2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
8 #include <linux/seq_file.h>
9 #include <linux/sysctl.h>
10 #include <linux/highmem.h>
11 #include <linux/mmu_notifier.h>
12 #include <linux/nodemask.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/compiler.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/mmdebug.h>
22 #include <linux/sched/signal.h>
23 #include <linux/rmap.h>
24 #include <linux/string_helpers.h>
25 #include <linux/swap.h>
26 #include <linux/swapops.h>
27 #include <linux/jhash.h>
30 #include <asm/pgtable.h>
34 #include <linux/hugetlb.h>
35 #include <linux/hugetlb_cgroup.h>
36 #include <linux/node.h>
37 #include <linux/userfaultfd_k.h>
40 int hugepages_treat_as_movable;
42 int hugetlb_max_hstate __read_mostly;
43 unsigned int default_hstate_idx;
44 struct hstate hstates[HUGE_MAX_HSTATE];
46 * Minimum page order among possible hugepage sizes, set to a proper value
49 static unsigned int minimum_order __read_mostly = UINT_MAX;
51 __initdata LIST_HEAD(huge_boot_pages);
53 /* for command line parsing */
54 static struct hstate * __initdata parsed_hstate;
55 static unsigned long __initdata default_hstate_max_huge_pages;
56 static unsigned long __initdata default_hstate_size;
57 static bool __initdata parsed_valid_hugepagesz = true;
60 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
61 * free_huge_pages, and surplus_huge_pages.
63 DEFINE_SPINLOCK(hugetlb_lock);
66 * Serializes faults on the same logical page. This is used to
67 * prevent spurious OOMs when the hugepage pool is fully utilized.
69 static int num_fault_mutexes;
70 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
72 static inline bool PageHugeFreed(struct page *head)
74 return page_private(head + 4) == -1UL;
77 static inline void SetPageHugeFreed(struct page *head)
79 set_page_private(head + 4, -1UL);
82 static inline void ClearPageHugeFreed(struct page *head)
84 set_page_private(head + 4, 0);
87 /* Forward declaration */
88 static int hugetlb_acct_memory(struct hstate *h, long delta);
90 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
92 bool free = (spool->count == 0) && (spool->used_hpages == 0);
94 spin_unlock(&spool->lock);
96 /* If no pages are used, and no other handles to the subpool
97 * remain, give up any reservations mased on minimum size and
100 if (spool->min_hpages != -1)
101 hugetlb_acct_memory(spool->hstate,
107 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
110 struct hugepage_subpool *spool;
112 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
116 spin_lock_init(&spool->lock);
118 spool->max_hpages = max_hpages;
120 spool->min_hpages = min_hpages;
122 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
126 spool->rsv_hpages = min_hpages;
131 void hugepage_put_subpool(struct hugepage_subpool *spool)
133 spin_lock(&spool->lock);
134 BUG_ON(!spool->count);
136 unlock_or_release_subpool(spool);
140 * Subpool accounting for allocating and reserving pages.
141 * Return -ENOMEM if there are not enough resources to satisfy the
142 * the request. Otherwise, return the number of pages by which the
143 * global pools must be adjusted (upward). The returned value may
144 * only be different than the passed value (delta) in the case where
145 * a subpool minimum size must be manitained.
147 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
155 spin_lock(&spool->lock);
157 if (spool->max_hpages != -1) { /* maximum size accounting */
158 if ((spool->used_hpages + delta) <= spool->max_hpages)
159 spool->used_hpages += delta;
166 /* minimum size accounting */
167 if (spool->min_hpages != -1 && spool->rsv_hpages) {
168 if (delta > spool->rsv_hpages) {
170 * Asking for more reserves than those already taken on
171 * behalf of subpool. Return difference.
173 ret = delta - spool->rsv_hpages;
174 spool->rsv_hpages = 0;
176 ret = 0; /* reserves already accounted for */
177 spool->rsv_hpages -= delta;
182 spin_unlock(&spool->lock);
187 * Subpool accounting for freeing and unreserving pages.
188 * Return the number of global page reservations that must be dropped.
189 * The return value may only be different than the passed value (delta)
190 * in the case where a subpool minimum size must be maintained.
192 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
200 spin_lock(&spool->lock);
202 if (spool->max_hpages != -1) /* maximum size accounting */
203 spool->used_hpages -= delta;
205 /* minimum size accounting */
206 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
207 if (spool->rsv_hpages + delta <= spool->min_hpages)
210 ret = spool->rsv_hpages + delta - spool->min_hpages;
212 spool->rsv_hpages += delta;
213 if (spool->rsv_hpages > spool->min_hpages)
214 spool->rsv_hpages = spool->min_hpages;
218 * If hugetlbfs_put_super couldn't free spool due to an outstanding
219 * quota reference, free it now.
221 unlock_or_release_subpool(spool);
226 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
228 return HUGETLBFS_SB(inode->i_sb)->spool;
231 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
233 return subpool_inode(file_inode(vma->vm_file));
237 * Region tracking -- allows tracking of reservations and instantiated pages
238 * across the pages in a mapping.
240 * The region data structures are embedded into a resv_map and protected
241 * by a resv_map's lock. The set of regions within the resv_map represent
242 * reservations for huge pages, or huge pages that have already been
243 * instantiated within the map. The from and to elements are huge page
244 * indicies into the associated mapping. from indicates the starting index
245 * of the region. to represents the first index past the end of the region.
247 * For example, a file region structure with from == 0 and to == 4 represents
248 * four huge pages in a mapping. It is important to note that the to element
249 * represents the first element past the end of the region. This is used in
250 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
252 * Interval notation of the form [from, to) will be used to indicate that
253 * the endpoint from is inclusive and to is exclusive.
256 struct list_head link;
262 * Add the huge page range represented by [f, t) to the reserve
263 * map. In the normal case, existing regions will be expanded
264 * to accommodate the specified range. Sufficient regions should
265 * exist for expansion due to the previous call to region_chg
266 * with the same range. However, it is possible that region_del
267 * could have been called after region_chg and modifed the map
268 * in such a way that no region exists to be expanded. In this
269 * case, pull a region descriptor from the cache associated with
270 * the map and use that for the new range.
272 * Return the number of new huge pages added to the map. This
273 * number is greater than or equal to zero.
275 static long region_add(struct resv_map *resv, long f, long t)
277 struct list_head *head = &resv->regions;
278 struct file_region *rg, *nrg, *trg;
281 spin_lock(&resv->lock);
282 /* Locate the region we are either in or before. */
283 list_for_each_entry(rg, head, link)
288 * If no region exists which can be expanded to include the
289 * specified range, the list must have been modified by an
290 * interleving call to region_del(). Pull a region descriptor
291 * from the cache and use it for this range.
293 if (&rg->link == head || t < rg->from) {
294 VM_BUG_ON(resv->region_cache_count <= 0);
296 resv->region_cache_count--;
297 nrg = list_first_entry(&resv->region_cache, struct file_region,
299 list_del(&nrg->link);
303 list_add(&nrg->link, rg->link.prev);
309 /* Round our left edge to the current segment if it encloses us. */
313 /* Check for and consume any regions we now overlap with. */
315 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
316 if (&rg->link == head)
321 /* If this area reaches higher then extend our area to
322 * include it completely. If this is not the first area
323 * which we intend to reuse, free it. */
327 /* Decrement return value by the deleted range.
328 * Another range will span this area so that by
329 * end of routine add will be >= zero
331 add -= (rg->to - rg->from);
337 add += (nrg->from - f); /* Added to beginning of region */
339 add += t - nrg->to; /* Added to end of region */
343 resv->adds_in_progress--;
344 spin_unlock(&resv->lock);
350 * Examine the existing reserve map and determine how many
351 * huge pages in the specified range [f, t) are NOT currently
352 * represented. This routine is called before a subsequent
353 * call to region_add that will actually modify the reserve
354 * map to add the specified range [f, t). region_chg does
355 * not change the number of huge pages represented by the
356 * map. However, if the existing regions in the map can not
357 * be expanded to represent the new range, a new file_region
358 * structure is added to the map as a placeholder. This is
359 * so that the subsequent region_add call will have all the
360 * regions it needs and will not fail.
362 * Upon entry, region_chg will also examine the cache of region descriptors
363 * associated with the map. If there are not enough descriptors cached, one
364 * will be allocated for the in progress add operation.
366 * Returns the number of huge pages that need to be added to the existing
367 * reservation map for the range [f, t). This number is greater or equal to
368 * zero. -ENOMEM is returned if a new file_region structure or cache entry
369 * is needed and can not be allocated.
371 static long region_chg(struct resv_map *resv, long f, long t)
373 struct list_head *head = &resv->regions;
374 struct file_region *rg, *nrg = NULL;
378 spin_lock(&resv->lock);
380 resv->adds_in_progress++;
383 * Check for sufficient descriptors in the cache to accommodate
384 * the number of in progress add operations.
386 if (resv->adds_in_progress > resv->region_cache_count) {
387 struct file_region *trg;
389 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
390 /* Must drop lock to allocate a new descriptor. */
391 resv->adds_in_progress--;
392 spin_unlock(&resv->lock);
394 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
400 spin_lock(&resv->lock);
401 list_add(&trg->link, &resv->region_cache);
402 resv->region_cache_count++;
406 /* Locate the region we are before or in. */
407 list_for_each_entry(rg, head, link)
411 /* If we are below the current region then a new region is required.
412 * Subtle, allocate a new region at the position but make it zero
413 * size such that we can guarantee to record the reservation. */
414 if (&rg->link == head || t < rg->from) {
416 resv->adds_in_progress--;
417 spin_unlock(&resv->lock);
418 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
424 INIT_LIST_HEAD(&nrg->link);
428 list_add(&nrg->link, rg->link.prev);
433 /* Round our left edge to the current segment if it encloses us. */
438 /* Check for and consume any regions we now overlap with. */
439 list_for_each_entry(rg, rg->link.prev, link) {
440 if (&rg->link == head)
445 /* We overlap with this area, if it extends further than
446 * us then we must extend ourselves. Account for its
447 * existing reservation. */
452 chg -= rg->to - rg->from;
456 spin_unlock(&resv->lock);
457 /* We already know we raced and no longer need the new region */
461 spin_unlock(&resv->lock);
466 * Abort the in progress add operation. The adds_in_progress field
467 * of the resv_map keeps track of the operations in progress between
468 * calls to region_chg and region_add. Operations are sometimes
469 * aborted after the call to region_chg. In such cases, region_abort
470 * is called to decrement the adds_in_progress counter.
472 * NOTE: The range arguments [f, t) are not needed or used in this
473 * routine. They are kept to make reading the calling code easier as
474 * arguments will match the associated region_chg call.
476 static void region_abort(struct resv_map *resv, long f, long t)
478 spin_lock(&resv->lock);
479 VM_BUG_ON(!resv->region_cache_count);
480 resv->adds_in_progress--;
481 spin_unlock(&resv->lock);
485 * Delete the specified range [f, t) from the reserve map. If the
486 * t parameter is LONG_MAX, this indicates that ALL regions after f
487 * should be deleted. Locate the regions which intersect [f, t)
488 * and either trim, delete or split the existing regions.
490 * Returns the number of huge pages deleted from the reserve map.
491 * In the normal case, the return value is zero or more. In the
492 * case where a region must be split, a new region descriptor must
493 * be allocated. If the allocation fails, -ENOMEM will be returned.
494 * NOTE: If the parameter t == LONG_MAX, then we will never split
495 * a region and possibly return -ENOMEM. Callers specifying
496 * t == LONG_MAX do not need to check for -ENOMEM error.
498 static long region_del(struct resv_map *resv, long f, long t)
500 struct list_head *head = &resv->regions;
501 struct file_region *rg, *trg;
502 struct file_region *nrg = NULL;
506 spin_lock(&resv->lock);
507 list_for_each_entry_safe(rg, trg, head, link) {
509 * Skip regions before the range to be deleted. file_region
510 * ranges are normally of the form [from, to). However, there
511 * may be a "placeholder" entry in the map which is of the form
512 * (from, to) with from == to. Check for placeholder entries
513 * at the beginning of the range to be deleted.
515 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
521 if (f > rg->from && t < rg->to) { /* Must split region */
523 * Check for an entry in the cache before dropping
524 * lock and attempting allocation.
527 resv->region_cache_count > resv->adds_in_progress) {
528 nrg = list_first_entry(&resv->region_cache,
531 list_del(&nrg->link);
532 resv->region_cache_count--;
536 spin_unlock(&resv->lock);
537 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
545 /* New entry for end of split region */
548 INIT_LIST_HEAD(&nrg->link);
550 /* Original entry is trimmed */
553 list_add(&nrg->link, &rg->link);
558 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
559 del += rg->to - rg->from;
565 if (f <= rg->from) { /* Trim beginning of region */
568 } else { /* Trim end of region */
574 spin_unlock(&resv->lock);
580 * A rare out of memory error was encountered which prevented removal of
581 * the reserve map region for a page. The huge page itself was free'ed
582 * and removed from the page cache. This routine will adjust the subpool
583 * usage count, and the global reserve count if needed. By incrementing
584 * these counts, the reserve map entry which could not be deleted will
585 * appear as a "reserved" entry instead of simply dangling with incorrect
588 void hugetlb_fix_reserve_counts(struct inode *inode)
590 struct hugepage_subpool *spool = subpool_inode(inode);
592 bool reserved = false;
594 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
595 if (rsv_adjust > 0) {
596 struct hstate *h = hstate_inode(inode);
598 if (!hugetlb_acct_memory(h, 1))
600 } else if (!rsv_adjust) {
605 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
609 * Count and return the number of huge pages in the reserve map
610 * that intersect with the range [f, t).
612 static long region_count(struct resv_map *resv, long f, long t)
614 struct list_head *head = &resv->regions;
615 struct file_region *rg;
618 spin_lock(&resv->lock);
619 /* Locate each segment we overlap with, and count that overlap. */
620 list_for_each_entry(rg, head, link) {
629 seg_from = max(rg->from, f);
630 seg_to = min(rg->to, t);
632 chg += seg_to - seg_from;
634 spin_unlock(&resv->lock);
640 * Convert the address within this vma to the page offset within
641 * the mapping, in pagecache page units; huge pages here.
643 static pgoff_t vma_hugecache_offset(struct hstate *h,
644 struct vm_area_struct *vma, unsigned long address)
646 return ((address - vma->vm_start) >> huge_page_shift(h)) +
647 (vma->vm_pgoff >> huge_page_order(h));
650 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
651 unsigned long address)
653 return vma_hugecache_offset(hstate_vma(vma), vma, address);
655 EXPORT_SYMBOL_GPL(linear_hugepage_index);
658 * Return the size of the pages allocated when backing a VMA. In the majority
659 * cases this will be same size as used by the page table entries.
661 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
663 struct hstate *hstate;
665 if (!is_vm_hugetlb_page(vma))
668 hstate = hstate_vma(vma);
670 return 1UL << huge_page_shift(hstate);
672 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
675 * Return the page size being used by the MMU to back a VMA. In the majority
676 * of cases, the page size used by the kernel matches the MMU size. On
677 * architectures where it differs, an architecture-specific version of this
678 * function is required.
680 #ifndef vma_mmu_pagesize
681 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
683 return vma_kernel_pagesize(vma);
688 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
689 * bits of the reservation map pointer, which are always clear due to
692 #define HPAGE_RESV_OWNER (1UL << 0)
693 #define HPAGE_RESV_UNMAPPED (1UL << 1)
694 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
697 * These helpers are used to track how many pages are reserved for
698 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
699 * is guaranteed to have their future faults succeed.
701 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
702 * the reserve counters are updated with the hugetlb_lock held. It is safe
703 * to reset the VMA at fork() time as it is not in use yet and there is no
704 * chance of the global counters getting corrupted as a result of the values.
706 * The private mapping reservation is represented in a subtly different
707 * manner to a shared mapping. A shared mapping has a region map associated
708 * with the underlying file, this region map represents the backing file
709 * pages which have ever had a reservation assigned which this persists even
710 * after the page is instantiated. A private mapping has a region map
711 * associated with the original mmap which is attached to all VMAs which
712 * reference it, this region map represents those offsets which have consumed
713 * reservation ie. where pages have been instantiated.
715 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
717 return (unsigned long)vma->vm_private_data;
720 static void set_vma_private_data(struct vm_area_struct *vma,
723 vma->vm_private_data = (void *)value;
726 struct resv_map *resv_map_alloc(void)
728 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
729 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
731 if (!resv_map || !rg) {
737 kref_init(&resv_map->refs);
738 spin_lock_init(&resv_map->lock);
739 INIT_LIST_HEAD(&resv_map->regions);
741 resv_map->adds_in_progress = 0;
743 INIT_LIST_HEAD(&resv_map->region_cache);
744 list_add(&rg->link, &resv_map->region_cache);
745 resv_map->region_cache_count = 1;
750 void resv_map_release(struct kref *ref)
752 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
753 struct list_head *head = &resv_map->region_cache;
754 struct file_region *rg, *trg;
756 /* Clear out any active regions before we release the map. */
757 region_del(resv_map, 0, LONG_MAX);
759 /* ... and any entries left in the cache */
760 list_for_each_entry_safe(rg, trg, head, link) {
765 VM_BUG_ON(resv_map->adds_in_progress);
770 static inline struct resv_map *inode_resv_map(struct inode *inode)
772 return inode->i_mapping->private_data;
775 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
777 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
778 if (vma->vm_flags & VM_MAYSHARE) {
779 struct address_space *mapping = vma->vm_file->f_mapping;
780 struct inode *inode = mapping->host;
782 return inode_resv_map(inode);
785 return (struct resv_map *)(get_vma_private_data(vma) &
790 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
792 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
793 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
795 set_vma_private_data(vma, (get_vma_private_data(vma) &
796 HPAGE_RESV_MASK) | (unsigned long)map);
799 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
801 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
802 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
804 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
807 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
809 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
811 return (get_vma_private_data(vma) & flag) != 0;
814 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
815 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
817 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
818 if (!(vma->vm_flags & VM_MAYSHARE))
819 vma->vm_private_data = (void *)0;
822 /* Returns true if the VMA has associated reserve pages */
823 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
825 if (vma->vm_flags & VM_NORESERVE) {
827 * This address is already reserved by other process(chg == 0),
828 * so, we should decrement reserved count. Without decrementing,
829 * reserve count remains after releasing inode, because this
830 * allocated page will go into page cache and is regarded as
831 * coming from reserved pool in releasing step. Currently, we
832 * don't have any other solution to deal with this situation
833 * properly, so add work-around here.
835 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
841 /* Shared mappings always use reserves */
842 if (vma->vm_flags & VM_MAYSHARE) {
844 * We know VM_NORESERVE is not set. Therefore, there SHOULD
845 * be a region map for all pages. The only situation where
846 * there is no region map is if a hole was punched via
847 * fallocate. In this case, there really are no reverves to
848 * use. This situation is indicated if chg != 0.
857 * Only the process that called mmap() has reserves for
860 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
862 * Like the shared case above, a hole punch or truncate
863 * could have been performed on the private mapping.
864 * Examine the value of chg to determine if reserves
865 * actually exist or were previously consumed.
866 * Very Subtle - The value of chg comes from a previous
867 * call to vma_needs_reserves(). The reserve map for
868 * private mappings has different (opposite) semantics
869 * than that of shared mappings. vma_needs_reserves()
870 * has already taken this difference in semantics into
871 * account. Therefore, the meaning of chg is the same
872 * as in the shared case above. Code could easily be
873 * combined, but keeping it separate draws attention to
874 * subtle differences.
885 static void enqueue_huge_page(struct hstate *h, struct page *page)
887 int nid = page_to_nid(page);
888 list_move(&page->lru, &h->hugepage_freelists[nid]);
889 h->free_huge_pages++;
890 h->free_huge_pages_node[nid]++;
891 SetPageHugeFreed(page);
894 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
898 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
899 if (!PageHWPoison(page))
902 * if 'non-isolated free hugepage' not found on the list,
903 * the allocation fails.
905 if (&h->hugepage_freelists[nid] == &page->lru)
907 list_move(&page->lru, &h->hugepage_activelist);
908 set_page_refcounted(page);
909 ClearPageHugeFreed(page);
910 h->free_huge_pages--;
911 h->free_huge_pages_node[nid]--;
915 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
918 unsigned int cpuset_mems_cookie;
919 struct zonelist *zonelist;
924 zonelist = node_zonelist(nid, gfp_mask);
927 cpuset_mems_cookie = read_mems_allowed_begin();
928 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
931 if (!cpuset_zone_allowed(zone, gfp_mask))
934 * no need to ask again on the same node. Pool is node rather than
937 if (zone_to_nid(zone) == node)
939 node = zone_to_nid(zone);
941 page = dequeue_huge_page_node_exact(h, node);
945 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
951 /* Movability of hugepages depends on migration support. */
952 static inline gfp_t htlb_alloc_mask(struct hstate *h)
954 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
955 return GFP_HIGHUSER_MOVABLE;
960 static struct page *dequeue_huge_page_vma(struct hstate *h,
961 struct vm_area_struct *vma,
962 unsigned long address, int avoid_reserve,
966 struct mempolicy *mpol;
968 nodemask_t *nodemask;
972 * A child process with MAP_PRIVATE mappings created by their parent
973 * have no page reserves. This check ensures that reservations are
974 * not "stolen". The child may still get SIGKILLed
976 if (!vma_has_reserves(vma, chg) &&
977 h->free_huge_pages - h->resv_huge_pages == 0)
980 /* If reserves cannot be used, ensure enough pages are in the pool */
981 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
984 gfp_mask = htlb_alloc_mask(h);
985 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
986 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
987 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
988 SetPagePrivate(page);
989 h->resv_huge_pages--;
1000 * common helper functions for hstate_next_node_to_{alloc|free}.
1001 * We may have allocated or freed a huge page based on a different
1002 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1003 * be outside of *nodes_allowed. Ensure that we use an allowed
1004 * node for alloc or free.
1006 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1008 nid = next_node_in(nid, *nodes_allowed);
1009 VM_BUG_ON(nid >= MAX_NUMNODES);
1014 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1016 if (!node_isset(nid, *nodes_allowed))
1017 nid = next_node_allowed(nid, nodes_allowed);
1022 * returns the previously saved node ["this node"] from which to
1023 * allocate a persistent huge page for the pool and advance the
1024 * next node from which to allocate, handling wrap at end of node
1027 static int hstate_next_node_to_alloc(struct hstate *h,
1028 nodemask_t *nodes_allowed)
1032 VM_BUG_ON(!nodes_allowed);
1034 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1035 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1041 * helper for free_pool_huge_page() - return the previously saved
1042 * node ["this node"] from which to free a huge page. Advance the
1043 * next node id whether or not we find a free huge page to free so
1044 * that the next attempt to free addresses the next node.
1046 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1050 VM_BUG_ON(!nodes_allowed);
1052 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1053 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1058 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1059 for (nr_nodes = nodes_weight(*mask); \
1061 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1064 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1065 for (nr_nodes = nodes_weight(*mask); \
1067 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1070 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1071 static void destroy_compound_gigantic_page(struct page *page,
1075 int nr_pages = 1 << order;
1076 struct page *p = page + 1;
1078 atomic_set(compound_mapcount_ptr(page), 0);
1079 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1080 clear_compound_head(p);
1081 set_page_refcounted(p);
1084 set_compound_order(page, 0);
1085 __ClearPageHead(page);
1088 static void free_gigantic_page(struct page *page, unsigned int order)
1090 free_contig_range(page_to_pfn(page), 1 << order);
1093 static int __alloc_gigantic_page(unsigned long start_pfn,
1094 unsigned long nr_pages, gfp_t gfp_mask)
1096 unsigned long end_pfn = start_pfn + nr_pages;
1097 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1101 static bool pfn_range_valid_gigantic(struct zone *z,
1102 unsigned long start_pfn, unsigned long nr_pages)
1104 unsigned long i, end_pfn = start_pfn + nr_pages;
1107 for (i = start_pfn; i < end_pfn; i++) {
1108 page = pfn_to_online_page(i);
1112 if (page_zone(page) != z)
1115 if (PageReserved(page))
1118 if (page_count(page) > 0)
1128 static bool zone_spans_last_pfn(const struct zone *zone,
1129 unsigned long start_pfn, unsigned long nr_pages)
1131 unsigned long last_pfn = start_pfn + nr_pages - 1;
1132 return zone_spans_pfn(zone, last_pfn);
1135 static struct page *alloc_gigantic_page(int nid, struct hstate *h)
1137 unsigned int order = huge_page_order(h);
1138 unsigned long nr_pages = 1 << order;
1139 unsigned long ret, pfn, flags;
1140 struct zonelist *zonelist;
1145 gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1146 zonelist = node_zonelist(nid, gfp_mask);
1147 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), NULL) {
1148 spin_lock_irqsave(&zone->lock, flags);
1150 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
1151 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
1152 if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) {
1154 * We release the zone lock here because
1155 * alloc_contig_range() will also lock the zone
1156 * at some point. If there's an allocation
1157 * spinning on this lock, it may win the race
1158 * and cause alloc_contig_range() to fail...
1160 spin_unlock_irqrestore(&zone->lock, flags);
1161 ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask);
1163 return pfn_to_page(pfn);
1164 spin_lock_irqsave(&zone->lock, flags);
1169 spin_unlock_irqrestore(&zone->lock, flags);
1175 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1176 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1178 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1182 page = alloc_gigantic_page(nid, h);
1184 prep_compound_gigantic_page(page, huge_page_order(h));
1185 prep_new_huge_page(h, page, nid);
1191 static int alloc_fresh_gigantic_page(struct hstate *h,
1192 nodemask_t *nodes_allowed)
1194 struct page *page = NULL;
1197 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1198 page = alloc_fresh_gigantic_page_node(h, node);
1206 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1207 static inline bool gigantic_page_supported(void) { return false; }
1208 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1209 static inline void destroy_compound_gigantic_page(struct page *page,
1210 unsigned int order) { }
1211 static inline int alloc_fresh_gigantic_page(struct hstate *h,
1212 nodemask_t *nodes_allowed) { return 0; }
1215 static void update_and_free_page(struct hstate *h, struct page *page)
1218 struct page *subpage = page;
1220 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1224 h->nr_huge_pages_node[page_to_nid(page)]--;
1225 for (i = 0; i < pages_per_huge_page(h);
1226 i++, subpage = mem_map_next(subpage, page, i)) {
1227 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1228 1 << PG_referenced | 1 << PG_dirty |
1229 1 << PG_active | 1 << PG_private |
1232 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1233 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1234 set_page_refcounted(page);
1235 if (hstate_is_gigantic(h)) {
1236 destroy_compound_gigantic_page(page, huge_page_order(h));
1237 free_gigantic_page(page, huge_page_order(h));
1239 __free_pages(page, huge_page_order(h));
1243 struct hstate *size_to_hstate(unsigned long size)
1247 for_each_hstate(h) {
1248 if (huge_page_size(h) == size)
1255 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1256 * to hstate->hugepage_activelist.)
1258 * This function can be called for tail pages, but never returns true for them.
1260 bool page_huge_active(struct page *page)
1262 return PageHeadHuge(page) && PagePrivate(&page[1]);
1265 /* never called for tail page */
1266 void set_page_huge_active(struct page *page)
1268 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1269 SetPagePrivate(&page[1]);
1272 static void clear_page_huge_active(struct page *page)
1274 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1275 ClearPagePrivate(&page[1]);
1278 void free_huge_page(struct page *page)
1281 * Can't pass hstate in here because it is called from the
1282 * compound page destructor.
1284 struct hstate *h = page_hstate(page);
1285 int nid = page_to_nid(page);
1286 struct hugepage_subpool *spool =
1287 (struct hugepage_subpool *)page_private(page);
1288 bool restore_reserve;
1290 set_page_private(page, 0);
1291 page->mapping = NULL;
1292 VM_BUG_ON_PAGE(page_count(page), page);
1293 VM_BUG_ON_PAGE(page_mapcount(page), page);
1294 restore_reserve = PagePrivate(page);
1295 ClearPagePrivate(page);
1298 * If PagePrivate() was set on page, page allocation consumed a
1299 * reservation. If the page was associated with a subpool, there
1300 * would have been a page reserved in the subpool before allocation
1301 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1302 * reservtion, do not call hugepage_subpool_put_pages() as this will
1303 * remove the reserved page from the subpool.
1305 if (!restore_reserve) {
1307 * A return code of zero implies that the subpool will be
1308 * under its minimum size if the reservation is not restored
1309 * after page is free. Therefore, force restore_reserve
1312 if (hugepage_subpool_put_pages(spool, 1) == 0)
1313 restore_reserve = true;
1316 spin_lock(&hugetlb_lock);
1317 clear_page_huge_active(page);
1318 hugetlb_cgroup_uncharge_page(hstate_index(h),
1319 pages_per_huge_page(h), page);
1320 if (restore_reserve)
1321 h->resv_huge_pages++;
1323 if (h->surplus_huge_pages_node[nid]) {
1324 /* remove the page from active list */
1325 list_del(&page->lru);
1326 update_and_free_page(h, page);
1327 h->surplus_huge_pages--;
1328 h->surplus_huge_pages_node[nid]--;
1330 arch_clear_hugepage_flags(page);
1331 enqueue_huge_page(h, page);
1333 spin_unlock(&hugetlb_lock);
1336 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1338 INIT_LIST_HEAD(&page->lru);
1339 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1340 spin_lock(&hugetlb_lock);
1341 set_hugetlb_cgroup(page, NULL);
1343 h->nr_huge_pages_node[nid]++;
1344 ClearPageHugeFreed(page);
1345 spin_unlock(&hugetlb_lock);
1346 put_page(page); /* free it into the hugepage allocator */
1349 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1352 int nr_pages = 1 << order;
1353 struct page *p = page + 1;
1355 /* we rely on prep_new_huge_page to set the destructor */
1356 set_compound_order(page, order);
1357 __ClearPageReserved(page);
1358 __SetPageHead(page);
1359 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1361 * For gigantic hugepages allocated through bootmem at
1362 * boot, it's safer to be consistent with the not-gigantic
1363 * hugepages and clear the PG_reserved bit from all tail pages
1364 * too. Otherwse drivers using get_user_pages() to access tail
1365 * pages may get the reference counting wrong if they see
1366 * PG_reserved set on a tail page (despite the head page not
1367 * having PG_reserved set). Enforcing this consistency between
1368 * head and tail pages allows drivers to optimize away a check
1369 * on the head page when they need know if put_page() is needed
1370 * after get_user_pages().
1372 __ClearPageReserved(p);
1373 set_page_count(p, 0);
1374 set_compound_head(p, page);
1376 atomic_set(compound_mapcount_ptr(page), -1);
1380 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1381 * transparent huge pages. See the PageTransHuge() documentation for more
1384 int PageHuge(struct page *page)
1386 if (!PageCompound(page))
1389 page = compound_head(page);
1390 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1392 EXPORT_SYMBOL_GPL(PageHuge);
1395 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1396 * normal or transparent huge pages.
1398 int PageHeadHuge(struct page *page_head)
1400 if (!PageHead(page_head))
1403 return get_compound_page_dtor(page_head) == free_huge_page;
1406 pgoff_t hugetlb_basepage_index(struct page *page)
1408 struct page *page_head = compound_head(page);
1409 pgoff_t index = page_index(page_head);
1410 unsigned long compound_idx;
1412 if (compound_order(page_head) >= MAX_ORDER)
1413 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1415 compound_idx = page - page_head;
1417 return (index << compound_order(page_head)) + compound_idx;
1420 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1424 page = __alloc_pages_node(nid,
1425 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1426 __GFP_RETRY_MAYFAIL|__GFP_NOWARN,
1427 huge_page_order(h));
1429 prep_new_huge_page(h, page, nid);
1435 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1441 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1442 page = alloc_fresh_huge_page_node(h, node);
1450 count_vm_event(HTLB_BUDDY_PGALLOC);
1452 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1458 * Free huge page from pool from next node to free.
1459 * Attempt to keep persistent huge pages more or less
1460 * balanced over allowed nodes.
1461 * Called with hugetlb_lock locked.
1463 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1469 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1471 * If we're returning unused surplus pages, only examine
1472 * nodes with surplus pages.
1474 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1475 !list_empty(&h->hugepage_freelists[node])) {
1477 list_entry(h->hugepage_freelists[node].next,
1479 list_del(&page->lru);
1480 h->free_huge_pages--;
1481 h->free_huge_pages_node[node]--;
1483 h->surplus_huge_pages--;
1484 h->surplus_huge_pages_node[node]--;
1486 update_and_free_page(h, page);
1496 * Dissolve a given free hugepage into free buddy pages. This function does
1497 * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1498 * number of free hugepages would be reduced below the number of reserved
1501 int dissolve_free_huge_page(struct page *page)
1506 spin_lock(&hugetlb_lock);
1507 if (PageHuge(page) && !page_count(page)) {
1508 struct page *head = compound_head(page);
1509 struct hstate *h = page_hstate(head);
1510 int nid = page_to_nid(head);
1511 if (h->free_huge_pages - h->resv_huge_pages == 0) {
1517 * We should make sure that the page is already on the free list
1518 * when it is dissolved.
1520 if (unlikely(!PageHugeFreed(head))) {
1521 spin_unlock(&hugetlb_lock);
1525 * Theoretically, we should return -EBUSY when we
1526 * encounter this race. In fact, we have a chance
1527 * to successfully dissolve the page if we do a
1528 * retry. Because the race window is quite small.
1529 * If we seize this opportunity, it is an optimization
1530 * for increasing the success rate of dissolving page.
1536 * Move PageHWPoison flag from head page to the raw error page,
1537 * which makes any subpages rather than the error page reusable.
1539 if (PageHWPoison(head) && page != head) {
1540 SetPageHWPoison(page);
1541 ClearPageHWPoison(head);
1543 list_del(&head->lru);
1544 h->free_huge_pages--;
1545 h->free_huge_pages_node[nid]--;
1546 h->max_huge_pages--;
1547 update_and_free_page(h, head);
1550 spin_unlock(&hugetlb_lock);
1555 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1556 * make specified memory blocks removable from the system.
1557 * Note that this will dissolve a free gigantic hugepage completely, if any
1558 * part of it lies within the given range.
1559 * Also note that if dissolve_free_huge_page() returns with an error, all
1560 * free hugepages that were dissolved before that error are lost.
1562 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1568 if (!hugepages_supported())
1571 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1572 page = pfn_to_page(pfn);
1573 if (PageHuge(page) && !page_count(page)) {
1574 rc = dissolve_free_huge_page(page);
1583 static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1584 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1586 int order = huge_page_order(h);
1588 gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
1589 if (nid == NUMA_NO_NODE)
1590 nid = numa_mem_id();
1591 return __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1594 static struct page *__alloc_buddy_huge_page(struct hstate *h, gfp_t gfp_mask,
1595 int nid, nodemask_t *nmask)
1600 if (hstate_is_gigantic(h))
1604 * Assume we will successfully allocate the surplus page to
1605 * prevent racing processes from causing the surplus to exceed
1608 * This however introduces a different race, where a process B
1609 * tries to grow the static hugepage pool while alloc_pages() is
1610 * called by process A. B will only examine the per-node
1611 * counters in determining if surplus huge pages can be
1612 * converted to normal huge pages in adjust_pool_surplus(). A
1613 * won't be able to increment the per-node counter, until the
1614 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1615 * no more huge pages can be converted from surplus to normal
1616 * state (and doesn't try to convert again). Thus, we have a
1617 * case where a surplus huge page exists, the pool is grown, and
1618 * the surplus huge page still exists after, even though it
1619 * should just have been converted to a normal huge page. This
1620 * does not leak memory, though, as the hugepage will be freed
1621 * once it is out of use. It also does not allow the counters to
1622 * go out of whack in adjust_pool_surplus() as we don't modify
1623 * the node values until we've gotten the hugepage and only the
1624 * per-node value is checked there.
1626 spin_lock(&hugetlb_lock);
1627 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1628 spin_unlock(&hugetlb_lock);
1632 h->surplus_huge_pages++;
1634 spin_unlock(&hugetlb_lock);
1636 page = __hugetlb_alloc_buddy_huge_page(h, gfp_mask, nid, nmask);
1638 spin_lock(&hugetlb_lock);
1640 INIT_LIST_HEAD(&page->lru);
1641 r_nid = page_to_nid(page);
1642 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1643 set_hugetlb_cgroup(page, NULL);
1645 * We incremented the global counters already
1647 h->nr_huge_pages_node[r_nid]++;
1648 h->surplus_huge_pages_node[r_nid]++;
1649 __count_vm_event(HTLB_BUDDY_PGALLOC);
1652 h->surplus_huge_pages--;
1653 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1655 spin_unlock(&hugetlb_lock);
1661 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1664 struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1665 struct vm_area_struct *vma, unsigned long addr)
1668 struct mempolicy *mpol;
1669 gfp_t gfp_mask = htlb_alloc_mask(h);
1671 nodemask_t *nodemask;
1673 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1674 page = __alloc_buddy_huge_page(h, gfp_mask, nid, nodemask);
1675 mpol_cond_put(mpol);
1681 * This allocation function is useful in the context where vma is irrelevant.
1682 * E.g. soft-offlining uses this function because it only cares physical
1683 * address of error page.
1685 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1687 gfp_t gfp_mask = htlb_alloc_mask(h);
1688 struct page *page = NULL;
1690 if (nid != NUMA_NO_NODE)
1691 gfp_mask |= __GFP_THISNODE;
1693 spin_lock(&hugetlb_lock);
1694 if (h->free_huge_pages - h->resv_huge_pages > 0)
1695 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1696 spin_unlock(&hugetlb_lock);
1699 page = __alloc_buddy_huge_page(h, gfp_mask, nid, NULL);
1705 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1708 gfp_t gfp_mask = htlb_alloc_mask(h);
1710 spin_lock(&hugetlb_lock);
1711 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1714 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1716 spin_unlock(&hugetlb_lock);
1720 spin_unlock(&hugetlb_lock);
1722 /* No reservations, try to overcommit */
1724 return __alloc_buddy_huge_page(h, gfp_mask, preferred_nid, nmask);
1728 * Increase the hugetlb pool such that it can accommodate a reservation
1731 static int gather_surplus_pages(struct hstate *h, int delta)
1733 struct list_head surplus_list;
1734 struct page *page, *tmp;
1736 int needed, allocated;
1737 bool alloc_ok = true;
1739 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1741 h->resv_huge_pages += delta;
1746 INIT_LIST_HEAD(&surplus_list);
1750 spin_unlock(&hugetlb_lock);
1751 for (i = 0; i < needed; i++) {
1752 page = __alloc_buddy_huge_page(h, htlb_alloc_mask(h),
1753 NUMA_NO_NODE, NULL);
1758 list_add(&page->lru, &surplus_list);
1764 * After retaking hugetlb_lock, we need to recalculate 'needed'
1765 * because either resv_huge_pages or free_huge_pages may have changed.
1767 spin_lock(&hugetlb_lock);
1768 needed = (h->resv_huge_pages + delta) -
1769 (h->free_huge_pages + allocated);
1774 * We were not able to allocate enough pages to
1775 * satisfy the entire reservation so we free what
1776 * we've allocated so far.
1781 * The surplus_list now contains _at_least_ the number of extra pages
1782 * needed to accommodate the reservation. Add the appropriate number
1783 * of pages to the hugetlb pool and free the extras back to the buddy
1784 * allocator. Commit the entire reservation here to prevent another
1785 * process from stealing the pages as they are added to the pool but
1786 * before they are reserved.
1788 needed += allocated;
1789 h->resv_huge_pages += delta;
1792 /* Free the needed pages to the hugetlb pool */
1793 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1797 * This page is now managed by the hugetlb allocator and has
1798 * no users -- drop the buddy allocator's reference.
1800 put_page_testzero(page);
1801 VM_BUG_ON_PAGE(page_count(page), page);
1802 enqueue_huge_page(h, page);
1805 spin_unlock(&hugetlb_lock);
1807 /* Free unnecessary surplus pages to the buddy allocator */
1808 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1810 spin_lock(&hugetlb_lock);
1816 * This routine has two main purposes:
1817 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1818 * in unused_resv_pages. This corresponds to the prior adjustments made
1819 * to the associated reservation map.
1820 * 2) Free any unused surplus pages that may have been allocated to satisfy
1821 * the reservation. As many as unused_resv_pages may be freed.
1823 * Called with hugetlb_lock held. However, the lock could be dropped (and
1824 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1825 * we must make sure nobody else can claim pages we are in the process of
1826 * freeing. Do this by ensuring resv_huge_page always is greater than the
1827 * number of huge pages we plan to free when dropping the lock.
1829 static void return_unused_surplus_pages(struct hstate *h,
1830 unsigned long unused_resv_pages)
1832 unsigned long nr_pages;
1834 /* Cannot return gigantic pages currently */
1835 if (hstate_is_gigantic(h))
1839 * Part (or even all) of the reservation could have been backed
1840 * by pre-allocated pages. Only free surplus pages.
1842 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1845 * We want to release as many surplus pages as possible, spread
1846 * evenly across all nodes with memory. Iterate across these nodes
1847 * until we can no longer free unreserved surplus pages. This occurs
1848 * when the nodes with surplus pages have no free pages.
1849 * free_pool_huge_page() will balance the the freed pages across the
1850 * on-line nodes with memory and will handle the hstate accounting.
1852 * Note that we decrement resv_huge_pages as we free the pages. If
1853 * we drop the lock, resv_huge_pages will still be sufficiently large
1854 * to cover subsequent pages we may free.
1856 while (nr_pages--) {
1857 h->resv_huge_pages--;
1858 unused_resv_pages--;
1859 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1861 cond_resched_lock(&hugetlb_lock);
1865 /* Fully uncommit the reservation */
1866 h->resv_huge_pages -= unused_resv_pages;
1871 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1872 * are used by the huge page allocation routines to manage reservations.
1874 * vma_needs_reservation is called to determine if the huge page at addr
1875 * within the vma has an associated reservation. If a reservation is
1876 * needed, the value 1 is returned. The caller is then responsible for
1877 * managing the global reservation and subpool usage counts. After
1878 * the huge page has been allocated, vma_commit_reservation is called
1879 * to add the page to the reservation map. If the page allocation fails,
1880 * the reservation must be ended instead of committed. vma_end_reservation
1881 * is called in such cases.
1883 * In the normal case, vma_commit_reservation returns the same value
1884 * as the preceding vma_needs_reservation call. The only time this
1885 * is not the case is if a reserve map was changed between calls. It
1886 * is the responsibility of the caller to notice the difference and
1887 * take appropriate action.
1889 * vma_add_reservation is used in error paths where a reservation must
1890 * be restored when a newly allocated huge page must be freed. It is
1891 * to be called after calling vma_needs_reservation to determine if a
1892 * reservation exists.
1894 enum vma_resv_mode {
1900 static long __vma_reservation_common(struct hstate *h,
1901 struct vm_area_struct *vma, unsigned long addr,
1902 enum vma_resv_mode mode)
1904 struct resv_map *resv;
1908 resv = vma_resv_map(vma);
1912 idx = vma_hugecache_offset(h, vma, addr);
1914 case VMA_NEEDS_RESV:
1915 ret = region_chg(resv, idx, idx + 1);
1917 case VMA_COMMIT_RESV:
1918 ret = region_add(resv, idx, idx + 1);
1921 region_abort(resv, idx, idx + 1);
1925 if (vma->vm_flags & VM_MAYSHARE)
1926 ret = region_add(resv, idx, idx + 1);
1928 region_abort(resv, idx, idx + 1);
1929 ret = region_del(resv, idx, idx + 1);
1936 if (vma->vm_flags & VM_MAYSHARE)
1938 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1940 * In most cases, reserves always exist for private mappings.
1941 * However, a file associated with mapping could have been
1942 * hole punched or truncated after reserves were consumed.
1943 * As subsequent fault on such a range will not use reserves.
1944 * Subtle - The reserve map for private mappings has the
1945 * opposite meaning than that of shared mappings. If NO
1946 * entry is in the reserve map, it means a reservation exists.
1947 * If an entry exists in the reserve map, it means the
1948 * reservation has already been consumed. As a result, the
1949 * return value of this routine is the opposite of the
1950 * value returned from reserve map manipulation routines above.
1958 return ret < 0 ? ret : 0;
1961 static long vma_needs_reservation(struct hstate *h,
1962 struct vm_area_struct *vma, unsigned long addr)
1964 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1967 static long vma_commit_reservation(struct hstate *h,
1968 struct vm_area_struct *vma, unsigned long addr)
1970 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1973 static void vma_end_reservation(struct hstate *h,
1974 struct vm_area_struct *vma, unsigned long addr)
1976 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1979 static long vma_add_reservation(struct hstate *h,
1980 struct vm_area_struct *vma, unsigned long addr)
1982 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1986 * This routine is called to restore a reservation on error paths. In the
1987 * specific error paths, a huge page was allocated (via alloc_huge_page)
1988 * and is about to be freed. If a reservation for the page existed,
1989 * alloc_huge_page would have consumed the reservation and set PagePrivate
1990 * in the newly allocated page. When the page is freed via free_huge_page,
1991 * the global reservation count will be incremented if PagePrivate is set.
1992 * However, free_huge_page can not adjust the reserve map. Adjust the
1993 * reserve map here to be consistent with global reserve count adjustments
1994 * to be made by free_huge_page.
1996 static void restore_reserve_on_error(struct hstate *h,
1997 struct vm_area_struct *vma, unsigned long address,
2000 if (unlikely(PagePrivate(page))) {
2001 long rc = vma_needs_reservation(h, vma, address);
2003 if (unlikely(rc < 0)) {
2005 * Rare out of memory condition in reserve map
2006 * manipulation. Clear PagePrivate so that
2007 * global reserve count will not be incremented
2008 * by free_huge_page. This will make it appear
2009 * as though the reservation for this page was
2010 * consumed. This may prevent the task from
2011 * faulting in the page at a later time. This
2012 * is better than inconsistent global huge page
2013 * accounting of reserve counts.
2015 ClearPagePrivate(page);
2017 rc = vma_add_reservation(h, vma, address);
2018 if (unlikely(rc < 0))
2020 * See above comment about rare out of
2023 ClearPagePrivate(page);
2025 vma_end_reservation(h, vma, address);
2029 struct page *alloc_huge_page(struct vm_area_struct *vma,
2030 unsigned long addr, int avoid_reserve)
2032 struct hugepage_subpool *spool = subpool_vma(vma);
2033 struct hstate *h = hstate_vma(vma);
2035 long map_chg, map_commit;
2038 struct hugetlb_cgroup *h_cg;
2040 idx = hstate_index(h);
2042 * Examine the region/reserve map to determine if the process
2043 * has a reservation for the page to be allocated. A return
2044 * code of zero indicates a reservation exists (no change).
2046 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2048 return ERR_PTR(-ENOMEM);
2051 * Processes that did not create the mapping will have no
2052 * reserves as indicated by the region/reserve map. Check
2053 * that the allocation will not exceed the subpool limit.
2054 * Allocations for MAP_NORESERVE mappings also need to be
2055 * checked against any subpool limit.
2057 if (map_chg || avoid_reserve) {
2058 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2060 vma_end_reservation(h, vma, addr);
2061 return ERR_PTR(-ENOSPC);
2065 * Even though there was no reservation in the region/reserve
2066 * map, there could be reservations associated with the
2067 * subpool that can be used. This would be indicated if the
2068 * return value of hugepage_subpool_get_pages() is zero.
2069 * However, if avoid_reserve is specified we still avoid even
2070 * the subpool reservations.
2076 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2078 goto out_subpool_put;
2080 spin_lock(&hugetlb_lock);
2082 * glb_chg is passed to indicate whether or not a page must be taken
2083 * from the global free pool (global change). gbl_chg == 0 indicates
2084 * a reservation exists for the allocation.
2086 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2088 spin_unlock(&hugetlb_lock);
2089 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
2091 goto out_uncharge_cgroup;
2092 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2093 SetPagePrivate(page);
2094 h->resv_huge_pages--;
2096 spin_lock(&hugetlb_lock);
2097 list_move(&page->lru, &h->hugepage_activelist);
2100 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2101 spin_unlock(&hugetlb_lock);
2103 set_page_private(page, (unsigned long)spool);
2105 map_commit = vma_commit_reservation(h, vma, addr);
2106 if (unlikely(map_chg > map_commit)) {
2108 * The page was added to the reservation map between
2109 * vma_needs_reservation and vma_commit_reservation.
2110 * This indicates a race with hugetlb_reserve_pages.
2111 * Adjust for the subpool count incremented above AND
2112 * in hugetlb_reserve_pages for the same page. Also,
2113 * the reservation count added in hugetlb_reserve_pages
2114 * no longer applies.
2118 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2119 hugetlb_acct_memory(h, -rsv_adjust);
2123 out_uncharge_cgroup:
2124 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2126 if (map_chg || avoid_reserve)
2127 hugepage_subpool_put_pages(spool, 1);
2128 vma_end_reservation(h, vma, addr);
2129 return ERR_PTR(-ENOSPC);
2133 * alloc_huge_page()'s wrapper which simply returns the page if allocation
2134 * succeeds, otherwise NULL. This function is called from new_vma_page(),
2135 * where no ERR_VALUE is expected to be returned.
2137 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
2138 unsigned long addr, int avoid_reserve)
2140 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
2146 int alloc_bootmem_huge_page(struct hstate *h)
2147 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2148 int __alloc_bootmem_huge_page(struct hstate *h)
2150 struct huge_bootmem_page *m;
2153 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2156 addr = memblock_virt_alloc_try_nid_nopanic(
2157 huge_page_size(h), huge_page_size(h),
2158 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
2161 * Use the beginning of the huge page to store the
2162 * huge_bootmem_page struct (until gather_bootmem
2163 * puts them into the mem_map).
2172 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2173 /* Put them into a private list first because mem_map is not up yet */
2174 list_add(&m->list, &huge_boot_pages);
2179 static void __init prep_compound_huge_page(struct page *page,
2182 if (unlikely(order > (MAX_ORDER - 1)))
2183 prep_compound_gigantic_page(page, order);
2185 prep_compound_page(page, order);
2188 /* Put bootmem huge pages into the standard lists after mem_map is up */
2189 static void __init gather_bootmem_prealloc(void)
2191 struct huge_bootmem_page *m;
2193 list_for_each_entry(m, &huge_boot_pages, list) {
2194 struct hstate *h = m->hstate;
2197 #ifdef CONFIG_HIGHMEM
2198 page = pfn_to_page(m->phys >> PAGE_SHIFT);
2199 memblock_free_late(__pa(m),
2200 sizeof(struct huge_bootmem_page));
2202 page = virt_to_page(m);
2204 WARN_ON(page_count(page) != 1);
2205 prep_compound_huge_page(page, h->order);
2206 WARN_ON(PageReserved(page));
2207 prep_new_huge_page(h, page, page_to_nid(page));
2209 * If we had gigantic hugepages allocated at boot time, we need
2210 * to restore the 'stolen' pages to totalram_pages in order to
2211 * fix confusing memory reports from free(1) and another
2212 * side-effects, like CommitLimit going negative.
2214 if (hstate_is_gigantic(h))
2215 adjust_managed_page_count(page, 1 << h->order);
2220 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2224 for (i = 0; i < h->max_huge_pages; ++i) {
2225 if (hstate_is_gigantic(h)) {
2226 if (!alloc_bootmem_huge_page(h))
2228 } else if (!alloc_fresh_huge_page(h,
2229 &node_states[N_MEMORY]))
2233 if (i < h->max_huge_pages) {
2236 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2237 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2238 h->max_huge_pages, buf, i);
2239 h->max_huge_pages = i;
2243 static void __init hugetlb_init_hstates(void)
2247 for_each_hstate(h) {
2248 if (minimum_order > huge_page_order(h))
2249 minimum_order = huge_page_order(h);
2251 /* oversize hugepages were init'ed in early boot */
2252 if (!hstate_is_gigantic(h))
2253 hugetlb_hstate_alloc_pages(h);
2255 VM_BUG_ON(minimum_order == UINT_MAX);
2258 static void __init report_hugepages(void)
2262 for_each_hstate(h) {
2265 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2266 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2267 buf, h->free_huge_pages);
2271 #ifdef CONFIG_HIGHMEM
2272 static void try_to_free_low(struct hstate *h, unsigned long count,
2273 nodemask_t *nodes_allowed)
2277 if (hstate_is_gigantic(h))
2280 for_each_node_mask(i, *nodes_allowed) {
2281 struct page *page, *next;
2282 struct list_head *freel = &h->hugepage_freelists[i];
2283 list_for_each_entry_safe(page, next, freel, lru) {
2284 if (count >= h->nr_huge_pages)
2286 if (PageHighMem(page))
2288 list_del(&page->lru);
2289 update_and_free_page(h, page);
2290 h->free_huge_pages--;
2291 h->free_huge_pages_node[page_to_nid(page)]--;
2296 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2297 nodemask_t *nodes_allowed)
2303 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2304 * balanced by operating on them in a round-robin fashion.
2305 * Returns 1 if an adjustment was made.
2307 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2312 VM_BUG_ON(delta != -1 && delta != 1);
2315 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2316 if (h->surplus_huge_pages_node[node])
2320 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2321 if (h->surplus_huge_pages_node[node] <
2322 h->nr_huge_pages_node[node])
2329 h->surplus_huge_pages += delta;
2330 h->surplus_huge_pages_node[node] += delta;
2334 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2335 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2336 nodemask_t *nodes_allowed)
2338 unsigned long min_count, ret;
2340 if (hstate_is_gigantic(h) && !gigantic_page_supported())
2341 return h->max_huge_pages;
2344 * Increase the pool size
2345 * First take pages out of surplus state. Then make up the
2346 * remaining difference by allocating fresh huge pages.
2348 * We might race with __alloc_buddy_huge_page() here and be unable
2349 * to convert a surplus huge page to a normal huge page. That is
2350 * not critical, though, it just means the overall size of the
2351 * pool might be one hugepage larger than it needs to be, but
2352 * within all the constraints specified by the sysctls.
2354 spin_lock(&hugetlb_lock);
2355 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2356 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2360 while (count > persistent_huge_pages(h)) {
2362 * If this allocation races such that we no longer need the
2363 * page, free_huge_page will handle it by freeing the page
2364 * and reducing the surplus.
2366 spin_unlock(&hugetlb_lock);
2368 /* yield cpu to avoid soft lockup */
2371 if (hstate_is_gigantic(h))
2372 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2374 ret = alloc_fresh_huge_page(h, nodes_allowed);
2375 spin_lock(&hugetlb_lock);
2379 /* Bail for signals. Probably ctrl-c from user */
2380 if (signal_pending(current))
2385 * Decrease the pool size
2386 * First return free pages to the buddy allocator (being careful
2387 * to keep enough around to satisfy reservations). Then place
2388 * pages into surplus state as needed so the pool will shrink
2389 * to the desired size as pages become free.
2391 * By placing pages into the surplus state independent of the
2392 * overcommit value, we are allowing the surplus pool size to
2393 * exceed overcommit. There are few sane options here. Since
2394 * __alloc_buddy_huge_page() is checking the global counter,
2395 * though, we'll note that we're not allowed to exceed surplus
2396 * and won't grow the pool anywhere else. Not until one of the
2397 * sysctls are changed, or the surplus pages go out of use.
2399 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2400 min_count = max(count, min_count);
2401 try_to_free_low(h, min_count, nodes_allowed);
2402 while (min_count < persistent_huge_pages(h)) {
2403 if (!free_pool_huge_page(h, nodes_allowed, 0))
2405 cond_resched_lock(&hugetlb_lock);
2407 while (count < persistent_huge_pages(h)) {
2408 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2412 ret = persistent_huge_pages(h);
2413 spin_unlock(&hugetlb_lock);
2417 #define HSTATE_ATTR_RO(_name) \
2418 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2420 #define HSTATE_ATTR(_name) \
2421 static struct kobj_attribute _name##_attr = \
2422 __ATTR(_name, 0644, _name##_show, _name##_store)
2424 static struct kobject *hugepages_kobj;
2425 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2427 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2429 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2433 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2434 if (hstate_kobjs[i] == kobj) {
2436 *nidp = NUMA_NO_NODE;
2440 return kobj_to_node_hstate(kobj, nidp);
2443 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2444 struct kobj_attribute *attr, char *buf)
2447 unsigned long nr_huge_pages;
2450 h = kobj_to_hstate(kobj, &nid);
2451 if (nid == NUMA_NO_NODE)
2452 nr_huge_pages = h->nr_huge_pages;
2454 nr_huge_pages = h->nr_huge_pages_node[nid];
2456 return sprintf(buf, "%lu\n", nr_huge_pages);
2459 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2460 struct hstate *h, int nid,
2461 unsigned long count, size_t len)
2464 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2466 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2471 if (nid == NUMA_NO_NODE) {
2473 * global hstate attribute
2475 if (!(obey_mempolicy &&
2476 init_nodemask_of_mempolicy(nodes_allowed))) {
2477 NODEMASK_FREE(nodes_allowed);
2478 nodes_allowed = &node_states[N_MEMORY];
2480 } else if (nodes_allowed) {
2482 * per node hstate attribute: adjust count to global,
2483 * but restrict alloc/free to the specified node.
2485 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2486 init_nodemask_of_node(nodes_allowed, nid);
2488 nodes_allowed = &node_states[N_MEMORY];
2490 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2492 if (nodes_allowed != &node_states[N_MEMORY])
2493 NODEMASK_FREE(nodes_allowed);
2497 NODEMASK_FREE(nodes_allowed);
2501 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2502 struct kobject *kobj, const char *buf,
2506 unsigned long count;
2510 err = kstrtoul(buf, 10, &count);
2514 h = kobj_to_hstate(kobj, &nid);
2515 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2518 static ssize_t nr_hugepages_show(struct kobject *kobj,
2519 struct kobj_attribute *attr, char *buf)
2521 return nr_hugepages_show_common(kobj, attr, buf);
2524 static ssize_t nr_hugepages_store(struct kobject *kobj,
2525 struct kobj_attribute *attr, const char *buf, size_t len)
2527 return nr_hugepages_store_common(false, kobj, buf, len);
2529 HSTATE_ATTR(nr_hugepages);
2534 * hstate attribute for optionally mempolicy-based constraint on persistent
2535 * huge page alloc/free.
2537 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2538 struct kobj_attribute *attr, char *buf)
2540 return nr_hugepages_show_common(kobj, attr, buf);
2543 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2544 struct kobj_attribute *attr, const char *buf, size_t len)
2546 return nr_hugepages_store_common(true, kobj, buf, len);
2548 HSTATE_ATTR(nr_hugepages_mempolicy);
2552 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2553 struct kobj_attribute *attr, char *buf)
2555 struct hstate *h = kobj_to_hstate(kobj, NULL);
2556 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2559 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2560 struct kobj_attribute *attr, const char *buf, size_t count)
2563 unsigned long input;
2564 struct hstate *h = kobj_to_hstate(kobj, NULL);
2566 if (hstate_is_gigantic(h))
2569 err = kstrtoul(buf, 10, &input);
2573 spin_lock(&hugetlb_lock);
2574 h->nr_overcommit_huge_pages = input;
2575 spin_unlock(&hugetlb_lock);
2579 HSTATE_ATTR(nr_overcommit_hugepages);
2581 static ssize_t free_hugepages_show(struct kobject *kobj,
2582 struct kobj_attribute *attr, char *buf)
2585 unsigned long free_huge_pages;
2588 h = kobj_to_hstate(kobj, &nid);
2589 if (nid == NUMA_NO_NODE)
2590 free_huge_pages = h->free_huge_pages;
2592 free_huge_pages = h->free_huge_pages_node[nid];
2594 return sprintf(buf, "%lu\n", free_huge_pages);
2596 HSTATE_ATTR_RO(free_hugepages);
2598 static ssize_t resv_hugepages_show(struct kobject *kobj,
2599 struct kobj_attribute *attr, char *buf)
2601 struct hstate *h = kobj_to_hstate(kobj, NULL);
2602 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2604 HSTATE_ATTR_RO(resv_hugepages);
2606 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2607 struct kobj_attribute *attr, char *buf)
2610 unsigned long surplus_huge_pages;
2613 h = kobj_to_hstate(kobj, &nid);
2614 if (nid == NUMA_NO_NODE)
2615 surplus_huge_pages = h->surplus_huge_pages;
2617 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2619 return sprintf(buf, "%lu\n", surplus_huge_pages);
2621 HSTATE_ATTR_RO(surplus_hugepages);
2623 static struct attribute *hstate_attrs[] = {
2624 &nr_hugepages_attr.attr,
2625 &nr_overcommit_hugepages_attr.attr,
2626 &free_hugepages_attr.attr,
2627 &resv_hugepages_attr.attr,
2628 &surplus_hugepages_attr.attr,
2630 &nr_hugepages_mempolicy_attr.attr,
2635 static const struct attribute_group hstate_attr_group = {
2636 .attrs = hstate_attrs,
2639 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2640 struct kobject **hstate_kobjs,
2641 const struct attribute_group *hstate_attr_group)
2644 int hi = hstate_index(h);
2646 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2647 if (!hstate_kobjs[hi])
2650 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2652 kobject_put(hstate_kobjs[hi]);
2653 hstate_kobjs[hi] = NULL;
2659 static void __init hugetlb_sysfs_init(void)
2664 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2665 if (!hugepages_kobj)
2668 for_each_hstate(h) {
2669 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2670 hstate_kobjs, &hstate_attr_group);
2672 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2679 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2680 * with node devices in node_devices[] using a parallel array. The array
2681 * index of a node device or _hstate == node id.
2682 * This is here to avoid any static dependency of the node device driver, in
2683 * the base kernel, on the hugetlb module.
2685 struct node_hstate {
2686 struct kobject *hugepages_kobj;
2687 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2689 static struct node_hstate node_hstates[MAX_NUMNODES];
2692 * A subset of global hstate attributes for node devices
2694 static struct attribute *per_node_hstate_attrs[] = {
2695 &nr_hugepages_attr.attr,
2696 &free_hugepages_attr.attr,
2697 &surplus_hugepages_attr.attr,
2701 static const struct attribute_group per_node_hstate_attr_group = {
2702 .attrs = per_node_hstate_attrs,
2706 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2707 * Returns node id via non-NULL nidp.
2709 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2713 for (nid = 0; nid < nr_node_ids; nid++) {
2714 struct node_hstate *nhs = &node_hstates[nid];
2716 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2717 if (nhs->hstate_kobjs[i] == kobj) {
2729 * Unregister hstate attributes from a single node device.
2730 * No-op if no hstate attributes attached.
2732 static void hugetlb_unregister_node(struct node *node)
2735 struct node_hstate *nhs = &node_hstates[node->dev.id];
2737 if (!nhs->hugepages_kobj)
2738 return; /* no hstate attributes */
2740 for_each_hstate(h) {
2741 int idx = hstate_index(h);
2742 if (nhs->hstate_kobjs[idx]) {
2743 kobject_put(nhs->hstate_kobjs[idx]);
2744 nhs->hstate_kobjs[idx] = NULL;
2748 kobject_put(nhs->hugepages_kobj);
2749 nhs->hugepages_kobj = NULL;
2754 * Register hstate attributes for a single node device.
2755 * No-op if attributes already registered.
2757 static void hugetlb_register_node(struct node *node)
2760 struct node_hstate *nhs = &node_hstates[node->dev.id];
2763 if (nhs->hugepages_kobj)
2764 return; /* already allocated */
2766 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2768 if (!nhs->hugepages_kobj)
2771 for_each_hstate(h) {
2772 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2774 &per_node_hstate_attr_group);
2776 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2777 h->name, node->dev.id);
2778 hugetlb_unregister_node(node);
2785 * hugetlb init time: register hstate attributes for all registered node
2786 * devices of nodes that have memory. All on-line nodes should have
2787 * registered their associated device by this time.
2789 static void __init hugetlb_register_all_nodes(void)
2793 for_each_node_state(nid, N_MEMORY) {
2794 struct node *node = node_devices[nid];
2795 if (node->dev.id == nid)
2796 hugetlb_register_node(node);
2800 * Let the node device driver know we're here so it can
2801 * [un]register hstate attributes on node hotplug.
2803 register_hugetlbfs_with_node(hugetlb_register_node,
2804 hugetlb_unregister_node);
2806 #else /* !CONFIG_NUMA */
2808 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2816 static void hugetlb_register_all_nodes(void) { }
2820 static int __init hugetlb_init(void)
2824 if (!hugepages_supported())
2827 if (!size_to_hstate(default_hstate_size)) {
2828 if (default_hstate_size != 0) {
2829 pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2830 default_hstate_size, HPAGE_SIZE);
2833 default_hstate_size = HPAGE_SIZE;
2834 if (!size_to_hstate(default_hstate_size))
2835 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2837 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2838 if (default_hstate_max_huge_pages) {
2839 if (!default_hstate.max_huge_pages)
2840 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2843 hugetlb_init_hstates();
2844 gather_bootmem_prealloc();
2847 hugetlb_sysfs_init();
2848 hugetlb_register_all_nodes();
2849 hugetlb_cgroup_file_init();
2852 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2854 num_fault_mutexes = 1;
2856 hugetlb_fault_mutex_table =
2857 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2858 BUG_ON(!hugetlb_fault_mutex_table);
2860 for (i = 0; i < num_fault_mutexes; i++)
2861 mutex_init(&hugetlb_fault_mutex_table[i]);
2864 subsys_initcall(hugetlb_init);
2866 /* Should be called on processing a hugepagesz=... option */
2867 void __init hugetlb_bad_size(void)
2869 parsed_valid_hugepagesz = false;
2872 void __init hugetlb_add_hstate(unsigned int order)
2877 if (size_to_hstate(PAGE_SIZE << order)) {
2878 pr_warn("hugepagesz= specified twice, ignoring\n");
2881 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2883 h = &hstates[hugetlb_max_hstate++];
2885 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2886 h->nr_huge_pages = 0;
2887 h->free_huge_pages = 0;
2888 for (i = 0; i < MAX_NUMNODES; ++i)
2889 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2890 INIT_LIST_HEAD(&h->hugepage_activelist);
2891 h->next_nid_to_alloc = first_memory_node;
2892 h->next_nid_to_free = first_memory_node;
2893 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2894 huge_page_size(h)/1024);
2899 static int __init hugetlb_nrpages_setup(char *s)
2902 static unsigned long *last_mhp;
2904 if (!parsed_valid_hugepagesz) {
2905 pr_warn("hugepages = %s preceded by "
2906 "an unsupported hugepagesz, ignoring\n", s);
2907 parsed_valid_hugepagesz = true;
2911 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2912 * so this hugepages= parameter goes to the "default hstate".
2914 else if (!hugetlb_max_hstate)
2915 mhp = &default_hstate_max_huge_pages;
2917 mhp = &parsed_hstate->max_huge_pages;
2919 if (mhp == last_mhp) {
2920 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2924 if (sscanf(s, "%lu", mhp) <= 0)
2928 * Global state is always initialized later in hugetlb_init.
2929 * But we need to allocate >= MAX_ORDER hstates here early to still
2930 * use the bootmem allocator.
2932 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2933 hugetlb_hstate_alloc_pages(parsed_hstate);
2939 __setup("hugepages=", hugetlb_nrpages_setup);
2941 static int __init hugetlb_default_setup(char *s)
2943 default_hstate_size = memparse(s, &s);
2946 __setup("default_hugepagesz=", hugetlb_default_setup);
2948 static unsigned int cpuset_mems_nr(unsigned int *array)
2951 unsigned int nr = 0;
2953 for_each_node_mask(node, cpuset_current_mems_allowed)
2959 #ifdef CONFIG_SYSCTL
2960 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
2961 void *buffer, size_t *length,
2962 loff_t *ppos, unsigned long *out)
2964 struct ctl_table dup_table;
2967 * In order to avoid races with __do_proc_doulongvec_minmax(), we
2968 * can duplicate the @table and alter the duplicate of it.
2971 dup_table.data = out;
2973 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
2976 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2977 struct ctl_table *table, int write,
2978 void __user *buffer, size_t *length, loff_t *ppos)
2980 struct hstate *h = &default_hstate;
2981 unsigned long tmp = h->max_huge_pages;
2984 if (!hugepages_supported())
2987 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
2993 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2994 NUMA_NO_NODE, tmp, *length);
2999 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3000 void __user *buffer, size_t *length, loff_t *ppos)
3003 return hugetlb_sysctl_handler_common(false, table, write,
3004 buffer, length, ppos);
3008 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3009 void __user *buffer, size_t *length, loff_t *ppos)
3011 return hugetlb_sysctl_handler_common(true, table, write,
3012 buffer, length, ppos);
3014 #endif /* CONFIG_NUMA */
3016 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3017 void __user *buffer,
3018 size_t *length, loff_t *ppos)
3020 struct hstate *h = &default_hstate;
3024 if (!hugepages_supported())
3027 tmp = h->nr_overcommit_huge_pages;
3029 if (write && hstate_is_gigantic(h))
3032 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3038 spin_lock(&hugetlb_lock);
3039 h->nr_overcommit_huge_pages = tmp;
3040 spin_unlock(&hugetlb_lock);
3046 #endif /* CONFIG_SYSCTL */
3048 void hugetlb_report_meminfo(struct seq_file *m)
3050 struct hstate *h = &default_hstate;
3051 if (!hugepages_supported())
3054 "HugePages_Total: %5lu\n"
3055 "HugePages_Free: %5lu\n"
3056 "HugePages_Rsvd: %5lu\n"
3057 "HugePages_Surp: %5lu\n"
3058 "Hugepagesize: %8lu kB\n",
3062 h->surplus_huge_pages,
3063 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3066 int hugetlb_report_node_meminfo(int nid, char *buf)
3068 struct hstate *h = &default_hstate;
3069 if (!hugepages_supported())
3072 "Node %d HugePages_Total: %5u\n"
3073 "Node %d HugePages_Free: %5u\n"
3074 "Node %d HugePages_Surp: %5u\n",
3075 nid, h->nr_huge_pages_node[nid],
3076 nid, h->free_huge_pages_node[nid],
3077 nid, h->surplus_huge_pages_node[nid]);
3080 void hugetlb_show_meminfo(void)
3085 if (!hugepages_supported())
3088 for_each_node_state(nid, N_MEMORY)
3090 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3092 h->nr_huge_pages_node[nid],
3093 h->free_huge_pages_node[nid],
3094 h->surplus_huge_pages_node[nid],
3095 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3098 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3100 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3101 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3104 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3105 unsigned long hugetlb_total_pages(void)
3108 unsigned long nr_total_pages = 0;
3111 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3112 return nr_total_pages;
3115 static int hugetlb_acct_memory(struct hstate *h, long delta)
3119 spin_lock(&hugetlb_lock);
3121 * When cpuset is configured, it breaks the strict hugetlb page
3122 * reservation as the accounting is done on a global variable. Such
3123 * reservation is completely rubbish in the presence of cpuset because
3124 * the reservation is not checked against page availability for the
3125 * current cpuset. Application can still potentially OOM'ed by kernel
3126 * with lack of free htlb page in cpuset that the task is in.
3127 * Attempt to enforce strict accounting with cpuset is almost
3128 * impossible (or too ugly) because cpuset is too fluid that
3129 * task or memory node can be dynamically moved between cpusets.
3131 * The change of semantics for shared hugetlb mapping with cpuset is
3132 * undesirable. However, in order to preserve some of the semantics,
3133 * we fall back to check against current free page availability as
3134 * a best attempt and hopefully to minimize the impact of changing
3135 * semantics that cpuset has.
3138 if (gather_surplus_pages(h, delta) < 0)
3141 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3142 return_unused_surplus_pages(h, delta);
3149 return_unused_surplus_pages(h, (unsigned long) -delta);
3152 spin_unlock(&hugetlb_lock);
3156 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3158 struct resv_map *resv = vma_resv_map(vma);
3161 * This new VMA should share its siblings reservation map if present.
3162 * The VMA will only ever have a valid reservation map pointer where
3163 * it is being copied for another still existing VMA. As that VMA
3164 * has a reference to the reservation map it cannot disappear until
3165 * after this open call completes. It is therefore safe to take a
3166 * new reference here without additional locking.
3168 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3169 kref_get(&resv->refs);
3172 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3174 struct hstate *h = hstate_vma(vma);
3175 struct resv_map *resv = vma_resv_map(vma);
3176 struct hugepage_subpool *spool = subpool_vma(vma);
3177 unsigned long reserve, start, end;
3180 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3183 start = vma_hugecache_offset(h, vma, vma->vm_start);
3184 end = vma_hugecache_offset(h, vma, vma->vm_end);
3186 reserve = (end - start) - region_count(resv, start, end);
3188 kref_put(&resv->refs, resv_map_release);
3192 * Decrement reserve counts. The global reserve count may be
3193 * adjusted if the subpool has a minimum size.
3195 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3196 hugetlb_acct_memory(h, -gbl_reserve);
3200 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3202 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3208 * We cannot handle pagefaults against hugetlb pages at all. They cause
3209 * handle_mm_fault() to try to instantiate regular-sized pages in the
3210 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3213 static int hugetlb_vm_op_fault(struct vm_fault *vmf)
3219 const struct vm_operations_struct hugetlb_vm_ops = {
3220 .fault = hugetlb_vm_op_fault,
3221 .open = hugetlb_vm_op_open,
3222 .close = hugetlb_vm_op_close,
3223 .split = hugetlb_vm_op_split,
3226 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3232 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3233 vma->vm_page_prot)));
3235 entry = huge_pte_wrprotect(mk_huge_pte(page,
3236 vma->vm_page_prot));
3238 entry = pte_mkyoung(entry);
3239 entry = pte_mkhuge(entry);
3240 entry = arch_make_huge_pte(entry, vma, page, writable);
3245 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3246 unsigned long address, pte_t *ptep)
3250 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3251 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3252 update_mmu_cache(vma, address, ptep);
3255 bool is_hugetlb_entry_migration(pte_t pte)
3259 if (huge_pte_none(pte) || pte_present(pte))
3261 swp = pte_to_swp_entry(pte);
3262 if (non_swap_entry(swp) && is_migration_entry(swp))
3268 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3272 if (huge_pte_none(pte) || pte_present(pte))
3274 swp = pte_to_swp_entry(pte);
3275 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3281 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3282 struct vm_area_struct *vma)
3284 pte_t *src_pte, *dst_pte, entry, dst_entry;
3285 struct page *ptepage;
3288 struct hstate *h = hstate_vma(vma);
3289 unsigned long sz = huge_page_size(h);
3290 unsigned long mmun_start; /* For mmu_notifiers */
3291 unsigned long mmun_end; /* For mmu_notifiers */
3294 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3296 mmun_start = vma->vm_start;
3297 mmun_end = vma->vm_end;
3299 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3301 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3302 spinlock_t *src_ptl, *dst_ptl;
3303 src_pte = huge_pte_offset(src, addr, sz);
3306 dst_pte = huge_pte_alloc(dst, addr, sz);
3313 * If the pagetables are shared don't copy or take references.
3314 * dst_pte == src_pte is the common case of src/dest sharing.
3316 * However, src could have 'unshared' and dst shares with
3317 * another vma. If dst_pte !none, this implies sharing.
3318 * Check here before taking page table lock, and once again
3319 * after taking the lock below.
3321 dst_entry = huge_ptep_get(dst_pte);
3322 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3325 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3326 src_ptl = huge_pte_lockptr(h, src, src_pte);
3327 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3328 entry = huge_ptep_get(src_pte);
3329 dst_entry = huge_ptep_get(dst_pte);
3330 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3332 * Skip if src entry none. Also, skip in the
3333 * unlikely case dst entry !none as this implies
3334 * sharing with another vma.
3337 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3338 is_hugetlb_entry_hwpoisoned(entry))) {
3339 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3341 if (is_write_migration_entry(swp_entry) && cow) {
3343 * COW mappings require pages in both
3344 * parent and child to be set to read.
3346 make_migration_entry_read(&swp_entry);
3347 entry = swp_entry_to_pte(swp_entry);
3348 set_huge_swap_pte_at(src, addr, src_pte,
3351 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3354 huge_ptep_set_wrprotect(src, addr, src_pte);
3355 mmu_notifier_invalidate_range(src, mmun_start,
3358 entry = huge_ptep_get(src_pte);
3359 ptepage = pte_page(entry);
3361 page_dup_rmap(ptepage, true);
3362 set_huge_pte_at(dst, addr, dst_pte, entry);
3363 hugetlb_count_add(pages_per_huge_page(h), dst);
3365 spin_unlock(src_ptl);
3366 spin_unlock(dst_ptl);
3370 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3375 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3376 unsigned long start, unsigned long end,
3377 struct page *ref_page)
3379 struct mm_struct *mm = vma->vm_mm;
3380 unsigned long address;
3385 struct hstate *h = hstate_vma(vma);
3386 unsigned long sz = huge_page_size(h);
3387 const unsigned long mmun_start = start; /* For mmu_notifiers */
3388 const unsigned long mmun_end = end; /* For mmu_notifiers */
3390 WARN_ON(!is_vm_hugetlb_page(vma));
3391 BUG_ON(start & ~huge_page_mask(h));
3392 BUG_ON(end & ~huge_page_mask(h));
3395 * This is a hugetlb vma, all the pte entries should point
3398 tlb_remove_check_page_size_change(tlb, sz);
3399 tlb_start_vma(tlb, vma);
3400 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3402 for (; address < end; address += sz) {
3403 ptep = huge_pte_offset(mm, address, sz);
3407 ptl = huge_pte_lock(h, mm, ptep);
3408 if (huge_pmd_unshare(mm, &address, ptep)) {
3413 pte = huge_ptep_get(ptep);
3414 if (huge_pte_none(pte)) {
3420 * Migrating hugepage or HWPoisoned hugepage is already
3421 * unmapped and its refcount is dropped, so just clear pte here.
3423 if (unlikely(!pte_present(pte))) {
3424 huge_pte_clear(mm, address, ptep, sz);
3429 page = pte_page(pte);
3431 * If a reference page is supplied, it is because a specific
3432 * page is being unmapped, not a range. Ensure the page we
3433 * are about to unmap is the actual page of interest.
3436 if (page != ref_page) {
3441 * Mark the VMA as having unmapped its page so that
3442 * future faults in this VMA will fail rather than
3443 * looking like data was lost
3445 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3448 pte = huge_ptep_get_and_clear(mm, address, ptep);
3449 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3450 if (huge_pte_dirty(pte))
3451 set_page_dirty(page);
3453 hugetlb_count_sub(pages_per_huge_page(h), mm);
3454 page_remove_rmap(page, true);
3457 tlb_remove_page_size(tlb, page, huge_page_size(h));
3459 * Bail out after unmapping reference page if supplied
3464 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3465 tlb_end_vma(tlb, vma);
3468 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3469 struct vm_area_struct *vma, unsigned long start,
3470 unsigned long end, struct page *ref_page)
3472 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3475 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3476 * test will fail on a vma being torn down, and not grab a page table
3477 * on its way out. We're lucky that the flag has such an appropriate
3478 * name, and can in fact be safely cleared here. We could clear it
3479 * before the __unmap_hugepage_range above, but all that's necessary
3480 * is to clear it before releasing the i_mmap_rwsem. This works
3481 * because in the context this is called, the VMA is about to be
3482 * destroyed and the i_mmap_rwsem is held.
3484 vma->vm_flags &= ~VM_MAYSHARE;
3487 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3488 unsigned long end, struct page *ref_page)
3490 struct mm_struct *mm;
3491 struct mmu_gather tlb;
3495 tlb_gather_mmu(&tlb, mm, start, end);
3496 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3497 tlb_finish_mmu(&tlb, start, end);
3501 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3502 * mappping it owns the reserve page for. The intention is to unmap the page
3503 * from other VMAs and let the children be SIGKILLed if they are faulting the
3506 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3507 struct page *page, unsigned long address)
3509 struct hstate *h = hstate_vma(vma);
3510 struct vm_area_struct *iter_vma;
3511 struct address_space *mapping;
3515 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3516 * from page cache lookup which is in HPAGE_SIZE units.
3518 address = address & huge_page_mask(h);
3519 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3521 mapping = vma->vm_file->f_mapping;
3524 * Take the mapping lock for the duration of the table walk. As
3525 * this mapping should be shared between all the VMAs,
3526 * __unmap_hugepage_range() is called as the lock is already held
3528 i_mmap_lock_write(mapping);
3529 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3530 /* Do not unmap the current VMA */
3531 if (iter_vma == vma)
3535 * Shared VMAs have their own reserves and do not affect
3536 * MAP_PRIVATE accounting but it is possible that a shared
3537 * VMA is using the same page so check and skip such VMAs.
3539 if (iter_vma->vm_flags & VM_MAYSHARE)
3543 * Unmap the page from other VMAs without their own reserves.
3544 * They get marked to be SIGKILLed if they fault in these
3545 * areas. This is because a future no-page fault on this VMA
3546 * could insert a zeroed page instead of the data existing
3547 * from the time of fork. This would look like data corruption
3549 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3550 unmap_hugepage_range(iter_vma, address,
3551 address + huge_page_size(h), page);
3553 i_mmap_unlock_write(mapping);
3557 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3558 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3559 * cannot race with other handlers or page migration.
3560 * Keep the pte_same checks anyway to make transition from the mutex easier.
3562 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3563 unsigned long address, pte_t *ptep,
3564 struct page *pagecache_page, spinlock_t *ptl)
3567 struct hstate *h = hstate_vma(vma);
3568 struct page *old_page, *new_page;
3569 int ret = 0, outside_reserve = 0;
3570 unsigned long mmun_start; /* For mmu_notifiers */
3571 unsigned long mmun_end; /* For mmu_notifiers */
3573 pte = huge_ptep_get(ptep);
3574 old_page = pte_page(pte);
3577 /* If no-one else is actually using this page, avoid the copy
3578 * and just make the page writable */
3579 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3580 page_move_anon_rmap(old_page, vma);
3581 set_huge_ptep_writable(vma, address, ptep);
3586 * If the process that created a MAP_PRIVATE mapping is about to
3587 * perform a COW due to a shared page count, attempt to satisfy
3588 * the allocation without using the existing reserves. The pagecache
3589 * page is used to determine if the reserve at this address was
3590 * consumed or not. If reserves were used, a partial faulted mapping
3591 * at the time of fork() could consume its reserves on COW instead
3592 * of the full address range.
3594 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3595 old_page != pagecache_page)
3596 outside_reserve = 1;
3601 * Drop page table lock as buddy allocator may be called. It will
3602 * be acquired again before returning to the caller, as expected.
3605 new_page = alloc_huge_page(vma, address, outside_reserve);
3607 if (IS_ERR(new_page)) {
3609 * If a process owning a MAP_PRIVATE mapping fails to COW,
3610 * it is due to references held by a child and an insufficient
3611 * huge page pool. To guarantee the original mappers
3612 * reliability, unmap the page from child processes. The child
3613 * may get SIGKILLed if it later faults.
3615 if (outside_reserve) {
3617 BUG_ON(huge_pte_none(pte));
3618 unmap_ref_private(mm, vma, old_page, address);
3619 BUG_ON(huge_pte_none(pte));
3621 ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3624 pte_same(huge_ptep_get(ptep), pte)))
3625 goto retry_avoidcopy;
3627 * race occurs while re-acquiring page table
3628 * lock, and our job is done.
3633 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3634 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3635 goto out_release_old;
3639 * When the original hugepage is shared one, it does not have
3640 * anon_vma prepared.
3642 if (unlikely(anon_vma_prepare(vma))) {
3644 goto out_release_all;
3647 copy_user_huge_page(new_page, old_page, address, vma,
3648 pages_per_huge_page(h));
3649 __SetPageUptodate(new_page);
3651 mmun_start = address & huge_page_mask(h);
3652 mmun_end = mmun_start + huge_page_size(h);
3653 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3656 * Retake the page table lock to check for racing updates
3657 * before the page tables are altered
3660 ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3662 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3663 ClearPagePrivate(new_page);
3666 huge_ptep_clear_flush(vma, address, ptep);
3667 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3668 set_huge_pte_at(mm, address, ptep,
3669 make_huge_pte(vma, new_page, 1));
3670 page_remove_rmap(old_page, true);
3671 hugepage_add_new_anon_rmap(new_page, vma, address);
3672 set_page_huge_active(new_page);
3673 /* Make the old page be freed below */
3674 new_page = old_page;
3677 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3679 restore_reserve_on_error(h, vma, address, new_page);
3684 spin_lock(ptl); /* Caller expects lock to be held */
3688 /* Return the pagecache page at a given address within a VMA */
3689 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3690 struct vm_area_struct *vma, unsigned long address)
3692 struct address_space *mapping;
3695 mapping = vma->vm_file->f_mapping;
3696 idx = vma_hugecache_offset(h, vma, address);
3698 return find_lock_page(mapping, idx);
3702 * Return whether there is a pagecache page to back given address within VMA.
3703 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3705 static bool hugetlbfs_pagecache_present(struct hstate *h,
3706 struct vm_area_struct *vma, unsigned long address)
3708 struct address_space *mapping;
3712 mapping = vma->vm_file->f_mapping;
3713 idx = vma_hugecache_offset(h, vma, address);
3715 page = find_get_page(mapping, idx);
3718 return page != NULL;
3721 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3724 struct inode *inode = mapping->host;
3725 struct hstate *h = hstate_inode(inode);
3726 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3730 ClearPagePrivate(page);
3733 * set page dirty so that it will not be removed from cache/file
3734 * by non-hugetlbfs specific code paths.
3736 set_page_dirty(page);
3738 spin_lock(&inode->i_lock);
3739 inode->i_blocks += blocks_per_huge_page(h);
3740 spin_unlock(&inode->i_lock);
3744 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3745 struct address_space *mapping, pgoff_t idx,
3746 unsigned long address, pte_t *ptep, unsigned int flags)
3748 struct hstate *h = hstate_vma(vma);
3749 int ret = VM_FAULT_SIGBUS;
3755 bool new_page = false;
3758 * Currently, we are forced to kill the process in the event the
3759 * original mapper has unmapped pages from the child due to a failed
3760 * COW. Warn that such a situation has occurred as it may not be obvious
3762 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3763 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3769 * Use page lock to guard against racing truncation
3770 * before we get page_table_lock.
3773 page = find_lock_page(mapping, idx);
3775 size = i_size_read(mapping->host) >> huge_page_shift(h);
3780 * Check for page in userfault range
3782 if (userfaultfd_missing(vma)) {
3784 struct vm_fault vmf = {
3789 * Hard to debug if it ends up being
3790 * used by a callee that assumes
3791 * something about the other
3792 * uninitialized fields... same as in
3798 * hugetlb_fault_mutex must be dropped before
3799 * handling userfault. Reacquire after handling
3800 * fault to make calling code simpler.
3802 hash = hugetlb_fault_mutex_hash(h, mapping, idx);
3803 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3804 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3805 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3809 page = alloc_huge_page(vma, address, 0);
3811 ret = PTR_ERR(page);
3815 ret = VM_FAULT_SIGBUS;
3818 clear_huge_page(page, address, pages_per_huge_page(h));
3819 __SetPageUptodate(page);
3822 if (vma->vm_flags & VM_MAYSHARE) {
3823 int err = huge_add_to_page_cache(page, mapping, idx);
3832 if (unlikely(anon_vma_prepare(vma))) {
3834 goto backout_unlocked;
3840 * If memory error occurs between mmap() and fault, some process
3841 * don't have hwpoisoned swap entry for errored virtual address.
3842 * So we need to block hugepage fault by PG_hwpoison bit check.
3844 if (unlikely(PageHWPoison(page))) {
3845 ret = VM_FAULT_HWPOISON_LARGE |
3846 VM_FAULT_SET_HINDEX(hstate_index(h));
3847 goto backout_unlocked;
3852 * If we are going to COW a private mapping later, we examine the
3853 * pending reservations for this page now. This will ensure that
3854 * any allocations necessary to record that reservation occur outside
3857 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3858 if (vma_needs_reservation(h, vma, address) < 0) {
3860 goto backout_unlocked;
3862 /* Just decrements count, does not deallocate */
3863 vma_end_reservation(h, vma, address);
3866 ptl = huge_pte_lock(h, mm, ptep);
3867 size = i_size_read(mapping->host) >> huge_page_shift(h);
3872 if (!huge_pte_none(huge_ptep_get(ptep)))
3876 ClearPagePrivate(page);
3877 hugepage_add_new_anon_rmap(page, vma, address);
3879 page_dup_rmap(page, true);
3880 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3881 && (vma->vm_flags & VM_SHARED)));
3882 set_huge_pte_at(mm, address, ptep, new_pte);
3884 hugetlb_count_add(pages_per_huge_page(h), mm);
3885 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3886 /* Optimization, do the COW without a second fault */
3887 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
3893 * Only make newly allocated pages active. Existing pages found
3894 * in the pagecache could be !page_huge_active() if they have been
3895 * isolated for migration.
3898 set_page_huge_active(page);
3908 restore_reserve_on_error(h, vma, address, page);
3914 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping,
3917 unsigned long key[2];
3920 key[0] = (unsigned long) mapping;
3923 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
3925 return hash & (num_fault_mutexes - 1);
3929 * For uniprocesor systems we always use a single mutex, so just
3930 * return 0 and avoid the hashing overhead.
3932 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping,
3939 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3940 unsigned long address, unsigned int flags)
3947 struct page *page = NULL;
3948 struct page *pagecache_page = NULL;
3949 struct hstate *h = hstate_vma(vma);
3950 struct address_space *mapping;
3951 int need_wait_lock = 0;
3953 address &= huge_page_mask(h);
3955 ptep = huge_pte_offset(mm, address, huge_page_size(h));
3957 entry = huge_ptep_get(ptep);
3958 if (unlikely(is_hugetlb_entry_migration(entry))) {
3959 migration_entry_wait_huge(vma, mm, ptep);
3961 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3962 return VM_FAULT_HWPOISON_LARGE |
3963 VM_FAULT_SET_HINDEX(hstate_index(h));
3965 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3967 return VM_FAULT_OOM;
3970 mapping = vma->vm_file->f_mapping;
3971 idx = vma_hugecache_offset(h, vma, address);
3974 * Serialize hugepage allocation and instantiation, so that we don't
3975 * get spurious allocation failures if two CPUs race to instantiate
3976 * the same page in the page cache.
3978 hash = hugetlb_fault_mutex_hash(h, mapping, idx);
3979 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3981 entry = huge_ptep_get(ptep);
3982 if (huge_pte_none(entry)) {
3983 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3990 * entry could be a migration/hwpoison entry at this point, so this
3991 * check prevents the kernel from going below assuming that we have
3992 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3993 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3996 if (!pte_present(entry))
4000 * If we are going to COW the mapping later, we examine the pending
4001 * reservations for this page now. This will ensure that any
4002 * allocations necessary to record that reservation occur outside the
4003 * spinlock. For private mappings, we also lookup the pagecache
4004 * page now as it is used to determine if a reservation has been
4007 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4008 if (vma_needs_reservation(h, vma, address) < 0) {
4012 /* Just decrements count, does not deallocate */
4013 vma_end_reservation(h, vma, address);
4015 if (!(vma->vm_flags & VM_MAYSHARE))
4016 pagecache_page = hugetlbfs_pagecache_page(h,
4020 ptl = huge_pte_lock(h, mm, ptep);
4022 /* Check for a racing update before calling hugetlb_cow */
4023 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4027 * hugetlb_cow() requires page locks of pte_page(entry) and
4028 * pagecache_page, so here we need take the former one
4029 * when page != pagecache_page or !pagecache_page.
4031 page = pte_page(entry);
4032 if (page != pagecache_page)
4033 if (!trylock_page(page)) {
4040 if (flags & FAULT_FLAG_WRITE) {
4041 if (!huge_pte_write(entry)) {
4042 ret = hugetlb_cow(mm, vma, address, ptep,
4043 pagecache_page, ptl);
4046 entry = huge_pte_mkdirty(entry);
4048 entry = pte_mkyoung(entry);
4049 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
4050 flags & FAULT_FLAG_WRITE))
4051 update_mmu_cache(vma, address, ptep);
4053 if (page != pagecache_page)
4059 if (pagecache_page) {
4060 unlock_page(pagecache_page);
4061 put_page(pagecache_page);
4064 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4066 * Generally it's safe to hold refcount during waiting page lock. But
4067 * here we just wait to defer the next page fault to avoid busy loop and
4068 * the page is not used after unlocked before returning from the current
4069 * page fault. So we are safe from accessing freed page, even if we wait
4070 * here without taking refcount.
4073 wait_on_page_locked(page);
4078 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4079 * modifications for huge pages.
4081 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4083 struct vm_area_struct *dst_vma,
4084 unsigned long dst_addr,
4085 unsigned long src_addr,
4086 struct page **pagep)
4088 struct address_space *mapping;
4091 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4092 struct hstate *h = hstate_vma(dst_vma);
4099 /* If a page already exists, then it's UFFDIO_COPY for
4100 * a non-missing case. Return -EEXIST.
4103 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
4108 page = alloc_huge_page(dst_vma, dst_addr, 0);
4114 ret = copy_huge_page_from_user(page,
4115 (const void __user *) src_addr,
4116 pages_per_huge_page(h), false);
4118 /* fallback to copy_from_user outside mmap_sem */
4119 if (unlikely(ret)) {
4122 /* don't free the page */
4131 * The memory barrier inside __SetPageUptodate makes sure that
4132 * preceding stores to the page contents become visible before
4133 * the set_pte_at() write.
4135 __SetPageUptodate(page);
4137 mapping = dst_vma->vm_file->f_mapping;
4138 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4141 * If shared, add to page cache
4144 size = i_size_read(mapping->host) >> huge_page_shift(h);
4147 goto out_release_nounlock;
4150 * Serialization between remove_inode_hugepages() and
4151 * huge_add_to_page_cache() below happens through the
4152 * hugetlb_fault_mutex_table that here must be hold by
4155 ret = huge_add_to_page_cache(page, mapping, idx);
4157 goto out_release_nounlock;
4160 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4164 * Recheck the i_size after holding PT lock to make sure not
4165 * to leave any page mapped (as page_mapped()) beyond the end
4166 * of the i_size (remove_inode_hugepages() is strict about
4167 * enforcing that). If we bail out here, we'll also leave a
4168 * page in the radix tree in the vm_shared case beyond the end
4169 * of the i_size, but remove_inode_hugepages() will take care
4170 * of it as soon as we drop the hugetlb_fault_mutex_table.
4172 size = i_size_read(mapping->host) >> huge_page_shift(h);
4175 goto out_release_unlock;
4178 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4179 goto out_release_unlock;
4182 page_dup_rmap(page, true);
4184 ClearPagePrivate(page);
4185 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4188 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4189 if (dst_vma->vm_flags & VM_WRITE)
4190 _dst_pte = huge_pte_mkdirty(_dst_pte);
4191 _dst_pte = pte_mkyoung(_dst_pte);
4193 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4195 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4196 dst_vma->vm_flags & VM_WRITE);
4197 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4199 /* No need to invalidate - it was non-present before */
4200 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4203 set_page_huge_active(page);
4213 out_release_nounlock:
4218 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4219 struct page **pages, struct vm_area_struct **vmas,
4220 unsigned long *position, unsigned long *nr_pages,
4221 long i, unsigned int flags, int *nonblocking)
4223 unsigned long pfn_offset;
4224 unsigned long vaddr = *position;
4225 unsigned long remainder = *nr_pages;
4226 struct hstate *h = hstate_vma(vma);
4229 while (vaddr < vma->vm_end && remainder) {
4231 spinlock_t *ptl = NULL;
4236 * If we have a pending SIGKILL, don't keep faulting pages and
4237 * potentially allocating memory.
4239 if (unlikely(fatal_signal_pending(current))) {
4245 * Some archs (sparc64, sh*) have multiple pte_ts to
4246 * each hugepage. We have to make sure we get the
4247 * first, for the page indexing below to work.
4249 * Note that page table lock is not held when pte is null.
4251 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4254 ptl = huge_pte_lock(h, mm, pte);
4255 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4258 * When coredumping, it suits get_dump_page if we just return
4259 * an error where there's an empty slot with no huge pagecache
4260 * to back it. This way, we avoid allocating a hugepage, and
4261 * the sparse dumpfile avoids allocating disk blocks, but its
4262 * huge holes still show up with zeroes where they need to be.
4264 if (absent && (flags & FOLL_DUMP) &&
4265 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4273 * We need call hugetlb_fault for both hugepages under migration
4274 * (in which case hugetlb_fault waits for the migration,) and
4275 * hwpoisoned hugepages (in which case we need to prevent the
4276 * caller from accessing to them.) In order to do this, we use
4277 * here is_swap_pte instead of is_hugetlb_entry_migration and
4278 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4279 * both cases, and because we can't follow correct pages
4280 * directly from any kind of swap entries.
4282 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4283 ((flags & FOLL_WRITE) &&
4284 !huge_pte_write(huge_ptep_get(pte)))) {
4286 unsigned int fault_flags = 0;
4290 if (flags & FOLL_WRITE)
4291 fault_flags |= FAULT_FLAG_WRITE;
4293 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4294 if (flags & FOLL_NOWAIT)
4295 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4296 FAULT_FLAG_RETRY_NOWAIT;
4297 if (flags & FOLL_TRIED) {
4298 VM_WARN_ON_ONCE(fault_flags &
4299 FAULT_FLAG_ALLOW_RETRY);
4300 fault_flags |= FAULT_FLAG_TRIED;
4302 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4303 if (ret & VM_FAULT_ERROR) {
4304 err = vm_fault_to_errno(ret, flags);
4308 if (ret & VM_FAULT_RETRY) {
4313 * VM_FAULT_RETRY must not return an
4314 * error, it will return zero
4317 * No need to update "position" as the
4318 * caller will not check it after
4319 * *nr_pages is set to 0.
4326 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4327 page = pte_page(huge_ptep_get(pte));
4330 * Instead of doing 'try_get_page()' below in the same_page
4331 * loop, just check the count once here.
4333 if (unlikely(page_count(page) <= 0)) {
4343 pages[i] = mem_map_offset(page, pfn_offset);
4354 if (vaddr < vma->vm_end && remainder &&
4355 pfn_offset < pages_per_huge_page(h)) {
4357 * We use pfn_offset to avoid touching the pageframes
4358 * of this compound page.
4364 *nr_pages = remainder;
4366 * setting position is actually required only if remainder is
4367 * not zero but it's faster not to add a "if (remainder)"
4375 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4377 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4380 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4383 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4384 unsigned long address, unsigned long end, pgprot_t newprot)
4386 struct mm_struct *mm = vma->vm_mm;
4387 unsigned long start = address;
4390 struct hstate *h = hstate_vma(vma);
4391 unsigned long pages = 0;
4393 BUG_ON(address >= end);
4394 flush_cache_range(vma, address, end);
4396 mmu_notifier_invalidate_range_start(mm, start, end);
4397 i_mmap_lock_write(vma->vm_file->f_mapping);
4398 for (; address < end; address += huge_page_size(h)) {
4400 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4403 ptl = huge_pte_lock(h, mm, ptep);
4404 if (huge_pmd_unshare(mm, &address, ptep)) {
4409 pte = huge_ptep_get(ptep);
4410 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4414 if (unlikely(is_hugetlb_entry_migration(pte))) {
4415 swp_entry_t entry = pte_to_swp_entry(pte);
4417 if (is_write_migration_entry(entry)) {
4420 make_migration_entry_read(&entry);
4421 newpte = swp_entry_to_pte(entry);
4422 set_huge_swap_pte_at(mm, address, ptep,
4423 newpte, huge_page_size(h));
4429 if (!huge_pte_none(pte)) {
4430 pte = huge_ptep_get_and_clear(mm, address, ptep);
4431 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
4432 pte = arch_make_huge_pte(pte, vma, NULL, 0);
4433 set_huge_pte_at(mm, address, ptep, pte);
4439 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4440 * may have cleared our pud entry and done put_page on the page table:
4441 * once we release i_mmap_rwsem, another task can do the final put_page
4442 * and that page table be reused and filled with junk.
4444 flush_hugetlb_tlb_range(vma, start, end);
4445 mmu_notifier_invalidate_range(mm, start, end);
4446 i_mmap_unlock_write(vma->vm_file->f_mapping);
4447 mmu_notifier_invalidate_range_end(mm, start, end);
4449 return pages << h->order;
4452 int hugetlb_reserve_pages(struct inode *inode,
4454 struct vm_area_struct *vma,
4455 vm_flags_t vm_flags)
4458 struct hstate *h = hstate_inode(inode);
4459 struct hugepage_subpool *spool = subpool_inode(inode);
4460 struct resv_map *resv_map;
4463 /* This should never happen */
4465 VM_WARN(1, "%s called with a negative range\n", __func__);
4470 * Only apply hugepage reservation if asked. At fault time, an
4471 * attempt will be made for VM_NORESERVE to allocate a page
4472 * without using reserves
4474 if (vm_flags & VM_NORESERVE)
4478 * Shared mappings base their reservation on the number of pages that
4479 * are already allocated on behalf of the file. Private mappings need
4480 * to reserve the full area even if read-only as mprotect() may be
4481 * called to make the mapping read-write. Assume !vma is a shm mapping
4483 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4484 resv_map = inode_resv_map(inode);
4486 chg = region_chg(resv_map, from, to);
4489 resv_map = resv_map_alloc();
4495 set_vma_resv_map(vma, resv_map);
4496 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4505 * There must be enough pages in the subpool for the mapping. If
4506 * the subpool has a minimum size, there may be some global
4507 * reservations already in place (gbl_reserve).
4509 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4510 if (gbl_reserve < 0) {
4516 * Check enough hugepages are available for the reservation.
4517 * Hand the pages back to the subpool if there are not
4519 ret = hugetlb_acct_memory(h, gbl_reserve);
4521 /* put back original number of pages, chg */
4522 (void)hugepage_subpool_put_pages(spool, chg);
4527 * Account for the reservations made. Shared mappings record regions
4528 * that have reservations as they are shared by multiple VMAs.
4529 * When the last VMA disappears, the region map says how much
4530 * the reservation was and the page cache tells how much of
4531 * the reservation was consumed. Private mappings are per-VMA and
4532 * only the consumed reservations are tracked. When the VMA
4533 * disappears, the original reservation is the VMA size and the
4534 * consumed reservations are stored in the map. Hence, nothing
4535 * else has to be done for private mappings here
4537 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4538 long add = region_add(resv_map, from, to);
4540 if (unlikely(chg > add)) {
4542 * pages in this range were added to the reserve
4543 * map between region_chg and region_add. This
4544 * indicates a race with alloc_huge_page. Adjust
4545 * the subpool and reserve counts modified above
4546 * based on the difference.
4550 rsv_adjust = hugepage_subpool_put_pages(spool,
4552 hugetlb_acct_memory(h, -rsv_adjust);
4557 if (!vma || vma->vm_flags & VM_MAYSHARE)
4558 /* Don't call region_abort if region_chg failed */
4560 region_abort(resv_map, from, to);
4561 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4562 kref_put(&resv_map->refs, resv_map_release);
4566 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4569 struct hstate *h = hstate_inode(inode);
4570 struct resv_map *resv_map = inode_resv_map(inode);
4572 struct hugepage_subpool *spool = subpool_inode(inode);
4576 chg = region_del(resv_map, start, end);
4578 * region_del() can fail in the rare case where a region
4579 * must be split and another region descriptor can not be
4580 * allocated. If end == LONG_MAX, it will not fail.
4586 spin_lock(&inode->i_lock);
4587 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4588 spin_unlock(&inode->i_lock);
4591 * If the subpool has a minimum size, the number of global
4592 * reservations to be released may be adjusted.
4594 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4595 hugetlb_acct_memory(h, -gbl_reserve);
4600 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4601 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4602 struct vm_area_struct *vma,
4603 unsigned long addr, pgoff_t idx)
4605 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4607 unsigned long sbase = saddr & PUD_MASK;
4608 unsigned long s_end = sbase + PUD_SIZE;
4610 /* Allow segments to share if only one is marked locked */
4611 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4612 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4615 * match the virtual addresses, permission and the alignment of the
4618 if (pmd_index(addr) != pmd_index(saddr) ||
4619 vm_flags != svm_flags ||
4620 sbase < svma->vm_start || svma->vm_end < s_end)
4626 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4628 unsigned long base = addr & PUD_MASK;
4629 unsigned long end = base + PUD_SIZE;
4632 * check on proper vm_flags and page table alignment
4634 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
4640 * Determine if start,end range within vma could be mapped by shared pmd.
4641 * If yes, adjust start and end to cover range associated with possible
4642 * shared pmd mappings.
4644 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4645 unsigned long *start, unsigned long *end)
4647 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
4648 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
4651 * vma need span at least one aligned PUD size and the start,end range
4652 * must at least partialy within it.
4654 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
4655 (*end <= v_start) || (*start >= v_end))
4658 /* Extend the range to be PUD aligned for a worst case scenario */
4659 if (*start > v_start)
4660 *start = ALIGN_DOWN(*start, PUD_SIZE);
4663 *end = ALIGN(*end, PUD_SIZE);
4667 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4668 * and returns the corresponding pte. While this is not necessary for the
4669 * !shared pmd case because we can allocate the pmd later as well, it makes the
4670 * code much cleaner. pmd allocation is essential for the shared case because
4671 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4672 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4673 * bad pmd for sharing.
4675 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4677 struct vm_area_struct *vma = find_vma(mm, addr);
4678 struct address_space *mapping = vma->vm_file->f_mapping;
4679 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4681 struct vm_area_struct *svma;
4682 unsigned long saddr;
4687 if (!vma_shareable(vma, addr))
4688 return (pte_t *)pmd_alloc(mm, pud, addr);
4690 i_mmap_lock_write(mapping);
4691 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4695 saddr = page_table_shareable(svma, vma, addr, idx);
4697 spte = huge_pte_offset(svma->vm_mm, saddr,
4698 vma_mmu_pagesize(svma));
4700 get_page(virt_to_page(spte));
4709 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4710 if (pud_none(*pud)) {
4711 pud_populate(mm, pud,
4712 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4715 put_page(virt_to_page(spte));
4719 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4720 i_mmap_unlock_write(mapping);
4725 * unmap huge page backed by shared pte.
4727 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4728 * indicated by page_count > 1, unmap is achieved by clearing pud and
4729 * decrementing the ref count. If count == 1, the pte page is not shared.
4731 * called with page table lock held.
4733 * returns: 1 successfully unmapped a shared pte page
4734 * 0 the underlying pte page is not shared, or it is the last user
4736 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4738 pgd_t *pgd = pgd_offset(mm, *addr);
4739 p4d_t *p4d = p4d_offset(pgd, *addr);
4740 pud_t *pud = pud_offset(p4d, *addr);
4742 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4743 if (page_count(virt_to_page(ptep)) == 1)
4747 put_page(virt_to_page(ptep));
4749 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4752 #define want_pmd_share() (1)
4753 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4754 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4759 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4764 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4765 unsigned long *start, unsigned long *end)
4768 #define want_pmd_share() (0)
4769 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4771 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4772 pte_t *huge_pte_alloc(struct mm_struct *mm,
4773 unsigned long addr, unsigned long sz)
4780 pgd = pgd_offset(mm, addr);
4781 p4d = p4d_alloc(mm, pgd, addr);
4784 pud = pud_alloc(mm, p4d, addr);
4786 if (sz == PUD_SIZE) {
4789 BUG_ON(sz != PMD_SIZE);
4790 if (want_pmd_share() && pud_none(*pud))
4791 pte = huge_pmd_share(mm, addr, pud);
4793 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4796 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4802 * huge_pte_offset() - Walk the page table to resolve the hugepage
4803 * entry at address @addr
4805 * Return: Pointer to page table or swap entry (PUD or PMD) for
4806 * address @addr, or NULL if a p*d_none() entry is encountered and the
4807 * size @sz doesn't match the hugepage size at this level of the page
4810 pte_t *huge_pte_offset(struct mm_struct *mm,
4811 unsigned long addr, unsigned long sz)
4815 pud_t *pud, pud_entry;
4816 pmd_t *pmd, pmd_entry;
4818 pgd = pgd_offset(mm, addr);
4819 if (!pgd_present(*pgd))
4821 p4d = p4d_offset(pgd, addr);
4822 if (!p4d_present(*p4d))
4825 pud = pud_offset(p4d, addr);
4826 pud_entry = READ_ONCE(*pud);
4827 if (sz != PUD_SIZE && pud_none(pud_entry))
4829 /* hugepage or swap? */
4830 if (pud_huge(pud_entry) || !pud_present(pud_entry))
4831 return (pte_t *)pud;
4833 pmd = pmd_offset(pud, addr);
4834 pmd_entry = READ_ONCE(*pmd);
4835 if (sz != PMD_SIZE && pmd_none(pmd_entry))
4837 /* hugepage or swap? */
4838 if (pmd_huge(pmd_entry) || !pmd_present(pmd_entry))
4839 return (pte_t *)pmd;
4844 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4847 * These functions are overwritable if your architecture needs its own
4850 struct page * __weak
4851 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4854 return ERR_PTR(-EINVAL);
4857 struct page * __weak
4858 follow_huge_pd(struct vm_area_struct *vma,
4859 unsigned long address, hugepd_t hpd, int flags, int pdshift)
4861 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4865 struct page * __weak
4866 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4867 pmd_t *pmd, int flags)
4869 struct page *page = NULL;
4873 ptl = pmd_lockptr(mm, pmd);
4876 * make sure that the address range covered by this pmd is not
4877 * unmapped from other threads.
4879 if (!pmd_huge(*pmd))
4881 pte = huge_ptep_get((pte_t *)pmd);
4882 if (pte_present(pte)) {
4883 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4884 if (flags & FOLL_GET)
4887 if (is_hugetlb_entry_migration(pte)) {
4889 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4893 * hwpoisoned entry is treated as no_page_table in
4894 * follow_page_mask().
4902 struct page * __weak
4903 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4904 pud_t *pud, int flags)
4906 if (flags & FOLL_GET)
4909 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4912 struct page * __weak
4913 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
4915 if (flags & FOLL_GET)
4918 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
4921 bool isolate_huge_page(struct page *page, struct list_head *list)
4925 spin_lock(&hugetlb_lock);
4926 if (!PageHeadHuge(page) || !page_huge_active(page) ||
4927 !get_page_unless_zero(page)) {
4931 clear_page_huge_active(page);
4932 list_move_tail(&page->lru, list);
4934 spin_unlock(&hugetlb_lock);
4938 void putback_active_hugepage(struct page *page)
4940 VM_BUG_ON_PAGE(!PageHead(page), page);
4941 spin_lock(&hugetlb_lock);
4942 set_page_huge_active(page);
4943 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4944 spin_unlock(&hugetlb_lock);