idx;
44 struct hstate hstates[HUGE_MAX_HSTATE];
46 * Minimum page order among possible hugepage sizes, set to a proper value
49 static unsigned int minimum_order __read_mostly = UINT_MAX;
51 __initdata LIST_HEAD(huge_boot_pages);
53 /* for command line parsing */
54 static struct hstate * __initdata parsed_hstate;
55 static unsigned long __initdata default_hstate_max_huge_pages;
56 static unsigned long __initdata default_hstate_size;
57 static bool __initdata parsed_valid_hugepagesz = true;
60 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
61 * free_huge_pages, and surplus_huge_pages.
63 DEFINE_SPINLOCK(hugetlb_lock);
66 * Serializes faults on the same logical page. This is used to
67 * prevent spurious OOMs when the hugepage pool is fully utilized.
69 static int num_fault_mutexes;
70 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
72 /* Forward declaration */
73 static int hugetlb_acct_memory(struct hstate *h, long delta);
75 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
77 bool free = (spool->count == 0) && (spool->used_hpages == 0);
79 spin_unlock(&spool->lock);
81 /* If no pages are used, and no other handles to the subpool
82 * remain, give up any reservations mased on minimum size and
85 if (spool->min_hpages != -1)
86 hugetlb_acct_memory(spool->hstate,
92 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
95 struct hugepage_subpool *spool;
97 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
101 spin_lock_init(&spool->lock);
103 spool->max_hpages = max_hpages;
105 spool->min_hpages = min_hpages;
107 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
111 spool->rsv_hpages = min_hpages;
116 void hugepage_put_subpool(struct hugepage_subpool *spool)
118 spin_lock(&spool->lock);
119 BUG_ON(!spool->count);
121 unlock_or_release_subpool(spool);
125 * Subpool accounting for allocating and reserving pages.
126 * Return -ENOMEM if there are not enough resources to satisfy the
127 * the request. Otherwise, return the number of pages by which the
128 * global pools must be adjusted (upward). The returned value may
129 * only be different than the passed value (delta) in the case where
130 * a subpool minimum size must be manitained.
132 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
140 spin_lock(&spool->lock);
142 if (spool->max_hpages != -1) { /* maximum size accounting */
143 if ((spool->used_hpages + delta) <= spool->max_hpages)
144 spool->used_hpages += delta;
151 /* minimum size accounting */
152 if (spool->min_hpages != -1 && spool->rsv_hpages) {
153 if (delta > spool->rsv_hpages) {
155 * Asking for more reserves than those already taken on
156 * behalf of subpool. Return difference.
158 ret = delta - spool->rsv_hpages;
159 spool->rsv_hpages = 0;
161 ret = 0; /* reserves already accounted for */
162 spool->rsv_hpages -= delta;
167 spin_unlock(&spool->lock);
172 * Subpool accounting for freeing and unreserving pages.
173 * Return the number of global page reservations that must be dropped.
174 * The return value may only be different than the passed value (delta)
175 * in the case where a subpool minimum size must be maintained.
177 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
185 spin_lock(&spool->lock);
187 if (spool->max_hpages != -1) /* maximum size accounting */
188 spool->used_hpages -= delta;
190 /* minimum size accounting */
191 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
192 if (spool->rsv_hpages + delta <= spool->min_hpages)
195 ret = spool->rsv_hpages + delta - spool->min_hpages;
197 spool->rsv_hpages += delta;
198 if (spool->rsv_hpages > spool->min_hpages)
199 spool->rsv_hpages = spool->min_hpages;
203 * If hugetlbfs_put_super couldn't free spool due to an outstanding
204 * quota reference, free it now.
206 unlock_or_release_subpool(spool);
211 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
213 return HUGETLBFS_SB(inode->i_sb)->spool;
216 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
218 return subpool_inode(file_inode(vma->vm_file));
222 * Region tracking -- allows tracking of reservations and instantiated pages
223 * across the pages in a mapping.
225 * The region data structures are embedded into a resv_map and protected
226 * by a resv_map's lock. The set of regions within the resv_map represent
227 * reservations for huge pages, or huge pages that have already been
228 * instantiated within the map. The from and to elements are huge page
229 * indicies into the associated mapping. from indicates the starting index
230 * of the region. to represents the first index past the end of the region.
232 * For example, a file region structure with from == 0 and to == 4 represents
233 * four huge pages in a mapping. It is important to note that the to element
234 * represents the first element past the end of the region. This is used in
235 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
237 * Interval notation of the form [from, to) will be used to indicate that
238 * the endpoint from is inclusive and to is exclusive.
241 struct list_head link;
247 * Add the huge page range represented by [f, t) to the reserve
248 * map. In the normal case, existing regions will be expanded
249 * to accommodate the specified range. Sufficient regions should
250 * exist for expansion due to the previous call to region_chg
251 * with the same range. However, it is possible that region_del
252 * could have been called after region_chg and modifed the map
253 * in such a way that no region exists to be expanded. In this
254 * case, pull a region descriptor from the cache associated with
255 * the map and use that for the new range.
257 * Return the number of new huge pages added to the map. This
258 * number is greater than or equal to zero.
260 static long region_add(struct resv_map *resv, long f, long t)
262 struct list_head *head = &resv->regions;
263 struct file_region *rg, *nrg, *trg;
266 spin_lock(&resv->lock);
267 /* Locate the region we are either in or before. */
268 list_for_each_entry(rg, head, link)
273 * If no region exists which can be expanded to include the
274 * specified range, the list must have been modified by an
275 * interleving call to region_del(). Pull a region descriptor
276 * from the cache and use it for this range.
278 if (&rg->link == head || t < rg->from) {
279 VM_BUG_ON(resv->region_cache_count <= 0);
281 resv->region_cache_count--;
282 nrg = list_first_entry(&resv->region_cache, struct file_region,
284 list_del(&nrg->link);
288 list_add(&nrg->link, rg->link.prev);
294 /* Round our left edge to the current segment if it encloses us. */
298 /* Check for and consume any regions we now overlap with. */
300 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
301 if (&rg->link == head)
306 /* If this area reaches higher then extend our area to
307 * include it completely. If this is not the first area
308 * which we intend to reuse, free it. */
312 /* Decrement return value by the deleted range.
313 * Another range will span this area so that by
314 * end of routine add will be >= zero
316 add -= (rg->to - rg->from);
322 add += (nrg->from - f); /* Added to beginning of region */
324 add += t - nrg->to; /* Added to end of region */
328 resv->adds_in_progress--;
329 spin_unlock(&resv->lock);
335 * Examine the existing reserve map and determine how many
336 * huge pages in the specified range [f, t) are NOT currently
337 * represented. This routine is called before a subsequent
338 * call to region_add that will actually modify the reserve
339 * map to add the specified range [f, t). region_chg does
340 * not change the number of huge pages represented by the
341 * map. However, if the existing regions in the map can not
342 * be expanded to represent the new range, a new file_region
343 * structure is added to the map as a placeholder. This is
344 * so that the subsequent region_add call will have all the
345 * regions it needs and will not fail.
347 * Upon entry, region_chg will also examine the cache of region descriptors
348 * associated with the map. If there are not enough descriptors cached, one
349 * will be allocated for the in progress add operation.
351 * Returns the number of huge pages that need to be added to the existing
352 * reservation map for the range [f, t). This number is greater or equal to
353 * zero. -ENOMEM is returned if a new file_region structure or cache entry
354 * is needed and can not be allocated.
356 static long region_chg(struct resv_map *resv, long f, long t)
358 struct list_head *head = &resv->regions;
359 struct file_region *rg, *nrg = NULL;
363 spin_lock(&resv->lock);
365 resv->adds_in_progress++;
368 * Check for sufficient descriptors in the cache to accommodate
369 * the number of in progress add operations.
371 if (resv->adds_in_progress > resv->region_cache_count) {
372 struct file_region *trg;
374 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
375 /* Must drop lock to allocate a new descriptor. */
376 resv->adds_in_progress--;
377 spin_unlock(&resv->lock);
379 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
385 spin_lock(&resv->lock);
386 list_add(&trg->link, &resv->region_cache);
387 resv->region_cache_count++;
391 /* Locate the region we are before or in. */
392 list_for_each_entry(rg, head, link)
396 /* If we are below the current region then a new region is required.
397 * Subtle, allocate a new region at the position but make it zero
398 * size such that we can guarantee to record the reservation. */
399 if (&rg->link == head || t < rg->from) {
401 resv->adds_in_progress--;
402 spin_unlock(&resv->lock);
403 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
409 INIT_LIST_HEAD(&nrg->link);
413 list_add(&nrg->link, rg->link.prev);
418 /* Round our left edge to the current segment if it encloses us. */
423 /* Check for and consume any regions we now overlap with. */
424 list_for_each_entry(rg, rg->link.prev, link) {
425 if (&rg->link == head)
430 /* We overlap with this area, if it extends further than
431 * us then we must extend ourselves. Account for its
432 * existing reservation. */
437 chg -= rg->to - rg->from;
441 spin_unlock(&resv->lock);
442 /* We already know we raced and no longer need the new region */
446 spin_unlock(&resv->lock);
451 * Abort the in progress add operation. The adds_in_progress field
452 * of the resv_map keeps track of the operations in progress between
453 * calls to region_chg and region_add. Operations are sometimes
454 * aborted after the call to region_chg. In such cases, region_abort
455 * is called to decrement the adds_in_progress counter.
457 * NOTE: The range arguments [f, t) are not needed or used in this
458 * routine. They are kept to make reading the calling code easier as
459 * arguments will match the associated region_chg call.
461 static void region_abort(struct resv_map *resv, long f, long t)
463 spin_lock(&resv->lock);
464 VM_BUG_ON(!resv->region_cache_count);
465 resv->adds_in_progress--;
466 spin_unlock(&resv->lock);
470 * Delete the specified range [f, t) from the reserve map. If the
471 * t parameter is LONG_MAX, this indicates that ALL regions after f
472 * should be deleted. Locate the regions which intersect [f, t)
473 * and either trim, delete or split the existing regions.
475 * Returns the number of huge pages deleted from the reserve map.
476 * In the normal case, the return value is zero or more. In the
477 * case where a region must be split, a new region descriptor must
478 * be allocated. If the allocation fails, -ENOMEM will be returned.
479 * NOTE: If the parameter t == LONG_MAX, then we will never split
480 * a region and possibly return -ENOMEM. Callers specifying
481 * t == LONG_MAX do not need to check for -ENOMEM error.
483 static long region_del(struct resv_map *resv, long f, long t)
485 struct list_head *head = &resv->regions;
486 struct file_region *rg, *trg;
487 struct file_region *nrg = NULL;
491 spin_lock(&resv->lock);
492 list_for_each_entry_safe(rg, trg, head, link) {
494 * Skip regions before the range to be deleted. file_region
495 * ranges are normally of the form [from, to). However, there
496 * may be a "placeholder" entry in the map which is of the form
497 * (from, to) with from == to. Check for placeholder entries
498 * at the beginning of the range to be deleted.
500 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
506 if (f > rg->from && t < rg->to) { /* Must split region */
508 * Check for an entry in the cache before dropping
509 * lock and attempting allocation.
512 resv->region_cache_count > resv->adds_in_progress) {
513 nrg = list_first_entry(&resv->region_cache,
516 list_del(&nrg->link);
517 resv->region_cache_count--;
521 spin_unlock(&resv->lock);
522 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
530 /* New entry for end of split region */
533 INIT_LIST_HEAD(&nrg->link);
535 /* Original entry is trimmed */
538 list_add(&nrg->link, &rg->link);
543 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
544 del += rg->to - rg->from;
550 if (f <= rg->from) { /* Trim beginning of region */
553 } else { /* Trim end of region */
559 spin_unlock(&resv->lock);
565 * A rare out of memory error was encountered which prevented removal of
566 * the reserve map region for a page. The huge page itself was free'ed
567 * and removed from the page cache. This routine will adjust the subpool
568 * usage count, and the global reserve count if needed. By incrementing
569 * these counts, the reserve map entry which could not be deleted will
570 * appear as a "reserved" entry instead of simply dangling with incorrect
573 void hugetlb_fix_reserve_counts(struct inode *inode)
575 struct hugepage_subpool *spool = subpool_inode(inode);
578 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
580 struct hstate *h = hstate_inode(inode);
582 hugetlb_acct_memory(h, 1);
587 * Count and return the number of huge pages in the reserve map
588 * that intersect with the range [f, t).
590 static long region_count(struct resv_map *resv, long f, long t)
592 struct list_head *head = &resv->regions;
593 struct file_region *rg;
596 spin_lock(&resv->lock);
597 /* Locate each segment we overlap with, and count that overlap. */
598 list_for_each_entry(rg, head, link) {
607 seg_from = max(rg->from, f);
608 seg_to = min(rg->to, t);
610 chg += seg_to - seg_from;
612 spin_unlock(&resv->lock);
618 * Convert the address within this vma to the page offset within
619 * the mapping, in pagecache page units; huge pages here.
621 static pgoff_t vma_hugecache_offset(struct hstate *h,
622 struct vm_area_struct *vma, unsigned long address)
624 return ((address - vma->vm_start) >> huge_page_shift(h)) +
625 (vma->vm_pgoff >> huge_page_order(h));
628 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
629 unsigned long address)
631 return vma_hugecache_offset(hstate_vma(vma), vma, address);
633 EXPORT_SYMBOL_GPL(linear_hugepage_index);
636 * Return the size of the pages allocated when backing a VMA. In the majority
637 * cases this will be same size as used by the page table entries.
639 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
641 if (vma->vm_ops && vma->vm_ops->pagesize)
642 return vma->vm_ops->pagesize(vma);
645 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
648 * Return the page size being used by the MMU to back a VMA. In the majority
649 * of cases, the page size used by the kernel matches the MMU size. On
650 * architectures where it differs, an architecture-specific 'strong'
651 * version of this symbol is required.
653 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
655 return vma_kernel_pagesize(vma);
659 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
660 * bits of the reservation map pointer, which are always clear due to
663 #define HPAGE_RESV_OWNER (1UL << 0)
664 #define HPAGE_RESV_UNMAPPED (1UL << 1)
665 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
668 * These helpers are used to track how many pages are reserved for
669 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
670 * is guaranteed to have their future faults succeed.
672 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
673 * the reserve counters are updated with the hugetlb_lock held. It is safe
674 * to reset the VMA at fork() time as it is not in use yet and there is no
675 * chance of the global counters getting corrupted as a result of the values.
677 * The private mapping reservation is represented in a subtly different
678 * manner to a shared mapping. A shared mapping has a region map associated
679 * with the underlying file, this region map represents the backing file
680 * pages which have ever had a reservation assigned which this persists even
681 * after the page is instantiated. A private mapping has a region map
682 * associated with the original mmap which is attached to all VMAs which
683 * reference it, this region map represents those offsets which have consumed
684 * reservation ie. where pages have been instantiated.
686 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
688 return (unsigned long)vma->vm_private_data;
691 static void set_vma_private_data(struct vm_area_struct *vma,
694 vma->vm_private_data = (void *)value;
697 struct resv_map *resv_map_alloc(void)
699 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
700 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
702 if (!resv_map || !rg) {
708 kref_init(&resv_map->refs);
709 spin_lock_init(&resv_map->lock);
710 INIT_LIST_HEAD(&resv_map->regions);
712 resv_map->adds_in_progress = 0;
714 INIT_LIST_HEAD(&resv_map->region_cache);
715 list_add(&rg->link, &resv_map->region_cache);
716 resv_map->region_cache_count = 1;
721 void resv_map_release(struct kref *ref)
723 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
724 struct list_head *head = &resv_map->region_cache;
725 struct file_region *rg, *trg;
727 /* Clear out any active regions before we release the map. */
728 region_del(resv_map, 0, LONG_MAX);
730 /* ... and any entries left in the cache */
731 list_for_each_entry_safe(rg, trg, head, link) {
736 VM_BUG_ON(resv_map->adds_in_progress);
741 static inline struct resv_map *inode_resv_map(struct inode *inode)
744 * At inode evict time, i_mapping may not point to the original
745 * address space within the inode. This original address space
746 * contains the pointer to the resv_map. So, always use the
747 * address space embedded within the inode.
748 * The VERY common case is inode->mapping == &inode->i_data but,
749 * this may not be true for device special inodes.
751 return (struct resv_map *)(&inode->i_data)->private_data;
754 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
756 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
757 if (vma->vm_flags & VM_MAYSHARE) {
758 struct address_space *mapping = vma->vm_file->f_mapping;
759 struct inode *inode = mapping->host;
761 return inode_resv_map(inode);
764 return (struct resv_map *)(get_vma_private_data(vma) &
769 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
771 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
772 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
774 set_vma_private_data(vma, (get_vma_private_data(vma) &
775 HPAGE_RESV_MASK) | (unsigned long)map);
778 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
780 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
781 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
783 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
786 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
788 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
790 return (get_vma_private_data(vma) & flag) != 0;
793 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
794 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
796 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
797 if (!(vma->vm_flags & VM_MAYSHARE))
798 vma->vm_private_data = (void *)0;
801 /* Returns true if the VMA has associated reserve pages */
802 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
804 if (vma->vm_flags & VM_NORESERVE) {
806 * This address is already reserved by other process(chg == 0),
807 * so, we should decrement reserved count. Without decrementing,
808 * reserve count remains after releasing inode, because this
809 * allocated page will go into page cache and is regarded as
810 * coming from reserved pool in releasing step. Currently, we
811 * don't have any other solution to deal with this situation
812 * properly, so add work-around here.
814 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
820 /* Shared mappings always use reserves */
821 if (vma->vm_flags & VM_MAYSHARE) {
823 * We know VM_NORESERVE is not set. Therefore, there SHOULD
824 * be a region map for all pages. The only situation where
825 * there is no region map is if a hole was punched via
826 * fallocate. In this case, there really are no reverves to
827 * use. This situation is indicated if chg != 0.
836 * Only the process that called mmap() has reserves for
839 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
841 * Like the shared case above, a hole punch or truncate
842 * could have been performed on the private mapping.
843 * Examine the value of chg to determine if reserves
844 * actually exist or were previously consumed.
845 * Very Subtle - The value of chg comes from a previous
846 * call to vma_needs_reserves(). The reserve map for
847 * private mappings has different (opposite) semantics
848 * than that of shared mappings. vma_needs_reserves()
849 * has already taken this difference in semantics into
850 * account. Therefore, the meaning of chg is the same
851 * as in the shared case above. Code could easily be
852 * combined, but keeping it separate draws attention to
853 * subtle differences.
864 static void enqueue_huge_page(struct hstate *h, struct page *page)
866 int nid = page_to_nid(page);
867 list_move(&page->lru, &h->hugepage_freelists[nid]);
868 h->free_huge_pages++;
869 h->free_huge_pages_node[nid]++;
872 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
876 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
877 if (!PageHWPoison(page))
880 * if 'non-isolated free hugepage' not found on the list,
881 * the allocation fails.
883 if (&h->hugepage_freelists[nid] == &page->lru)
885 list_move(&page->lru, &h->hugepage_activelist);
886 set_page_refcounted(page);
887 h->free_huge_pages--;
888 h->free_huge_pages_node[nid]--;
892 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
895 unsigned int cpuset_mems_cookie;
896 struct zonelist *zonelist;
899 int node = NUMA_NO_NODE;
901 zonelist = node_zonelist(nid, gfp_mask);
904 cpuset_mems_cookie = read_mems_allowed_begin();
905 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
908 if (!cpuset_zone_allowed(zone, gfp_mask))
911 * no need to ask again on the same node. Pool is node rather than
914 if (zone_to_nid(zone) == node)
916 node = zone_to_nid(zone);
918 page = dequeue_huge_page_node_exact(h, node);
922 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
928 /* Movability of hugepages depends on migration support. */
929 static inline gfp_t htlb_alloc_mask(struct hstate *h)
931 if (hugepage_movable_supported(h))
932 return GFP_HIGHUSER_MOVABLE;
937 static struct page *dequeue_huge_page_vma(struct hstate *h,
938 struct vm_area_struct *vma,
939 unsigned long address, int avoid_reserve,
943 struct mempolicy *mpol;
945 nodemask_t *nodemask;
949 * A child process with MAP_PRIVATE mappings created by their parent
950 * have no page reserves. This check ensures that reservations are
951 * not "stolen". The child may still get SIGKILLed
953 if (!vma_has_reserves(vma, chg) &&
954 h->free_huge_pages - h->resv_huge_pages == 0)
957 /* If reserves cannot be used, ensure enough pages are in the pool */
958 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
961 gfp_mask = htlb_alloc_mask(h);
962 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
963 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
964 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
965 SetPagePrivate(page);
966 h->resv_huge_pages--;
977 * common helper functions for hstate_next_node_to_{alloc|free}.
978 * We may have allocated or freed a huge page based on a different
979 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
980 * be outside of *nodes_allowed. Ensure that we use an allowed
981 * node for alloc or free.
983 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
985 nid = next_node_in(nid, *nodes_allowed);
986 VM_BUG_ON(nid >= MAX_NUMNODES);
991 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
993 if (!node_isset(nid, *nodes_allowed))
994 nid = next_node_allowed(nid, nodes_allowed);
999 * returns the previously saved node ["this node"] from which to
1000 * allocate a persistent huge page for the pool and advance the
1001 * next node from which to allocate, handling wrap at end of node
1004 static int hstate_next_node_to_alloc(struct hstate *h,
1005 nodemask_t *nodes_allowed)
1009 VM_BUG_ON(!nodes_allowed);
1011 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1012 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1018 * helper for free_pool_huge_page() - return the previously saved
1019 * node ["this node"] from which to free a huge page. Advance the
1020 * next node id whether or not we find a free huge page to free so
1021 * that the next attempt to free addresses the next node.
1023 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1027 VM_BUG_ON(!nodes_allowed);
1029 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1030 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1035 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1036 for (nr_nodes = nodes_weight(*mask); \
1038 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1041 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1042 for (nr_nodes = nodes_weight(*mask); \
1044 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1047 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1048 static void destroy_compound_gigantic_page(struct page *page,
1052 int nr_pages = 1 << order;
1053 struct page *p = page + 1;
1055 atomic_set(compound_mapcount_ptr(page), 0);
1056 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1057 clear_compound_head(p);
1058 set_page_refcounted(p);
1061 set_compound_order(page, 0);
1062 __ClearPageHead(page);
1065 static void free_gigantic_page(struct page *page, unsigned int order)
1067 free_contig_range(page_to_pfn(page), 1 << order);
1070 #ifdef CONFIG_CONTIG_ALLOC
1071 static int __alloc_gigantic_page(unsigned long start_pfn,
1072 unsigned long nr_pages, gfp_t gfp_mask)
1074 unsigned long end_pfn = start_pfn + nr_pages;
1075 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1079 static bool pfn_range_valid_gigantic(struct zone *z,
1080 unsigned long start_pfn, unsigned long nr_pages)
1082 unsigned long i, end_pfn = start_pfn + nr_pages;
1085 for (i = start_pfn; i < end_pfn; i++) {
1089 page = pfn_to_page(i);
1091 if (page_zone(page) != z)
1094 if (PageReserved(page))
1097 if (page_count(page) > 0)
1107 static bool zone_spans_last_pfn(const struct zone *zone,
1108 unsigned long start_pfn, unsigned long nr_pages)
1110 unsigned long last_pfn = start_pfn + nr_pages - 1;
1111 return zone_spans_pfn(zone, last_pfn);
1114 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1115 int nid, nodemask_t *nodemask)
1117 unsigned int order = huge_page_order(h);
1118 unsigned long nr_pages = 1 << order;
1119 unsigned long ret, pfn, flags;
1120 struct zonelist *zonelist;
1124 zonelist = node_zonelist(nid, gfp_mask);
1125 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) {
1126 spin_lock_irqsave(&zone->lock, flags);
1128 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
1129 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
1130 if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) {
1132 * We release the zone lock here because
1133 * alloc_contig_range() will also lock the zone
1134 * at some point. If there's an allocation
1135 * spinning on this lock, it may win the race
1136 * and cause alloc_contig_range() to fail...
1138 spin_unlock_irqrestore(&zone->lock, flags);
1139 ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask);
1141 return pfn_to_page(pfn);
1142 spin_lock_irqsave(&zone->lock, flags);
1147 spin_unlock_irqrestore(&zone->lock, flags);
1153 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1154 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1155 #else /* !CONFIG_CONTIG_ALLOC */
1156 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1157 int nid, nodemask_t *nodemask)
1161 #endif /* CONFIG_CONTIG_ALLOC */
1163 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1164 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1165 int nid, nodemask_t *nodemask)
1169 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1170 static inline void destroy_compound_gigantic_page(struct page *page,
1171 unsigned int order) { }
1174 static void update_and_free_page(struct hstate *h, struct page *page)
1178 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1182 h->nr_huge_pages_node[page_to_nid(page)]--;
1183 for (i = 0; i < pages_per_huge_page(h); i++) {
1184 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1185 1 << PG_referenced | 1 << PG_dirty |
1186 1 << PG_active | 1 << PG_private |
1189 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1190 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1191 set_page_refcounted(page);
1192 if (hstate_is_gigantic(h)) {
1193 destroy_compound_gigantic_page(page, huge_page_order(h));
1194 free_gigantic_page(page, huge_page_order(h));
1196 __free_pages(page, huge_page_order(h));
1200 struct hstate *size_to_hstate(unsigned long size)
1204 for_each_hstate(h) {
1205 if (huge_page_size(h) == size)
1212 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1213 * to hstate->hugepage_activelist.)
1215 * This function can be called for tail pages, but never returns true for them.
1217 bool page_huge_active(struct page *page)
1219 VM_BUG_ON_PAGE(!PageHuge(page), page);
1220 return PageHead(page) && PagePrivate(&page[1]);
1223 /* never called for tail page */
1224 static void set_page_huge_active(struct page *page)
1226 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1227 SetPagePrivate(&page[1]);
1230 static void clear_page_huge_active(struct page *page)
1232 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1233 ClearPagePrivate(&page[1]);
1237 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1240 static inline bool PageHugeTemporary(struct page *page)
1242 if (!PageHuge(page))
1245 return (unsigned long)page[2].mapping == -1U;
1248 static inline void SetPageHugeTemporary(struct page *page)
1250 page[2].mapping = (void *)-1U;
1253 static inline void ClearPageHugeTemporary(struct page *page)
1255 page[2].mapping = NULL;
1258 void free_huge_page(struct page *page)
1261 * Can't pass hstate in here because it is called from the
1262 * compound page destructor.
1264 struct hstate *h = page_hstate(page);
1265 int nid = page_to_nid(page);
1266 struct hugepage_subpool *spool =
1267 (struct hugepage_subpool *)page_private(page);
1268 bool restore_reserve;
1270 VM_BUG_ON_PAGE(page_count(page), page);
1271 VM_BUG_ON_PAGE(page_mapcount(page), page);
1273 set_page_private(page, 0);
1274 page->mapping = NULL;
1275 restore_reserve = PagePrivate(page);
1276 ClearPagePrivate(page);
1279 * If PagePrivate() was set on page, page allocation consumed a
1280 * reservation. If the page was associated with a subpool, there
1281 * would have been a page reserved in the subpool before allocation
1282 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1283 * reservtion, do not call hugepage_subpool_put_pages() as this will
1284 * remove the reserved page from the subpool.
1286 if (!restore_reserve) {
1288 * A return code of zero implies that the subpool will be
1289 * under its minimum size if the reservation is not restored
1290 * after page is free. Therefore, force restore_reserve
1293 if (hugepage_subpool_put_pages(spool, 1) == 0)
1294 restore_reserve = true;
1297 spin_lock(&hugetlb_lock);
1298 clear_page_huge_active(page);
1299 hugetlb_cgroup_uncharge_page(hstate_index(h),
1300 pages_per_huge_page(h), page);
1301 if (restore_reserve)
1302 h->resv_huge_pages++;
1304 if (PageHugeTemporary(page)) {
1305 list_del(&page->lru);
1306 ClearPageHugeTemporary(page);
1307 update_and_free_page(h, page);
1308 } else if (h->surplus_huge_pages_node[nid]) {
1309 /* remove the page from active list */
1310 list_del(&page->lru);
1311 update_and_free_page(h, page);
1312 h->surplus_huge_pages--;
1313 h->surplus_huge_pages_node[nid]--;
1315 arch_clear_hugepage_flags(page);
1316 enqueue_huge_page(h, page);
1318 spin_unlock(&hugetlb_lock);
1321 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1323 INIT_LIST_HEAD(&page->lru);
1324 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1325 spin_lock(&hugetlb_lock);
1326 set_hugetlb_cgroup(page, NULL);
1328 h->nr_huge_pages_node[nid]++;
1329 spin_unlock(&hugetlb_lock);
1332 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1335 int nr_pages = 1 << order;
1336 struct page *p = page + 1;
1338 /* we rely on prep_new_huge_page to set the destructor */
1339 set_compound_order(page, order);
1340 __ClearPageReserved(page);
1341 __SetPageHead(page);
1342 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1344 * For gigantic hugepages allocated through bootmem at
1345 * boot, it's safer to be consistent with the not-gigantic
1346 * hugepages and clear the PG_reserved bit from all tail pages
1347 * too. Otherwse drivers using get_user_pages() to access tail
1348 * pages may get the reference counting wrong if they see
1349 * PG_reserved set on a tail page (despite the head page not
1350 * having PG_reserved set). Enforcing this consistency between
1351 * head and tail pages allows drivers to optimize away a check
1352 * on the head page when they need know if put_page() is needed
1353 * after get_user_pages().
1355 __ClearPageReserved(p);
1356 set_page_count(p, 0);
1357 set_compound_head(p, page);
1359 atomic_set(compound_mapcount_ptr(page), -1);
1363 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1364 * transparent huge pages. See the PageTransHuge() documentation for more
1367 int PageHuge(struct page *page)
1369 if (!PageCompound(page))
1372 page = compound_head(page);
1373 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1375 EXPORT_SYMBOL_GPL(PageHuge);
1378 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1379 * normal or transparent huge pages.
1381 int PageHeadHuge(struct page *page_head)
1383 if (!PageHead(page_head))
1386 return get_compound_page_dtor(page_head) == free_huge_page;
1389 pgoff_t __basepage_index(struct page *page)
1391 struct page *page_head = compound_head(page);
1392 pgoff_t index = page_index(page_head);
1393 unsigned long compound_idx;
1395 if (!PageHuge(page_head))
1396 return page_index(page);
1398 if (compound_order(page_head) >= MAX_ORDER)
1399 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1401 compound_idx = page - page_head;
1403 return (index << compound_order(page_head)) + compound_idx;
1406 static struct page *alloc_buddy_huge_page(struct hstate *h,
1407 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1409 int order = huge_page_order(h);
1412 gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
1413 if (nid == NUMA_NO_NODE)
1414 nid = numa_mem_id();
1415 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1417 __count_vm_event(HTLB_BUDDY_PGALLOC);
1419 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1425 * Common helper to allocate a fresh hugetlb page. All specific allocators
1426 * should use this function to get new hugetlb pages
1428 static struct page *alloc_fresh_huge_page(struct hstate *h,
1429 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1433 if (hstate_is_gigantic(h))
1434 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1436 page = alloc_buddy_huge_page(h, gfp_mask,
1441 if (hstate_is_gigantic(h))
1442 prep_compound_gigantic_page(page, huge_page_order(h));
1443 prep_new_huge_page(h, page, page_to_nid(page));
1449 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1452 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1456 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1458 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1459 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed);
1467 put_page(page); /* free it into the hugepage allocator */
1473 * Free huge page from pool from next node to free.
1474 * Attempt to keep persistent huge pages more or less
1475 * balanced over allowed nodes.
1476 * Called with hugetlb_lock locked.
1478 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1484 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1486 * If we're returning unused surplus pages, only examine
1487 * nodes with surplus pages.
1489 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1490 !list_empty(&h->hugepage_freelists[node])) {
1492 list_entry(h->hugepage_freelists[node].next,
1494 list_del(&page->lru);
1495 h->free_huge_pages--;
1496 h->free_huge_pages_node[node]--;
1498 h->surplus_huge_pages--;
1499 h->surplus_huge_pages_node[node]--;
1501 update_and_free_page(h, page);
1511 * Dissolve a given free hugepage into free buddy pages. This function does
1512 * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1513 * dissolution fails because a give page is not a free hugepage, or because
1514 * free hugepages are fully reserved.
1516 int dissolve_free_huge_page(struct page *page)
1520 spin_lock(&hugetlb_lock);
1521 if (PageHuge(page) && !page_count(page)) {
1522 struct page *head = compound_head(page);
1523 struct hstate *h = page_hstate(head);
1524 int nid = page_to_nid(head);
1525 if (h->free_huge_pages - h->resv_huge_pages == 0)
1528 * Move PageHWPoison flag from head page to the raw error page,
1529 * which makes any subpages rather than the error page reusable.
1531 if (PageHWPoison(head) && page != head) {
1532 SetPageHWPoison(page);
1533 ClearPageHWPoison(head);
1535 list_del(&head->lru);
1536 h->free_huge_pages--;
1537 h->free_huge_pages_node[nid]--;
1538 h->max_huge_pages--;
1539 update_and_free_page(h, head);
1543 spin_unlock(&hugetlb_lock);
1548 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1549 * make specified memory blocks removable from the system.
1550 * Note that this will dissolve a free gigantic hugepage completely, if any
1551 * part of it lies within the given range.
1552 * Also note that if dissolve_free_huge_page() returns with an error, all
1553 * free hugepages that were dissolved before that error are lost.
1555 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1561 if (!hugepages_supported())
1564 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1565 page = pfn_to_page(pfn);
1566 if (PageHuge(page) && !page_count(page)) {
1567 rc = dissolve_free_huge_page(page);
1577 * Allocates a fresh surplus page from the page allocator.
1579 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1580 int nid, nodemask_t *nmask)
1582 struct page *page = NULL;
1584 if (hstate_is_gigantic(h))
1587 spin_lock(&hugetlb_lock);
1588 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1590 spin_unlock(&hugetlb_lock);
1592 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1596 spin_lock(&hugetlb_lock);
1598 * We could have raced with the pool size change.
1599 * Double check that and simply deallocate the new page
1600 * if we would end up overcommiting the surpluses. Abuse
1601 * temporary page to workaround the nasty free_huge_page
1604 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1605 SetPageHugeTemporary(page);
1606 spin_unlock(&hugetlb_lock);
1610 h->surplus_huge_pages++;
1611 h->surplus_huge_pages_node[page_to_nid(page)]++;
1615 spin_unlock(&hugetlb_lock);
1620 struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1621 int nid, nodemask_t *nmask)
1625 if (hstate_is_gigantic(h))
1628 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1633 * We do not account these pages as surplus because they are only
1634 * temporary and will be released properly on the last reference
1636 SetPageHugeTemporary(page);
1642 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1645 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1646 struct vm_area_struct *vma, unsigned long addr)
1649 struct mempolicy *mpol;
1650 gfp_t gfp_mask = htlb_alloc_mask(h);
1652 nodemask_t *nodemask;
1654 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1655 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1656 mpol_cond_put(mpol);
1661 /* page migration callback function */
1662 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1664 gfp_t gfp_mask = htlb_alloc_mask(h);
1665 struct page *page = NULL;
1667 if (nid != NUMA_NO_NODE)
1668 gfp_mask |= __GFP_THISNODE;
1670 spin_lock(&hugetlb_lock);
1671 if (h->free_huge_pages - h->resv_huge_pages > 0)
1672 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1673 spin_unlock(&hugetlb_lock);
1676 page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1681 /* page migration callback function */
1682 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1685 gfp_t gfp_mask = htlb_alloc_mask(h);
1687 spin_lock(&hugetlb_lock);
1688 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1691 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1693 spin_unlock(&hugetlb_lock);
1697 spin_unlock(&hugetlb_lock);
1699 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1702 /* mempolicy aware migration callback */
1703 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1704 unsigned long address)
1706 struct mempolicy *mpol;
1707 nodemask_t *nodemask;
1712 gfp_mask = htlb_alloc_mask(h);
1713 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1714 page = alloc_huge_page_nodemask(h, node, nodemask);
1715 mpol_cond_put(mpol);
1721 * Increase the hugetlb pool such that it can accommodate a reservation
1724 static int gather_surplus_pages(struct hstate *h, int delta)
1726 struct list_head surplus_list;
1727 struct page *page, *tmp;
1729 int needed, allocated;
1730 bool alloc_ok = true;
1732 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1734 h->resv_huge_pages += delta;
1739 INIT_LIST_HEAD(&surplus_list);
1743 spin_unlock(&hugetlb_lock);
1744 for (i = 0; i < needed; i++) {
1745 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1746 NUMA_NO_NODE, NULL);
1751 list_add(&page->lru, &surplus_list);
1757 * After retaking hugetlb_lock, we need to recalculate 'needed'
1758 * because either resv_huge_pages or free_huge_pages may have changed.
1760 spin_lock(&hugetlb_lock);
1761 needed = (h->resv_huge_pages + delta) -
1762 (h->free_huge_pages + allocated);
1767 * We were not able to allocate enough pages to
1768 * satisfy the entire reservation so we free what
1769 * we've allocated so far.
1774 * The surplus_list now contains _at_least_ the number of extra pages
1775 * needed to accommodate the reservation. Add the appropriate number
1776 * of pages to the hugetlb pool and free the extras back to the buddy
1777 * allocator. Commit the entire reservation here to prevent another
1778 * process from stealing the pages as they are added to the pool but
1779 * before they are reserved.
1781 needed += allocated;
1782 h->resv_huge_pages += delta;
1785 /* Free the needed pages to the hugetlb pool */
1786 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1790 * This page is now managed by the hugetlb allocator and has
1791 * no users -- drop the buddy allocator's reference.
1793 put_page_testzero(page);
1794 VM_BUG_ON_PAGE(page_count(page), page);
1795 enqueue_huge_page(h, page);
1798 spin_unlock(&hugetlb_lock);
1800 /* Free unnecessary surplus pages to the buddy allocator */
1801 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1803 spin_lock(&hugetlb_lock);
1809 * This routine has two main purposes:
1810 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1811 * in unused_resv_pages. This corresponds to the prior adjustments made
1812 * to the associated reservation map.
1813 * 2) Free any unused surplus pages that may have been allocated to satisfy
1814 * the reservation. As many as unused_resv_pages may be freed.
1816 * Called with hugetlb_lock held. However, the lock could be dropped (and
1817 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1818 * we must make sure nobody else can claim pages we are in the process of
1819 * freeing. Do this by ensuring resv_huge_page always is greater than the
1820 * number of huge pages we plan to free when dropping the lock.
1822 static void return_unused_surplus_pages(struct hstate *h,
1823 unsigned long unused_resv_pages)
1825 unsigned long nr_pages;
1827 /* Cannot return gigantic pages currently */
1828 if (hstate_is_gigantic(h))
1832 * Part (or even all) of the reservation could have been backed
1833 * by pre-allocated pages. Only free surplus pages.
1835 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1838 * We want to release as many surplus pages as possible, spread
1839 * evenly across all nodes with memory. Iterate across these nodes
1840 * until we can no longer free unreserved surplus pages. This occurs
1841 * when the nodes with surplus pages have no free pages.
1842 * free_pool_huge_page() will balance the the freed pages across the
1843 * on-line nodes with memory and will handle the hstate accounting.
1845 * Note that we decrement resv_huge_pages as we free the pages. If
1846 * we drop the lock, resv_huge_pages will still be sufficiently large
1847 * to cover subsequent pages we may free.
1849 while (nr_pages--) {
1850 h->resv_huge_pages--;
1851 unused_resv_pages--;
1852 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1854 cond_resched_lock(&hugetlb_lock);
1858 /* Fully uncommit the reservation */
1859 h->resv_huge_pages -= unused_resv_pages;
1864 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1865 * are used by the huge page allocation routines to manage reservations.
1867 * vma_needs_reservation is called to determine if the huge page at addr
1868 * within the vma has an associated reservation. If a reservation is
1869 * needed, the value 1 is returned. The caller is then responsible for
1870 * managing the global reservation and subpool usage counts. After
1871 * the huge page has been allocated, vma_commit_reservation is called
1872 * to add the page to the reservation map. If the page allocation fails,
1873 * the reservation must be ended instead of committed. vma_end_reservation
1874 * is called in such cases.
1876 * In the normal case, vma_commit_reservation returns the same value
1877 * as the preceding vma_needs_reservation call. The only time this
1878 * is not the case is if a reserve map was changed between calls. It
1879 * is the responsibility of the caller to notice the difference and
1880 * take appropriate action.
1882 * vma_add_reservation is used in error paths where a reservation must
1883 * be restored when a newly allocated huge page must be freed. It is
1884 * to be called after calling vma_needs_reservation to determine if a
1885 * reservation exists.
1887 enum vma_resv_mode {
1893 static long __vma_reservation_common(struct hstate *h,
1894 struct vm_area_struct *vma, unsigned long addr,
1895 enum vma_resv_mode mode)
1897 struct resv_map *resv;
1901 resv = vma_resv_map(vma);
1905 idx = vma_hugecache_offset(h, vma, addr);
1907 case VMA_NEEDS_RESV:
1908 ret = region_chg(resv, idx, idx + 1);
1910 case VMA_COMMIT_RESV:
1911 ret = region_add(resv, idx, idx + 1);
1914 region_abort(resv, idx, idx + 1);
1918 if (vma->vm_flags & VM_MAYSHARE)
1919 ret = region_add(resv, idx, idx + 1);
1921 region_abort(resv, idx, idx + 1);
1922 ret = region_del(resv, idx, idx + 1);
1929 if (vma->vm_flags & VM_MAYSHARE)
1931 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1933 * In most cases, reserves always exist for private mappings.
1934 * However, a file associated with mapping could have been
1935 * hole punched or truncated after reserves were consumed.
1936 * As subsequent fault on such a range will not use reserves.
1937 * Subtle - The reserve map for private mappings has the
1938 * opposite meaning than that of shared mappings. If NO
1939 * entry is in the reserve map, it means a reservation exists.
1940 * If an entry exists in the reserve map, it means the
1941 * reservation has already been consumed. As a result, the
1942 * return value of this routine is the opposite of the
1943 * value returned from reserve map manipulation routines above.
1951 return ret < 0 ? ret : 0;
1954 static long vma_needs_reservation(struct hstate *h,
1955 struct vm_area_struct *vma, unsigned long addr)
1957 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1960 static long vma_commit_reservation(struct hstate *h,
1961 struct vm_area_struct *vma, unsigned long addr)
1963 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1966 static void vma_end_reservation(struct hstate *h,
1967 struct vm_area_struct *vma, unsigned long addr)
1969 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1972 static long vma_add_reservation(struct hstate *h,
1973 struct vm_area_struct *vma, unsigned long addr)
1975 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1979 * This routine is called to restore a reservation on error paths. In the
1980 * specific error paths, a huge page was allocated (via alloc_huge_page)
1981 * and is about to be freed. If a reservation for the page existed,
1982 * alloc_huge_page would have consumed the reservation and set PagePrivate
1983 * in the newly allocated page. When the page is freed via free_huge_page,
1984 * the global reservation count will be incremented if PagePrivate is set.
1985 * However, free_huge_page can not adjust the reserve map. Adjust the
1986 * reserve map here to be consistent with global reserve count adjustments
1987 * to be made by free_huge_page.
1989 static void restore_reserve_on_error(struct hstate *h,
1990 struct vm_area_struct *vma, unsigned long address,
1993 if (unlikely(PagePrivate(page))) {
1994 long rc = vma_needs_reservation(h, vma, address);
1996 if (unlikely(rc < 0)) {
1998 * Rare out of memory condition in reserve map
1999 * manipulation. Clear PagePrivate so that
2000 * global reserve count will not be incremented
2001 * by free_huge_page. This will make it appear
2002 * as though the reservation for this page was
2003 * consumed. This may prevent the task from
2004 * faulting in the page at a later time. This
2005 * is better than inconsistent global huge page
2006 * accounting of reserve counts.
2008 ClearPagePrivate(page);
2010 rc = vma_add_reservation(h, vma, address);
2011 if (unlikely(rc < 0))
2013 * See above comment about rare out of
2016 ClearPagePrivate(page);
2018 vma_end_reservation(h, vma, address);
2022 struct page *alloc_huge_page(struct vm_area_struct *vma,
2023 unsigned long addr, int avoid_reserve)
2025 struct hugepage_subpool *spool = subpool_vma(vma);
2026 struct hstate *h = hstate_vma(vma);
2028 long map_chg, map_commit;
2031 struct hugetlb_cgroup *h_cg;
2033 idx = hstate_index(h);
2035 * Examine the region/reserve map to determine if the process
2036 * has a reservation for the page to be allocated. A return
2037 * code of zero indicates a reservation exists (no change).
2039 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2041 return ERR_PTR(-ENOMEM);
2044 * Processes that did not create the mapping will have no
2045 * reserves as indicated by the region/reserve map. Check
2046 * that the allocation will not exceed the subpool limit.
2047 * Allocations for MAP_NORESERVE mappings also need to be
2048 * checked against any subpool limit.
2050 if (map_chg || avoid_reserve) {
2051 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2053 vma_end_reservation(h, vma, addr);
2054 return ERR_PTR(-ENOSPC);
2058 * Even though there was no reservation in the region/reserve
2059 * map, there could be reservations associated with the
2060 * subpool that can be used. This would be indicated if the
2061 * return value of hugepage_subpool_get_pages() is zero.
2062 * However, if avoid_reserve is specified we still avoid even
2063 * the subpool reservations.
2069 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2071 goto out_subpool_put;
2073 spin_lock(&hugetlb_lock);
2075 * glb_chg is passed to indicate whether or not a page must be taken
2076 * from the global free pool (global change). gbl_chg == 0 indicates
2077 * a reservation exists for the allocation.
2079 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2081 spin_unlock(&hugetlb_lock);
2082 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2084 goto out_uncharge_cgroup;
2085 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2086 SetPagePrivate(page);
2087 h->resv_huge_pages--;
2089 spin_lock(&hugetlb_lock);
2090 list_move(&page->lru, &h->hugepage_activelist);
2093 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2094 spin_unlock(&hugetlb_lock);
2096 set_page_private(page, (unsigned long)spool);
2098 map_commit = vma_commit_reservation(h, vma, addr);
2099 if (unlikely(map_chg > map_commit)) {
2101 * The page was added to the reservation map between
2102 * vma_needs_reservation and vma_commit_reservation.
2103 * This indicates a race with hugetlb_reserve_pages.
2104 * Adjust for the subpool count incremented above AND
2105 * in hugetlb_reserve_pages for the same page. Also,
2106 * the reservation count added in hugetlb_reserve_pages
2107 * no longer applies.
2111 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2112 hugetlb_acct_memory(h, -rsv_adjust);
2116 out_uncharge_cgroup:
2117 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2119 if (map_chg || avoid_reserve)
2120 hugepage_subpool_put_pages(spool, 1);
2121 vma_end_reservation(h, vma, addr);
2122 return ERR_PTR(-ENOSPC);
2125 int alloc_bootmem_huge_page(struct hstate *h)
2126 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2127 int __alloc_bootmem_huge_page(struct hstate *h)
2129 struct huge_bootmem_page *m;
2132 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2135 addr = memblock_alloc_try_nid_raw(
2136 huge_page_size(h), huge_page_size(h),
2137 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2140 * Use the beginning of the huge page to store the
2141 * huge_bootmem_page struct (until gather_bootmem
2142 * puts them into the mem_map).
2151 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2152 /* Put them into a private list first because mem_map is not up yet */
2153 INIT_LIST_HEAD(&m->list);
2154 list_add(&m->list, &huge_boot_pages);
2159 static void __init prep_compound_huge_page(struct page *page,
2162 if (unlikely(order > (MAX_ORDER - 1)))
2163 prep_compound_gigantic_page(page, order);
2165 prep_compound_page(page, order);
2168 /* Put bootmem huge pages into the standard lists after mem_map is up */
2169 static void __init gather_bootmem_prealloc(void)
2171 struct huge_bootmem_page *m;
2173 list_for_each_entry(m, &huge_boot_pages, list) {
2174 struct page *page = virt_to_page(m);
2175 struct hstate *h = m->hstate;
2177 WARN_ON(page_count(page) != 1);
2178 prep_compound_huge_page(page, h->order);
2179 WARN_ON(PageReserved(page));
2180 prep_new_huge_page(h, page, page_to_nid(page));
2181 put_page(page); /* free it into the hugepage allocator */
2184 * If we had gigantic hugepages allocated at boot time, we need
2185 * to restore the 'stolen' pages to totalram_pages in order to
2186 * fix confusing memory reports from free(1) and another
2187 * side-effects, like CommitLimit going negative.
2189 if (hstate_is_gigantic(h))
2190 adjust_managed_page_count(page, 1 << h->order);
2195 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2199 for (i = 0; i < h->max_huge_pages; ++i) {
2200 if (hstate_is_gigantic(h)) {
2201 if (!alloc_bootmem_huge_page(h))
2203 } else if (!alloc_pool_huge_page(h,
2204 &node_states[N_MEMORY]))
2208 if (i < h->max_huge_pages) {
2211 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2212 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2213 h->max_huge_pages, buf, i);
2214 h->max_huge_pages = i;
2218 static void __init hugetlb_init_hstates(void)
2222 for_each_hstate(h) {
2223 if (minimum_order > huge_page_order(h))
2224 minimum_order = huge_page_order(h);
2226 /* oversize hugepages were init'ed in early boot */
2227 if (!hstate_is_gigantic(h))
2228 hugetlb_hstate_alloc_pages(h);
2230 VM_BUG_ON(minimum_order == UINT_MAX);
2233 static void __init report_hugepages(void)
2237 for_each_hstate(h) {
2240 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2241 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2242 buf, h->free_huge_pages);
2246 #ifdef CONFIG_HIGHMEM
2247 static void try_to_free_low(struct hstate *h, unsigned long count,
2248 nodemask_t *nodes_allowed)
2252 if (hstate_is_gigantic(h))
2255 for_each_node_mask(i, *nodes_allowed) {
2256 struct page *page, *next;
2257 struct list_head *freel = &h->hugepage_freelists[i];
2258 list_for_each_entry_safe(page, next, freel, lru) {
2259 if (count >= h->nr_huge_pages)
2261 if (PageHighMem(page))
2263 list_del(&page->lru);
2264 update_and_free_page(h, page);
2265 h->free_huge_pages--;
2266 h->free_huge_pages_node[page_to_nid(page)]--;
2271 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2272 nodemask_t *nodes_allowed)
2278 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2279 * balanced by operating on them in a round-robin fashion.
2280 * Returns 1 if an adjustment was made.
2282 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2287 VM_BUG_ON(delta != -1 && delta != 1);
2290 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2291 if (h->surplus_huge_pages_node[node])
2295 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2296 if (h->surplus_huge_pages_node[node] <
2297 h->nr_huge_pages_node[node])
2304 h->surplus_huge_pages += delta;
2305 h->surplus_huge_pages_node[node] += delta;
2309 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2310 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2311 nodemask_t *nodes_allowed)
2313 unsigned long min_count, ret;
2315 spin_lock(&hugetlb_lock);
2318 * Check for a node specific request.
2319 * Changing node specific huge page count may require a corresponding
2320 * change to the global count. In any case, the passed node mask
2321 * (nodes_allowed) will restrict alloc/free to the specified node.
2323 if (nid != NUMA_NO_NODE) {
2324 unsigned long old_count = count;
2326 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2328 * User may have specified a large count value which caused the
2329 * above calculation to overflow. In this case, they wanted
2330 * to allocate as many huge pages as possible. Set count to
2331 * largest possible value to align with their intention.
2333 if (count < old_count)
2338 * Gigantic pages runtime allocation depend on the capability for large
2339 * page range allocation.
2340 * If the system does not provide this feature, return an error when
2341 * the user tries to allocate gigantic pages but let the user free the
2342 * boottime allocated gigantic pages.
2344 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2345 if (count > persistent_huge_pages(h)) {
2346 spin_unlock(&hugetlb_lock);
2349 /* Fall through to decrease pool */
2353 * Increase the pool size
2354 * First take pages out of surplus state. Then make up the
2355 * remaining difference by allocating fresh huge pages.
2357 * We might race with alloc_surplus_huge_page() here and be unable
2358 * to convert a surplus huge page to a normal huge page. That is
2359 * not critical, though, it just means the overall size of the
2360 * pool might be one hugepage larger than it needs to be, but
2361 * within all the constraints specified by the sysctls.
2363 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2364 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2368 while (count > persistent_huge_pages(h)) {
2370 * If this allocation races such that we no longer need the
2371 * page, free_huge_page will handle it by freeing the page
2372 * and reducing the surplus.
2374 spin_unlock(&hugetlb_lock);
2376 /* yield cpu to avoid soft lockup */
2379 ret = alloc_pool_huge_page(h, nodes_allowed);
2380 spin_lock(&hugetlb_lock);
2384 /* Bail for signals. Probably ctrl-c from user */
2385 if (signal_pending(current))
2390 * Decrease the pool size
2391 * First return free pages to the buddy allocator (being careful
2392 * to keep enough around to satisfy reservations). Then place
2393 * pages into surplus state as needed so the pool will shrink
2394 * to the desired size as pages become free.
2396 * By placing pages into the surplus state independent of the
2397 * overcommit value, we are allowing the surplus pool size to
2398 * exceed overcommit. There are few sane options here. Since
2399 * alloc_surplus_huge_page() is checking the global counter,
2400 * though, we'll note that we're not allowed to exceed surplus
2401 * and won't grow the pool anywhere else. Not until one of the
2402 * sysctls are changed, or the surplus pages go out of use.
2404 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2405 min_count = max(count, min_count);
2406 try_to_free_low(h, min_count, nodes_allowed);
2407 while (min_count < persistent_huge_pages(h)) {
2408 if (!free_pool_huge_page(h, nodes_allowed, 0))
2410 cond_resched_lock(&hugetlb_lock);
2412 while (count < persistent_huge_pages(h)) {
2413 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2417 h->max_huge_pages = persistent_huge_pages(h);
2418 spin_unlock(&hugetlb_lock);
2423 #define HSTATE_ATTR_RO(_name) \
2424 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2426 #define HSTATE_ATTR(_name) \
2427 static struct kobj_attribute _name##_attr = \
2428 __ATTR(_name, 0644, _name##_show, _name##_store)
2430 static struct kobject *hugepages_kobj;
2431 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2433 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2435 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2439 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2440 if (hstate_kobjs[i] == kobj) {
2442 *nidp = NUMA_NO_NODE;
2446 return kobj_to_node_hstate(kobj, nidp);
2449 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2450 struct kobj_attribute *attr, char *buf)
2453 unsigned long nr_huge_pages;
2456 h = kobj_to_hstate(kobj, &nid);
2457 if (nid == NUMA_NO_NODE)
2458 nr_huge_pages = h->nr_huge_pages;
2460 nr_huge_pages = h->nr_huge_pages_node[nid];
2462 return sprintf(buf, "%lu\n", nr_huge_pages);
2465 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2466 struct hstate *h, int nid,
2467 unsigned long count, size_t len)
2470 nodemask_t nodes_allowed, *n_mask;
2472 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2475 if (nid == NUMA_NO_NODE) {
2477 * global hstate attribute
2479 if (!(obey_mempolicy &&
2480 init_nodemask_of_mempolicy(&nodes_allowed)))
2481 n_mask = &node_states[N_MEMORY];
2483 n_mask = &nodes_allowed;
2486 * Node specific request. count adjustment happens in
2487 * set_max_huge_pages() after acquiring hugetlb_lock.
2489 init_nodemask_of_node(&nodes_allowed, nid);
2490 n_mask = &nodes_allowed;
2493 err = set_max_huge_pages(h, count, nid, n_mask);
2495 return err ? err : len;
2498 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2499 struct kobject *kobj, const char *buf,
2503 unsigned long count;
2507 err = kstrtoul(buf, 10, &count);
2511 h = kobj_to_hstate(kobj, &nid);
2512 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2515 static ssize_t nr_hugepages_show(struct kobject *kobj,
2516 struct kobj_attribute *attr, char *buf)
2518 return nr_hugepages_show_common(kobj, attr, buf);
2521 static ssize_t nr_hugepages_store(struct kobject *kobj,
2522 struct kobj_attribute *attr, const char *buf, size_t len)
2524 return nr_hugepages_store_common(false, kobj, buf, len);
2526 HSTATE_ATTR(nr_hugepages);
2531 * hstate attribute for optionally mempolicy-based constraint on persistent
2532 * huge page alloc/free.
2534 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2535 struct kobj_attribute *attr, char *buf)
2537 return nr_hugepages_show_common(kobj, attr, buf);
2540 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2541 struct kobj_attribute *attr, const char *buf, size_t len)
2543 return nr_hugepages_store_common(true, kobj, buf, len);
2545 HSTATE_ATTR(nr_hugepages_mempolicy);
2549 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2550 struct kobj_attribute *attr, char *buf)
2552 struct hstate *h = kobj_to_hstate(kobj, NULL);
2553 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2556 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2557 struct kobj_attribute *attr, const char *buf, size_t count)
2560 unsigned long input;
2561 struct hstate *h = kobj_to_hstate(kobj, NULL);
2563 if (hstate_is_gigantic(h))
2566 err = kstrtoul(buf, 10, &input);
2570 spin_lock(&hugetlb_lock);
2571 h->nr_overcommit_huge_pages = input;
2572 spin_unlock(&hugetlb_lock);
2576 HSTATE_ATTR(nr_overcommit_hugepages);
2578 static ssize_t free_hugepages_show(struct kobject *kobj,
2579 struct kobj_attribute *attr, char *buf)
2582 unsigned long free_huge_pages;
2585 h = kobj_to_hstate(kobj, &nid);
2586 if (nid == NUMA_NO_NODE)
2587 free_huge_pages = h->free_huge_pages;
2589 free_huge_pages = h->free_huge_pages_node[nid];
2591 return sprintf(buf, "%lu\n", free_huge_pages);
2593 HSTATE_ATTR_RO(free_hugepages);
2595 static ssize_t resv_hugepages_show(struct kobject *kobj,
2596 struct kobj_attribute *attr, char *buf)
2598 struct hstate *h = kobj_to_hstate(kobj, NULL);
2599 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2601 HSTATE_ATTR_RO(resv_hugepages);
2603 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2604 struct kobj_attribute *attr, char *buf)
2607 unsigned long surplus_huge_pages;
2610 h = kobj_to_hstate(kobj, &nid);
2611 if (nid == NUMA_NO_NODE)
2612 surplus_huge_pages = h->surplus_huge_pages;
2614 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2616 return sprintf(buf, "%lu\n", surplus_huge_pages);
2618 HSTATE_ATTR_RO(surplus_hugepages);
2620 static struct attribute *hstate_attrs[] = {
2621 &nr_hugepages_attr.attr,
2622 &nr_overcommit_hugepages_attr.attr,
2623 &free_hugepages_attr.attr,
2624 &resv_hugepages_attr.attr,
2625 &surplus_hugepages_attr.attr,
2627 &nr_hugepages_mempolicy_attr.attr,
2632 static const struct attribute_group hstate_attr_group = {
2633 .attrs = hstate_attrs,
2636 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2637 struct kobject **hstate_kobjs,
2638 const struct attribute_group *hstate_attr_group)
2641 int hi = hstate_index(h);
2643 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2644 if (!hstate_kobjs[hi])
2647 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2649 kobject_put(hstate_kobjs[hi]);
2654 static void __init hugetlb_sysfs_init(void)
2659 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2660 if (!hugepages_kobj)
2663 for_each_hstate(h) {
2664 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2665 hstate_kobjs, &hstate_attr_group);
2667 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2674 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2675 * with node devices in node_devices[] using a parallel array. The array
2676 * index of a node device or _hstate == node id.
2677 * This is here to avoid any static dependency of the node device driver, in
2678 * the base kernel, on the hugetlb module.
2680 struct node_hstate {
2681 struct kobject *hugepages_kobj;
2682 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2684 static struct node_hstate node_hstates[MAX_NUMNODES];
2687 * A subset of global hstate attributes for node devices
2689 static struct attribute *per_node_hstate_attrs[] = {
2690 &nr_hugepages_attr.attr,
2691 &free_hugepages_attr.attr,
2692 &surplus_hugepages_attr.attr,
2696 static const struct attribute_group per_node_hstate_attr_group = {
2697 .attrs = per_node_hstate_attrs,
2701 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2702 * Returns node id via non-NULL nidp.
2704 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2708 for (nid = 0; nid < nr_node_ids; nid++) {
2709 struct node_hstate *nhs = &node_hstates[nid];
2711 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2712 if (nhs->hstate_kobjs[i] == kobj) {
2724 * Unregister hstate attributes from a single node device.
2725 * No-op if no hstate attributes attached.
2727 static void hugetlb_unregister_node(struct node *node)
2730 struct node_hstate *nhs = &node_hstates[node->dev.id];
2732 if (!nhs->hugepages_kobj)
2733 return; /* no hstate attributes */
2735 for_each_hstate(h) {
2736 int idx = hstate_index(h);
2737 if (nhs->hstate_kobjs[idx]) {
2738 kobject_put(nhs->hstate_kobjs[idx]);
2739 nhs->hstate_kobjs[idx] = NULL;
2743 kobject_put(nhs->hugepages_kobj);
2744 nhs->hugepages_kobj = NULL;
2749 * Register hstate attributes for a single node device.
2750 * No-op if attributes already registered.
2752 static void hugetlb_register_node(struct node *node)
2755 struct node_hstate *nhs = &node_hstates[node->dev.id];
2758 if (nhs->hugepages_kobj)
2759 return; /* already allocated */
2761 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2763 if (!nhs->hugepages_kobj)
2766 for_each_hstate(h) {
2767 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2769 &per_node_hstate_attr_group);
2771 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2772 h->name, node->dev.id);
2773 hugetlb_unregister_node(node);
2780 * hugetlb init time: register hstate attributes for all registered node
2781 * devices of nodes that have memory. All on-line nodes should have
2782 * registered their associated device by this time.
2784 static void __init hugetlb_register_all_nodes(void)
2788 for_each_node_state(nid, N_MEMORY) {
2789 struct node *node = node_devices[nid];
2790 if (node->dev.id == nid)
2791 hugetlb_register_node(node);
2795 * Let the node device driver know we're here so it can
2796 * [un]register hstate attributes on node hotplug.
2798 register_hugetlbfs_with_node(hugetlb_register_node,
2799 hugetlb_unregister_node);
2801 #else /* !CONFIG_NUMA */
2803 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2811 static void hugetlb_register_all_nodes(void) { }
2815 static int __init hugetlb_init(void)
2819 if (!hugepages_supported())
2822 if (!size_to_hstate(default_hstate_size)) {
2823 if (default_hstate_size != 0) {
2824 pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2825 default_hstate_size, HPAGE_SIZE);
2828 default_hstate_size = HPAGE_SIZE;
2829 if (!size_to_hstate(default_hstate_size))
2830 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2832 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2833 if (default_hstate_max_huge_pages) {
2834 if (!default_hstate.max_huge_pages)
2835 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2838 hugetlb_init_hstates();
2839 gather_bootmem_prealloc();
2842 hugetlb_sysfs_init();
2843 hugetlb_register_all_nodes();
2844 hugetlb_cgroup_file_init();
2847 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2849 num_fault_mutexes = 1;
2851 hugetlb_fault_mutex_table =
2852 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
2854 BUG_ON(!hugetlb_fault_mutex_table);
2856 for (i = 0; i < num_fault_mutexes; i++)
2857 mutex_init(&hugetlb_fault_mutex_table[i]);
2860 subsys_initcall(hugetlb_init);
2862 /* Should be called on processing a hugepagesz=... option */
2863 void __init hugetlb_bad_size(void)
2865 parsed_valid_hugepagesz = false;
2868 void __init hugetlb_add_hstate(unsigned int order)
2873 if (size_to_hstate(PAGE_SIZE << order)) {
2874 pr_warn("hugepagesz= specified twice, ignoring\n");
2877 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2879 h = &hstates[hugetlb_max_hstate++];
2881 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2882 h->nr_huge_pages = 0;
2883 h->free_huge_pages = 0;
2884 for (i = 0; i < MAX_NUMNODES; ++i)
2885 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2886 INIT_LIST_HEAD(&h->hugepage_activelist);
2887 h->next_nid_to_alloc = first_memory_node;
2888 h->next_nid_to_free = first_memory_node;
2889 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2890 huge_page_size(h)/1024);
2895 static int __init hugetlb_nrpages_setup(char *s)
2898 static unsigned long *last_mhp;
2900 if (!parsed_valid_hugepagesz) {
2901 pr_warn("hugepages = %s preceded by "
2902 "an unsupported hugepagesz, ignoring\n", s);
2903 parsed_valid_hugepagesz = true;
2907 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2908 * so this hugepages= parameter goes to the "default hstate".
2910 else if (!hugetlb_max_hstate)
2911 mhp = &default_hstate_max_huge_pages;
2913 mhp = &parsed_hstate->max_huge_pages;
2915 if (mhp == last_mhp) {
2916 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2920 if (sscanf(s, "%lu", mhp) <= 0)
2924 * Global state is always initialized later in hugetlb_init.
2925 * But we need to allocate >= MAX_ORDER hstates here early to still
2926 * use the bootmem allocator.
2928 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2929 hugetlb_hstate_alloc_pages(parsed_hstate);
2935 __setup("hugepages=", hugetlb_nrpages_setup);
2937 static int __init hugetlb_default_setup(char *s)
2939 default_hstate_size = memparse(s, &s);
2942 __setup("default_hugepagesz=", hugetlb_default_setup);
2944 static unsigned int cpuset_mems_nr(unsigned int *array)
2947 unsigned int nr = 0;
2949 for_each_node_mask(node, cpuset_current_mems_allowed)
2955 #ifdef CONFIG_SYSCTL
2956 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2957 struct ctl_table *table, int write,
2958 void __user *buffer, size_t *length, loff_t *ppos)
2960 struct hstate *h = &default_hstate;
2961 unsigned long tmp = h->max_huge_pages;
2964 if (!hugepages_supported())
2968 table->maxlen = sizeof(unsigned long);
2969 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2974 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2975 NUMA_NO_NODE, tmp, *length);
2980 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2981 void __user *buffer, size_t *length, loff_t *ppos)
2984 return hugetlb_sysctl_handler_common(false, table, write,
2985 buffer, length, ppos);
2989 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2990 void __user *buffer, size_t *length, loff_t *ppos)
2992 return hugetlb_sysctl_handler_common(true, table, write,
2993 buffer, length, ppos);
2995 #endif /* CONFIG_NUMA */
2997 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2998 void __user *buffer,
2999 size_t *length, loff_t *ppos)
3001 struct hstate *h = &default_hstate;
3005 if (!hugepages_supported())
3008 tmp = h->nr_overcommit_huge_pages;
3010 if (write && hstate_is_gigantic(h))
3014 table->maxlen = sizeof(unsigned long);
3015 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3020 spin_lock(&hugetlb_lock);
3021 h->nr_overcommit_huge_pages = tmp;
3022 spin_unlock(&hugetlb_lock);
3028 #endif /* CONFIG_SYSCTL */
3030 void hugetlb_report_meminfo(struct seq_file *m)
3033 unsigned long total = 0;
3035 if (!hugepages_supported())
3038 for_each_hstate(h) {
3039 unsigned long count = h->nr_huge_pages;
3041 total += (PAGE_SIZE << huge_page_order(h)) * count;
3043 if (h == &default_hstate)
3045 "HugePages_Total: %5lu\n"
3046 "HugePages_Free: %5lu\n"
3047 "HugePages_Rsvd: %5lu\n"
3048 "HugePages_Surp: %5lu\n"
3049 "Hugepagesize: %8lu kB\n",
3053 h->surplus_huge_pages,
3054 (PAGE_SIZE << huge_page_order(h)) / 1024);
3057 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3060 int hugetlb_report_node_meminfo(int nid, char *buf)
3062 struct hstate *h = &default_hstate;
3063 if (!hugepages_supported())
3066 "Node %d HugePages_Total: %5u\n"
3067 "Node %d HugePages_Free: %5u\n"
3068 "Node %d HugePages_Surp: %5u\n",
3069 nid, h->nr_huge_pages_node[nid],
3070 nid, h->free_huge_pages_node[nid],
3071 nid, h->surplus_huge_pages_node[nid]);
3074 void hugetlb_show_meminfo(void)
3079 if (!hugepages_supported())
3082 for_each_node_state(nid, N_MEMORY)
3084 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3086 h->nr_huge_pages_node[nid],
3087 h->free_huge_pages_node[nid],
3088 h->surplus_huge_pages_node[nid],
3089 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3092 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3094 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3095 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3098 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3099 unsigned long hugetlb_total_pages(void)
3102 unsigned long nr_total_pages = 0;
3105 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3106 return nr_total_pages;
3109 static int hugetlb_acct_memory(struct hstate *h, long delta)
3113 spin_lock(&hugetlb_lock);
3115 * When cpuset is configured, it breaks the strict hugetlb page
3116 * reservation as the accounting is done on a global variable. Such
3117 * reservation is completely rubbish in the presence of cpuset because
3118 * the reservation is not checked against page availability for the
3119 * current cpuset. Application can still potentially OOM'ed by kernel
3120 * with lack of free htlb page in cpuset that the task is in.
3121 * Attempt to enforce strict accounting with cpuset is almost
3122 * impossible (or too ugly) because cpuset is too fluid that
3123 * task or memory node can be dynamically moved between cpusets.
3125 * The change of semantics for shared hugetlb mapping with cpuset is
3126 * undesirable. However, in order to preserve some of the semantics,
3127 * we fall back to check against current free page availability as
3128 * a best attempt and hopefully to minimize the impact of changing
3129 * semantics that cpuset has.
3132 if (gather_surplus_pages(h, delta) < 0)
3135 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3136 return_unused_surplus_pages(h, delta);
3143 return_unused_surplus_pages(h, (unsigned long) -delta);
3146 spin_unlock(&hugetlb_lock);
3150 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3152 struct resv_map *resv = vma_resv_map(vma);
3155 * This new VMA should share its siblings reservation map if present.
3156 * The VMA will only ever have a valid reservation map pointer where
3157 * it is being copied for another still existing VMA. As that VMA
3158 * has a reference to the reservation map it cannot disappear until
3159 * after this open call completes. It is therefore safe to take a
3160 * new reference here without additional locking.
3162 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3163 kref_get(&resv->refs);
3166 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3168 struct hstate *h = hstate_vma(vma);
3169 struct resv_map *resv = vma_resv_map(vma);
3170 struct hugepage_subpool *spool = subpool_vma(vma);
3171 unsigned long reserve, start, end;
3174 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3177 start = vma_hugecache_offset(h, vma, vma->vm_start);
3178 end = vma_hugecache_offset(h, vma, vma->vm_end);
3180 reserve = (end - start) - region_count(resv, start, end);
3182 kref_put(&resv->refs, resv_map_release);
3186 * Decrement reserve counts. The global reserve count may be
3187 * adjusted if the subpool has a minimum size.
3189 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3190 hugetlb_acct_memory(h, -gbl_reserve);
3194 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3196 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3201 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3203 struct hstate *hstate = hstate_vma(vma);
3205 return 1UL << huge_page_shift(hstate);
3209 * We cannot handle pagefaults against hugetlb pages at all. They cause
3210 * handle_mm_fault() to try to instantiate regular-sized pages in the
3211 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3214 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3221 * When a new function is introduced to vm_operations_struct and added
3222 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3223 * This is because under System V memory model, mappings created via
3224 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3225 * their original vm_ops are overwritten with shm_vm_ops.
3227 const struct vm_operations_struct hugetlb_vm_ops = {
3228 .fault = hugetlb_vm_op_fault,
3229 .open = hugetlb_vm_op_open,
3230 .close = hugetlb_vm_op_close,
3231 .split = hugetlb_vm_op_split,
3232 .pagesize = hugetlb_vm_op_pagesize,
3235 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3241 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3242 vma->vm_page_prot)));
3244 entry = huge_pte_wrprotect(mk_huge_pte(page,
3245 vma->vm_page_prot));
3247 entry = pte_mkyoung(entry);
3248 entry = pte_mkhuge(entry);
3249 entry = arch_make_huge_pte(entry, vma, page, writable);
3254 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3255 unsigned long address, pte_t *ptep)
3259 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3260 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3261 update_mmu_cache(vma, address, ptep);
3264 bool is_hugetlb_entry_migration(pte_t pte)
3268 if (huge_pte_none(pte) || pte_present(pte))
3270 swp = pte_to_swp_entry(pte);
3271 if (non_swap_entry(swp) && is_migration_entry(swp))
3277 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3281 if (huge_pte_none(pte) || pte_present(pte))
3283 swp = pte_to_swp_entry(pte);
3284 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3290 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3291 struct vm_area_struct *vma)
3293 pte_t *src_pte, *dst_pte, entry, dst_entry;
3294 struct page *ptepage;
3297 struct hstate *h = hstate_vma(vma);
3298 unsigned long sz = huge_page_size(h);
3299 struct mmu_notifier_range range;
3302 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3305 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3308 mmu_notifier_invalidate_range_start(&range);
3311 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3312 spinlock_t *src_ptl, *dst_ptl;
3313 src_pte = huge_pte_offset(src, addr, sz);
3316 dst_pte = huge_pte_alloc(dst, addr, sz);
3323 * If the pagetables are shared don't copy or take references.
3324 * dst_pte == src_pte is the common case of src/dest sharing.
3326 * However, src could have 'unshared' and dst shares with
3327 * another vma. If dst_pte !none, this implies sharing.
3328 * Check here before taking page table lock, and once again
3329 * after taking the lock below.
3331 dst_entry = huge_ptep_get(dst_pte);
3332 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3335 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3336 src_ptl = huge_pte_lockptr(h, src, src_pte);
3337 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3338 entry = huge_ptep_get(src_pte);
3339 dst_entry = huge_ptep_get(dst_pte);
3340 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3342 * Skip if src entry none. Also, skip in the
3343 * unlikely case dst entry !none as this implies
3344 * sharing with another vma.
3347 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3348 is_hugetlb_entry_hwpoisoned(entry))) {
3349 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3351 if (is_write_migration_entry(swp_entry) && cow) {
3353 * COW mappings require pages in both
3354 * parent and child to be set to read.
3356 make_migration_entry_read(&swp_entry);
3357 entry = swp_entry_to_pte(swp_entry);
3358 set_huge_swap_pte_at(src, addr, src_pte,
3361 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3365 * No need to notify as we are downgrading page
3366 * table protection not changing it to point
3369 * See Documentation/vm/mmu_notifier.rst
3371 huge_ptep_set_wrprotect(src, addr, src_pte);
3373 entry = huge_ptep_get(src_pte);
3374 ptepage = pte_page(entry);
3376 page_dup_rmap(ptepage, true);
3377 set_huge_pte_at(dst, addr, dst_pte, entry);
3378 hugetlb_count_add(pages_per_huge_page(h), dst);
3380 spin_unlock(src_ptl);
3381 spin_unlock(dst_ptl);
3385 mmu_notifier_invalidate_range_end(&range);
3390 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3391 unsigned long start, unsigned long end,
3392 struct page *ref_page)
3394 struct mm_struct *mm = vma->vm_mm;
3395 unsigned long address;
3400 struct hstate *h = hstate_vma(vma);
3401 unsigned long sz = huge_page_size(h);
3402 struct mmu_notifier_range range;
3404 WARN_ON(!is_vm_hugetlb_page(vma));
3405 BUG_ON(start & ~huge_page_mask(h));
3406 BUG_ON(end & ~huge_page_mask(h));
3409 * This is a hugetlb vma, all the pte entries should point
3412 tlb_change_page_size(tlb, sz);
3413 tlb_start_vma(tlb, vma);
3416 * If sharing possible, alert mmu notifiers of worst case.
3418 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3420 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3421 mmu_notifier_invalidate_range_start(&range);
3423 for (; address < end; address += sz) {
3424 ptep = huge_pte_offset(mm, address, sz);
3428 ptl = huge_pte_lock(h, mm, ptep);
3429 if (huge_pmd_unshare(mm, &address, ptep)) {
3432 * We just unmapped a page of PMDs by clearing a PUD.
3433 * The caller's TLB flush range should cover this area.
3438 pte = huge_ptep_get(ptep);
3439 if (huge_pte_none(pte)) {
3445 * Migrating hugepage or HWPoisoned hugepage is already
3446 * unmapped and its refcount is dropped, so just clear pte here.
3448 if (unlikely(!pte_present(pte))) {
3449 huge_pte_clear(mm, address, ptep, sz);
3454 page = pte_page(pte);
3456 * If a reference page is supplied, it is because a specific
3457 * page is being unmapped, not a range. Ensure the page we
3458 * are about to unmap is the actual page of interest.
3461 if (page != ref_page) {
3466 * Mark the VMA as having unmapped its page so that
3467 * future faults in this VMA will fail rather than
3468 * looking like data was lost
3470 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3473 pte = huge_ptep_get_and_clear(mm, address, ptep);
3474 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3475 if (huge_pte_dirty(pte))
3476 set_page_dirty(page);
3478 hugetlb_count_sub(pages_per_huge_page(h), mm);
3479 page_remove_rmap(page, true);
3482 tlb_remove_page_size(tlb, page, huge_page_size(h));
3484 * Bail out after unmapping reference page if supplied
3489 mmu_notifier_invalidate_range_end(&range);
3490 tlb_end_vma(tlb, vma);
3493 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3494 struct vm_area_struct *vma, unsigned long start,
3495 unsigned long end, struct page *ref_page)
3497 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3500 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3501 * test will fail on a vma being torn down, and not grab a page table
3502 * on its way out. We're lucky that the flag has such an appropriate
3503 * name, and can in fact be safely cleared here. We could clear it
3504 * before the __unmap_hugepage_range above, but all that's necessary
3505 * is to clear it before releasing the i_mmap_rwsem. This works
3506 * because in the context this is called, the VMA is about to be
3507 * destroyed and the i_mmap_rwsem is held.
3509 vma->vm_flags &= ~VM_MAYSHARE;
3512 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3513 unsigned long end, struct page *ref_page)
3515 struct mm_struct *mm;
3516 struct mmu_gather tlb;
3517 unsigned long tlb_start = start;
3518 unsigned long tlb_end = end;
3521 * If shared PMDs were possibly used within this vma range, adjust
3522 * start/end for worst case tlb flushing.
3523 * Note that we can not be sure if PMDs are shared until we try to
3524 * unmap pages. However, we want to make sure TLB flushing covers
3525 * the largest possible range.
3527 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
3531 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
3532 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3533 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
3537 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3538 * mappping it owns the reserve page for. The intention is to unmap the page
3539 * from other VMAs and let the children be SIGKILLed if they are faulting the
3542 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3543 struct page *page, unsigned long address)
3545 struct hstate *h = hstate_vma(vma);
3546 struct vm_area_struct *iter_vma;
3547 struct address_space *mapping;
3551 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3552 * from page cache lookup which is in HPAGE_SIZE units.
3554 address = address & huge_page_mask(h);
3555 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3557 mapping = vma->vm_file->f_mapping;
3560 * Take the mapping lock for the duration of the table walk. As
3561 * this mapping should be shared between all the VMAs,
3562 * __unmap_hugepage_range() is called as the lock is already held
3564 i_mmap_lock_write(mapping);
3565 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3566 /* Do not unmap the current VMA */
3567 if (iter_vma == vma)
3571 * Shared VMAs have their own reserves and do not affect
3572 * MAP_PRIVATE accounting but it is possible that a shared
3573 * VMA is using the same page so check and skip such VMAs.
3575 if (iter_vma->vm_flags & VM_MAYSHARE)
3579 * Unmap the page from other VMAs without their own reserves.
3580 * They get marked to be SIGKILLed if they fault in these
3581 * areas. This is because a future no-page fault on this VMA
3582 * could insert a zeroed page instead of the data existing
3583 * from the time of fork. This would look like data corruption
3585 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3586 unmap_hugepage_range(iter_vma, address,
3587 address + huge_page_size(h), page);
3589 i_mmap_unlock_write(mapping);
3593 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3594 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3595 * cannot race with other handlers or page migration.
3596 * Keep the pte_same checks anyway to make transition from the mutex easier.
3598 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3599 unsigned long address, pte_t *ptep,
3600 struct page *pagecache_page, spinlock_t *ptl)
3603 struct hstate *h = hstate_vma(vma);
3604 struct page *old_page, *new_page;
3605 int outside_reserve = 0;
3607 unsigned long haddr = address & huge_page_mask(h);
3608 struct mmu_notifier_range range;
3610 pte = huge_ptep_get(ptep);
3611 old_page = pte_page(pte);
3614 /* If no-one else is actually using this page, avoid the copy
3615 * and just make the page writable */
3616 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3617 page_move_anon_rmap(old_page, vma);
3618 set_huge_ptep_writable(vma, haddr, ptep);
3623 * If the process that created a MAP_PRIVATE mapping is about to
3624 * perform a COW due to a shared page count, attempt to satisfy
3625 * the allocation without using the existing reserves. The pagecache
3626 * page is used to determine if the reserve at this address was
3627 * consumed or not. If reserves were used, a partial faulted mapping
3628 * at the time of fork() could consume its reserves on COW instead
3629 * of the full address range.
3631 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3632 old_page != pagecache_page)
3633 outside_reserve = 1;
3638 * Drop page table lock as buddy allocator may be called. It will
3639 * be acquired again before returning to the caller, as expected.
3642 new_page = alloc_huge_page(vma, haddr, outside_reserve);
3644 if (IS_ERR(new_page)) {
3646 * If a process owning a MAP_PRIVATE mapping fails to COW,
3647 * it is due to references held by a child and an insufficient
3648 * huge page pool. To guarantee the original mappers
3649 * reliability, unmap the page from child processes. The child
3650 * may get SIGKILLed if it later faults.
3652 if (outside_reserve) {
3654 BUG_ON(huge_pte_none(pte));
3655 unmap_ref_private(mm, vma, old_page, haddr);
3656 BUG_ON(huge_pte_none(pte));
3658 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3660 pte_same(huge_ptep_get(ptep), pte)))
3661 goto retry_avoidcopy;
3663 * race occurs while re-acquiring page table
3664 * lock, and our job is done.
3669 ret = vmf_error(PTR_ERR(new_page));
3670 goto out_release_old;
3674 * When the original hugepage is shared one, it does not have
3675 * anon_vma prepared.
3677 if (unlikely(anon_vma_prepare(vma))) {
3679 goto out_release_all;
3682 copy_user_huge_page(new_page, old_page, address, vma,
3683 pages_per_huge_page(h));
3684 __SetPageUptodate(new_page);
3686 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
3687 haddr + huge_page_size(h));
3688 mmu_notifier_invalidate_range_start(&range);
3691 * Retake the page table lock to check for racing updates
3692 * before the page tables are altered
3695 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3696 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3697 ClearPagePrivate(new_page);
3700 huge_ptep_clear_flush(vma, haddr, ptep);
3701 mmu_notifier_invalidate_range(mm, range.start, range.end);
3702 set_huge_pte_at(mm, haddr, ptep,
3703 make_huge_pte(vma, new_page, 1));
3704 page_remove_rmap(old_page, true);
3705 hugepage_add_new_anon_rmap(new_page, vma, haddr);
3706 set_page_huge_active(new_page);
3707 /* Make the old page be freed below */
3708 new_page = old_page;
3711 mmu_notifier_invalidate_range_end(&range);
3713 restore_reserve_on_error(h, vma, haddr, new_page);
3718 spin_lock(ptl); /* Caller expects lock to be held */
3722 /* Return the pagecache page at a given address within a VMA */
3723 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3724 struct vm_area_struct *vma, unsigned long address)
3726 struct address_space *mapping;
3729 mapping = vma->vm_file->f_mapping;
3730 idx = vma_hugecache_offset(h, vma, address);
3732 return find_lock_page(mapping, idx);
3736 * Return whether there is a pagecache page to back given address within VMA.
3737 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3739 static bool hugetlbfs_pagecache_present(struct hstate *h,
3740 struct vm_area_struct *vma, unsigned long address)
3742 struct address_space *mapping;
3746 mapping = vma->vm_file->f_mapping;
3747 idx = vma_hugecache_offset(h, vma, address);
3749 page = find_get_page(mapping, idx);
3752 return page != NULL;
3755 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3758 struct inode *inode = mapping->host;
3759 struct hstate *h = hstate_inode(inode);
3760 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3764 ClearPagePrivate(page);
3767 * set page dirty so that it will not be removed from cache/file
3768 * by non-hugetlbfs specific code paths.
3770 set_page_dirty(page);
3772 spin_lock(&inode->i_lock);
3773 inode->i_blocks += blocks_per_huge_page(h);
3774 spin_unlock(&inode->i_lock);
3778 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
3779 struct vm_area_struct *vma,
3780 struct address_space *mapping, pgoff_t idx,
3781 unsigned long address, pte_t *ptep, unsigned int flags)
3783 struct hstate *h = hstate_vma(vma);
3784 vm_fault_t ret = VM_FAULT_SIGBUS;
3790 unsigned long haddr = address & huge_page_mask(h);
3791 bool new_page = false;
3794 * Currently, we are forced to kill the process in the event the
3795 * original mapper has unmapped pages from the child due to a failed
3796 * COW. Warn that such a situation has occurred as it may not be obvious
3798 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3799 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3805 * Use page lock to guard against racing truncation
3806 * before we get page_table_lock.
3809 page = find_lock_page(mapping, idx);
3811 size = i_size_read(mapping->host) >> huge_page_shift(h);
3816 * Check for page in userfault range
3818 if (userfaultfd_missing(vma)) {
3820 struct vm_fault vmf = {
3825 * Hard to debug if it ends up being
3826 * used by a callee that assumes
3827 * something about the other
3828 * uninitialized fields... same as in
3834 * hugetlb_fault_mutex must be dropped before
3835 * handling userfault. Reacquire after handling
3836 * fault to make calling code simpler.
3838 hash = hugetlb_fault_mutex_hash(h, mapping, idx, haddr);
3839 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3840 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3841 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3845 page = alloc_huge_page(vma, haddr, 0);
3847 ret = vmf_error(PTR_ERR(page));
3850 clear_huge_page(page, address, pages_per_huge_page(h));
3851 __SetPageUptodate(page);
3854 if (vma->vm_flags & VM_MAYSHARE) {
3855 int err = huge_add_to_page_cache(page, mapping, idx);
3864 if (unlikely(anon_vma_prepare(vma))) {
3866 goto backout_unlocked;
3872 * If memory error occurs between mmap() and fault, some process
3873 * don't have hwpoisoned swap entry for errored virtual address.
3874 * So we need to block hugepage fault by PG_hwpoison bit check.
3876 if (unlikely(PageHWPoison(page))) {
3877 ret = VM_FAULT_HWPOISON |
3878 VM_FAULT_SET_HINDEX(hstate_index(h));
3879 goto backout_unlocked;
3884 * If we are going to COW a private mapping later, we examine the
3885 * pending reservations for this page now. This will ensure that
3886 * any allocations necessary to record that reservation occur outside
3889 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3890 if (vma_needs_reservation(h, vma, haddr) < 0) {
3892 goto backout_unlocked;
3894 /* Just decrements count, does not deallocate */
3895 vma_end_reservation(h, vma, haddr);
3898 ptl = huge_pte_lock(h, mm, ptep);
3899 size = i_size_read(mapping->host) >> huge_page_shift(h);
3904 if (!huge_pte_none(huge_ptep_get(ptep)))
3908 ClearPagePrivate(page);
3909 hugepage_add_new_anon_rmap(page, vma, haddr);
3911 page_dup_rmap(page, true);
3912 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3913 && (vma->vm_flags & VM_SHARED)));
3914 set_huge_pte_at(mm, haddr, ptep, new_pte);
3916 hugetlb_count_add(pages_per_huge_page(h), mm);
3917 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3918 /* Optimization, do the COW without a second fault */
3919 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
3925 * Only make newly allocated pages active. Existing pages found
3926 * in the pagecache could be !page_huge_active() if they have been
3927 * isolated for migration.
3930 set_page_huge_active(page);
3940 restore_reserve_on_error(h, vma, haddr, page);
3946 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping,
3947 pgoff_t idx, unsigned long address)
3949 unsigned long key[2];
3952 key[0] = (unsigned long) mapping;
3955 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3957 return hash & (num_fault_mutexes - 1);
3961 * For uniprocesor systems we always use a single mutex, so just
3962 * return 0 and avoid the hashing overhead.
3964 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping,
3965 pgoff_t idx, unsigned long address)
3971 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3972 unsigned long address, unsigned int flags)
3979 struct page *page = NULL;
3980 struct page *pagecache_page = NULL;
3981 struct hstate *h = hstate_vma(vma);
3982 struct address_space *mapping;
3983 int need_wait_lock = 0;
3984 unsigned long haddr = address & huge_page_mask(h);
3986 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3988 entry = huge_ptep_get(ptep);
3989 if (unlikely(is_hugetlb_entry_migration(entry))) {
3990 migration_entry_wait_huge(vma, mm, ptep);
3992 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3993 return VM_FAULT_HWPOISON_LARGE |
3994 VM_FAULT_SET_HINDEX(hstate_index(h));
3996 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
3998 return VM_FAULT_OOM;
4001 mapping = vma->vm_file->f_mapping;
4002 idx = vma_hugecache_offset(h, vma, haddr);
4005 * Serialize hugepage allocation and instantiation, so that we don't
4006 * get spurious allocation failures if two CPUs race to instantiate
4007 * the same page in the page cache.
4009 hash = hugetlb_fault_mutex_hash(h, mapping, idx, haddr);
4010 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4012 entry = huge_ptep_get(ptep);
4013 if (huge_pte_none(entry)) {
4014 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4021 * entry could be a migration/hwpoison entry at this point, so this
4022 * check prevents the kernel from going below assuming that we have
4023 * a active hugepage in pagecache. This goto expects the 2nd page fault,
4024 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
4027 if (!pte_present(entry))
4031 * If we are going to COW the mapping later, we examine the pending
4032 * reservations for this page now. This will ensure that any
4033 * allocations necessary to record that reservation occur outside the
4034 * spinlock. For private mappings, we also lookup the pagecache
4035 * page now as it is used to determine if a reservation has been
4038 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4039 if (vma_needs_reservation(h, vma, haddr) < 0) {
4043 /* Just decrements count, does not deallocate */
4044 vma_end_reservation(h, vma, haddr);
4046 if (!(vma->vm_flags & VM_MAYSHARE))
4047 pagecache_page = hugetlbfs_pagecache_page(h,
4051 ptl = huge_pte_lock(h, mm, ptep);
4053 /* Check for a racing update before calling hugetlb_cow */
4054 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4058 * hugetlb_cow() requires page locks of pte_page(entry) and
4059 * pagecache_page, so here we need take the former one
4060 * when page != pagecache_page or !pagecache_page.
4062 page = pte_page(entry);
4063 if (page != pagecache_page)
4064 if (!trylock_page(page)) {
4071 if (flags & FAULT_FLAG_WRITE) {
4072 if (!huge_pte_write(entry)) {
4073 ret = hugetlb_cow(mm, vma, address, ptep,
4074 pagecache_page, ptl);
4077 entry = huge_pte_mkdirty(entry);
4079 entry = pte_mkyoung(entry);
4080 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4081 flags & FAULT_FLAG_WRITE))
4082 update_mmu_cache(vma, haddr, ptep);
4084 if (page != pagecache_page)
4090 if (pagecache_page) {
4091 unlock_page(pagecache_page);
4092 put_page(pagecache_page);
4095 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4097 * Generally it's safe to hold refcount during waiting page lock. But
4098 * here we just wait to defer the next page fault to avoid busy loop and
4099 * the page is not used after unlocked before returning from the current
4100 * page fault. So we are safe from accessing freed page, even if we wait
4101 * here without taking refcount.
4104 wait_on_page_locked(page);
4109 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4110 * modifications for huge pages.
4112 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4114 struct vm_area_struct *dst_vma,
4115 unsigned long dst_addr,
4116 unsigned long src_addr,
4117 struct page **pagep)
4119 struct address_space *mapping;
4122 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4123 struct hstate *h = hstate_vma(dst_vma);
4131 page = alloc_huge_page(dst_vma, dst_addr, 0);
4135 ret = copy_huge_page_from_user(page,
4136 (const void __user *) src_addr,
4137 pages_per_huge_page(h), false);
4139 /* fallback to copy_from_user outside mmap_sem */
4140 if (unlikely(ret)) {
4143 /* don't free the page */
4152 * The memory barrier inside __SetPageUptodate makes sure that
4153 * preceding stores to the page contents become visible before
4154 * the set_pte_at() write.
4156 __SetPageUptodate(page);
4158 mapping = dst_vma->vm_file->f_mapping;
4159 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4162 * If shared, add to page cache
4165 size = i_size_read(mapping->host) >> huge_page_shift(h);
4168 goto out_release_nounlock;
4171 * Serialization between remove_inode_hugepages() and
4172 * huge_add_to_page_cache() below happens through the
4173 * hugetlb_fault_mutex_table that here must be hold by
4176 ret = huge_add_to_page_cache(page, mapping, idx);
4178 goto out_release_nounlock;
4181 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4185 * Recheck the i_size after holding PT lock to make sure not
4186 * to leave any page mapped (as page_mapped()) beyond the end
4187 * of the i_size (remove_inode_hugepages() is strict about
4188 * enforcing that). If we bail out here, we'll also leave a
4189 * page in the radix tree in the vm_shared case beyond the end
4190 * of the i_size, but remove_inode_hugepages() will take care
4191 * of it as soon as we drop the hugetlb_fault_mutex_table.
4193 size = i_size_read(mapping->host) >> huge_page_shift(h);
4196 goto out_release_unlock;
4199 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4200 goto out_release_unlock;
4203 page_dup_rmap(page, true);
4205 ClearPagePrivate(page);
4206 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4209 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4210 if (dst_vma->vm_flags & VM_WRITE)
4211 _dst_pte = huge_pte_mkdirty(_dst_pte);
4212 _dst_pte = pte_mkyoung(_dst_pte);
4214 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4216 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4217 dst_vma->vm_flags & VM_WRITE);
4218 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4220 /* No need to invalidate - it was non-present before */
4221 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4224 set_page_huge_active(page);
4234 out_release_nounlock:
4239 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4240 struct page **pages, struct vm_area_struct **vmas,
4241 unsigned long *position, unsigned long *nr_pages,
4242 long i, unsigned int flags, int *nonblocking)
4244 unsigned long pfn_offset;
4245 unsigned long vaddr = *position;
4246 unsigned long remainder = *nr_pages;
4247 struct hstate *h = hstate_vma(vma);
4250 while (vaddr < vma->vm_end && remainder) {
4252 spinlock_t *ptl = NULL;
4257 * If we have a pending SIGKILL, don't keep faulting pages and
4258 * potentially allocating memory.
4260 if (fatal_signal_pending(current)) {
4266 * Some archs (sparc64, sh*) have multiple pte_ts to
4267 * each hugepage. We have to make sure we get the
4268 * first, for the page indexing below to work.
4270 * Note that page table lock is not held when pte is null.
4272 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4275 ptl = huge_pte_lock(h, mm, pte);
4276 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4279 * When coredumping, it suits get_dump_page if we just return
4280 * an error where there's an empty slot with no huge pagecache
4281 * to back it. This way, we avoid allocating a hugepage, and
4282 * the sparse dumpfile avoids allocating disk blocks, but its
4283 * huge holes still show up with zeroes where they need to be.
4285 if (absent && (flags & FOLL_DUMP) &&
4286 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4294 * We need call hugetlb_fault for both hugepages under migration
4295 * (in which case hugetlb_fault waits for the migration,) and
4296 * hwpoisoned hugepages (in which case we need to prevent the
4297 * caller from accessing to them.) In order to do this, we use
4298 * here is_swap_pte instead of is_hugetlb_entry_migration and
4299 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4300 * both cases, and because we can't follow correct pages
4301 * directly from any kind of swap entries.
4303 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4304 ((flags & FOLL_WRITE) &&
4305 !huge_pte_write(huge_ptep_get(pte)))) {
4307 unsigned int fault_flags = 0;
4311 if (flags & FOLL_WRITE)
4312 fault_flags |= FAULT_FLAG_WRITE;
4314 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4315 if (flags & FOLL_NOWAIT)
4316 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4317 FAULT_FLAG_RETRY_NOWAIT;
4318 if (flags & FOLL_TRIED) {
4319 VM_WARN_ON_ONCE(fault_flags &
4320 FAULT_FLAG_ALLOW_RETRY);
4321 fault_flags |= FAULT_FLAG_TRIED;
4323 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4324 if (ret & VM_FAULT_ERROR) {
4325 err = vm_fault_to_errno(ret, flags);
4329 if (ret & VM_FAULT_RETRY) {
4331 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4335 * VM_FAULT_RETRY must not return an
4336 * error, it will return zero
4339 * No need to update "position" as the
4340 * caller will not check it after
4341 * *nr_pages is set to 0.
4348 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4349 page = pte_page(huge_ptep_get(pte));
4352 * Instead of doing 'try_get_page()' below in the same_page
4353 * loop, just check the count once here.
4355 if (unlikely(page_count(page) <= 0)) {
4365 pages[i] = mem_map_offset(page, pfn_offset);
4376 if (vaddr < vma->vm_end && remainder &&
4377 pfn_offset < pages_per_huge_page(h)) {
4379 * We use pfn_offset to avoid touching the pageframes
4380 * of this compound page.
4386 *nr_pages = remainder;
4388 * setting position is actually required only if remainder is
4389 * not zero but it's faster not to add a "if (remainder)"
4397 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4399 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4402 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4405 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4406 unsigned long address, unsigned long end, pgprot_t newprot)
4408 struct mm_struct *mm = vma->vm_mm;
4409 unsigned long start = address;
4412 struct hstate *h = hstate_vma(vma);
4413 unsigned long pages = 0;
4414 bool shared_pmd = false;
4415 struct mmu_notifier_range range;
4418 * In the case of shared PMDs, the area to flush could be beyond
4419 * start/end. Set range.start/range.end to cover the maximum possible
4420 * range if PMD sharing is possible.
4422 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
4423 0, vma, mm, start, end);
4424 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4426 BUG_ON(address >= end);
4427 flush_cache_range(vma, range.start, range.end);
4429 mmu_notifier_invalidate_range_start(&range);
4430 i_mmap_lock_write(vma->vm_file->f_mapping);
4431 for (; address < end; address += huge_page_size(h)) {
4433 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4436 ptl = huge_pte_lock(h, mm, ptep);
4437 if (huge_pmd_unshare(mm, &address, ptep)) {
4443 pte = huge_ptep_get(ptep);
4444 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4448 if (unlikely(is_hugetlb_entry_migration(pte))) {
4449 swp_entry_t entry = pte_to_swp_entry(pte);
4451 if (is_write_migration_entry(entry)) {
4454 make_migration_entry_read(&entry);
4455 newpte = swp_entry_to_pte(entry);
4456 set_huge_swap_pte_at(mm, address, ptep,
4457 newpte, huge_page_size(h));
4463 if (!huge_pte_none(pte)) {
4466 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
4467 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
4468 pte = arch_make_huge_pte(pte, vma, NULL, 0);
4469 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
4475 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4476 * may have cleared our pud entry and done put_page on the page table:
4477 * once we release i_mmap_rwsem, another task can do the final put_page
4478 * and that page table be reused and filled with junk. If we actually
4479 * did unshare a page of pmds, flush the range corresponding to the pud.
4482 flush_hugetlb_tlb_range(vma, range.start, range.end);
4484 flush_hugetlb_tlb_range(vma, start, end);
4486 * No need to call mmu_notifier_invalidate_range() we are downgrading
4487 * page table protection not changing it to point to a new page.
4489 * See Documentation/vm/mmu_notifier.rst
4491 i_mmap_unlock_write(vma->vm_file->f_mapping);
4492 mmu_notifier_invalidate_range_end(&range);
4494 return pages << h->order;
4497 int hugetlb_reserve_pages(struct inode *inode,
4499 struct vm_area_struct *vma,
4500 vm_flags_t vm_flags)
4503 struct hstate *h = hstate_inode(inode);
4504 struct hugepage_subpool *spool = subpool_inode(inode);
4505 struct resv_map *resv_map;
4508 /* This should never happen */
4510 VM_WARN(1, "%s called with a negative range\n", __func__);
4515 * Only apply hugepage reservation if asked. At fault time, an
4516 * attempt will be made for VM_NORESERVE to allocate a page
4517 * without using reserves
4519 if (vm_flags & VM_NORESERVE)
4523 * Shared mappings base their reservation on the number of pages that
4524 * are already allocated on behalf of the file. Private mappings need
4525 * to reserve the full area even if read-only as mprotect() may be
4526 * called to make the mapping read-write. Assume !vma is a shm mapping
4528 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4530 * resv_map can not be NULL as hugetlb_reserve_pages is only
4531 * called for inodes for which resv_maps were created (see
4532 * hugetlbfs_get_inode).
4534 resv_map = inode_resv_map(inode);
4536 chg = region_chg(resv_map, from, to);
4539 resv_map = resv_map_alloc();
4545 set_vma_resv_map(vma, resv_map);
4546 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4555 * There must be enough pages in the subpool for the mapping. If
4556 * the subpool has a minimum size, there may be some global
4557 * reservations already in place (gbl_reserve).
4559 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4560 if (gbl_reserve < 0) {
4566 * Check enough hugepages are available for the reservation.
4567 * Hand the pages back to the subpool if there are not
4569 ret = hugetlb_acct_memory(h, gbl_reserve);
4571 /* put back original number of pages, chg */
4572 (void)hugepage_subpool_put_pages(spool, chg);
4577 * Account for the reservations made. Shared mappings record regions
4578 * that have reservations as they are shared by multiple VMAs.
4579 * When the last VMA disappears, the region map says how much
4580 * the reservation was and the page cache tells how much of
4581 * the reservation was consumed. Private mappings are per-VMA and
4582 * only the consumed reservations are tracked. When the VMA
4583 * disappears, the original reservation is the VMA size and the
4584 * consumed reservations are stored in the map. Hence, nothing
4585 * else has to be done for private mappings here
4587 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4588 long add = region_add(resv_map, from, to);
4590 if (unlikely(chg > add)) {
4592 * pages in this range were added to the reserve
4593 * map between region_chg and region_add. This
4594 * indicates a race with alloc_huge_page. Adjust
4595 * the subpool and reserve counts modified above
4596 * based on the difference.
4600 rsv_adjust = hugepage_subpool_put_pages(spool,
4602 hugetlb_acct_memory(h, -rsv_adjust);
4607 if (!vma || vma->vm_flags & VM_MAYSHARE)
4608 /* Don't call region_abort if region_chg failed */
4610 region_abort(resv_map, from, to);
4611 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4612 kref_put(&resv_map->refs, resv_map_release);
4616 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4619 struct hstate *h = hstate_inode(inode);
4620 struct resv_map *resv_map = inode_resv_map(inode);
4622 struct hugepage_subpool *spool = subpool_inode(inode);
4626 * Since this routine can be called in the evict inode path for all
4627 * hugetlbfs inodes, resv_map could be NULL.
4630 chg = region_del(resv_map, start, end);
4632 * region_del() can fail in the rare case where a region
4633 * must be split and another region descriptor can not be
4634 * allocated. If end == LONG_MAX, it will not fail.
4640 spin_lock(&inode->i_lock);
4641 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4642 spin_unlock(&inode->i_lock);
4645 * If the subpool has a minimum size, the number of global
4646 * reservations to be released may be adjusted.
4648 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4649 hugetlb_acct_memory(h, -gbl_reserve);
4654 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4655 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4656 struct vm_area_struct *vma,
4657 unsigned long addr, pgoff_t idx)
4659 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4661 unsigned long sbase = saddr & PUD_MASK;
4662 unsigned long s_end = sbase + PUD_SIZE;
4664 /* Allow segments to share if only one is marked locked */
4665 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4666 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4669 * match the virtual addresses, permission and the alignment of the
4672 if (pmd_index(addr) != pmd_index(saddr) ||
4673 vm_flags != svm_flags ||
4674 sbase < svma->vm_start || svma->vm_end < s_end)
4680 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4682 unsigned long base = addr & PUD_MASK;
4683 unsigned long end = base + PUD_SIZE;
4686 * check on proper vm_flags and page table alignment
4688 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
4694 * Determine if start,end range within vma could be mapped by shared pmd.
4695 * If yes, adjust start and end to cover range associated with possible
4696 * shared pmd mappings.
4698 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4699 unsigned long *start, unsigned long *end)
4701 unsigned long check_addr = *start;
4703 if (!(vma->vm_flags & VM_MAYSHARE))
4706 for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
4707 unsigned long a_start = check_addr & PUD_MASK;
4708 unsigned long a_end = a_start + PUD_SIZE;
4711 * If sharing is possible, adjust start/end if necessary.
4713 if (range_in_vma(vma, a_start, a_end)) {
4714 if (a_start < *start)
4723 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4724 * and returns the corresponding pte. While this is not necessary for the
4725 * !shared pmd case because we can allocate the pmd later as well, it makes the
4726 * code much cleaner. pmd allocation is essential for the shared case because
4727 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4728 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4729 * bad pmd for sharing.
4731 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4733 struct vm_area_struct *vma = find_vma(mm, addr);
4734 struct address_space *mapping = vma->vm_file->f_mapping;
4735 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4737 struct vm_area_struct *svma;
4738 unsigned long saddr;
4743 if (!vma_shareable(vma, addr))
4744 return (pte_t *)pmd_alloc(mm, pud, addr);
4746 i_mmap_lock_write(mapping);
4747 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4751 saddr = page_table_shareable(svma, vma, addr, idx);
4753 spte = huge_pte_offset(svma->vm_mm, saddr,
4754 vma_mmu_pagesize(svma));
4756 get_page(virt_to_page(spte));
4765 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4766 if (pud_none(*pud)) {
4767 pud_populate(mm, pud,
4768 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4771 put_page(virt_to_page(spte));
4775 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4776 i_mmap_unlock_write(mapping);
4781 * unmap huge page backed by shared pte.
4783 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4784 * indicated by page_count > 1, unmap is achieved by clearing pud and
4785 * decrementing the ref count. If count == 1, the pte page is not shared.
4787 * called with page table lock held.
4789 * returns: 1 successfully unmapped a shared pte page
4790 * 0 the underlying pte page is not shared, or it is the last user
4792 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4794 pgd_t *pgd = pgd_offset(mm, *addr);
4795 p4d_t *p4d = p4d_offset(pgd, *addr);
4796 pud_t *pud = pud_offset(p4d, *addr);
4798 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4799 if (page_count(virt_to_page(ptep)) == 1)
4803 put_page(virt_to_page(ptep));
4805 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4808 #define want_pmd_share() (1)
4809 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4810 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4815 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4820 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4821 unsigned long *start, unsigned long *end)
4824 #define want_pmd_share() (0)
4825 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4827 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4828 pte_t *huge_pte_alloc(struct mm_struct *mm,
4829 unsigned long addr, unsigned long sz)
4836 pgd = pgd_offset(mm, addr);
4837 p4d = p4d_alloc(mm, pgd, addr);
4840 pud = pud_alloc(mm, p4d, addr);
4842 if (sz == PUD_SIZE) {
4845 BUG_ON(sz != PMD_SIZE);
4846 if (want_pmd_share() && pud_none(*pud))
4847 pte = huge_pmd_share(mm, addr, pud);
4849 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4852 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4858 * huge_pte_offset() - Walk the page table to resolve the hugepage
4859 * entry at address @addr
4861 * Return: Pointer to page table or swap entry (PUD or PMD) for
4862 * address @addr, or NULL if a p*d_none() entry is encountered and the
4863 * size @sz doesn't match the hugepage size at this level of the page
4866 pte_t *huge_pte_offset(struct mm_struct *mm,
4867 unsigned long addr, unsigned long sz)
4874 pgd = pgd_offset(mm, addr);
4875 if (!pgd_present(*pgd))
4877 p4d = p4d_offset(pgd, addr);
4878 if (!p4d_present(*p4d))
4881 pud = pud_offset(p4d, addr);
4882 if (sz != PUD_SIZE && pud_none(*pud))
4884 /* hugepage or swap? */
4885 if (pud_huge(*pud) || !pud_present(*pud))
4886 return (pte_t *)pud;
4888 pmd = pmd_offset(pud, addr);
4889 if (sz != PMD_SIZE && pmd_none(*pmd))
4891 /* hugepage or swap? */
4892 if (pmd_huge(*pmd) || !pmd_present(*pmd))
4893 return (pte_t *)pmd;
4898 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4901 * These functions are overwritable if your architecture needs its own
4904 struct page * __weak
4905 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4908 return ERR_PTR(-EINVAL);
4911 struct page * __weak
4912 follow_huge_pd(struct vm_area_struct *vma,
4913 unsigned long address, hugepd_t hpd, int flags, int pdshift)
4915 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4919 struct page * __weak
4920 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4921 pmd_t *pmd, int flags)
4923 struct page *page = NULL;
4927 ptl = pmd_lockptr(mm, pmd);
4930 * make sure that the address range covered by this pmd is not
4931 * unmapped from other threads.
4933 if (!pmd_huge(*pmd))
4935 pte = huge_ptep_get((pte_t *)pmd);
4936 if (pte_present(pte)) {
4937 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4938 if (flags & FOLL_GET)
4941 if (is_hugetlb_entry_migration(pte)) {
4943 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4947 * hwpoisoned entry is treated as no_page_table in
4948 * follow_page_mask().
4956 struct page * __weak
4957 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4958 pud_t *pud, int flags)
4960 if (flags & FOLL_GET)
4963 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4966 struct page * __weak
4967 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
4969 if (flags & FOLL_GET)
4972 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
4975 bool isolate_huge_page(struct page *page, struct list_head *list)
4979 VM_BUG_ON_PAGE(!PageHead(page), page);
4980 spin_lock(&hugetlb_lock);
4981 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4985 clear_page_huge_active(page);
4986 list_move_tail(&page->lru, list);
4988 spin_unlock(&hugetlb_lock);
4992 void putback_active_hugepage(struct page *page)
4994 VM_BUG_ON_PAGE(!PageHead(page), page);
4995 spin_lock(&hugetlb_lock);
4996 set_page_huge_active(page);
4997 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4998 spin_unlock(&hugetlb_lock);
5002 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5004 struct hstate *h = page_hstate(oldpage);
5006 hugetlb_cgroup_migrate(oldpage, newpage);
5007 set_page_owner_migrate_reason(newpage, reason);
5010 * transfer temporary state of the new huge page. This is
5011 * reverse to other transitions because the newpage is going to
5012 * be final while the old one will be freed so it takes over
5013 * the temporary status.
5015 * Also note that we have to transfer the per-node surplus state
5016 * here as well otherwise the global surplus count will not match
5019 if (PageHugeTemporary(newpage)) {
5020 int old_nid = page_to_nid(oldpage);
5021 int new_nid = page_to_nid(newpage);
5023 SetPageHugeTemporary(oldpage);
5024 ClearPageHugeTemporary(newpage);
5026 spin_lock(&hugetlb_lock);
5027 if (h->surplus_huge_pages_node[old_nid]) {
5028 h->surplus_huge_pages_node[old_nid]--;
5029 h->surplus_huge_pages_node[new_nid]++;
5031 spin_unlock(&hugetlb_lock);