2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
8 #include <linux/seq_file.h>
9 #include <linux/sysctl.h>
10 #include <linux/highmem.h>
11 #include <linux/mmu_notifier.h>
12 #include <linux/nodemask.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/compiler.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/rmap.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
24 #include <linux/page-isolation.h>
25 #include <linux/jhash.h>
28 #include <asm/pgtable.h>
32 #include <linux/hugetlb.h>
33 #include <linux/hugetlb_cgroup.h>
34 #include <linux/node.h>
37 int hugepages_treat_as_movable;
39 int hugetlb_max_hstate __read_mostly;
40 unsigned int default_hstate_idx;
41 struct hstate hstates[HUGE_MAX_HSTATE];
43 * Minimum page order among possible hugepage sizes, set to a proper value
46 static unsigned int minimum_order __read_mostly = UINT_MAX;
48 __initdata LIST_HEAD(huge_boot_pages);
50 /* for command line parsing */
51 static struct hstate * __initdata parsed_hstate;
52 static unsigned long __initdata default_hstate_max_huge_pages;
53 static unsigned long __initdata default_hstate_size;
54 static bool __initdata parsed_valid_hugepagesz = true;
57 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
58 * free_huge_pages, and surplus_huge_pages.
60 DEFINE_SPINLOCK(hugetlb_lock);
63 * Serializes faults on the same logical page. This is used to
64 * prevent spurious OOMs when the hugepage pool is fully utilized.
66 static int num_fault_mutexes;
67 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
69 /* Forward declaration */
70 static int hugetlb_acct_memory(struct hstate *h, long delta);
72 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
74 bool free = (spool->count == 0) && (spool->used_hpages == 0);
76 spin_unlock(&spool->lock);
78 /* If no pages are used, and no other handles to the subpool
79 * remain, give up any reservations mased on minimum size and
82 if (spool->min_hpages != -1)
83 hugetlb_acct_memory(spool->hstate,
89 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
92 struct hugepage_subpool *spool;
94 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
98 spin_lock_init(&spool->lock);
100 spool->max_hpages = max_hpages;
102 spool->min_hpages = min_hpages;
104 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
108 spool->rsv_hpages = min_hpages;
113 void hugepage_put_subpool(struct hugepage_subpool *spool)
115 spin_lock(&spool->lock);
116 BUG_ON(!spool->count);
118 unlock_or_release_subpool(spool);
122 * Subpool accounting for allocating and reserving pages.
123 * Return -ENOMEM if there are not enough resources to satisfy the
124 * the request. Otherwise, return the number of pages by which the
125 * global pools must be adjusted (upward). The returned value may
126 * only be different than the passed value (delta) in the case where
127 * a subpool minimum size must be manitained.
129 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
137 spin_lock(&spool->lock);
139 if (spool->max_hpages != -1) { /* maximum size accounting */
140 if ((spool->used_hpages + delta) <= spool->max_hpages)
141 spool->used_hpages += delta;
148 /* minimum size accounting */
149 if (spool->min_hpages != -1 && spool->rsv_hpages) {
150 if (delta > spool->rsv_hpages) {
152 * Asking for more reserves than those already taken on
153 * behalf of subpool. Return difference.
155 ret = delta - spool->rsv_hpages;
156 spool->rsv_hpages = 0;
158 ret = 0; /* reserves already accounted for */
159 spool->rsv_hpages -= delta;
164 spin_unlock(&spool->lock);
169 * Subpool accounting for freeing and unreserving pages.
170 * Return the number of global page reservations that must be dropped.
171 * The return value may only be different than the passed value (delta)
172 * in the case where a subpool minimum size must be maintained.
174 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
182 spin_lock(&spool->lock);
184 if (spool->max_hpages != -1) /* maximum size accounting */
185 spool->used_hpages -= delta;
187 /* minimum size accounting */
188 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
189 if (spool->rsv_hpages + delta <= spool->min_hpages)
192 ret = spool->rsv_hpages + delta - spool->min_hpages;
194 spool->rsv_hpages += delta;
195 if (spool->rsv_hpages > spool->min_hpages)
196 spool->rsv_hpages = spool->min_hpages;
200 * If hugetlbfs_put_super couldn't free spool due to an outstanding
201 * quota reference, free it now.
203 unlock_or_release_subpool(spool);
208 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
210 return HUGETLBFS_SB(inode->i_sb)->spool;
213 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
215 return subpool_inode(file_inode(vma->vm_file));
219 * Region tracking -- allows tracking of reservations and instantiated pages
220 * across the pages in a mapping.
222 * The region data structures are embedded into a resv_map and protected
223 * by a resv_map's lock. The set of regions within the resv_map represent
224 * reservations for huge pages, or huge pages that have already been
225 * instantiated within the map. The from and to elements are huge page
226 * indicies into the associated mapping. from indicates the starting index
227 * of the region. to represents the first index past the end of the region.
229 * For example, a file region structure with from == 0 and to == 4 represents
230 * four huge pages in a mapping. It is important to note that the to element
231 * represents the first element past the end of the region. This is used in
232 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
234 * Interval notation of the form [from, to) will be used to indicate that
235 * the endpoint from is inclusive and to is exclusive.
238 struct list_head link;
244 * Add the huge page range represented by [f, t) to the reserve
245 * map. In the normal case, existing regions will be expanded
246 * to accommodate the specified range. Sufficient regions should
247 * exist for expansion due to the previous call to region_chg
248 * with the same range. However, it is possible that region_del
249 * could have been called after region_chg and modifed the map
250 * in such a way that no region exists to be expanded. In this
251 * case, pull a region descriptor from the cache associated with
252 * the map and use that for the new range.
254 * Return the number of new huge pages added to the map. This
255 * number is greater than or equal to zero.
257 static long region_add(struct resv_map *resv, long f, long t)
259 struct list_head *head = &resv->regions;
260 struct file_region *rg, *nrg, *trg;
263 spin_lock(&resv->lock);
264 /* Locate the region we are either in or before. */
265 list_for_each_entry(rg, head, link)
270 * If no region exists which can be expanded to include the
271 * specified range, the list must have been modified by an
272 * interleving call to region_del(). Pull a region descriptor
273 * from the cache and use it for this range.
275 if (&rg->link == head || t < rg->from) {
276 VM_BUG_ON(resv->region_cache_count <= 0);
278 resv->region_cache_count--;
279 nrg = list_first_entry(&resv->region_cache, struct file_region,
281 list_del(&nrg->link);
285 list_add(&nrg->link, rg->link.prev);
291 /* Round our left edge to the current segment if it encloses us. */
295 /* Check for and consume any regions we now overlap with. */
297 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
298 if (&rg->link == head)
303 /* If this area reaches higher then extend our area to
304 * include it completely. If this is not the first area
305 * which we intend to reuse, free it. */
309 /* Decrement return value by the deleted range.
310 * Another range will span this area so that by
311 * end of routine add will be >= zero
313 add -= (rg->to - rg->from);
319 add += (nrg->from - f); /* Added to beginning of region */
321 add += t - nrg->to; /* Added to end of region */
325 resv->adds_in_progress--;
326 spin_unlock(&resv->lock);
332 * Examine the existing reserve map and determine how many
333 * huge pages in the specified range [f, t) are NOT currently
334 * represented. This routine is called before a subsequent
335 * call to region_add that will actually modify the reserve
336 * map to add the specified range [f, t). region_chg does
337 * not change the number of huge pages represented by the
338 * map. However, if the existing regions in the map can not
339 * be expanded to represent the new range, a new file_region
340 * structure is added to the map as a placeholder. This is
341 * so that the subsequent region_add call will have all the
342 * regions it needs and will not fail.
344 * Upon entry, region_chg will also examine the cache of region descriptors
345 * associated with the map. If there are not enough descriptors cached, one
346 * will be allocated for the in progress add operation.
348 * Returns the number of huge pages that need to be added to the existing
349 * reservation map for the range [f, t). This number is greater or equal to
350 * zero. -ENOMEM is returned if a new file_region structure or cache entry
351 * is needed and can not be allocated.
353 static long region_chg(struct resv_map *resv, long f, long t)
355 struct list_head *head = &resv->regions;
356 struct file_region *rg, *nrg = NULL;
360 spin_lock(&resv->lock);
362 resv->adds_in_progress++;
365 * Check for sufficient descriptors in the cache to accommodate
366 * the number of in progress add operations.
368 if (resv->adds_in_progress > resv->region_cache_count) {
369 struct file_region *trg;
371 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
372 /* Must drop lock to allocate a new descriptor. */
373 resv->adds_in_progress--;
374 spin_unlock(&resv->lock);
376 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
382 spin_lock(&resv->lock);
383 list_add(&trg->link, &resv->region_cache);
384 resv->region_cache_count++;
388 /* Locate the region we are before or in. */
389 list_for_each_entry(rg, head, link)
393 /* If we are below the current region then a new region is required.
394 * Subtle, allocate a new region at the position but make it zero
395 * size such that we can guarantee to record the reservation. */
396 if (&rg->link == head || t < rg->from) {
398 resv->adds_in_progress--;
399 spin_unlock(&resv->lock);
400 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
406 INIT_LIST_HEAD(&nrg->link);
410 list_add(&nrg->link, rg->link.prev);
415 /* Round our left edge to the current segment if it encloses us. */
420 /* Check for and consume any regions we now overlap with. */
421 list_for_each_entry(rg, rg->link.prev, link) {
422 if (&rg->link == head)
427 /* We overlap with this area, if it extends further than
428 * us then we must extend ourselves. Account for its
429 * existing reservation. */
434 chg -= rg->to - rg->from;
438 spin_unlock(&resv->lock);
439 /* We already know we raced and no longer need the new region */
443 spin_unlock(&resv->lock);
448 * Abort the in progress add operation. The adds_in_progress field
449 * of the resv_map keeps track of the operations in progress between
450 * calls to region_chg and region_add. Operations are sometimes
451 * aborted after the call to region_chg. In such cases, region_abort
452 * is called to decrement the adds_in_progress counter.
454 * NOTE: The range arguments [f, t) are not needed or used in this
455 * routine. They are kept to make reading the calling code easier as
456 * arguments will match the associated region_chg call.
458 static void region_abort(struct resv_map *resv, long f, long t)
460 spin_lock(&resv->lock);
461 VM_BUG_ON(!resv->region_cache_count);
462 resv->adds_in_progress--;
463 spin_unlock(&resv->lock);
467 * Delete the specified range [f, t) from the reserve map. If the
468 * t parameter is LONG_MAX, this indicates that ALL regions after f
469 * should be deleted. Locate the regions which intersect [f, t)
470 * and either trim, delete or split the existing regions.
472 * Returns the number of huge pages deleted from the reserve map.
473 * In the normal case, the return value is zero or more. In the
474 * case where a region must be split, a new region descriptor must
475 * be allocated. If the allocation fails, -ENOMEM will be returned.
476 * NOTE: If the parameter t == LONG_MAX, then we will never split
477 * a region and possibly return -ENOMEM. Callers specifying
478 * t == LONG_MAX do not need to check for -ENOMEM error.
480 static long region_del(struct resv_map *resv, long f, long t)
482 struct list_head *head = &resv->regions;
483 struct file_region *rg, *trg;
484 struct file_region *nrg = NULL;
488 spin_lock(&resv->lock);
489 list_for_each_entry_safe(rg, trg, head, link) {
491 * Skip regions before the range to be deleted. file_region
492 * ranges are normally of the form [from, to). However, there
493 * may be a "placeholder" entry in the map which is of the form
494 * (from, to) with from == to. Check for placeholder entries
495 * at the beginning of the range to be deleted.
497 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
503 if (f > rg->from && t < rg->to) { /* Must split region */
505 * Check for an entry in the cache before dropping
506 * lock and attempting allocation.
509 resv->region_cache_count > resv->adds_in_progress) {
510 nrg = list_first_entry(&resv->region_cache,
513 list_del(&nrg->link);
514 resv->region_cache_count--;
518 spin_unlock(&resv->lock);
519 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
527 /* New entry for end of split region */
530 INIT_LIST_HEAD(&nrg->link);
532 /* Original entry is trimmed */
535 list_add(&nrg->link, &rg->link);
540 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
541 del += rg->to - rg->from;
547 if (f <= rg->from) { /* Trim beginning of region */
550 } else { /* Trim end of region */
556 spin_unlock(&resv->lock);
562 * A rare out of memory error was encountered which prevented removal of
563 * the reserve map region for a page. The huge page itself was free'ed
564 * and removed from the page cache. This routine will adjust the subpool
565 * usage count, and the global reserve count if needed. By incrementing
566 * these counts, the reserve map entry which could not be deleted will
567 * appear as a "reserved" entry instead of simply dangling with incorrect
570 void hugetlb_fix_reserve_counts(struct inode *inode, bool restore_reserve)
572 struct hugepage_subpool *spool = subpool_inode(inode);
575 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
576 if (restore_reserve && rsv_adjust) {
577 struct hstate *h = hstate_inode(inode);
579 hugetlb_acct_memory(h, 1);
584 * Count and return the number of huge pages in the reserve map
585 * that intersect with the range [f, t).
587 static long region_count(struct resv_map *resv, long f, long t)
589 struct list_head *head = &resv->regions;
590 struct file_region *rg;
593 spin_lock(&resv->lock);
594 /* Locate each segment we overlap with, and count that overlap. */
595 list_for_each_entry(rg, head, link) {
604 seg_from = max(rg->from, f);
605 seg_to = min(rg->to, t);
607 chg += seg_to - seg_from;
609 spin_unlock(&resv->lock);
615 * Convert the address within this vma to the page offset within
616 * the mapping, in pagecache page units; huge pages here.
618 static pgoff_t vma_hugecache_offset(struct hstate *h,
619 struct vm_area_struct *vma, unsigned long address)
621 return ((address - vma->vm_start) >> huge_page_shift(h)) +
622 (vma->vm_pgoff >> huge_page_order(h));
625 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
626 unsigned long address)
628 return vma_hugecache_offset(hstate_vma(vma), vma, address);
632 * Return the size of the pages allocated when backing a VMA. In the majority
633 * cases this will be same size as used by the page table entries.
635 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
637 struct hstate *hstate;
639 if (!is_vm_hugetlb_page(vma))
642 hstate = hstate_vma(vma);
644 return 1UL << huge_page_shift(hstate);
646 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
649 * Return the page size being used by the MMU to back a VMA. In the majority
650 * of cases, the page size used by the kernel matches the MMU size. On
651 * architectures where it differs, an architecture-specific version of this
652 * function is required.
654 #ifndef vma_mmu_pagesize
655 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
657 return vma_kernel_pagesize(vma);
662 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
663 * bits of the reservation map pointer, which are always clear due to
666 #define HPAGE_RESV_OWNER (1UL << 0)
667 #define HPAGE_RESV_UNMAPPED (1UL << 1)
668 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
671 * These helpers are used to track how many pages are reserved for
672 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
673 * is guaranteed to have their future faults succeed.
675 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
676 * the reserve counters are updated with the hugetlb_lock held. It is safe
677 * to reset the VMA at fork() time as it is not in use yet and there is no
678 * chance of the global counters getting corrupted as a result of the values.
680 * The private mapping reservation is represented in a subtly different
681 * manner to a shared mapping. A shared mapping has a region map associated
682 * with the underlying file, this region map represents the backing file
683 * pages which have ever had a reservation assigned which this persists even
684 * after the page is instantiated. A private mapping has a region map
685 * associated with the original mmap which is attached to all VMAs which
686 * reference it, this region map represents those offsets which have consumed
687 * reservation ie. where pages have been instantiated.
689 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
691 return (unsigned long)vma->vm_private_data;
694 static void set_vma_private_data(struct vm_area_struct *vma,
697 vma->vm_private_data = (void *)value;
700 struct resv_map *resv_map_alloc(void)
702 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
703 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
705 if (!resv_map || !rg) {
711 kref_init(&resv_map->refs);
712 spin_lock_init(&resv_map->lock);
713 INIT_LIST_HEAD(&resv_map->regions);
715 resv_map->adds_in_progress = 0;
717 INIT_LIST_HEAD(&resv_map->region_cache);
718 list_add(&rg->link, &resv_map->region_cache);
719 resv_map->region_cache_count = 1;
724 void resv_map_release(struct kref *ref)
726 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
727 struct list_head *head = &resv_map->region_cache;
728 struct file_region *rg, *trg;
730 /* Clear out any active regions before we release the map. */
731 region_del(resv_map, 0, LONG_MAX);
733 /* ... and any entries left in the cache */
734 list_for_each_entry_safe(rg, trg, head, link) {
739 VM_BUG_ON(resv_map->adds_in_progress);
744 static inline struct resv_map *inode_resv_map(struct inode *inode)
746 return inode->i_mapping->private_data;
749 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
751 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
752 if (vma->vm_flags & VM_MAYSHARE) {
753 struct address_space *mapping = vma->vm_file->f_mapping;
754 struct inode *inode = mapping->host;
756 return inode_resv_map(inode);
759 return (struct resv_map *)(get_vma_private_data(vma) &
764 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
766 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
767 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
769 set_vma_private_data(vma, (get_vma_private_data(vma) &
770 HPAGE_RESV_MASK) | (unsigned long)map);
773 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
775 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
776 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
778 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
781 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
783 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
785 return (get_vma_private_data(vma) & flag) != 0;
788 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
789 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
791 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
792 if (!(vma->vm_flags & VM_MAYSHARE))
793 vma->vm_private_data = (void *)0;
796 /* Returns true if the VMA has associated reserve pages */
797 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
799 if (vma->vm_flags & VM_NORESERVE) {
801 * This address is already reserved by other process(chg == 0),
802 * so, we should decrement reserved count. Without decrementing,
803 * reserve count remains after releasing inode, because this
804 * allocated page will go into page cache and is regarded as
805 * coming from reserved pool in releasing step. Currently, we
806 * don't have any other solution to deal with this situation
807 * properly, so add work-around here.
809 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
815 /* Shared mappings always use reserves */
816 if (vma->vm_flags & VM_MAYSHARE) {
818 * We know VM_NORESERVE is not set. Therefore, there SHOULD
819 * be a region map for all pages. The only situation where
820 * there is no region map is if a hole was punched via
821 * fallocate. In this case, there really are no reverves to
822 * use. This situation is indicated if chg != 0.
831 * Only the process that called mmap() has reserves for
834 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
840 static void enqueue_huge_page(struct hstate *h, struct page *page)
842 int nid = page_to_nid(page);
843 list_move(&page->lru, &h->hugepage_freelists[nid]);
844 h->free_huge_pages++;
845 h->free_huge_pages_node[nid]++;
848 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
852 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
853 if (!is_migrate_isolate_page(page))
856 * if 'non-isolated free hugepage' not found on the list,
857 * the allocation fails.
859 if (&h->hugepage_freelists[nid] == &page->lru)
861 list_move(&page->lru, &h->hugepage_activelist);
862 set_page_refcounted(page);
863 h->free_huge_pages--;
864 h->free_huge_pages_node[nid]--;
868 /* Movability of hugepages depends on migration support. */
869 static inline gfp_t htlb_alloc_mask(struct hstate *h)
871 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
872 return GFP_HIGHUSER_MOVABLE;
877 static struct page *dequeue_huge_page_vma(struct hstate *h,
878 struct vm_area_struct *vma,
879 unsigned long address, int avoid_reserve,
882 struct page *page = NULL;
883 struct mempolicy *mpol;
884 nodemask_t *nodemask;
885 struct zonelist *zonelist;
888 unsigned int cpuset_mems_cookie;
891 * A child process with MAP_PRIVATE mappings created by their parent
892 * have no page reserves. This check ensures that reservations are
893 * not "stolen". The child may still get SIGKILLed
895 if (!vma_has_reserves(vma, chg) &&
896 h->free_huge_pages - h->resv_huge_pages == 0)
899 /* If reserves cannot be used, ensure enough pages are in the pool */
900 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
904 cpuset_mems_cookie = read_mems_allowed_begin();
905 zonelist = huge_zonelist(vma, address,
906 htlb_alloc_mask(h), &mpol, &nodemask);
908 for_each_zone_zonelist_nodemask(zone, z, zonelist,
909 MAX_NR_ZONES - 1, nodemask) {
910 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
911 page = dequeue_huge_page_node(h, zone_to_nid(zone));
915 if (!vma_has_reserves(vma, chg))
918 SetPagePrivate(page);
919 h->resv_huge_pages--;
926 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
935 * common helper functions for hstate_next_node_to_{alloc|free}.
936 * We may have allocated or freed a huge page based on a different
937 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
938 * be outside of *nodes_allowed. Ensure that we use an allowed
939 * node for alloc or free.
941 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
943 nid = next_node_in(nid, *nodes_allowed);
944 VM_BUG_ON(nid >= MAX_NUMNODES);
949 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
951 if (!node_isset(nid, *nodes_allowed))
952 nid = next_node_allowed(nid, nodes_allowed);
957 * returns the previously saved node ["this node"] from which to
958 * allocate a persistent huge page for the pool and advance the
959 * next node from which to allocate, handling wrap at end of node
962 static int hstate_next_node_to_alloc(struct hstate *h,
963 nodemask_t *nodes_allowed)
967 VM_BUG_ON(!nodes_allowed);
969 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
970 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
976 * helper for free_pool_huge_page() - return the previously saved
977 * node ["this node"] from which to free a huge page. Advance the
978 * next node id whether or not we find a free huge page to free so
979 * that the next attempt to free addresses the next node.
981 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
985 VM_BUG_ON(!nodes_allowed);
987 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
988 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
993 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
994 for (nr_nodes = nodes_weight(*mask); \
996 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
999 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1000 for (nr_nodes = nodes_weight(*mask); \
1002 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1005 #if defined(CONFIG_X86_64) && ((defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA))
1006 static void destroy_compound_gigantic_page(struct page *page,
1010 int nr_pages = 1 << order;
1011 struct page *p = page + 1;
1013 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1014 clear_compound_head(p);
1015 set_page_refcounted(p);
1018 set_compound_order(page, 0);
1019 __ClearPageHead(page);
1022 static void free_gigantic_page(struct page *page, unsigned int order)
1024 free_contig_range(page_to_pfn(page), 1 << order);
1027 static int __alloc_gigantic_page(unsigned long start_pfn,
1028 unsigned long nr_pages)
1030 unsigned long end_pfn = start_pfn + nr_pages;
1031 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1034 static bool pfn_range_valid_gigantic(struct zone *z,
1035 unsigned long start_pfn, unsigned long nr_pages)
1037 unsigned long i, end_pfn = start_pfn + nr_pages;
1040 for (i = start_pfn; i < end_pfn; i++) {
1044 page = pfn_to_page(i);
1046 if (page_zone(page) != z)
1049 if (PageReserved(page))
1052 if (page_count(page) > 0)
1062 static bool zone_spans_last_pfn(const struct zone *zone,
1063 unsigned long start_pfn, unsigned long nr_pages)
1065 unsigned long last_pfn = start_pfn + nr_pages - 1;
1066 return zone_spans_pfn(zone, last_pfn);
1069 static struct page *alloc_gigantic_page(int nid, unsigned int order)
1071 unsigned long nr_pages = 1 << order;
1072 unsigned long ret, pfn, flags;
1075 z = NODE_DATA(nid)->node_zones;
1076 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
1077 spin_lock_irqsave(&z->lock, flags);
1079 pfn = ALIGN(z->zone_start_pfn, nr_pages);
1080 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
1081 if (pfn_range_valid_gigantic(z, pfn, nr_pages)) {
1083 * We release the zone lock here because
1084 * alloc_contig_range() will also lock the zone
1085 * at some point. If there's an allocation
1086 * spinning on this lock, it may win the race
1087 * and cause alloc_contig_range() to fail...
1089 spin_unlock_irqrestore(&z->lock, flags);
1090 ret = __alloc_gigantic_page(pfn, nr_pages);
1092 return pfn_to_page(pfn);
1093 spin_lock_irqsave(&z->lock, flags);
1098 spin_unlock_irqrestore(&z->lock, flags);
1104 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1105 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1107 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1111 page = alloc_gigantic_page(nid, huge_page_order(h));
1113 prep_compound_gigantic_page(page, huge_page_order(h));
1114 prep_new_huge_page(h, page, nid);
1120 static int alloc_fresh_gigantic_page(struct hstate *h,
1121 nodemask_t *nodes_allowed)
1123 struct page *page = NULL;
1126 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1127 page = alloc_fresh_gigantic_page_node(h, node);
1135 static inline bool gigantic_page_supported(void) { return true; }
1137 static inline bool gigantic_page_supported(void) { return false; }
1138 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1139 static inline void destroy_compound_gigantic_page(struct page *page,
1140 unsigned int order) { }
1141 static inline int alloc_fresh_gigantic_page(struct hstate *h,
1142 nodemask_t *nodes_allowed) { return 0; }
1145 static void update_and_free_page(struct hstate *h, struct page *page)
1149 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1153 h->nr_huge_pages_node[page_to_nid(page)]--;
1154 for (i = 0; i < pages_per_huge_page(h); i++) {
1155 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1156 1 << PG_referenced | 1 << PG_dirty |
1157 1 << PG_active | 1 << PG_private |
1160 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1161 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1162 set_page_refcounted(page);
1163 if (hstate_is_gigantic(h)) {
1164 destroy_compound_gigantic_page(page, huge_page_order(h));
1165 free_gigantic_page(page, huge_page_order(h));
1167 __free_pages(page, huge_page_order(h));
1171 struct hstate *size_to_hstate(unsigned long size)
1175 for_each_hstate(h) {
1176 if (huge_page_size(h) == size)
1183 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1184 * to hstate->hugepage_activelist.)
1186 * This function can be called for tail pages, but never returns true for them.
1188 bool page_huge_active(struct page *page)
1190 VM_BUG_ON_PAGE(!PageHuge(page), page);
1191 return PageHead(page) && PagePrivate(&page[1]);
1194 /* never called for tail page */
1195 static void set_page_huge_active(struct page *page)
1197 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1198 SetPagePrivate(&page[1]);
1201 static void clear_page_huge_active(struct page *page)
1203 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1204 ClearPagePrivate(&page[1]);
1207 void free_huge_page(struct page *page)
1210 * Can't pass hstate in here because it is called from the
1211 * compound page destructor.
1213 struct hstate *h = page_hstate(page);
1214 int nid = page_to_nid(page);
1215 struct hugepage_subpool *spool =
1216 (struct hugepage_subpool *)page_private(page);
1217 bool restore_reserve;
1219 set_page_private(page, 0);
1220 page->mapping = NULL;
1221 VM_BUG_ON_PAGE(page_count(page), page);
1222 VM_BUG_ON_PAGE(page_mapcount(page), page);
1223 restore_reserve = PagePrivate(page);
1224 ClearPagePrivate(page);
1227 * A return code of zero implies that the subpool will be under its
1228 * minimum size if the reservation is not restored after page is free.
1229 * Therefore, force restore_reserve operation.
1231 if (hugepage_subpool_put_pages(spool, 1) == 0)
1232 restore_reserve = true;
1234 spin_lock(&hugetlb_lock);
1235 clear_page_huge_active(page);
1236 hugetlb_cgroup_uncharge_page(hstate_index(h),
1237 pages_per_huge_page(h), page);
1238 if (restore_reserve)
1239 h->resv_huge_pages++;
1241 if (h->surplus_huge_pages_node[nid]) {
1242 /* remove the page from active list */
1243 list_del(&page->lru);
1244 update_and_free_page(h, page);
1245 h->surplus_huge_pages--;
1246 h->surplus_huge_pages_node[nid]--;
1248 arch_clear_hugepage_flags(page);
1249 enqueue_huge_page(h, page);
1251 spin_unlock(&hugetlb_lock);
1254 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1256 INIT_LIST_HEAD(&page->lru);
1257 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1258 spin_lock(&hugetlb_lock);
1259 set_hugetlb_cgroup(page, NULL);
1261 h->nr_huge_pages_node[nid]++;
1262 spin_unlock(&hugetlb_lock);
1263 put_page(page); /* free it into the hugepage allocator */
1266 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1269 int nr_pages = 1 << order;
1270 struct page *p = page + 1;
1272 /* we rely on prep_new_huge_page to set the destructor */
1273 set_compound_order(page, order);
1274 __ClearPageReserved(page);
1275 __SetPageHead(page);
1276 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1278 * For gigantic hugepages allocated through bootmem at
1279 * boot, it's safer to be consistent with the not-gigantic
1280 * hugepages and clear the PG_reserved bit from all tail pages
1281 * too. Otherwse drivers using get_user_pages() to access tail
1282 * pages may get the reference counting wrong if they see
1283 * PG_reserved set on a tail page (despite the head page not
1284 * having PG_reserved set). Enforcing this consistency between
1285 * head and tail pages allows drivers to optimize away a check
1286 * on the head page when they need know if put_page() is needed
1287 * after get_user_pages().
1289 __ClearPageReserved(p);
1290 set_page_count(p, 0);
1291 set_compound_head(p, page);
1293 atomic_set(compound_mapcount_ptr(page), -1);
1297 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1298 * transparent huge pages. See the PageTransHuge() documentation for more
1301 int PageHuge(struct page *page)
1303 if (!PageCompound(page))
1306 page = compound_head(page);
1307 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1309 EXPORT_SYMBOL_GPL(PageHuge);
1312 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1313 * normal or transparent huge pages.
1315 int PageHeadHuge(struct page *page_head)
1317 if (!PageHead(page_head))
1320 return get_compound_page_dtor(page_head) == free_huge_page;
1323 pgoff_t __basepage_index(struct page *page)
1325 struct page *page_head = compound_head(page);
1326 pgoff_t index = page_index(page_head);
1327 unsigned long compound_idx;
1329 if (!PageHuge(page_head))
1330 return page_index(page);
1332 if (compound_order(page_head) >= MAX_ORDER)
1333 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1335 compound_idx = page - page_head;
1337 return (index << compound_order(page_head)) + compound_idx;
1340 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1344 page = __alloc_pages_node(nid,
1345 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1346 __GFP_REPEAT|__GFP_NOWARN,
1347 huge_page_order(h));
1349 prep_new_huge_page(h, page, nid);
1355 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1361 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1362 page = alloc_fresh_huge_page_node(h, node);
1370 count_vm_event(HTLB_BUDDY_PGALLOC);
1372 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1378 * Free huge page from pool from next node to free.
1379 * Attempt to keep persistent huge pages more or less
1380 * balanced over allowed nodes.
1381 * Called with hugetlb_lock locked.
1383 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1389 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1391 * If we're returning unused surplus pages, only examine
1392 * nodes with surplus pages.
1394 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1395 !list_empty(&h->hugepage_freelists[node])) {
1397 list_entry(h->hugepage_freelists[node].next,
1399 list_del(&page->lru);
1400 h->free_huge_pages--;
1401 h->free_huge_pages_node[node]--;
1403 h->surplus_huge_pages--;
1404 h->surplus_huge_pages_node[node]--;
1406 update_and_free_page(h, page);
1416 * Dissolve a given free hugepage into free buddy pages. This function does
1417 * nothing for in-use (including surplus) hugepages.
1419 static void dissolve_free_huge_page(struct page *page)
1421 spin_lock(&hugetlb_lock);
1422 if (PageHuge(page) && !page_count(page)) {
1423 struct hstate *h = page_hstate(page);
1424 int nid = page_to_nid(page);
1425 list_del(&page->lru);
1426 h->free_huge_pages--;
1427 h->free_huge_pages_node[nid]--;
1428 update_and_free_page(h, page);
1430 spin_unlock(&hugetlb_lock);
1434 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1435 * make specified memory blocks removable from the system.
1436 * Note that start_pfn should aligned with (minimum) hugepage size.
1438 void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1442 if (!hugepages_supported())
1445 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order));
1446 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
1447 dissolve_free_huge_page(pfn_to_page(pfn));
1451 * There are 3 ways this can get called:
1452 * 1. With vma+addr: we use the VMA's memory policy
1453 * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge
1454 * page from any node, and let the buddy allocator itself figure
1456 * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page
1457 * strictly from 'nid'
1459 static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1460 struct vm_area_struct *vma, unsigned long addr, int nid)
1462 int order = huge_page_order(h);
1463 gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN;
1464 unsigned int cpuset_mems_cookie;
1467 * We need a VMA to get a memory policy. If we do not
1468 * have one, we use the 'nid' argument.
1470 * The mempolicy stuff below has some non-inlined bits
1471 * and calls ->vm_ops. That makes it hard to optimize at
1472 * compile-time, even when NUMA is off and it does
1473 * nothing. This helps the compiler optimize it out.
1475 if (!IS_ENABLED(CONFIG_NUMA) || !vma) {
1477 * If a specific node is requested, make sure to
1478 * get memory from there, but only when a node
1479 * is explicitly specified.
1481 if (nid != NUMA_NO_NODE)
1482 gfp |= __GFP_THISNODE;
1484 * Make sure to call something that can handle
1487 return alloc_pages_node(nid, gfp, order);
1491 * OK, so we have a VMA. Fetch the mempolicy and try to
1492 * allocate a huge page with it. We will only reach this
1493 * when CONFIG_NUMA=y.
1497 struct mempolicy *mpol;
1498 struct zonelist *zl;
1499 nodemask_t *nodemask;
1501 cpuset_mems_cookie = read_mems_allowed_begin();
1502 zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask);
1503 mpol_cond_put(mpol);
1504 page = __alloc_pages_nodemask(gfp, order, zl, nodemask);
1507 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1513 * There are two ways to allocate a huge page:
1514 * 1. When you have a VMA and an address (like a fault)
1515 * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1517 * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in
1518 * this case which signifies that the allocation should be done with
1519 * respect for the VMA's memory policy.
1521 * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1522 * implies that memory policies will not be taken in to account.
1524 static struct page *__alloc_buddy_huge_page(struct hstate *h,
1525 struct vm_area_struct *vma, unsigned long addr, int nid)
1530 if (hstate_is_gigantic(h))
1534 * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1535 * This makes sure the caller is picking _one_ of the modes with which
1536 * we can call this function, not both.
1538 if (vma || (addr != -1)) {
1539 VM_WARN_ON_ONCE(addr == -1);
1540 VM_WARN_ON_ONCE(nid != NUMA_NO_NODE);
1543 * Assume we will successfully allocate the surplus page to
1544 * prevent racing processes from causing the surplus to exceed
1547 * This however introduces a different race, where a process B
1548 * tries to grow the static hugepage pool while alloc_pages() is
1549 * called by process A. B will only examine the per-node
1550 * counters in determining if surplus huge pages can be
1551 * converted to normal huge pages in adjust_pool_surplus(). A
1552 * won't be able to increment the per-node counter, until the
1553 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1554 * no more huge pages can be converted from surplus to normal
1555 * state (and doesn't try to convert again). Thus, we have a
1556 * case where a surplus huge page exists, the pool is grown, and
1557 * the surplus huge page still exists after, even though it
1558 * should just have been converted to a normal huge page. This
1559 * does not leak memory, though, as the hugepage will be freed
1560 * once it is out of use. It also does not allow the counters to
1561 * go out of whack in adjust_pool_surplus() as we don't modify
1562 * the node values until we've gotten the hugepage and only the
1563 * per-node value is checked there.
1565 spin_lock(&hugetlb_lock);
1566 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1567 spin_unlock(&hugetlb_lock);
1571 h->surplus_huge_pages++;
1573 spin_unlock(&hugetlb_lock);
1575 page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid);
1577 spin_lock(&hugetlb_lock);
1579 INIT_LIST_HEAD(&page->lru);
1580 r_nid = page_to_nid(page);
1581 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1582 set_hugetlb_cgroup(page, NULL);
1584 * We incremented the global counters already
1586 h->nr_huge_pages_node[r_nid]++;
1587 h->surplus_huge_pages_node[r_nid]++;
1588 __count_vm_event(HTLB_BUDDY_PGALLOC);
1591 h->surplus_huge_pages--;
1592 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1594 spin_unlock(&hugetlb_lock);
1600 * Allocate a huge page from 'nid'. Note, 'nid' may be
1601 * NUMA_NO_NODE, which means that it may be allocated
1605 struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid)
1607 unsigned long addr = -1;
1609 return __alloc_buddy_huge_page(h, NULL, addr, nid);
1613 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1616 struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1617 struct vm_area_struct *vma, unsigned long addr)
1619 return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE);
1623 * This allocation function is useful in the context where vma is irrelevant.
1624 * E.g. soft-offlining uses this function because it only cares physical
1625 * address of error page.
1627 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1629 struct page *page = NULL;
1631 spin_lock(&hugetlb_lock);
1632 if (h->free_huge_pages - h->resv_huge_pages > 0)
1633 page = dequeue_huge_page_node(h, nid);
1634 spin_unlock(&hugetlb_lock);
1637 page = __alloc_buddy_huge_page_no_mpol(h, nid);
1643 * Increase the hugetlb pool such that it can accommodate a reservation
1646 static int gather_surplus_pages(struct hstate *h, int delta)
1648 struct list_head surplus_list;
1649 struct page *page, *tmp;
1651 int needed, allocated;
1652 bool alloc_ok = true;
1654 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1656 h->resv_huge_pages += delta;
1661 INIT_LIST_HEAD(&surplus_list);
1665 spin_unlock(&hugetlb_lock);
1666 for (i = 0; i < needed; i++) {
1667 page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE);
1672 list_add(&page->lru, &surplus_list);
1677 * After retaking hugetlb_lock, we need to recalculate 'needed'
1678 * because either resv_huge_pages or free_huge_pages may have changed.
1680 spin_lock(&hugetlb_lock);
1681 needed = (h->resv_huge_pages + delta) -
1682 (h->free_huge_pages + allocated);
1687 * We were not able to allocate enough pages to
1688 * satisfy the entire reservation so we free what
1689 * we've allocated so far.
1694 * The surplus_list now contains _at_least_ the number of extra pages
1695 * needed to accommodate the reservation. Add the appropriate number
1696 * of pages to the hugetlb pool and free the extras back to the buddy
1697 * allocator. Commit the entire reservation here to prevent another
1698 * process from stealing the pages as they are added to the pool but
1699 * before they are reserved.
1701 needed += allocated;
1702 h->resv_huge_pages += delta;
1705 /* Free the needed pages to the hugetlb pool */
1706 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1710 * This page is now managed by the hugetlb allocator and has
1711 * no users -- drop the buddy allocator's reference.
1713 put_page_testzero(page);
1714 VM_BUG_ON_PAGE(page_count(page), page);
1715 enqueue_huge_page(h, page);
1718 spin_unlock(&hugetlb_lock);
1720 /* Free unnecessary surplus pages to the buddy allocator */
1721 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1723 spin_lock(&hugetlb_lock);
1729 * When releasing a hugetlb pool reservation, any surplus pages that were
1730 * allocated to satisfy the reservation must be explicitly freed if they were
1732 * Called with hugetlb_lock held.
1734 static void return_unused_surplus_pages(struct hstate *h,
1735 unsigned long unused_resv_pages)
1737 unsigned long nr_pages;
1739 /* Uncommit the reservation */
1740 h->resv_huge_pages -= unused_resv_pages;
1742 /* Cannot return gigantic pages currently */
1743 if (hstate_is_gigantic(h))
1746 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1749 * We want to release as many surplus pages as possible, spread
1750 * evenly across all nodes with memory. Iterate across these nodes
1751 * until we can no longer free unreserved surplus pages. This occurs
1752 * when the nodes with surplus pages have no free pages.
1753 * free_pool_huge_page() will balance the the freed pages across the
1754 * on-line nodes with memory and will handle the hstate accounting.
1756 while (nr_pages--) {
1757 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1759 cond_resched_lock(&hugetlb_lock);
1765 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1766 * are used by the huge page allocation routines to manage reservations.
1768 * vma_needs_reservation is called to determine if the huge page at addr
1769 * within the vma has an associated reservation. If a reservation is
1770 * needed, the value 1 is returned. The caller is then responsible for
1771 * managing the global reservation and subpool usage counts. After
1772 * the huge page has been allocated, vma_commit_reservation is called
1773 * to add the page to the reservation map. If the page allocation fails,
1774 * the reservation must be ended instead of committed. vma_end_reservation
1775 * is called in such cases.
1777 * In the normal case, vma_commit_reservation returns the same value
1778 * as the preceding vma_needs_reservation call. The only time this
1779 * is not the case is if a reserve map was changed between calls. It
1780 * is the responsibility of the caller to notice the difference and
1781 * take appropriate action.
1783 enum vma_resv_mode {
1788 static long __vma_reservation_common(struct hstate *h,
1789 struct vm_area_struct *vma, unsigned long addr,
1790 enum vma_resv_mode mode)
1792 struct resv_map *resv;
1796 resv = vma_resv_map(vma);
1800 idx = vma_hugecache_offset(h, vma, addr);
1802 case VMA_NEEDS_RESV:
1803 ret = region_chg(resv, idx, idx + 1);
1805 case VMA_COMMIT_RESV:
1806 ret = region_add(resv, idx, idx + 1);
1809 region_abort(resv, idx, idx + 1);
1816 if (vma->vm_flags & VM_MAYSHARE)
1819 return ret < 0 ? ret : 0;
1822 static long vma_needs_reservation(struct hstate *h,
1823 struct vm_area_struct *vma, unsigned long addr)
1825 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1828 static long vma_commit_reservation(struct hstate *h,
1829 struct vm_area_struct *vma, unsigned long addr)
1831 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1834 static void vma_end_reservation(struct hstate *h,
1835 struct vm_area_struct *vma, unsigned long addr)
1837 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1840 struct page *alloc_huge_page(struct vm_area_struct *vma,
1841 unsigned long addr, int avoid_reserve)
1843 struct hugepage_subpool *spool = subpool_vma(vma);
1844 struct hstate *h = hstate_vma(vma);
1846 long map_chg, map_commit;
1849 struct hugetlb_cgroup *h_cg;
1851 idx = hstate_index(h);
1853 * Examine the region/reserve map to determine if the process
1854 * has a reservation for the page to be allocated. A return
1855 * code of zero indicates a reservation exists (no change).
1857 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
1859 return ERR_PTR(-ENOMEM);
1862 * Processes that did not create the mapping will have no
1863 * reserves as indicated by the region/reserve map. Check
1864 * that the allocation will not exceed the subpool limit.
1865 * Allocations for MAP_NORESERVE mappings also need to be
1866 * checked against any subpool limit.
1868 if (map_chg || avoid_reserve) {
1869 gbl_chg = hugepage_subpool_get_pages(spool, 1);
1871 vma_end_reservation(h, vma, addr);
1872 return ERR_PTR(-ENOSPC);
1876 * Even though there was no reservation in the region/reserve
1877 * map, there could be reservations associated with the
1878 * subpool that can be used. This would be indicated if the
1879 * return value of hugepage_subpool_get_pages() is zero.
1880 * However, if avoid_reserve is specified we still avoid even
1881 * the subpool reservations.
1887 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1889 goto out_subpool_put;
1891 spin_lock(&hugetlb_lock);
1893 * glb_chg is passed to indicate whether or not a page must be taken
1894 * from the global free pool (global change). gbl_chg == 0 indicates
1895 * a reservation exists for the allocation.
1897 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
1899 spin_unlock(&hugetlb_lock);
1900 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
1902 goto out_uncharge_cgroup;
1903 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
1904 SetPagePrivate(page);
1905 h->resv_huge_pages--;
1907 spin_lock(&hugetlb_lock);
1908 list_move(&page->lru, &h->hugepage_activelist);
1911 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1912 spin_unlock(&hugetlb_lock);
1914 set_page_private(page, (unsigned long)spool);
1916 map_commit = vma_commit_reservation(h, vma, addr);
1917 if (unlikely(map_chg > map_commit)) {
1919 * The page was added to the reservation map between
1920 * vma_needs_reservation and vma_commit_reservation.
1921 * This indicates a race with hugetlb_reserve_pages.
1922 * Adjust for the subpool count incremented above AND
1923 * in hugetlb_reserve_pages for the same page. Also,
1924 * the reservation count added in hugetlb_reserve_pages
1925 * no longer applies.
1929 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
1930 hugetlb_acct_memory(h, -rsv_adjust);
1934 out_uncharge_cgroup:
1935 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1937 if (map_chg || avoid_reserve)
1938 hugepage_subpool_put_pages(spool, 1);
1939 vma_end_reservation(h, vma, addr);
1940 return ERR_PTR(-ENOSPC);
1944 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1945 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1946 * where no ERR_VALUE is expected to be returned.
1948 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1949 unsigned long addr, int avoid_reserve)
1951 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1957 int __weak alloc_bootmem_huge_page(struct hstate *h)
1959 struct huge_bootmem_page *m;
1962 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1965 addr = memblock_virt_alloc_try_nid_nopanic(
1966 huge_page_size(h), huge_page_size(h),
1967 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1970 * Use the beginning of the huge page to store the
1971 * huge_bootmem_page struct (until gather_bootmem
1972 * puts them into the mem_map).
1981 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
1982 /* Put them into a private list first because mem_map is not up yet */
1983 list_add(&m->list, &huge_boot_pages);
1988 static void __init prep_compound_huge_page(struct page *page,
1991 if (unlikely(order > (MAX_ORDER - 1)))
1992 prep_compound_gigantic_page(page, order);
1994 prep_compound_page(page, order);
1997 /* Put bootmem huge pages into the standard lists after mem_map is up */
1998 static void __init gather_bootmem_prealloc(void)
2000 struct huge_bootmem_page *m;
2002 list_for_each_entry(m, &huge_boot_pages, list) {
2003 struct hstate *h = m->hstate;
2006 #ifdef CONFIG_HIGHMEM
2007 page = pfn_to_page(m->phys >> PAGE_SHIFT);
2008 memblock_free_late(__pa(m),
2009 sizeof(struct huge_bootmem_page));
2011 page = virt_to_page(m);
2013 WARN_ON(page_count(page) != 1);
2014 prep_compound_huge_page(page, h->order);
2015 WARN_ON(PageReserved(page));
2016 prep_new_huge_page(h, page, page_to_nid(page));
2018 * If we had gigantic hugepages allocated at boot time, we need
2019 * to restore the 'stolen' pages to totalram_pages in order to
2020 * fix confusing memory reports from free(1) and another
2021 * side-effects, like CommitLimit going negative.
2023 if (hstate_is_gigantic(h))
2024 adjust_managed_page_count(page, 1 << h->order);
2028 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2032 for (i = 0; i < h->max_huge_pages; ++i) {
2033 if (hstate_is_gigantic(h)) {
2034 if (!alloc_bootmem_huge_page(h))
2036 } else if (!alloc_fresh_huge_page(h,
2037 &node_states[N_MEMORY]))
2040 h->max_huge_pages = i;
2043 static void __init hugetlb_init_hstates(void)
2047 for_each_hstate(h) {
2048 if (minimum_order > huge_page_order(h))
2049 minimum_order = huge_page_order(h);
2051 /* oversize hugepages were init'ed in early boot */
2052 if (!hstate_is_gigantic(h))
2053 hugetlb_hstate_alloc_pages(h);
2055 VM_BUG_ON(minimum_order == UINT_MAX);
2058 static char * __init memfmt(char *buf, unsigned long n)
2060 if (n >= (1UL << 30))
2061 sprintf(buf, "%lu GB", n >> 30);
2062 else if (n >= (1UL << 20))
2063 sprintf(buf, "%lu MB", n >> 20);
2065 sprintf(buf, "%lu KB", n >> 10);
2069 static void __init report_hugepages(void)
2073 for_each_hstate(h) {
2075 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2076 memfmt(buf, huge_page_size(h)),
2077 h->free_huge_pages);
2081 #ifdef CONFIG_HIGHMEM
2082 static void try_to_free_low(struct hstate *h, unsigned long count,
2083 nodemask_t *nodes_allowed)
2087 if (hstate_is_gigantic(h))
2090 for_each_node_mask(i, *nodes_allowed) {
2091 struct page *page, *next;
2092 struct list_head *freel = &h->hugepage_freelists[i];
2093 list_for_each_entry_safe(page, next, freel, lru) {
2094 if (count >= h->nr_huge_pages)
2096 if (PageHighMem(page))
2098 list_del(&page->lru);
2099 update_and_free_page(h, page);
2100 h->free_huge_pages--;
2101 h->free_huge_pages_node[page_to_nid(page)]--;
2106 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2107 nodemask_t *nodes_allowed)
2113 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2114 * balanced by operating on them in a round-robin fashion.
2115 * Returns 1 if an adjustment was made.
2117 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2122 VM_BUG_ON(delta != -1 && delta != 1);
2125 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2126 if (h->surplus_huge_pages_node[node])
2130 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2131 if (h->surplus_huge_pages_node[node] <
2132 h->nr_huge_pages_node[node])
2139 h->surplus_huge_pages += delta;
2140 h->surplus_huge_pages_node[node] += delta;
2144 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2145 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2146 nodemask_t *nodes_allowed)
2148 unsigned long min_count, ret;
2150 if (hstate_is_gigantic(h) && !gigantic_page_supported())
2151 return h->max_huge_pages;
2154 * Increase the pool size
2155 * First take pages out of surplus state. Then make up the
2156 * remaining difference by allocating fresh huge pages.
2158 * We might race with __alloc_buddy_huge_page() here and be unable
2159 * to convert a surplus huge page to a normal huge page. That is
2160 * not critical, though, it just means the overall size of the
2161 * pool might be one hugepage larger than it needs to be, but
2162 * within all the constraints specified by the sysctls.
2164 spin_lock(&hugetlb_lock);
2165 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2166 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2170 while (count > persistent_huge_pages(h)) {
2172 * If this allocation races such that we no longer need the
2173 * page, free_huge_page will handle it by freeing the page
2174 * and reducing the surplus.
2176 spin_unlock(&hugetlb_lock);
2177 if (hstate_is_gigantic(h))
2178 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2180 ret = alloc_fresh_huge_page(h, nodes_allowed);
2181 spin_lock(&hugetlb_lock);
2185 /* Bail for signals. Probably ctrl-c from user */
2186 if (signal_pending(current))
2191 * Decrease the pool size
2192 * First return free pages to the buddy allocator (being careful
2193 * to keep enough around to satisfy reservations). Then place
2194 * pages into surplus state as needed so the pool will shrink
2195 * to the desired size as pages become free.
2197 * By placing pages into the surplus state independent of the
2198 * overcommit value, we are allowing the surplus pool size to
2199 * exceed overcommit. There are few sane options here. Since
2200 * __alloc_buddy_huge_page() is checking the global counter,
2201 * though, we'll note that we're not allowed to exceed surplus
2202 * and won't grow the pool anywhere else. Not until one of the
2203 * sysctls are changed, or the surplus pages go out of use.
2205 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2206 min_count = max(count, min_count);
2207 try_to_free_low(h, min_count, nodes_allowed);
2208 while (min_count < persistent_huge_pages(h)) {
2209 if (!free_pool_huge_page(h, nodes_allowed, 0))
2211 cond_resched_lock(&hugetlb_lock);
2213 while (count < persistent_huge_pages(h)) {
2214 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2218 ret = persistent_huge_pages(h);
2219 spin_unlock(&hugetlb_lock);
2223 #define HSTATE_ATTR_RO(_name) \
2224 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2226 #define HSTATE_ATTR(_name) \
2227 static struct kobj_attribute _name##_attr = \
2228 __ATTR(_name, 0644, _name##_show, _name##_store)
2230 static struct kobject *hugepages_kobj;
2231 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2233 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2235 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2239 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2240 if (hstate_kobjs[i] == kobj) {
2242 *nidp = NUMA_NO_NODE;
2246 return kobj_to_node_hstate(kobj, nidp);
2249 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2250 struct kobj_attribute *attr, char *buf)
2253 unsigned long nr_huge_pages;
2256 h = kobj_to_hstate(kobj, &nid);
2257 if (nid == NUMA_NO_NODE)
2258 nr_huge_pages = h->nr_huge_pages;
2260 nr_huge_pages = h->nr_huge_pages_node[nid];
2262 return sprintf(buf, "%lu\n", nr_huge_pages);
2265 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2266 struct hstate *h, int nid,
2267 unsigned long count, size_t len)
2270 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2272 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2277 if (nid == NUMA_NO_NODE) {
2279 * global hstate attribute
2281 if (!(obey_mempolicy &&
2282 init_nodemask_of_mempolicy(nodes_allowed))) {
2283 NODEMASK_FREE(nodes_allowed);
2284 nodes_allowed = &node_states[N_MEMORY];
2286 } else if (nodes_allowed) {
2288 * per node hstate attribute: adjust count to global,
2289 * but restrict alloc/free to the specified node.
2291 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2292 init_nodemask_of_node(nodes_allowed, nid);
2294 nodes_allowed = &node_states[N_MEMORY];
2296 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2298 if (nodes_allowed != &node_states[N_MEMORY])
2299 NODEMASK_FREE(nodes_allowed);
2303 NODEMASK_FREE(nodes_allowed);
2307 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2308 struct kobject *kobj, const char *buf,
2312 unsigned long count;
2316 err = kstrtoul(buf, 10, &count);
2320 h = kobj_to_hstate(kobj, &nid);
2321 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2324 static ssize_t nr_hugepages_show(struct kobject *kobj,
2325 struct kobj_attribute *attr, char *buf)
2327 return nr_hugepages_show_common(kobj, attr, buf);
2330 static ssize_t nr_hugepages_store(struct kobject *kobj,
2331 struct kobj_attribute *attr, const char *buf, size_t len)
2333 return nr_hugepages_store_common(false, kobj, buf, len);
2335 HSTATE_ATTR(nr_hugepages);
2340 * hstate attribute for optionally mempolicy-based constraint on persistent
2341 * huge page alloc/free.
2343 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2344 struct kobj_attribute *attr, char *buf)
2346 return nr_hugepages_show_common(kobj, attr, buf);
2349 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2350 struct kobj_attribute *attr, const char *buf, size_t len)
2352 return nr_hugepages_store_common(true, kobj, buf, len);
2354 HSTATE_ATTR(nr_hugepages_mempolicy);
2358 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2359 struct kobj_attribute *attr, char *buf)
2361 struct hstate *h = kobj_to_hstate(kobj, NULL);
2362 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2365 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2366 struct kobj_attribute *attr, const char *buf, size_t count)
2369 unsigned long input;
2370 struct hstate *h = kobj_to_hstate(kobj, NULL);
2372 if (hstate_is_gigantic(h))
2375 err = kstrtoul(buf, 10, &input);
2379 spin_lock(&hugetlb_lock);
2380 h->nr_overcommit_huge_pages = input;
2381 spin_unlock(&hugetlb_lock);
2385 HSTATE_ATTR(nr_overcommit_hugepages);
2387 static ssize_t free_hugepages_show(struct kobject *kobj,
2388 struct kobj_attribute *attr, char *buf)
2391 unsigned long free_huge_pages;
2394 h = kobj_to_hstate(kobj, &nid);
2395 if (nid == NUMA_NO_NODE)
2396 free_huge_pages = h->free_huge_pages;
2398 free_huge_pages = h->free_huge_pages_node[nid];
2400 return sprintf(buf, "%lu\n", free_huge_pages);
2402 HSTATE_ATTR_RO(free_hugepages);
2404 static ssize_t resv_hugepages_show(struct kobject *kobj,
2405 struct kobj_attribute *attr, char *buf)
2407 struct hstate *h = kobj_to_hstate(kobj, NULL);
2408 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2410 HSTATE_ATTR_RO(resv_hugepages);
2412 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2413 struct kobj_attribute *attr, char *buf)
2416 unsigned long surplus_huge_pages;
2419 h = kobj_to_hstate(kobj, &nid);
2420 if (nid == NUMA_NO_NODE)
2421 surplus_huge_pages = h->surplus_huge_pages;
2423 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2425 return sprintf(buf, "%lu\n", surplus_huge_pages);
2427 HSTATE_ATTR_RO(surplus_hugepages);
2429 static struct attribute *hstate_attrs[] = {
2430 &nr_hugepages_attr.attr,
2431 &nr_overcommit_hugepages_attr.attr,
2432 &free_hugepages_attr.attr,
2433 &resv_hugepages_attr.attr,
2434 &surplus_hugepages_attr.attr,
2436 &nr_hugepages_mempolicy_attr.attr,
2441 static struct attribute_group hstate_attr_group = {
2442 .attrs = hstate_attrs,
2445 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2446 struct kobject **hstate_kobjs,
2447 struct attribute_group *hstate_attr_group)
2450 int hi = hstate_index(h);
2452 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2453 if (!hstate_kobjs[hi])
2456 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2458 kobject_put(hstate_kobjs[hi]);
2463 static void __init hugetlb_sysfs_init(void)
2468 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2469 if (!hugepages_kobj)
2472 for_each_hstate(h) {
2473 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2474 hstate_kobjs, &hstate_attr_group);
2476 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2483 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2484 * with node devices in node_devices[] using a parallel array. The array
2485 * index of a node device or _hstate == node id.
2486 * This is here to avoid any static dependency of the node device driver, in
2487 * the base kernel, on the hugetlb module.
2489 struct node_hstate {
2490 struct kobject *hugepages_kobj;
2491 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2493 static struct node_hstate node_hstates[MAX_NUMNODES];
2496 * A subset of global hstate attributes for node devices
2498 static struct attribute *per_node_hstate_attrs[] = {
2499 &nr_hugepages_attr.attr,
2500 &free_hugepages_attr.attr,
2501 &surplus_hugepages_attr.attr,
2505 static struct attribute_group per_node_hstate_attr_group = {
2506 .attrs = per_node_hstate_attrs,
2510 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2511 * Returns node id via non-NULL nidp.
2513 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2517 for (nid = 0; nid < nr_node_ids; nid++) {
2518 struct node_hstate *nhs = &node_hstates[nid];
2520 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2521 if (nhs->hstate_kobjs[i] == kobj) {
2533 * Unregister hstate attributes from a single node device.
2534 * No-op if no hstate attributes attached.
2536 static void hugetlb_unregister_node(struct node *node)
2539 struct node_hstate *nhs = &node_hstates[node->dev.id];
2541 if (!nhs->hugepages_kobj)
2542 return; /* no hstate attributes */
2544 for_each_hstate(h) {
2545 int idx = hstate_index(h);
2546 if (nhs->hstate_kobjs[idx]) {
2547 kobject_put(nhs->hstate_kobjs[idx]);
2548 nhs->hstate_kobjs[idx] = NULL;
2552 kobject_put(nhs->hugepages_kobj);
2553 nhs->hugepages_kobj = NULL;
2558 * Register hstate attributes for a single node device.
2559 * No-op if attributes already registered.
2561 static void hugetlb_register_node(struct node *node)
2564 struct node_hstate *nhs = &node_hstates[node->dev.id];
2567 if (nhs->hugepages_kobj)
2568 return; /* already allocated */
2570 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2572 if (!nhs->hugepages_kobj)
2575 for_each_hstate(h) {
2576 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2578 &per_node_hstate_attr_group);
2580 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2581 h->name, node->dev.id);
2582 hugetlb_unregister_node(node);
2589 * hugetlb init time: register hstate attributes for all registered node
2590 * devices of nodes that have memory. All on-line nodes should have
2591 * registered their associated device by this time.
2593 static void __init hugetlb_register_all_nodes(void)
2597 for_each_node_state(nid, N_MEMORY) {
2598 struct node *node = node_devices[nid];
2599 if (node->dev.id == nid)
2600 hugetlb_register_node(node);
2604 * Let the node device driver know we're here so it can
2605 * [un]register hstate attributes on node hotplug.
2607 register_hugetlbfs_with_node(hugetlb_register_node,
2608 hugetlb_unregister_node);
2610 #else /* !CONFIG_NUMA */
2612 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2620 static void hugetlb_register_all_nodes(void) { }
2624 static int __init hugetlb_init(void)
2628 if (!hugepages_supported())
2631 if (!size_to_hstate(default_hstate_size)) {
2632 default_hstate_size = HPAGE_SIZE;
2633 if (!size_to_hstate(default_hstate_size))
2634 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2636 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2637 if (default_hstate_max_huge_pages) {
2638 if (!default_hstate.max_huge_pages)
2639 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2642 hugetlb_init_hstates();
2643 gather_bootmem_prealloc();
2646 hugetlb_sysfs_init();
2647 hugetlb_register_all_nodes();
2648 hugetlb_cgroup_file_init();
2651 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2653 num_fault_mutexes = 1;
2655 hugetlb_fault_mutex_table =
2656 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2657 BUG_ON(!hugetlb_fault_mutex_table);
2659 for (i = 0; i < num_fault_mutexes; i++)
2660 mutex_init(&hugetlb_fault_mutex_table[i]);
2663 subsys_initcall(hugetlb_init);
2665 /* Should be called on processing a hugepagesz=... option */
2666 void __init hugetlb_bad_size(void)
2668 parsed_valid_hugepagesz = false;
2671 void __init hugetlb_add_hstate(unsigned int order)
2676 if (size_to_hstate(PAGE_SIZE << order)) {
2677 pr_warn("hugepagesz= specified twice, ignoring\n");
2680 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2682 h = &hstates[hugetlb_max_hstate++];
2684 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2685 h->nr_huge_pages = 0;
2686 h->free_huge_pages = 0;
2687 for (i = 0; i < MAX_NUMNODES; ++i)
2688 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2689 INIT_LIST_HEAD(&h->hugepage_activelist);
2690 h->next_nid_to_alloc = first_memory_node;
2691 h->next_nid_to_free = first_memory_node;
2692 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2693 huge_page_size(h)/1024);
2698 static int __init hugetlb_nrpages_setup(char *s)
2701 static unsigned long *last_mhp;
2703 if (!parsed_valid_hugepagesz) {
2704 pr_warn("hugepages = %s preceded by "
2705 "an unsupported hugepagesz, ignoring\n", s);
2706 parsed_valid_hugepagesz = true;
2710 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2711 * so this hugepages= parameter goes to the "default hstate".
2713 else if (!hugetlb_max_hstate)
2714 mhp = &default_hstate_max_huge_pages;
2716 mhp = &parsed_hstate->max_huge_pages;
2718 if (mhp == last_mhp) {
2719 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2723 if (sscanf(s, "%lu", mhp) <= 0)
2727 * Global state is always initialized later in hugetlb_init.
2728 * But we need to allocate >= MAX_ORDER hstates here early to still
2729 * use the bootmem allocator.
2731 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2732 hugetlb_hstate_alloc_pages(parsed_hstate);
2738 __setup("hugepages=", hugetlb_nrpages_setup);
2740 static int __init hugetlb_default_setup(char *s)
2742 default_hstate_size = memparse(s, &s);
2745 __setup("default_hugepagesz=", hugetlb_default_setup);
2747 static unsigned int cpuset_mems_nr(unsigned int *array)
2750 unsigned int nr = 0;
2752 for_each_node_mask(node, cpuset_current_mems_allowed)
2758 #ifdef CONFIG_SYSCTL
2759 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2760 struct ctl_table *table, int write,
2761 void __user *buffer, size_t *length, loff_t *ppos)
2763 struct hstate *h = &default_hstate;
2764 unsigned long tmp = h->max_huge_pages;
2767 if (!hugepages_supported())
2771 table->maxlen = sizeof(unsigned long);
2772 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2777 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2778 NUMA_NO_NODE, tmp, *length);
2783 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2784 void __user *buffer, size_t *length, loff_t *ppos)
2787 return hugetlb_sysctl_handler_common(false, table, write,
2788 buffer, length, ppos);
2792 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2793 void __user *buffer, size_t *length, loff_t *ppos)
2795 return hugetlb_sysctl_handler_common(true, table, write,
2796 buffer, length, ppos);
2798 #endif /* CONFIG_NUMA */
2800 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2801 void __user *buffer,
2802 size_t *length, loff_t *ppos)
2804 struct hstate *h = &default_hstate;
2808 if (!hugepages_supported())
2811 tmp = h->nr_overcommit_huge_pages;
2813 if (write && hstate_is_gigantic(h))
2817 table->maxlen = sizeof(unsigned long);
2818 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2823 spin_lock(&hugetlb_lock);
2824 h->nr_overcommit_huge_pages = tmp;
2825 spin_unlock(&hugetlb_lock);
2831 #endif /* CONFIG_SYSCTL */
2833 void hugetlb_report_meminfo(struct seq_file *m)
2835 struct hstate *h = &default_hstate;
2836 if (!hugepages_supported())
2839 "HugePages_Total: %5lu\n"
2840 "HugePages_Free: %5lu\n"
2841 "HugePages_Rsvd: %5lu\n"
2842 "HugePages_Surp: %5lu\n"
2843 "Hugepagesize: %8lu kB\n",
2847 h->surplus_huge_pages,
2848 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2851 int hugetlb_report_node_meminfo(int nid, char *buf)
2853 struct hstate *h = &default_hstate;
2854 if (!hugepages_supported())
2857 "Node %d HugePages_Total: %5u\n"
2858 "Node %d HugePages_Free: %5u\n"
2859 "Node %d HugePages_Surp: %5u\n",
2860 nid, h->nr_huge_pages_node[nid],
2861 nid, h->free_huge_pages_node[nid],
2862 nid, h->surplus_huge_pages_node[nid]);
2865 void hugetlb_show_meminfo(void)
2870 if (!hugepages_supported())
2873 for_each_node_state(nid, N_MEMORY)
2875 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2877 h->nr_huge_pages_node[nid],
2878 h->free_huge_pages_node[nid],
2879 h->surplus_huge_pages_node[nid],
2880 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2883 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
2885 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
2886 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
2889 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2890 unsigned long hugetlb_total_pages(void)
2893 unsigned long nr_total_pages = 0;
2896 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2897 return nr_total_pages;
2900 static int hugetlb_acct_memory(struct hstate *h, long delta)
2904 spin_lock(&hugetlb_lock);
2906 * When cpuset is configured, it breaks the strict hugetlb page
2907 * reservation as the accounting is done on a global variable. Such
2908 * reservation is completely rubbish in the presence of cpuset because
2909 * the reservation is not checked against page availability for the
2910 * current cpuset. Application can still potentially OOM'ed by kernel
2911 * with lack of free htlb page in cpuset that the task is in.
2912 * Attempt to enforce strict accounting with cpuset is almost
2913 * impossible (or too ugly) because cpuset is too fluid that
2914 * task or memory node can be dynamically moved between cpusets.
2916 * The change of semantics for shared hugetlb mapping with cpuset is
2917 * undesirable. However, in order to preserve some of the semantics,
2918 * we fall back to check against current free page availability as
2919 * a best attempt and hopefully to minimize the impact of changing
2920 * semantics that cpuset has.
2923 if (gather_surplus_pages(h, delta) < 0)
2926 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2927 return_unused_surplus_pages(h, delta);
2934 return_unused_surplus_pages(h, (unsigned long) -delta);
2937 spin_unlock(&hugetlb_lock);
2941 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2943 struct resv_map *resv = vma_resv_map(vma);
2946 * This new VMA should share its siblings reservation map if present.
2947 * The VMA will only ever have a valid reservation map pointer where
2948 * it is being copied for another still existing VMA. As that VMA
2949 * has a reference to the reservation map it cannot disappear until
2950 * after this open call completes. It is therefore safe to take a
2951 * new reference here without additional locking.
2953 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2954 kref_get(&resv->refs);
2957 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2959 struct hstate *h = hstate_vma(vma);
2960 struct resv_map *resv = vma_resv_map(vma);
2961 struct hugepage_subpool *spool = subpool_vma(vma);
2962 unsigned long reserve, start, end;
2965 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2968 start = vma_hugecache_offset(h, vma, vma->vm_start);
2969 end = vma_hugecache_offset(h, vma, vma->vm_end);
2971 reserve = (end - start) - region_count(resv, start, end);
2973 kref_put(&resv->refs, resv_map_release);
2977 * Decrement reserve counts. The global reserve count may be
2978 * adjusted if the subpool has a minimum size.
2980 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
2981 hugetlb_acct_memory(h, -gbl_reserve);
2986 * We cannot handle pagefaults against hugetlb pages at all. They cause
2987 * handle_mm_fault() to try to instantiate regular-sized pages in the
2988 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2991 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2997 const struct vm_operations_struct hugetlb_vm_ops = {
2998 .fault = hugetlb_vm_op_fault,
2999 .open = hugetlb_vm_op_open,
3000 .close = hugetlb_vm_op_close,
3003 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3009 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3010 vma->vm_page_prot)));
3012 entry = huge_pte_wrprotect(mk_huge_pte(page,
3013 vma->vm_page_prot));
3015 entry = pte_mkyoung(entry);
3016 entry = pte_mkhuge(entry);
3017 entry = arch_make_huge_pte(entry, vma, page, writable);
3022 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3023 unsigned long address, pte_t *ptep)
3027 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3028 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3029 update_mmu_cache(vma, address, ptep);
3032 static int is_hugetlb_entry_migration(pte_t pte)
3036 if (huge_pte_none(pte) || pte_present(pte))
3038 swp = pte_to_swp_entry(pte);
3039 if (non_swap_entry(swp) && is_migration_entry(swp))
3045 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3049 if (huge_pte_none(pte) || pte_present(pte))
3051 swp = pte_to_swp_entry(pte);
3052 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3058 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3059 struct vm_area_struct *vma)
3061 pte_t *src_pte, *dst_pte, entry;
3062 struct page *ptepage;
3065 struct hstate *h = hstate_vma(vma);
3066 unsigned long sz = huge_page_size(h);
3067 unsigned long mmun_start; /* For mmu_notifiers */
3068 unsigned long mmun_end; /* For mmu_notifiers */
3071 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3073 mmun_start = vma->vm_start;
3074 mmun_end = vma->vm_end;
3076 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3078 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3079 spinlock_t *src_ptl, *dst_ptl;
3080 src_pte = huge_pte_offset(src, addr);
3083 dst_pte = huge_pte_alloc(dst, addr, sz);
3089 /* If the pagetables are shared don't copy or take references */
3090 if (dst_pte == src_pte)
3093 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3094 src_ptl = huge_pte_lockptr(h, src, src_pte);
3095 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3096 entry = huge_ptep_get(src_pte);
3097 if (huge_pte_none(entry)) { /* skip none entry */
3099 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3100 is_hugetlb_entry_hwpoisoned(entry))) {
3101 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3103 if (is_write_migration_entry(swp_entry) && cow) {
3105 * COW mappings require pages in both
3106 * parent and child to be set to read.
3108 make_migration_entry_read(&swp_entry);
3109 entry = swp_entry_to_pte(swp_entry);
3110 set_huge_pte_at(src, addr, src_pte, entry);
3112 set_huge_pte_at(dst, addr, dst_pte, entry);
3115 huge_ptep_set_wrprotect(src, addr, src_pte);
3116 mmu_notifier_invalidate_range(src, mmun_start,
3119 entry = huge_ptep_get(src_pte);
3120 ptepage = pte_page(entry);
3122 page_dup_rmap(ptepage, true);
3123 set_huge_pte_at(dst, addr, dst_pte, entry);
3124 hugetlb_count_add(pages_per_huge_page(h), dst);
3126 spin_unlock(src_ptl);
3127 spin_unlock(dst_ptl);
3131 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3136 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3137 unsigned long start, unsigned long end,
3138 struct page *ref_page)
3140 int force_flush = 0;
3141 struct mm_struct *mm = vma->vm_mm;
3142 unsigned long address;
3147 struct hstate *h = hstate_vma(vma);
3148 unsigned long sz = huge_page_size(h);
3149 const unsigned long mmun_start = start; /* For mmu_notifiers */
3150 const unsigned long mmun_end = end; /* For mmu_notifiers */
3152 WARN_ON(!is_vm_hugetlb_page(vma));
3153 BUG_ON(start & ~huge_page_mask(h));
3154 BUG_ON(end & ~huge_page_mask(h));
3156 tlb_start_vma(tlb, vma);
3157 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3160 for (; address < end; address += sz) {
3161 ptep = huge_pte_offset(mm, address);
3165 ptl = huge_pte_lock(h, mm, ptep);
3166 if (huge_pmd_unshare(mm, &address, ptep))
3169 pte = huge_ptep_get(ptep);
3170 if (huge_pte_none(pte))
3174 * Migrating hugepage or HWPoisoned hugepage is already
3175 * unmapped and its refcount is dropped, so just clear pte here.
3177 if (unlikely(!pte_present(pte))) {
3178 huge_pte_clear(mm, address, ptep);
3182 page = pte_page(pte);
3184 * If a reference page is supplied, it is because a specific
3185 * page is being unmapped, not a range. Ensure the page we
3186 * are about to unmap is the actual page of interest.
3189 if (page != ref_page)
3193 * Mark the VMA as having unmapped its page so that
3194 * future faults in this VMA will fail rather than
3195 * looking like data was lost
3197 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3200 pte = huge_ptep_get_and_clear(mm, address, ptep);
3201 tlb_remove_tlb_entry(tlb, ptep, address);
3202 if (huge_pte_dirty(pte))
3203 set_page_dirty(page);
3205 hugetlb_count_sub(pages_per_huge_page(h), mm);
3206 page_remove_rmap(page, true);
3207 force_flush = !__tlb_remove_page(tlb, page);
3213 /* Bail out after unmapping reference page if supplied */
3222 * mmu_gather ran out of room to batch pages, we break out of
3223 * the PTE lock to avoid doing the potential expensive TLB invalidate
3224 * and page-free while holding it.
3229 if (address < end && !ref_page)
3232 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3233 tlb_end_vma(tlb, vma);
3236 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3237 struct vm_area_struct *vma, unsigned long start,
3238 unsigned long end, struct page *ref_page)
3240 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3243 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3244 * test will fail on a vma being torn down, and not grab a page table
3245 * on its way out. We're lucky that the flag has such an appropriate
3246 * name, and can in fact be safely cleared here. We could clear it
3247 * before the __unmap_hugepage_range above, but all that's necessary
3248 * is to clear it before releasing the i_mmap_rwsem. This works
3249 * because in the context this is called, the VMA is about to be
3250 * destroyed and the i_mmap_rwsem is held.
3252 vma->vm_flags &= ~VM_MAYSHARE;
3255 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3256 unsigned long end, struct page *ref_page)
3258 struct mm_struct *mm;
3259 struct mmu_gather tlb;
3263 tlb_gather_mmu(&tlb, mm, start, end);
3264 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3265 tlb_finish_mmu(&tlb, start, end);
3269 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3270 * mappping it owns the reserve page for. The intention is to unmap the page
3271 * from other VMAs and let the children be SIGKILLed if they are faulting the
3274 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3275 struct page *page, unsigned long address)
3277 struct hstate *h = hstate_vma(vma);
3278 struct vm_area_struct *iter_vma;
3279 struct address_space *mapping;
3283 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3284 * from page cache lookup which is in HPAGE_SIZE units.
3286 address = address & huge_page_mask(h);
3287 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3289 mapping = file_inode(vma->vm_file)->i_mapping;
3292 * Take the mapping lock for the duration of the table walk. As
3293 * this mapping should be shared between all the VMAs,
3294 * __unmap_hugepage_range() is called as the lock is already held
3296 i_mmap_lock_write(mapping);
3297 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3298 /* Do not unmap the current VMA */
3299 if (iter_vma == vma)
3303 * Shared VMAs have their own reserves and do not affect
3304 * MAP_PRIVATE accounting but it is possible that a shared
3305 * VMA is using the same page so check and skip such VMAs.
3307 if (iter_vma->vm_flags & VM_MAYSHARE)
3311 * Unmap the page from other VMAs without their own reserves.
3312 * They get marked to be SIGKILLed if they fault in these
3313 * areas. This is because a future no-page fault on this VMA
3314 * could insert a zeroed page instead of the data existing
3315 * from the time of fork. This would look like data corruption
3317 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3318 unmap_hugepage_range(iter_vma, address,
3319 address + huge_page_size(h), page);
3321 i_mmap_unlock_write(mapping);
3325 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3326 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3327 * cannot race with other handlers or page migration.
3328 * Keep the pte_same checks anyway to make transition from the mutex easier.
3330 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3331 unsigned long address, pte_t *ptep, pte_t pte,
3332 struct page *pagecache_page, spinlock_t *ptl)
3334 struct hstate *h = hstate_vma(vma);
3335 struct page *old_page, *new_page;
3336 int ret = 0, outside_reserve = 0;
3337 unsigned long mmun_start; /* For mmu_notifiers */
3338 unsigned long mmun_end; /* For mmu_notifiers */
3340 old_page = pte_page(pte);
3343 /* If no-one else is actually using this page, avoid the copy
3344 * and just make the page writable */
3345 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3346 page_move_anon_rmap(old_page, vma, address);
3347 set_huge_ptep_writable(vma, address, ptep);
3352 * If the process that created a MAP_PRIVATE mapping is about to
3353 * perform a COW due to a shared page count, attempt to satisfy
3354 * the allocation without using the existing reserves. The pagecache
3355 * page is used to determine if the reserve at this address was
3356 * consumed or not. If reserves were used, a partial faulted mapping
3357 * at the time of fork() could consume its reserves on COW instead
3358 * of the full address range.
3360 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3361 old_page != pagecache_page)
3362 outside_reserve = 1;
3367 * Drop page table lock as buddy allocator may be called. It will
3368 * be acquired again before returning to the caller, as expected.
3371 new_page = alloc_huge_page(vma, address, outside_reserve);
3373 if (IS_ERR(new_page)) {
3375 * If a process owning a MAP_PRIVATE mapping fails to COW,
3376 * it is due to references held by a child and an insufficient
3377 * huge page pool. To guarantee the original mappers
3378 * reliability, unmap the page from child processes. The child
3379 * may get SIGKILLed if it later faults.
3381 if (outside_reserve) {
3383 BUG_ON(huge_pte_none(pte));
3384 unmap_ref_private(mm, vma, old_page, address);
3385 BUG_ON(huge_pte_none(pte));
3387 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3389 pte_same(huge_ptep_get(ptep), pte)))
3390 goto retry_avoidcopy;
3392 * race occurs while re-acquiring page table
3393 * lock, and our job is done.
3398 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3399 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3400 goto out_release_old;
3404 * When the original hugepage is shared one, it does not have
3405 * anon_vma prepared.
3407 if (unlikely(anon_vma_prepare(vma))) {
3409 goto out_release_all;
3412 copy_user_huge_page(new_page, old_page, address, vma,
3413 pages_per_huge_page(h));
3414 __SetPageUptodate(new_page);
3415 set_page_huge_active(new_page);
3417 mmun_start = address & huge_page_mask(h);
3418 mmun_end = mmun_start + huge_page_size(h);
3419 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3422 * Retake the page table lock to check for racing updates
3423 * before the page tables are altered
3426 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3427 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3428 ClearPagePrivate(new_page);
3431 huge_ptep_clear_flush(vma, address, ptep);
3432 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3433 set_huge_pte_at(mm, address, ptep,
3434 make_huge_pte(vma, new_page, 1));
3435 page_remove_rmap(old_page, true);
3436 hugepage_add_new_anon_rmap(new_page, vma, address);
3437 /* Make the old page be freed below */
3438 new_page = old_page;
3441 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3447 spin_lock(ptl); /* Caller expects lock to be held */
3451 /* Return the pagecache page at a given address within a VMA */
3452 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3453 struct vm_area_struct *vma, unsigned long address)
3455 struct address_space *mapping;
3458 mapping = vma->vm_file->f_mapping;
3459 idx = vma_hugecache_offset(h, vma, address);
3461 return find_lock_page(mapping, idx);
3465 * Return whether there is a pagecache page to back given address within VMA.
3466 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3468 static bool hugetlbfs_pagecache_present(struct hstate *h,
3469 struct vm_area_struct *vma, unsigned long address)
3471 struct address_space *mapping;
3475 mapping = vma->vm_file->f_mapping;
3476 idx = vma_hugecache_offset(h, vma, address);
3478 page = find_get_page(mapping, idx);
3481 return page != NULL;
3484 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3487 struct inode *inode = mapping->host;
3488 struct hstate *h = hstate_inode(inode);
3489 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3493 ClearPagePrivate(page);
3495 spin_lock(&inode->i_lock);
3496 inode->i_blocks += blocks_per_huge_page(h);
3497 spin_unlock(&inode->i_lock);
3501 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3502 struct address_space *mapping, pgoff_t idx,
3503 unsigned long address, pte_t *ptep, unsigned int flags)
3505 struct hstate *h = hstate_vma(vma);
3506 int ret = VM_FAULT_SIGBUS;
3514 * Currently, we are forced to kill the process in the event the
3515 * original mapper has unmapped pages from the child due to a failed
3516 * COW. Warn that such a situation has occurred as it may not be obvious
3518 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3519 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3525 * Use page lock to guard against racing truncation
3526 * before we get page_table_lock.
3529 page = find_lock_page(mapping, idx);
3531 size = i_size_read(mapping->host) >> huge_page_shift(h);
3534 page = alloc_huge_page(vma, address, 0);
3536 ret = PTR_ERR(page);
3540 ret = VM_FAULT_SIGBUS;
3543 clear_huge_page(page, address, pages_per_huge_page(h));
3544 __SetPageUptodate(page);
3545 set_page_huge_active(page);
3547 if (vma->vm_flags & VM_MAYSHARE) {
3548 int err = huge_add_to_page_cache(page, mapping, idx);
3557 if (unlikely(anon_vma_prepare(vma))) {
3559 goto backout_unlocked;
3565 * If memory error occurs between mmap() and fault, some process
3566 * don't have hwpoisoned swap entry for errored virtual address.
3567 * So we need to block hugepage fault by PG_hwpoison bit check.
3569 if (unlikely(PageHWPoison(page))) {
3570 ret = VM_FAULT_HWPOISON |
3571 VM_FAULT_SET_HINDEX(hstate_index(h));
3572 goto backout_unlocked;
3577 * If we are going to COW a private mapping later, we examine the
3578 * pending reservations for this page now. This will ensure that
3579 * any allocations necessary to record that reservation occur outside
3582 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3583 if (vma_needs_reservation(h, vma, address) < 0) {
3585 goto backout_unlocked;
3587 /* Just decrements count, does not deallocate */
3588 vma_end_reservation(h, vma, address);
3591 ptl = huge_pte_lockptr(h, mm, ptep);
3593 size = i_size_read(mapping->host) >> huge_page_shift(h);
3598 if (!huge_pte_none(huge_ptep_get(ptep)))
3602 ClearPagePrivate(page);
3603 hugepage_add_new_anon_rmap(page, vma, address);
3605 page_dup_rmap(page, true);
3606 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3607 && (vma->vm_flags & VM_SHARED)));
3608 set_huge_pte_at(mm, address, ptep, new_pte);
3610 hugetlb_count_add(pages_per_huge_page(h), mm);
3611 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3612 /* Optimization, do the COW without a second fault */
3613 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
3630 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3631 struct vm_area_struct *vma,
3632 struct address_space *mapping,
3633 pgoff_t idx, unsigned long address)
3635 unsigned long key[2];
3638 if (vma->vm_flags & VM_SHARED) {
3639 key[0] = (unsigned long) mapping;
3642 key[0] = (unsigned long) mm;
3643 key[1] = address >> huge_page_shift(h);
3646 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3648 return hash & (num_fault_mutexes - 1);
3652 * For uniprocesor systems we always use a single mutex, so just
3653 * return 0 and avoid the hashing overhead.
3655 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3656 struct vm_area_struct *vma,
3657 struct address_space *mapping,
3658 pgoff_t idx, unsigned long address)
3664 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3665 unsigned long address, unsigned int flags)
3672 struct page *page = NULL;
3673 struct page *pagecache_page = NULL;
3674 struct hstate *h = hstate_vma(vma);
3675 struct address_space *mapping;
3676 int need_wait_lock = 0;
3678 address &= huge_page_mask(h);
3680 ptep = huge_pte_offset(mm, address);
3682 entry = huge_ptep_get(ptep);
3683 if (unlikely(is_hugetlb_entry_migration(entry))) {
3684 migration_entry_wait_huge(vma, mm, ptep);
3686 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3687 return VM_FAULT_HWPOISON_LARGE |
3688 VM_FAULT_SET_HINDEX(hstate_index(h));
3690 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3692 return VM_FAULT_OOM;
3695 mapping = vma->vm_file->f_mapping;
3696 idx = vma_hugecache_offset(h, vma, address);
3699 * Serialize hugepage allocation and instantiation, so that we don't
3700 * get spurious allocation failures if two CPUs race to instantiate
3701 * the same page in the page cache.
3703 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3704 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3706 entry = huge_ptep_get(ptep);
3707 if (huge_pte_none(entry)) {
3708 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3715 * entry could be a migration/hwpoison entry at this point, so this
3716 * check prevents the kernel from going below assuming that we have
3717 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3718 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3721 if (!pte_present(entry))
3725 * If we are going to COW the mapping later, we examine the pending
3726 * reservations for this page now. This will ensure that any
3727 * allocations necessary to record that reservation occur outside the
3728 * spinlock. For private mappings, we also lookup the pagecache
3729 * page now as it is used to determine if a reservation has been
3732 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3733 if (vma_needs_reservation(h, vma, address) < 0) {
3737 /* Just decrements count, does not deallocate */
3738 vma_end_reservation(h, vma, address);
3740 if (!(vma->vm_flags & VM_MAYSHARE))
3741 pagecache_page = hugetlbfs_pagecache_page(h,
3745 ptl = huge_pte_lock(h, mm, ptep);
3747 /* Check for a racing update before calling hugetlb_cow */
3748 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3752 * hugetlb_cow() requires page locks of pte_page(entry) and
3753 * pagecache_page, so here we need take the former one
3754 * when page != pagecache_page or !pagecache_page.
3756 page = pte_page(entry);
3757 if (page != pagecache_page)
3758 if (!trylock_page(page)) {
3765 if (flags & FAULT_FLAG_WRITE) {
3766 if (!huge_pte_write(entry)) {
3767 ret = hugetlb_cow(mm, vma, address, ptep, entry,
3768 pagecache_page, ptl);
3771 entry = huge_pte_mkdirty(entry);
3773 entry = pte_mkyoung(entry);
3774 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3775 flags & FAULT_FLAG_WRITE))
3776 update_mmu_cache(vma, address, ptep);
3778 if (page != pagecache_page)
3784 if (pagecache_page) {
3785 unlock_page(pagecache_page);
3786 put_page(pagecache_page);
3789 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3791 * Generally it's safe to hold refcount during waiting page lock. But
3792 * here we just wait to defer the next page fault to avoid busy loop and
3793 * the page is not used after unlocked before returning from the current
3794 * page fault. So we are safe from accessing freed page, even if we wait
3795 * here without taking refcount.
3798 wait_on_page_locked(page);
3802 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3803 struct page **pages, struct vm_area_struct **vmas,
3804 unsigned long *position, unsigned long *nr_pages,
3805 long i, unsigned int flags)
3807 unsigned long pfn_offset;
3808 unsigned long vaddr = *position;
3809 unsigned long remainder = *nr_pages;
3810 struct hstate *h = hstate_vma(vma);
3812 while (vaddr < vma->vm_end && remainder) {
3814 spinlock_t *ptl = NULL;
3819 * If we have a pending SIGKILL, don't keep faulting pages and
3820 * potentially allocating memory.
3822 if (unlikely(fatal_signal_pending(current))) {
3828 * Some archs (sparc64, sh*) have multiple pte_ts to
3829 * each hugepage. We have to make sure we get the
3830 * first, for the page indexing below to work.
3832 * Note that page table lock is not held when pte is null.
3834 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3836 ptl = huge_pte_lock(h, mm, pte);
3837 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3840 * When coredumping, it suits get_dump_page if we just return
3841 * an error where there's an empty slot with no huge pagecache
3842 * to back it. This way, we avoid allocating a hugepage, and
3843 * the sparse dumpfile avoids allocating disk blocks, but its
3844 * huge holes still show up with zeroes where they need to be.
3846 if (absent && (flags & FOLL_DUMP) &&
3847 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3855 * We need call hugetlb_fault for both hugepages under migration
3856 * (in which case hugetlb_fault waits for the migration,) and
3857 * hwpoisoned hugepages (in which case we need to prevent the
3858 * caller from accessing to them.) In order to do this, we use
3859 * here is_swap_pte instead of is_hugetlb_entry_migration and
3860 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3861 * both cases, and because we can't follow correct pages
3862 * directly from any kind of swap entries.
3864 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3865 ((flags & FOLL_WRITE) &&
3866 !huge_pte_write(huge_ptep_get(pte)))) {
3871 ret = hugetlb_fault(mm, vma, vaddr,
3872 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3873 if (!(ret & VM_FAULT_ERROR))
3880 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3881 page = pte_page(huge_ptep_get(pte));
3884 pages[i] = mem_map_offset(page, pfn_offset);
3895 if (vaddr < vma->vm_end && remainder &&
3896 pfn_offset < pages_per_huge_page(h)) {
3898 * We use pfn_offset to avoid touching the pageframes
3899 * of this compound page.
3905 *nr_pages = remainder;
3908 return i ? i : -EFAULT;
3911 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3912 unsigned long address, unsigned long end, pgprot_t newprot)
3914 struct mm_struct *mm = vma->vm_mm;
3915 unsigned long start = address;
3918 struct hstate *h = hstate_vma(vma);
3919 unsigned long pages = 0;
3921 BUG_ON(address >= end);
3922 flush_cache_range(vma, address, end);
3924 mmu_notifier_invalidate_range_start(mm, start, end);
3925 i_mmap_lock_write(vma->vm_file->f_mapping);
3926 for (; address < end; address += huge_page_size(h)) {
3928 ptep = huge_pte_offset(mm, address);
3931 ptl = huge_pte_lock(h, mm, ptep);
3932 if (huge_pmd_unshare(mm, &address, ptep)) {
3937 pte = huge_ptep_get(ptep);
3938 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3942 if (unlikely(is_hugetlb_entry_migration(pte))) {
3943 swp_entry_t entry = pte_to_swp_entry(pte);
3945 if (is_write_migration_entry(entry)) {
3948 make_migration_entry_read(&entry);
3949 newpte = swp_entry_to_pte(entry);
3950 set_huge_pte_at(mm, address, ptep, newpte);
3956 if (!huge_pte_none(pte)) {
3957 pte = huge_ptep_get_and_clear(mm, address, ptep);
3958 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3959 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3960 set_huge_pte_at(mm, address, ptep, pte);
3966 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
3967 * may have cleared our pud entry and done put_page on the page table:
3968 * once we release i_mmap_rwsem, another task can do the final put_page
3969 * and that page table be reused and filled with junk.
3971 flush_tlb_range(vma, start, end);
3972 mmu_notifier_invalidate_range(mm, start, end);
3973 i_mmap_unlock_write(vma->vm_file->f_mapping);
3974 mmu_notifier_invalidate_range_end(mm, start, end);
3976 return pages << h->order;
3979 int hugetlb_reserve_pages(struct inode *inode,
3981 struct vm_area_struct *vma,
3982 vm_flags_t vm_flags)
3985 struct hstate *h = hstate_inode(inode);
3986 struct hugepage_subpool *spool = subpool_inode(inode);
3987 struct resv_map *resv_map;
3991 * Only apply hugepage reservation if asked. At fault time, an
3992 * attempt will be made for VM_NORESERVE to allocate a page
3993 * without using reserves
3995 if (vm_flags & VM_NORESERVE)
3999 * Shared mappings base their reservation on the number of pages that
4000 * are already allocated on behalf of the file. Private mappings need
4001 * to reserve the full area even if read-only as mprotect() may be
4002 * called to make the mapping read-write. Assume !vma is a shm mapping
4004 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4005 resv_map = inode_resv_map(inode);
4007 chg = region_chg(resv_map, from, to);
4010 resv_map = resv_map_alloc();
4016 set_vma_resv_map(vma, resv_map);
4017 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4026 * There must be enough pages in the subpool for the mapping. If
4027 * the subpool has a minimum size, there may be some global
4028 * reservations already in place (gbl_reserve).
4030 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4031 if (gbl_reserve < 0) {
4037 * Check enough hugepages are available for the reservation.
4038 * Hand the pages back to the subpool if there are not
4040 ret = hugetlb_acct_memory(h, gbl_reserve);
4042 /* put back original number of pages, chg */
4043 (void)hugepage_subpool_put_pages(spool, chg);
4048 * Account for the reservations made. Shared mappings record regions
4049 * that have reservations as they are shared by multiple VMAs.
4050 * When the last VMA disappears, the region map says how much
4051 * the reservation was and the page cache tells how much of
4052 * the reservation was consumed. Private mappings are per-VMA and
4053 * only the consumed reservations are tracked. When the VMA
4054 * disappears, the original reservation is the VMA size and the
4055 * consumed reservations are stored in the map. Hence, nothing
4056 * else has to be done for private mappings here
4058 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4059 long add = region_add(resv_map, from, to);
4061 if (unlikely(chg > add)) {
4063 * pages in this range were added to the reserve
4064 * map between region_chg and region_add. This
4065 * indicates a race with alloc_huge_page. Adjust
4066 * the subpool and reserve counts modified above
4067 * based on the difference.
4071 rsv_adjust = hugepage_subpool_put_pages(spool,
4073 hugetlb_acct_memory(h, -rsv_adjust);
4078 if (!vma || vma->vm_flags & VM_MAYSHARE)
4079 region_abort(resv_map, from, to);
4080 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4081 kref_put(&resv_map->refs, resv_map_release);
4085 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4088 struct hstate *h = hstate_inode(inode);
4089 struct resv_map *resv_map = inode_resv_map(inode);
4091 struct hugepage_subpool *spool = subpool_inode(inode);
4095 chg = region_del(resv_map, start, end);
4097 * region_del() can fail in the rare case where a region
4098 * must be split and another region descriptor can not be
4099 * allocated. If end == LONG_MAX, it will not fail.
4105 spin_lock(&inode->i_lock);
4106 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4107 spin_unlock(&inode->i_lock);
4110 * If the subpool has a minimum size, the number of global
4111 * reservations to be released may be adjusted.
4113 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4114 hugetlb_acct_memory(h, -gbl_reserve);
4119 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4120 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4121 struct vm_area_struct *vma,
4122 unsigned long addr, pgoff_t idx)
4124 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4126 unsigned long sbase = saddr & PUD_MASK;
4127 unsigned long s_end = sbase + PUD_SIZE;
4129 /* Allow segments to share if only one is marked locked */
4130 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4131 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4134 * match the virtual addresses, permission and the alignment of the
4137 if (pmd_index(addr) != pmd_index(saddr) ||
4138 vm_flags != svm_flags ||
4139 sbase < svma->vm_start || svma->vm_end < s_end)
4145 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4147 unsigned long base = addr & PUD_MASK;
4148 unsigned long end = base + PUD_SIZE;
4151 * check on proper vm_flags and page table alignment
4153 if (vma->vm_flags & VM_MAYSHARE &&
4154 vma->vm_start <= base && end <= vma->vm_end)
4160 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4161 * and returns the corresponding pte. While this is not necessary for the
4162 * !shared pmd case because we can allocate the pmd later as well, it makes the
4163 * code much cleaner. pmd allocation is essential for the shared case because
4164 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4165 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4166 * bad pmd for sharing.
4168 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4170 struct vm_area_struct *vma = find_vma(mm, addr);
4171 struct address_space *mapping = vma->vm_file->f_mapping;
4172 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4174 struct vm_area_struct *svma;
4175 unsigned long saddr;
4180 if (!vma_shareable(vma, addr))
4181 return (pte_t *)pmd_alloc(mm, pud, addr);
4183 i_mmap_lock_write(mapping);
4184 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4188 saddr = page_table_shareable(svma, vma, addr, idx);
4190 spte = huge_pte_offset(svma->vm_mm, saddr);
4193 get_page(virt_to_page(spte));
4202 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
4204 if (pud_none(*pud)) {
4205 pud_populate(mm, pud,
4206 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4208 put_page(virt_to_page(spte));
4213 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4214 i_mmap_unlock_write(mapping);
4219 * unmap huge page backed by shared pte.
4221 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4222 * indicated by page_count > 1, unmap is achieved by clearing pud and
4223 * decrementing the ref count. If count == 1, the pte page is not shared.
4225 * called with page table lock held.
4227 * returns: 1 successfully unmapped a shared pte page
4228 * 0 the underlying pte page is not shared, or it is the last user
4230 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4232 pgd_t *pgd = pgd_offset(mm, *addr);
4233 pud_t *pud = pud_offset(pgd, *addr);
4235 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4236 if (page_count(virt_to_page(ptep)) == 1)
4240 put_page(virt_to_page(ptep));
4242 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4245 #define want_pmd_share() (1)
4246 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4247 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4252 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4256 #define want_pmd_share() (0)
4257 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4259 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4260 pte_t *huge_pte_alloc(struct mm_struct *mm,
4261 unsigned long addr, unsigned long sz)
4267 pgd = pgd_offset(mm, addr);
4268 pud = pud_alloc(mm, pgd, addr);
4270 if (sz == PUD_SIZE) {
4273 BUG_ON(sz != PMD_SIZE);
4274 if (want_pmd_share() && pud_none(*pud))
4275 pte = huge_pmd_share(mm, addr, pud);
4277 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4280 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
4285 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
4291 pgd = pgd_offset(mm, addr);
4292 if (pgd_present(*pgd)) {
4293 pud = pud_offset(pgd, addr);
4294 if (pud_present(*pud)) {
4296 return (pte_t *)pud;
4297 pmd = pmd_offset(pud, addr);
4300 return (pte_t *) pmd;
4303 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4306 * These functions are overwritable if your architecture needs its own
4309 struct page * __weak
4310 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4313 return ERR_PTR(-EINVAL);
4316 struct page * __weak
4317 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4318 pmd_t *pmd, int flags)
4320 struct page *page = NULL;
4323 ptl = pmd_lockptr(mm, pmd);
4326 * make sure that the address range covered by this pmd is not
4327 * unmapped from other threads.
4329 if (!pmd_huge(*pmd))
4331 if (pmd_present(*pmd)) {
4332 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4333 if (flags & FOLL_GET)
4336 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
4338 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4342 * hwpoisoned entry is treated as no_page_table in
4343 * follow_page_mask().
4351 struct page * __weak
4352 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4353 pud_t *pud, int flags)
4355 if (flags & FOLL_GET)
4358 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4361 #ifdef CONFIG_MEMORY_FAILURE
4364 * This function is called from memory failure code.
4365 * Assume the caller holds page lock of the head page.
4367 int dequeue_hwpoisoned_huge_page(struct page *hpage)
4369 struct hstate *h = page_hstate(hpage);
4370 int nid = page_to_nid(hpage);
4373 spin_lock(&hugetlb_lock);
4375 * Just checking !page_huge_active is not enough, because that could be
4376 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4378 if (!page_huge_active(hpage) && !page_count(hpage)) {
4380 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4381 * but dangling hpage->lru can trigger list-debug warnings
4382 * (this happens when we call unpoison_memory() on it),
4383 * so let it point to itself with list_del_init().
4385 list_del_init(&hpage->lru);
4386 set_page_refcounted(hpage);
4387 h->free_huge_pages--;
4388 h->free_huge_pages_node[nid]--;
4391 spin_unlock(&hugetlb_lock);
4396 bool isolate_huge_page(struct page *page, struct list_head *list)
4400 VM_BUG_ON_PAGE(!PageHead(page), page);
4401 spin_lock(&hugetlb_lock);
4402 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4406 clear_page_huge_active(page);
4407 list_move_tail(&page->lru, list);
4409 spin_unlock(&hugetlb_lock);
4413 void putback_active_hugepage(struct page *page)
4415 VM_BUG_ON_PAGE(!PageHead(page), page);
4416 spin_lock(&hugetlb_lock);
4417 set_page_huge_active(page);
4418 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4419 spin_unlock(&hugetlb_lock);