2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
8 #include <linux/seq_file.h>
9 #include <linux/sysctl.h>
10 #include <linux/highmem.h>
11 #include <linux/mmu_notifier.h>
12 #include <linux/nodemask.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/compiler.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/memblock.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/mmdebug.h>
22 #include <linux/sched/signal.h>
23 #include <linux/rmap.h>
24 #include <linux/string_helpers.h>
25 #include <linux/swap.h>
26 #include <linux/swapops.h>
27 #include <linux/jhash.h>
28 #include <linux/numa.h>
31 #include <asm/pgtable.h>
35 #include <linux/hugetlb.h>
36 #include <linux/hugetlb_cgroup.h>
37 #include <linux/node.h>
38 #include <linux/userfaultfd_k.h>
39 #include <linux/page_owner.h>
42 int hugetlb_max_hstate __read_mostly;
43 unsigned int default_hstate_idx;
44 struct hstate hstates[HUGE_MAX_HSTATE];
46 * Minimum page order among possible hugepage sizes, set to a proper value
49 static unsigned int minimum_order __read_mostly = UINT_MAX;
51 __initdata LIST_HEAD(huge_boot_pages);
53 /* for command line parsing */
54 static struct hstate * __initdata parsed_hstate;
55 static unsigned long __initdata default_hstate_max_huge_pages;
56 static unsigned long __initdata default_hstate_size;
57 static bool __initdata parsed_valid_hugepagesz = true;
60 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
61 * free_huge_pages, and surplus_huge_pages.
63 DEFINE_SPINLOCK(hugetlb_lock);
66 * Serializes faults on the same logical page. This is used to
67 * prevent spurious OOMs when the hugepage pool is fully utilized.
69 static int num_fault_mutexes;
70 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
72 /* Forward declaration */
73 static int hugetlb_acct_memory(struct hstate *h, long delta);
75 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
77 bool free = (spool->count == 0) && (spool->used_hpages == 0);
79 spin_unlock(&spool->lock);
81 /* If no pages are used, and no other handles to the subpool
82 * remain, give up any reservations mased on minimum size and
85 if (spool->min_hpages != -1)
86 hugetlb_acct_memory(spool->hstate,
92 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
95 struct hugepage_subpool *spool;
97 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
101 spin_lock_init(&spool->lock);
103 spool->max_hpages = max_hpages;
105 spool->min_hpages = min_hpages;
107 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
111 spool->rsv_hpages = min_hpages;
116 void hugepage_put_subpool(struct hugepage_subpool *spool)
118 spin_lock(&spool->lock);
119 BUG_ON(!spool->count);
121 unlock_or_release_subpool(spool);
125 * Subpool accounting for allocating and reserving pages.
126 * Return -ENOMEM if there are not enough resources to satisfy the
127 * the request. Otherwise, return the number of pages by which the
128 * global pools must be adjusted (upward). The returned value may
129 * only be different than the passed value (delta) in the case where
130 * a subpool minimum size must be manitained.
132 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
140 spin_lock(&spool->lock);
142 if (spool->max_hpages != -1) { /* maximum size accounting */
143 if ((spool->used_hpages + delta) <= spool->max_hpages)
144 spool->used_hpages += delta;
151 /* minimum size accounting */
152 if (spool->min_hpages != -1 && spool->rsv_hpages) {
153 if (delta > spool->rsv_hpages) {
155 * Asking for more reserves than those already taken on
156 * behalf of subpool. Return difference.
158 ret = delta - spool->rsv_hpages;
159 spool->rsv_hpages = 0;
161 ret = 0; /* reserves already accounted for */
162 spool->rsv_hpages -= delta;
167 spin_unlock(&spool->lock);
172 * Subpool accounting for freeing and unreserving pages.
173 * Return the number of global page reservations that must be dropped.
174 * The return value may only be different than the passed value (delta)
175 * in the case where a subpool minimum size must be maintained.
177 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
185 spin_lock(&spool->lock);
187 if (spool->max_hpages != -1) /* maximum size accounting */
188 spool->used_hpages -= delta;
190 /* minimum size accounting */
191 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
192 if (spool->rsv_hpages + delta <= spool->min_hpages)
195 ret = spool->rsv_hpages + delta - spool->min_hpages;
197 spool->rsv_hpages += delta;
198 if (spool->rsv_hpages > spool->min_hpages)
199 spool->rsv_hpages = spool->min_hpages;
203 * If hugetlbfs_put_super couldn't free spool due to an outstanding
204 * quota reference, free it now.
206 unlock_or_release_subpool(spool);
211 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
213 return HUGETLBFS_SB(inode->i_sb)->spool;
216 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
218 return subpool_inode(file_inode(vma->vm_file));
222 * Region tracking -- allows tracking of reservations and instantiated pages
223 * across the pages in a mapping.
225 * The region data structures are embedded into a resv_map and protected
226 * by a resv_map's lock. The set of regions within the resv_map represent
227 * reservations for huge pages, or huge pages that have already been
228 * instantiated within the map. The from and to elements are huge page
229 * indicies into the associated mapping. from indicates the starting index
230 * of the region. to represents the first index past the end of the region.
232 * For example, a file region structure with from == 0 and to == 4 represents
233 * four huge pages in a mapping. It is important to note that the to element
234 * represents the first element past the end of the region. This is used in
235 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
237 * Interval notation of the form [from, to) will be used to indicate that
238 * the endpoint from is inclusive and to is exclusive.
241 struct list_head link;
247 * Add the huge page range represented by [f, t) to the reserve
248 * map. In the normal case, existing regions will be expanded
249 * to accommodate the specified range. Sufficient regions should
250 * exist for expansion due to the previous call to region_chg
251 * with the same range. However, it is possible that region_del
252 * could have been called after region_chg and modifed the map
253 * in such a way that no region exists to be expanded. In this
254 * case, pull a region descriptor from the cache associated with
255 * the map and use that for the new range.
257 * Return the number of new huge pages added to the map. This
258 * number is greater than or equal to zero.
260 static long region_add(struct resv_map *resv, long f, long t)
262 struct list_head *head = &resv->regions;
263 struct file_region *rg, *nrg, *trg;
266 spin_lock(&resv->lock);
267 /* Locate the region we are either in or before. */
268 list_for_each_entry(rg, head, link)
273 * If no region exists which can be expanded to include the
274 * specified range, the list must have been modified by an
275 * interleving call to region_del(). Pull a region descriptor
276 * from the cache and use it for this range.
278 if (&rg->link == head || t < rg->from) {
279 VM_BUG_ON(resv->region_cache_count <= 0);
281 resv->region_cache_count--;
282 nrg = list_first_entry(&resv->region_cache, struct file_region,
284 list_del(&nrg->link);
288 list_add(&nrg->link, rg->link.prev);
294 /* Round our left edge to the current segment if it encloses us. */
298 /* Check for and consume any regions we now overlap with. */
300 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
301 if (&rg->link == head)
306 /* If this area reaches higher then extend our area to
307 * include it completely. If this is not the first area
308 * which we intend to reuse, free it. */
312 /* Decrement return value by the deleted range.
313 * Another range will span this area so that by
314 * end of routine add will be >= zero
316 add -= (rg->to - rg->from);
322 add += (nrg->from - f); /* Added to beginning of region */
324 add += t - nrg->to; /* Added to end of region */
328 resv->adds_in_progress--;
329 spin_unlock(&resv->lock);
335 * Examine the existing reserve map and determine how many
336 * huge pages in the specified range [f, t) are NOT currently
337 * represented. This routine is called before a subsequent
338 * call to region_add that will actually modify the reserve
339 * map to add the specified range [f, t). region_chg does
340 * not change the number of huge pages represented by the
341 * map. However, if the existing regions in the map can not
342 * be expanded to represent the new range, a new file_region
343 * structure is added to the map as a placeholder. This is
344 * so that the subsequent region_add call will have all the
345 * regions it needs and will not fail.
347 * Upon entry, region_chg will also examine the cache of region descriptors
348 * associated with the map. If there are not enough descriptors cached, one
349 * will be allocated for the in progress add operation.
351 * Returns the number of huge pages that need to be added to the existing
352 * reservation map for the range [f, t). This number is greater or equal to
353 * zero. -ENOMEM is returned if a new file_region structure or cache entry
354 * is needed and can not be allocated.
356 static long region_chg(struct resv_map *resv, long f, long t)
358 struct list_head *head = &resv->regions;
359 struct file_region *rg, *nrg = NULL;
363 spin_lock(&resv->lock);
365 resv->adds_in_progress++;
368 * Check for sufficient descriptors in the cache to accommodate
369 * the number of in progress add operations.
371 if (resv->adds_in_progress > resv->region_cache_count) {
372 struct file_region *trg;
374 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
375 /* Must drop lock to allocate a new descriptor. */
376 resv->adds_in_progress--;
377 spin_unlock(&resv->lock);
379 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
385 spin_lock(&resv->lock);
386 list_add(&trg->link, &resv->region_cache);
387 resv->region_cache_count++;
391 /* Locate the region we are before or in. */
392 list_for_each_entry(rg, head, link)
396 /* If we are below the current region then a new region is required.
397 * Subtle, allocate a new region at the position but make it zero
398 * size such that we can guarantee to record the reservation. */
399 if (&rg->link == head || t < rg->from) {
401 resv->adds_in_progress--;
402 spin_unlock(&resv->lock);
403 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
409 INIT_LIST_HEAD(&nrg->link);
413 list_add(&nrg->link, rg->link.prev);
418 /* Round our left edge to the current segment if it encloses us. */
423 /* Check for and consume any regions we now overlap with. */
424 list_for_each_entry(rg, rg->link.prev, link) {
425 if (&rg->link == head)
430 /* We overlap with this area, if it extends further than
431 * us then we must extend ourselves. Account for its
432 * existing reservation. */
437 chg -= rg->to - rg->from;
441 spin_unlock(&resv->lock);
442 /* We already know we raced and no longer need the new region */
446 spin_unlock(&resv->lock);
451 * Abort the in progress add operation. The adds_in_progress field
452 * of the resv_map keeps track of the operations in progress between
453 * calls to region_chg and region_add. Operations are sometimes
454 * aborted after the call to region_chg. In such cases, region_abort
455 * is called to decrement the adds_in_progress counter.
457 * NOTE: The range arguments [f, t) are not needed or used in this
458 * routine. They are kept to make reading the calling code easier as
459 * arguments will match the associated region_chg call.
461 static void region_abort(struct resv_map *resv, long f, long t)
463 spin_lock(&resv->lock);
464 VM_BUG_ON(!resv->region_cache_count);
465 resv->adds_in_progress--;
466 spin_unlock(&resv->lock);
470 * Delete the specified range [f, t) from the reserve map. If the
471 * t parameter is LONG_MAX, this indicates that ALL regions after f
472 * should be deleted. Locate the regions which intersect [f, t)
473 * and either trim, delete or split the existing regions.
475 * Returns the number of huge pages deleted from the reserve map.
476 * In the normal case, the return value is zero or more. In the
477 * case where a region must be split, a new region descriptor must
478 * be allocated. If the allocation fails, -ENOMEM will be returned.
479 * NOTE: If the parameter t == LONG_MAX, then we will never split
480 * a region and possibly return -ENOMEM. Callers specifying
481 * t == LONG_MAX do not need to check for -ENOMEM error.
483 static long region_del(struct resv_map *resv, long f, long t)
485 struct list_head *head = &resv->regions;
486 struct file_region *rg, *trg;
487 struct file_region *nrg = NULL;
491 spin_lock(&resv->lock);
492 list_for_each_entry_safe(rg, trg, head, link) {
494 * Skip regions before the range to be deleted. file_region
495 * ranges are normally of the form [from, to). However, there
496 * may be a "placeholder" entry in the map which is of the form
497 * (from, to) with from == to. Check for placeholder entries
498 * at the beginning of the range to be deleted.
500 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
506 if (f > rg->from && t < rg->to) { /* Must split region */
508 * Check for an entry in the cache before dropping
509 * lock and attempting allocation.
512 resv->region_cache_count > resv->adds_in_progress) {
513 nrg = list_first_entry(&resv->region_cache,
516 list_del(&nrg->link);
517 resv->region_cache_count--;
521 spin_unlock(&resv->lock);
522 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
530 /* New entry for end of split region */
533 INIT_LIST_HEAD(&nrg->link);
535 /* Original entry is trimmed */
538 list_add(&nrg->link, &rg->link);
543 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
544 del += rg->to - rg->from;
550 if (f <= rg->from) { /* Trim beginning of region */
553 } else { /* Trim end of region */
559 spin_unlock(&resv->lock);
565 * A rare out of memory error was encountered which prevented removal of
566 * the reserve map region for a page. The huge page itself was free'ed
567 * and removed from the page cache. This routine will adjust the subpool
568 * usage count, and the global reserve count if needed. By incrementing
569 * these counts, the reserve map entry which could not be deleted will
570 * appear as a "reserved" entry instead of simply dangling with incorrect
573 void hugetlb_fix_reserve_counts(struct inode *inode)
575 struct hugepage_subpool *spool = subpool_inode(inode);
578 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
580 struct hstate *h = hstate_inode(inode);
582 hugetlb_acct_memory(h, 1);
587 * Count and return the number of huge pages in the reserve map
588 * that intersect with the range [f, t).
590 static long region_count(struct resv_map *resv, long f, long t)
592 struct list_head *head = &resv->regions;
593 struct file_region *rg;
596 spin_lock(&resv->lock);
597 /* Locate each segment we overlap with, and count that overlap. */
598 list_for_each_entry(rg, head, link) {
607 seg_from = max(rg->from, f);
608 seg_to = min(rg->to, t);
610 chg += seg_to - seg_from;
612 spin_unlock(&resv->lock);
618 * Convert the address within this vma to the page offset within
619 * the mapping, in pagecache page units; huge pages here.
621 static pgoff_t vma_hugecache_offset(struct hstate *h,
622 struct vm_area_struct *vma, unsigned long address)
624 return ((address - vma->vm_start) >> huge_page_shift(h)) +
625 (vma->vm_pgoff >> huge_page_order(h));
628 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
629 unsigned long address)
631 return vma_hugecache_offset(hstate_vma(vma), vma, address);
633 EXPORT_SYMBOL_GPL(linear_hugepage_index);
636 * Return the size of the pages allocated when backing a VMA. In the majority
637 * cases this will be same size as used by the page table entries.
639 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
641 if (vma->vm_ops && vma->vm_ops->pagesize)
642 return vma->vm_ops->pagesize(vma);
645 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
648 * Return the page size being used by the MMU to back a VMA. In the majority
649 * of cases, the page size used by the kernel matches the MMU size. On
650 * architectures where it differs, an architecture-specific 'strong'
651 * version of this symbol is required.
653 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
655 return vma_kernel_pagesize(vma);
659 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
660 * bits of the reservation map pointer, which are always clear due to
663 #define HPAGE_RESV_OWNER (1UL << 0)
664 #define HPAGE_RESV_UNMAPPED (1UL << 1)
665 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
668 * These helpers are used to track how many pages are reserved for
669 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
670 * is guaranteed to have their future faults succeed.
672 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
673 * the reserve counters are updated with the hugetlb_lock held. It is safe
674 * to reset the VMA at fork() time as it is not in use yet and there is no
675 * chance of the global counters getting corrupted as a result of the values.
677 * The private mapping reservation is represented in a subtly different
678 * manner to a shared mapping. A shared mapping has a region map associated
679 * with the underlying file, this region map represents the backing file
680 * pages which have ever had a reservation assigned which this persists even
681 * after the page is instantiated. A private mapping has a region map
682 * associated with the original mmap which is attached to all VMAs which
683 * reference it, this region map represents those offsets which have consumed
684 * reservation ie. where pages have been instantiated.
686 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
688 return (unsigned long)vma->vm_private_data;
691 static void set_vma_private_data(struct vm_area_struct *vma,
694 vma->vm_private_data = (void *)value;
697 struct resv_map *resv_map_alloc(void)
699 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
700 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
702 if (!resv_map || !rg) {
708 kref_init(&resv_map->refs);
709 spin_lock_init(&resv_map->lock);
710 INIT_LIST_HEAD(&resv_map->regions);
712 resv_map->adds_in_progress = 0;
714 INIT_LIST_HEAD(&resv_map->region_cache);
715 list_add(&rg->link, &resv_map->region_cache);
716 resv_map->region_cache_count = 1;
721 void resv_map_release(struct kref *ref)
723 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
724 struct list_head *head = &resv_map->region_cache;
725 struct file_region *rg, *trg;
727 /* Clear out any active regions before we release the map. */
728 region_del(resv_map, 0, LONG_MAX);
730 /* ... and any entries left in the cache */
731 list_for_each_entry_safe(rg, trg, head, link) {
736 VM_BUG_ON(resv_map->adds_in_progress);
741 static inline struct resv_map *inode_resv_map(struct inode *inode)
743 return inode->i_mapping->private_data;
746 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
748 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
749 if (vma->vm_flags & VM_MAYSHARE) {
750 struct address_space *mapping = vma->vm_file->f_mapping;
751 struct inode *inode = mapping->host;
753 return inode_resv_map(inode);
756 return (struct resv_map *)(get_vma_private_data(vma) &
761 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
763 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
764 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
766 set_vma_private_data(vma, (get_vma_private_data(vma) &
767 HPAGE_RESV_MASK) | (unsigned long)map);
770 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
772 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
773 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
775 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
778 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
780 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
782 return (get_vma_private_data(vma) & flag) != 0;
785 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
786 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
788 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
789 if (!(vma->vm_flags & VM_MAYSHARE))
790 vma->vm_private_data = (void *)0;
793 /* Returns true if the VMA has associated reserve pages */
794 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
796 if (vma->vm_flags & VM_NORESERVE) {
798 * This address is already reserved by other process(chg == 0),
799 * so, we should decrement reserved count. Without decrementing,
800 * reserve count remains after releasing inode, because this
801 * allocated page will go into page cache and is regarded as
802 * coming from reserved pool in releasing step. Currently, we
803 * don't have any other solution to deal with this situation
804 * properly, so add work-around here.
806 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
812 /* Shared mappings always use reserves */
813 if (vma->vm_flags & VM_MAYSHARE) {
815 * We know VM_NORESERVE is not set. Therefore, there SHOULD
816 * be a region map for all pages. The only situation where
817 * there is no region map is if a hole was punched via
818 * fallocate. In this case, there really are no reverves to
819 * use. This situation is indicated if chg != 0.
828 * Only the process that called mmap() has reserves for
831 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
833 * Like the shared case above, a hole punch or truncate
834 * could have been performed on the private mapping.
835 * Examine the value of chg to determine if reserves
836 * actually exist or were previously consumed.
837 * Very Subtle - The value of chg comes from a previous
838 * call to vma_needs_reserves(). The reserve map for
839 * private mappings has different (opposite) semantics
840 * than that of shared mappings. vma_needs_reserves()
841 * has already taken this difference in semantics into
842 * account. Therefore, the meaning of chg is the same
843 * as in the shared case above. Code could easily be
844 * combined, but keeping it separate draws attention to
845 * subtle differences.
856 static void enqueue_huge_page(struct hstate *h, struct page *page)
858 int nid = page_to_nid(page);
859 list_move(&page->lru, &h->hugepage_freelists[nid]);
860 h->free_huge_pages++;
861 h->free_huge_pages_node[nid]++;
864 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
868 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
869 if (!PageHWPoison(page))
872 * if 'non-isolated free hugepage' not found on the list,
873 * the allocation fails.
875 if (&h->hugepage_freelists[nid] == &page->lru)
877 list_move(&page->lru, &h->hugepage_activelist);
878 set_page_refcounted(page);
879 h->free_huge_pages--;
880 h->free_huge_pages_node[nid]--;
884 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
887 unsigned int cpuset_mems_cookie;
888 struct zonelist *zonelist;
891 int node = NUMA_NO_NODE;
893 zonelist = node_zonelist(nid, gfp_mask);
896 cpuset_mems_cookie = read_mems_allowed_begin();
897 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
900 if (!cpuset_zone_allowed(zone, gfp_mask))
903 * no need to ask again on the same node. Pool is node rather than
906 if (zone_to_nid(zone) == node)
908 node = zone_to_nid(zone);
910 page = dequeue_huge_page_node_exact(h, node);
914 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
920 /* Movability of hugepages depends on migration support. */
921 static inline gfp_t htlb_alloc_mask(struct hstate *h)
923 if (hugepage_movable_supported(h))
924 return GFP_HIGHUSER_MOVABLE;
929 static struct page *dequeue_huge_page_vma(struct hstate *h,
930 struct vm_area_struct *vma,
931 unsigned long address, int avoid_reserve,
935 struct mempolicy *mpol;
937 nodemask_t *nodemask;
941 * A child process with MAP_PRIVATE mappings created by their parent
942 * have no page reserves. This check ensures that reservations are
943 * not "stolen". The child may still get SIGKILLed
945 if (!vma_has_reserves(vma, chg) &&
946 h->free_huge_pages - h->resv_huge_pages == 0)
949 /* If reserves cannot be used, ensure enough pages are in the pool */
950 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
953 gfp_mask = htlb_alloc_mask(h);
954 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
955 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
956 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
957 SetPagePrivate(page);
958 h->resv_huge_pages--;
969 * common helper functions for hstate_next_node_to_{alloc|free}.
970 * We may have allocated or freed a huge page based on a different
971 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
972 * be outside of *nodes_allowed. Ensure that we use an allowed
973 * node for alloc or free.
975 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
977 nid = next_node_in(nid, *nodes_allowed);
978 VM_BUG_ON(nid >= MAX_NUMNODES);
983 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
985 if (!node_isset(nid, *nodes_allowed))
986 nid = next_node_allowed(nid, nodes_allowed);
991 * returns the previously saved node ["this node"] from which to
992 * allocate a persistent huge page for the pool and advance the
993 * next node from which to allocate, handling wrap at end of node
996 static int hstate_next_node_to_alloc(struct hstate *h,
997 nodemask_t *nodes_allowed)
1001 VM_BUG_ON(!nodes_allowed);
1003 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1004 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1010 * helper for free_pool_huge_page() - return the previously saved
1011 * node ["this node"] from which to free a huge page. Advance the
1012 * next node id whether or not we find a free huge page to free so
1013 * that the next attempt to free addresses the next node.
1015 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1019 VM_BUG_ON(!nodes_allowed);
1021 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1022 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1027 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1028 for (nr_nodes = nodes_weight(*mask); \
1030 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1033 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1034 for (nr_nodes = nodes_weight(*mask); \
1036 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1039 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1040 static void destroy_compound_gigantic_page(struct page *page,
1044 int nr_pages = 1 << order;
1045 struct page *p = page + 1;
1047 atomic_set(compound_mapcount_ptr(page), 0);
1048 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1049 clear_compound_head(p);
1050 set_page_refcounted(p);
1053 set_compound_order(page, 0);
1054 __ClearPageHead(page);
1057 static void free_gigantic_page(struct page *page, unsigned int order)
1059 free_contig_range(page_to_pfn(page), 1 << order);
1062 #ifdef CONFIG_CONTIG_ALLOC
1063 static int __alloc_gigantic_page(unsigned long start_pfn,
1064 unsigned long nr_pages, gfp_t gfp_mask)
1066 unsigned long end_pfn = start_pfn + nr_pages;
1067 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1071 static bool pfn_range_valid_gigantic(struct zone *z,
1072 unsigned long start_pfn, unsigned long nr_pages)
1074 unsigned long i, end_pfn = start_pfn + nr_pages;
1077 for (i = start_pfn; i < end_pfn; i++) {
1081 page = pfn_to_page(i);
1083 if (page_zone(page) != z)
1086 if (PageReserved(page))
1089 if (page_count(page) > 0)
1099 static bool zone_spans_last_pfn(const struct zone *zone,
1100 unsigned long start_pfn, unsigned long nr_pages)
1102 unsigned long last_pfn = start_pfn + nr_pages - 1;
1103 return zone_spans_pfn(zone, last_pfn);
1106 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1107 int nid, nodemask_t *nodemask)
1109 unsigned int order = huge_page_order(h);
1110 unsigned long nr_pages = 1 << order;
1111 unsigned long ret, pfn, flags;
1112 struct zonelist *zonelist;
1116 zonelist = node_zonelist(nid, gfp_mask);
1117 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) {
1118 spin_lock_irqsave(&zone->lock, flags);
1120 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
1121 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
1122 if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) {
1124 * We release the zone lock here because
1125 * alloc_contig_range() will also lock the zone
1126 * at some point. If there's an allocation
1127 * spinning on this lock, it may win the race
1128 * and cause alloc_contig_range() to fail...
1130 spin_unlock_irqrestore(&zone->lock, flags);
1131 ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask);
1133 return pfn_to_page(pfn);
1134 spin_lock_irqsave(&zone->lock, flags);
1139 spin_unlock_irqrestore(&zone->lock, flags);
1145 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1146 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1147 #else /* !CONFIG_CONTIG_ALLOC */
1148 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1149 int nid, nodemask_t *nodemask)
1153 #endif /* CONFIG_CONTIG_ALLOC */
1155 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1156 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1157 int nid, nodemask_t *nodemask)
1161 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1162 static inline void destroy_compound_gigantic_page(struct page *page,
1163 unsigned int order) { }
1166 static void update_and_free_page(struct hstate *h, struct page *page)
1170 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1174 h->nr_huge_pages_node[page_to_nid(page)]--;
1175 for (i = 0; i < pages_per_huge_page(h); i++) {
1176 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1177 1 << PG_referenced | 1 << PG_dirty |
1178 1 << PG_active | 1 << PG_private |
1181 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1182 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1183 set_page_refcounted(page);
1184 if (hstate_is_gigantic(h)) {
1185 destroy_compound_gigantic_page(page, huge_page_order(h));
1186 free_gigantic_page(page, huge_page_order(h));
1188 __free_pages(page, huge_page_order(h));
1192 struct hstate *size_to_hstate(unsigned long size)
1196 for_each_hstate(h) {
1197 if (huge_page_size(h) == size)
1204 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1205 * to hstate->hugepage_activelist.)
1207 * This function can be called for tail pages, but never returns true for them.
1209 bool page_huge_active(struct page *page)
1211 VM_BUG_ON_PAGE(!PageHuge(page), page);
1212 return PageHead(page) && PagePrivate(&page[1]);
1215 /* never called for tail page */
1216 static void set_page_huge_active(struct page *page)
1218 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1219 SetPagePrivate(&page[1]);
1222 static void clear_page_huge_active(struct page *page)
1224 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1225 ClearPagePrivate(&page[1]);
1229 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1232 static inline bool PageHugeTemporary(struct page *page)
1234 if (!PageHuge(page))
1237 return (unsigned long)page[2].mapping == -1U;
1240 static inline void SetPageHugeTemporary(struct page *page)
1242 page[2].mapping = (void *)-1U;
1245 static inline void ClearPageHugeTemporary(struct page *page)
1247 page[2].mapping = NULL;
1250 void free_huge_page(struct page *page)
1253 * Can't pass hstate in here because it is called from the
1254 * compound page destructor.
1256 struct hstate *h = page_hstate(page);
1257 int nid = page_to_nid(page);
1258 struct hugepage_subpool *spool =
1259 (struct hugepage_subpool *)page_private(page);
1260 bool restore_reserve;
1262 VM_BUG_ON_PAGE(page_count(page), page);
1263 VM_BUG_ON_PAGE(page_mapcount(page), page);
1265 set_page_private(page, 0);
1266 page->mapping = NULL;
1267 restore_reserve = PagePrivate(page);
1268 ClearPagePrivate(page);
1271 * If PagePrivate() was set on page, page allocation consumed a
1272 * reservation. If the page was associated with a subpool, there
1273 * would have been a page reserved in the subpool before allocation
1274 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1275 * reservtion, do not call hugepage_subpool_put_pages() as this will
1276 * remove the reserved page from the subpool.
1278 if (!restore_reserve) {
1280 * A return code of zero implies that the subpool will be
1281 * under its minimum size if the reservation is not restored
1282 * after page is free. Therefore, force restore_reserve
1285 if (hugepage_subpool_put_pages(spool, 1) == 0)
1286 restore_reserve = true;
1289 spin_lock(&hugetlb_lock);
1290 clear_page_huge_active(page);
1291 hugetlb_cgroup_uncharge_page(hstate_index(h),
1292 pages_per_huge_page(h), page);
1293 if (restore_reserve)
1294 h->resv_huge_pages++;
1296 if (PageHugeTemporary(page)) {
1297 list_del(&page->lru);
1298 ClearPageHugeTemporary(page);
1299 update_and_free_page(h, page);
1300 } else if (h->surplus_huge_pages_node[nid]) {
1301 /* remove the page from active list */
1302 list_del(&page->lru);
1303 update_and_free_page(h, page);
1304 h->surplus_huge_pages--;
1305 h->surplus_huge_pages_node[nid]--;
1307 arch_clear_hugepage_flags(page);
1308 enqueue_huge_page(h, page);
1310 spin_unlock(&hugetlb_lock);
1313 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1315 INIT_LIST_HEAD(&page->lru);
1316 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1317 spin_lock(&hugetlb_lock);
1318 set_hugetlb_cgroup(page, NULL);
1320 h->nr_huge_pages_node[nid]++;
1321 spin_unlock(&hugetlb_lock);
1324 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1327 int nr_pages = 1 << order;
1328 struct page *p = page + 1;
1330 /* we rely on prep_new_huge_page to set the destructor */
1331 set_compound_order(page, order);
1332 __ClearPageReserved(page);
1333 __SetPageHead(page);
1334 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1336 * For gigantic hugepages allocated through bootmem at
1337 * boot, it's safer to be consistent with the not-gigantic
1338 * hugepages and clear the PG_reserved bit from all tail pages
1339 * too. Otherwse drivers using get_user_pages() to access tail
1340 * pages may get the reference counting wrong if they see
1341 * PG_reserved set on a tail page (despite the head page not
1342 * having PG_reserved set). Enforcing this consistency between
1343 * head and tail pages allows drivers to optimize away a check
1344 * on the head page when they need know if put_page() is needed
1345 * after get_user_pages().
1347 __ClearPageReserved(p);
1348 set_page_count(p, 0);
1349 set_compound_head(p, page);
1351 atomic_set(compound_mapcount_ptr(page), -1);
1355 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1356 * transparent huge pages. See the PageTransHuge() documentation for more
1359 int PageHuge(struct page *page)
1361 if (!PageCompound(page))
1364 page = compound_head(page);
1365 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1367 EXPORT_SYMBOL_GPL(PageHuge);
1370 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1371 * normal or transparent huge pages.
1373 int PageHeadHuge(struct page *page_head)
1375 if (!PageHead(page_head))
1378 return get_compound_page_dtor(page_head) == free_huge_page;
1381 pgoff_t __basepage_index(struct page *page)
1383 struct page *page_head = compound_head(page);
1384 pgoff_t index = page_index(page_head);
1385 unsigned long compound_idx;
1387 if (!PageHuge(page_head))
1388 return page_index(page);
1390 if (compound_order(page_head) >= MAX_ORDER)
1391 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1393 compound_idx = page - page_head;
1395 return (index << compound_order(page_head)) + compound_idx;
1398 static struct page *alloc_buddy_huge_page(struct hstate *h,
1399 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1401 int order = huge_page_order(h);
1404 gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
1405 if (nid == NUMA_NO_NODE)
1406 nid = numa_mem_id();
1407 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1409 __count_vm_event(HTLB_BUDDY_PGALLOC);
1411 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1417 * Common helper to allocate a fresh hugetlb page. All specific allocators
1418 * should use this function to get new hugetlb pages
1420 static struct page *alloc_fresh_huge_page(struct hstate *h,
1421 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1425 if (hstate_is_gigantic(h))
1426 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1428 page = alloc_buddy_huge_page(h, gfp_mask,
1433 if (hstate_is_gigantic(h))
1434 prep_compound_gigantic_page(page, huge_page_order(h));
1435 prep_new_huge_page(h, page, page_to_nid(page));
1441 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1444 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1448 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1450 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1451 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed);
1459 put_page(page); /* free it into the hugepage allocator */
1465 * Free huge page from pool from next node to free.
1466 * Attempt to keep persistent huge pages more or less
1467 * balanced over allowed nodes.
1468 * Called with hugetlb_lock locked.
1470 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1476 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1478 * If we're returning unused surplus pages, only examine
1479 * nodes with surplus pages.
1481 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1482 !list_empty(&h->hugepage_freelists[node])) {
1484 list_entry(h->hugepage_freelists[node].next,
1486 list_del(&page->lru);
1487 h->free_huge_pages--;
1488 h->free_huge_pages_node[node]--;
1490 h->surplus_huge_pages--;
1491 h->surplus_huge_pages_node[node]--;
1493 update_and_free_page(h, page);
1503 * Dissolve a given free hugepage into free buddy pages. This function does
1504 * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1505 * dissolution fails because a give page is not a free hugepage, or because
1506 * free hugepages are fully reserved.
1508 int dissolve_free_huge_page(struct page *page)
1512 spin_lock(&hugetlb_lock);
1513 if (PageHuge(page) && !page_count(page)) {
1514 struct page *head = compound_head(page);
1515 struct hstate *h = page_hstate(head);
1516 int nid = page_to_nid(head);
1517 if (h->free_huge_pages - h->resv_huge_pages == 0)
1520 * Move PageHWPoison flag from head page to the raw error page,
1521 * which makes any subpages rather than the error page reusable.
1523 if (PageHWPoison(head) && page != head) {
1524 SetPageHWPoison(page);
1525 ClearPageHWPoison(head);
1527 list_del(&head->lru);
1528 h->free_huge_pages--;
1529 h->free_huge_pages_node[nid]--;
1530 h->max_huge_pages--;
1531 update_and_free_page(h, head);
1535 spin_unlock(&hugetlb_lock);
1540 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1541 * make specified memory blocks removable from the system.
1542 * Note that this will dissolve a free gigantic hugepage completely, if any
1543 * part of it lies within the given range.
1544 * Also note that if dissolve_free_huge_page() returns with an error, all
1545 * free hugepages that were dissolved before that error are lost.
1547 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1553 if (!hugepages_supported())
1556 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1557 page = pfn_to_page(pfn);
1558 if (PageHuge(page) && !page_count(page)) {
1559 rc = dissolve_free_huge_page(page);
1569 * Allocates a fresh surplus page from the page allocator.
1571 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1572 int nid, nodemask_t *nmask)
1574 struct page *page = NULL;
1576 if (hstate_is_gigantic(h))
1579 spin_lock(&hugetlb_lock);
1580 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1582 spin_unlock(&hugetlb_lock);
1584 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1588 spin_lock(&hugetlb_lock);
1590 * We could have raced with the pool size change.
1591 * Double check that and simply deallocate the new page
1592 * if we would end up overcommiting the surpluses. Abuse
1593 * temporary page to workaround the nasty free_huge_page
1596 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1597 SetPageHugeTemporary(page);
1598 spin_unlock(&hugetlb_lock);
1602 h->surplus_huge_pages++;
1603 h->surplus_huge_pages_node[page_to_nid(page)]++;
1607 spin_unlock(&hugetlb_lock);
1612 struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1613 int nid, nodemask_t *nmask)
1617 if (hstate_is_gigantic(h))
1620 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1625 * We do not account these pages as surplus because they are only
1626 * temporary and will be released properly on the last reference
1628 SetPageHugeTemporary(page);
1634 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1637 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1638 struct vm_area_struct *vma, unsigned long addr)
1641 struct mempolicy *mpol;
1642 gfp_t gfp_mask = htlb_alloc_mask(h);
1644 nodemask_t *nodemask;
1646 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1647 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1648 mpol_cond_put(mpol);
1653 /* page migration callback function */
1654 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1656 gfp_t gfp_mask = htlb_alloc_mask(h);
1657 struct page *page = NULL;
1659 if (nid != NUMA_NO_NODE)
1660 gfp_mask |= __GFP_THISNODE;
1662 spin_lock(&hugetlb_lock);
1663 if (h->free_huge_pages - h->resv_huge_pages > 0)
1664 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1665 spin_unlock(&hugetlb_lock);
1668 page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1673 /* page migration callback function */
1674 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1677 gfp_t gfp_mask = htlb_alloc_mask(h);
1679 spin_lock(&hugetlb_lock);
1680 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1683 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1685 spin_unlock(&hugetlb_lock);
1689 spin_unlock(&hugetlb_lock);
1691 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1694 /* mempolicy aware migration callback */
1695 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1696 unsigned long address)
1698 struct mempolicy *mpol;
1699 nodemask_t *nodemask;
1704 gfp_mask = htlb_alloc_mask(h);
1705 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1706 page = alloc_huge_page_nodemask(h, node, nodemask);
1707 mpol_cond_put(mpol);
1713 * Increase the hugetlb pool such that it can accommodate a reservation
1716 static int gather_surplus_pages(struct hstate *h, int delta)
1718 struct list_head surplus_list;
1719 struct page *page, *tmp;
1721 int needed, allocated;
1722 bool alloc_ok = true;
1724 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1726 h->resv_huge_pages += delta;
1731 INIT_LIST_HEAD(&surplus_list);
1735 spin_unlock(&hugetlb_lock);
1736 for (i = 0; i < needed; i++) {
1737 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1738 NUMA_NO_NODE, NULL);
1743 list_add(&page->lru, &surplus_list);
1749 * After retaking hugetlb_lock, we need to recalculate 'needed'
1750 * because either resv_huge_pages or free_huge_pages may have changed.
1752 spin_lock(&hugetlb_lock);
1753 needed = (h->resv_huge_pages + delta) -
1754 (h->free_huge_pages + allocated);
1759 * We were not able to allocate enough pages to
1760 * satisfy the entire reservation so we free what
1761 * we've allocated so far.
1766 * The surplus_list now contains _at_least_ the number of extra pages
1767 * needed to accommodate the reservation. Add the appropriate number
1768 * of pages to the hugetlb pool and free the extras back to the buddy
1769 * allocator. Commit the entire reservation here to prevent another
1770 * process from stealing the pages as they are added to the pool but
1771 * before they are reserved.
1773 needed += allocated;
1774 h->resv_huge_pages += delta;
1777 /* Free the needed pages to the hugetlb pool */
1778 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1782 * This page is now managed by the hugetlb allocator and has
1783 * no users -- drop the buddy allocator's reference.
1785 put_page_testzero(page);
1786 VM_BUG_ON_PAGE(page_count(page), page);
1787 enqueue_huge_page(h, page);
1790 spin_unlock(&hugetlb_lock);
1792 /* Free unnecessary surplus pages to the buddy allocator */
1793 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1795 spin_lock(&hugetlb_lock);
1801 * This routine has two main purposes:
1802 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1803 * in unused_resv_pages. This corresponds to the prior adjustments made
1804 * to the associated reservation map.
1805 * 2) Free any unused surplus pages that may have been allocated to satisfy
1806 * the reservation. As many as unused_resv_pages may be freed.
1808 * Called with hugetlb_lock held. However, the lock could be dropped (and
1809 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1810 * we must make sure nobody else can claim pages we are in the process of
1811 * freeing. Do this by ensuring resv_huge_page always is greater than the
1812 * number of huge pages we plan to free when dropping the lock.
1814 static void return_unused_surplus_pages(struct hstate *h,
1815 unsigned long unused_resv_pages)
1817 unsigned long nr_pages;
1819 /* Cannot return gigantic pages currently */
1820 if (hstate_is_gigantic(h))
1824 * Part (or even all) of the reservation could have been backed
1825 * by pre-allocated pages. Only free surplus pages.
1827 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1830 * We want to release as many surplus pages as possible, spread
1831 * evenly across all nodes with memory. Iterate across these nodes
1832 * until we can no longer free unreserved surplus pages. This occurs
1833 * when the nodes with surplus pages have no free pages.
1834 * free_pool_huge_page() will balance the the freed pages across the
1835 * on-line nodes with memory and will handle the hstate accounting.
1837 * Note that we decrement resv_huge_pages as we free the pages. If
1838 * we drop the lock, resv_huge_pages will still be sufficiently large
1839 * to cover subsequent pages we may free.
1841 while (nr_pages--) {
1842 h->resv_huge_pages--;
1843 unused_resv_pages--;
1844 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1846 cond_resched_lock(&hugetlb_lock);
1850 /* Fully uncommit the reservation */
1851 h->resv_huge_pages -= unused_resv_pages;
1856 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1857 * are used by the huge page allocation routines to manage reservations.
1859 * vma_needs_reservation is called to determine if the huge page at addr
1860 * within the vma has an associated reservation. If a reservation is
1861 * needed, the value 1 is returned. The caller is then responsible for
1862 * managing the global reservation and subpool usage counts. After
1863 * the huge page has been allocated, vma_commit_reservation is called
1864 * to add the page to the reservation map. If the page allocation fails,
1865 * the reservation must be ended instead of committed. vma_end_reservation
1866 * is called in such cases.
1868 * In the normal case, vma_commit_reservation returns the same value
1869 * as the preceding vma_needs_reservation call. The only time this
1870 * is not the case is if a reserve map was changed between calls. It
1871 * is the responsibility of the caller to notice the difference and
1872 * take appropriate action.
1874 * vma_add_reservation is used in error paths where a reservation must
1875 * be restored when a newly allocated huge page must be freed. It is
1876 * to be called after calling vma_needs_reservation to determine if a
1877 * reservation exists.
1879 enum vma_resv_mode {
1885 static long __vma_reservation_common(struct hstate *h,
1886 struct vm_area_struct *vma, unsigned long addr,
1887 enum vma_resv_mode mode)
1889 struct resv_map *resv;
1893 resv = vma_resv_map(vma);
1897 idx = vma_hugecache_offset(h, vma, addr);
1899 case VMA_NEEDS_RESV:
1900 ret = region_chg(resv, idx, idx + 1);
1902 case VMA_COMMIT_RESV:
1903 ret = region_add(resv, idx, idx + 1);
1906 region_abort(resv, idx, idx + 1);
1910 if (vma->vm_flags & VM_MAYSHARE)
1911 ret = region_add(resv, idx, idx + 1);
1913 region_abort(resv, idx, idx + 1);
1914 ret = region_del(resv, idx, idx + 1);
1921 if (vma->vm_flags & VM_MAYSHARE)
1923 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1925 * In most cases, reserves always exist for private mappings.
1926 * However, a file associated with mapping could have been
1927 * hole punched or truncated after reserves were consumed.
1928 * As subsequent fault on such a range will not use reserves.
1929 * Subtle - The reserve map for private mappings has the
1930 * opposite meaning than that of shared mappings. If NO
1931 * entry is in the reserve map, it means a reservation exists.
1932 * If an entry exists in the reserve map, it means the
1933 * reservation has already been consumed. As a result, the
1934 * return value of this routine is the opposite of the
1935 * value returned from reserve map manipulation routines above.
1943 return ret < 0 ? ret : 0;
1946 static long vma_needs_reservation(struct hstate *h,
1947 struct vm_area_struct *vma, unsigned long addr)
1949 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1952 static long vma_commit_reservation(struct hstate *h,
1953 struct vm_area_struct *vma, unsigned long addr)
1955 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1958 static void vma_end_reservation(struct hstate *h,
1959 struct vm_area_struct *vma, unsigned long addr)
1961 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1964 static long vma_add_reservation(struct hstate *h,
1965 struct vm_area_struct *vma, unsigned long addr)
1967 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1971 * This routine is called to restore a reservation on error paths. In the
1972 * specific error paths, a huge page was allocated (via alloc_huge_page)
1973 * and is about to be freed. If a reservation for the page existed,
1974 * alloc_huge_page would have consumed the reservation and set PagePrivate
1975 * in the newly allocated page. When the page is freed via free_huge_page,
1976 * the global reservation count will be incremented if PagePrivate is set.
1977 * However, free_huge_page can not adjust the reserve map. Adjust the
1978 * reserve map here to be consistent with global reserve count adjustments
1979 * to be made by free_huge_page.
1981 static void restore_reserve_on_error(struct hstate *h,
1982 struct vm_area_struct *vma, unsigned long address,
1985 if (unlikely(PagePrivate(page))) {
1986 long rc = vma_needs_reservation(h, vma, address);
1988 if (unlikely(rc < 0)) {
1990 * Rare out of memory condition in reserve map
1991 * manipulation. Clear PagePrivate so that
1992 * global reserve count will not be incremented
1993 * by free_huge_page. This will make it appear
1994 * as though the reservation for this page was
1995 * consumed. This may prevent the task from
1996 * faulting in the page at a later time. This
1997 * is better than inconsistent global huge page
1998 * accounting of reserve counts.
2000 ClearPagePrivate(page);
2002 rc = vma_add_reservation(h, vma, address);
2003 if (unlikely(rc < 0))
2005 * See above comment about rare out of
2008 ClearPagePrivate(page);
2010 vma_end_reservation(h, vma, address);
2014 struct page *alloc_huge_page(struct vm_area_struct *vma,
2015 unsigned long addr, int avoid_reserve)
2017 struct hugepage_subpool *spool = subpool_vma(vma);
2018 struct hstate *h = hstate_vma(vma);
2020 long map_chg, map_commit;
2023 struct hugetlb_cgroup *h_cg;
2025 idx = hstate_index(h);
2027 * Examine the region/reserve map to determine if the process
2028 * has a reservation for the page to be allocated. A return
2029 * code of zero indicates a reservation exists (no change).
2031 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2033 return ERR_PTR(-ENOMEM);
2036 * Processes that did not create the mapping will have no
2037 * reserves as indicated by the region/reserve map. Check
2038 * that the allocation will not exceed the subpool limit.
2039 * Allocations for MAP_NORESERVE mappings also need to be
2040 * checked against any subpool limit.
2042 if (map_chg || avoid_reserve) {
2043 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2045 vma_end_reservation(h, vma, addr);
2046 return ERR_PTR(-ENOSPC);
2050 * Even though there was no reservation in the region/reserve
2051 * map, there could be reservations associated with the
2052 * subpool that can be used. This would be indicated if the
2053 * return value of hugepage_subpool_get_pages() is zero.
2054 * However, if avoid_reserve is specified we still avoid even
2055 * the subpool reservations.
2061 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2063 goto out_subpool_put;
2065 spin_lock(&hugetlb_lock);
2067 * glb_chg is passed to indicate whether or not a page must be taken
2068 * from the global free pool (global change). gbl_chg == 0 indicates
2069 * a reservation exists for the allocation.
2071 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2073 spin_unlock(&hugetlb_lock);
2074 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2076 goto out_uncharge_cgroup;
2077 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2078 SetPagePrivate(page);
2079 h->resv_huge_pages--;
2081 spin_lock(&hugetlb_lock);
2082 list_move(&page->lru, &h->hugepage_activelist);
2085 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2086 spin_unlock(&hugetlb_lock);
2088 set_page_private(page, (unsigned long)spool);
2090 map_commit = vma_commit_reservation(h, vma, addr);
2091 if (unlikely(map_chg > map_commit)) {
2093 * The page was added to the reservation map between
2094 * vma_needs_reservation and vma_commit_reservation.
2095 * This indicates a race with hugetlb_reserve_pages.
2096 * Adjust for the subpool count incremented above AND
2097 * in hugetlb_reserve_pages for the same page. Also,
2098 * the reservation count added in hugetlb_reserve_pages
2099 * no longer applies.
2103 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2104 hugetlb_acct_memory(h, -rsv_adjust);
2108 out_uncharge_cgroup:
2109 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2111 if (map_chg || avoid_reserve)
2112 hugepage_subpool_put_pages(spool, 1);
2113 vma_end_reservation(h, vma, addr);
2114 return ERR_PTR(-ENOSPC);
2117 int alloc_bootmem_huge_page(struct hstate *h)
2118 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2119 int __alloc_bootmem_huge_page(struct hstate *h)
2121 struct huge_bootmem_page *m;
2124 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2127 addr = memblock_alloc_try_nid_raw(
2128 huge_page_size(h), huge_page_size(h),
2129 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2132 * Use the beginning of the huge page to store the
2133 * huge_bootmem_page struct (until gather_bootmem
2134 * puts them into the mem_map).
2143 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2144 /* Put them into a private list first because mem_map is not up yet */
2145 INIT_LIST_HEAD(&m->list);
2146 list_add(&m->list, &huge_boot_pages);
2151 static void __init prep_compound_huge_page(struct page *page,
2154 if (unlikely(order > (MAX_ORDER - 1)))
2155 prep_compound_gigantic_page(page, order);
2157 prep_compound_page(page, order);
2160 /* Put bootmem huge pages into the standard lists after mem_map is up */
2161 static void __init gather_bootmem_prealloc(void)
2163 struct huge_bootmem_page *m;
2165 list_for_each_entry(m, &huge_boot_pages, list) {
2166 struct page *page = virt_to_page(m);
2167 struct hstate *h = m->hstate;
2169 WARN_ON(page_count(page) != 1);
2170 prep_compound_huge_page(page, h->order);
2171 WARN_ON(PageReserved(page));
2172 prep_new_huge_page(h, page, page_to_nid(page));
2173 put_page(page); /* free it into the hugepage allocator */
2176 * If we had gigantic hugepages allocated at boot time, we need
2177 * to restore the 'stolen' pages to totalram_pages in order to
2178 * fix confusing memory reports from free(1) and another
2179 * side-effects, like CommitLimit going negative.
2181 if (hstate_is_gigantic(h))
2182 adjust_managed_page_count(page, 1 << h->order);
2187 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2191 for (i = 0; i < h->max_huge_pages; ++i) {
2192 if (hstate_is_gigantic(h)) {
2193 if (!alloc_bootmem_huge_page(h))
2195 } else if (!alloc_pool_huge_page(h,
2196 &node_states[N_MEMORY]))
2200 if (i < h->max_huge_pages) {
2203 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2204 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2205 h->max_huge_pages, buf, i);
2206 h->max_huge_pages = i;
2210 static void __init hugetlb_init_hstates(void)
2214 for_each_hstate(h) {
2215 if (minimum_order > huge_page_order(h))
2216 minimum_order = huge_page_order(h);
2218 /* oversize hugepages were init'ed in early boot */
2219 if (!hstate_is_gigantic(h))
2220 hugetlb_hstate_alloc_pages(h);
2222 VM_BUG_ON(minimum_order == UINT_MAX);
2225 static void __init report_hugepages(void)
2229 for_each_hstate(h) {
2232 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2233 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2234 buf, h->free_huge_pages);
2238 #ifdef CONFIG_HIGHMEM
2239 static void try_to_free_low(struct hstate *h, unsigned long count,
2240 nodemask_t *nodes_allowed)
2244 if (hstate_is_gigantic(h))
2247 for_each_node_mask(i, *nodes_allowed) {
2248 struct page *page, *next;
2249 struct list_head *freel = &h->hugepage_freelists[i];
2250 list_for_each_entry_safe(page, next, freel, lru) {
2251 if (count >= h->nr_huge_pages)
2253 if (PageHighMem(page))
2255 list_del(&page->lru);
2256 update_and_free_page(h, page);
2257 h->free_huge_pages--;
2258 h->free_huge_pages_node[page_to_nid(page)]--;
2263 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2264 nodemask_t *nodes_allowed)
2270 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2271 * balanced by operating on them in a round-robin fashion.
2272 * Returns 1 if an adjustment was made.
2274 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2279 VM_BUG_ON(delta != -1 && delta != 1);
2282 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2283 if (h->surplus_huge_pages_node[node])
2287 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2288 if (h->surplus_huge_pages_node[node] <
2289 h->nr_huge_pages_node[node])
2296 h->surplus_huge_pages += delta;
2297 h->surplus_huge_pages_node[node] += delta;
2301 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2302 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2303 nodemask_t *nodes_allowed)
2305 unsigned long min_count, ret;
2307 spin_lock(&hugetlb_lock);
2310 * Check for a node specific request.
2311 * Changing node specific huge page count may require a corresponding
2312 * change to the global count. In any case, the passed node mask
2313 * (nodes_allowed) will restrict alloc/free to the specified node.
2315 if (nid != NUMA_NO_NODE) {
2316 unsigned long old_count = count;
2318 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2320 * User may have specified a large count value which caused the
2321 * above calculation to overflow. In this case, they wanted
2322 * to allocate as many huge pages as possible. Set count to
2323 * largest possible value to align with their intention.
2325 if (count < old_count)
2330 * Gigantic pages runtime allocation depend on the capability for large
2331 * page range allocation.
2332 * If the system does not provide this feature, return an error when
2333 * the user tries to allocate gigantic pages but let the user free the
2334 * boottime allocated gigantic pages.
2336 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2337 if (count > persistent_huge_pages(h)) {
2338 spin_unlock(&hugetlb_lock);
2341 /* Fall through to decrease pool */
2345 * Increase the pool size
2346 * First take pages out of surplus state. Then make up the
2347 * remaining difference by allocating fresh huge pages.
2349 * We might race with alloc_surplus_huge_page() here and be unable
2350 * to convert a surplus huge page to a normal huge page. That is
2351 * not critical, though, it just means the overall size of the
2352 * pool might be one hugepage larger than it needs to be, but
2353 * within all the constraints specified by the sysctls.
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 ret = alloc_pool_huge_page(h, nodes_allowed);
2372 spin_lock(&hugetlb_lock);
2376 /* Bail for signals. Probably ctrl-c from user */
2377 if (signal_pending(current))
2382 * Decrease the pool size
2383 * First return free pages to the buddy allocator (being careful
2384 * to keep enough around to satisfy reservations). Then place
2385 * pages into surplus state as needed so the pool will shrink
2386 * to the desired size as pages become free.
2388 * By placing pages into the surplus state independent of the
2389 * overcommit value, we are allowing the surplus pool size to
2390 * exceed overcommit. There are few sane options here. Since
2391 * alloc_surplus_huge_page() is checking the global counter,
2392 * though, we'll note that we're not allowed to exceed surplus
2393 * and won't grow the pool anywhere else. Not until one of the
2394 * sysctls are changed, or the surplus pages go out of use.
2396 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2397 min_count = max(count, min_count);
2398 try_to_free_low(h, min_count, nodes_allowed);
2399 while (min_count < persistent_huge_pages(h)) {
2400 if (!free_pool_huge_page(h, nodes_allowed, 0))
2402 cond_resched_lock(&hugetlb_lock);
2404 while (count < persistent_huge_pages(h)) {
2405 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2409 h->max_huge_pages = persistent_huge_pages(h);
2410 spin_unlock(&hugetlb_lock);
2415 #define HSTATE_ATTR_RO(_name) \
2416 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2418 #define HSTATE_ATTR(_name) \
2419 static struct kobj_attribute _name##_attr = \
2420 __ATTR(_name, 0644, _name##_show, _name##_store)
2422 static struct kobject *hugepages_kobj;
2423 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2425 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2427 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2431 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2432 if (hstate_kobjs[i] == kobj) {
2434 *nidp = NUMA_NO_NODE;
2438 return kobj_to_node_hstate(kobj, nidp);
2441 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2442 struct kobj_attribute *attr, char *buf)
2445 unsigned long nr_huge_pages;
2448 h = kobj_to_hstate(kobj, &nid);
2449 if (nid == NUMA_NO_NODE)
2450 nr_huge_pages = h->nr_huge_pages;
2452 nr_huge_pages = h->nr_huge_pages_node[nid];
2454 return sprintf(buf, "%lu\n", nr_huge_pages);
2457 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2458 struct hstate *h, int nid,
2459 unsigned long count, size_t len)
2462 nodemask_t nodes_allowed, *n_mask;
2464 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2467 if (nid == NUMA_NO_NODE) {
2469 * global hstate attribute
2471 if (!(obey_mempolicy &&
2472 init_nodemask_of_mempolicy(&nodes_allowed)))
2473 n_mask = &node_states[N_MEMORY];
2475 n_mask = &nodes_allowed;
2478 * Node specific request. count adjustment happens in
2479 * set_max_huge_pages() after acquiring hugetlb_lock.
2481 init_nodemask_of_node(&nodes_allowed, nid);
2482 n_mask = &nodes_allowed;
2485 err = set_max_huge_pages(h, count, nid, n_mask);
2487 return err ? err : len;
2490 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2491 struct kobject *kobj, const char *buf,
2495 unsigned long count;
2499 err = kstrtoul(buf, 10, &count);
2503 h = kobj_to_hstate(kobj, &nid);
2504 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2507 static ssize_t nr_hugepages_show(struct kobject *kobj,
2508 struct kobj_attribute *attr, char *buf)
2510 return nr_hugepages_show_common(kobj, attr, buf);
2513 static ssize_t nr_hugepages_store(struct kobject *kobj,
2514 struct kobj_attribute *attr, const char *buf, size_t len)
2516 return nr_hugepages_store_common(false, kobj, buf, len);
2518 HSTATE_ATTR(nr_hugepages);
2523 * hstate attribute for optionally mempolicy-based constraint on persistent
2524 * huge page alloc/free.
2526 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2527 struct kobj_attribute *attr, char *buf)
2529 return nr_hugepages_show_common(kobj, attr, buf);
2532 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2533 struct kobj_attribute *attr, const char *buf, size_t len)
2535 return nr_hugepages_store_common(true, kobj, buf, len);
2537 HSTATE_ATTR(nr_hugepages_mempolicy);
2541 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2542 struct kobj_attribute *attr, char *buf)
2544 struct hstate *h = kobj_to_hstate(kobj, NULL);
2545 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2548 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2549 struct kobj_attribute *attr, const char *buf, size_t count)
2552 unsigned long input;
2553 struct hstate *h = kobj_to_hstate(kobj, NULL);
2555 if (hstate_is_gigantic(h))
2558 err = kstrtoul(buf, 10, &input);
2562 spin_lock(&hugetlb_lock);
2563 h->nr_overcommit_huge_pages = input;
2564 spin_unlock(&hugetlb_lock);
2568 HSTATE_ATTR(nr_overcommit_hugepages);
2570 static ssize_t free_hugepages_show(struct kobject *kobj,
2571 struct kobj_attribute *attr, char *buf)
2574 unsigned long free_huge_pages;
2577 h = kobj_to_hstate(kobj, &nid);
2578 if (nid == NUMA_NO_NODE)
2579 free_huge_pages = h->free_huge_pages;
2581 free_huge_pages = h->free_huge_pages_node[nid];
2583 return sprintf(buf, "%lu\n", free_huge_pages);
2585 HSTATE_ATTR_RO(free_hugepages);
2587 static ssize_t resv_hugepages_show(struct kobject *kobj,
2588 struct kobj_attribute *attr, char *buf)
2590 struct hstate *h = kobj_to_hstate(kobj, NULL);
2591 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2593 HSTATE_ATTR_RO(resv_hugepages);
2595 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2596 struct kobj_attribute *attr, char *buf)
2599 unsigned long surplus_huge_pages;
2602 h = kobj_to_hstate(kobj, &nid);
2603 if (nid == NUMA_NO_NODE)
2604 surplus_huge_pages = h->surplus_huge_pages;
2606 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2608 return sprintf(buf, "%lu\n", surplus_huge_pages);
2610 HSTATE_ATTR_RO(surplus_hugepages);
2612 static struct attribute *hstate_attrs[] = {
2613 &nr_hugepages_attr.attr,
2614 &nr_overcommit_hugepages_attr.attr,
2615 &free_hugepages_attr.attr,
2616 &resv_hugepages_attr.attr,
2617 &surplus_hugepages_attr.attr,
2619 &nr_hugepages_mempolicy_attr.attr,
2624 static const struct attribute_group hstate_attr_group = {
2625 .attrs = hstate_attrs,
2628 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2629 struct kobject **hstate_kobjs,
2630 const struct attribute_group *hstate_attr_group)
2633 int hi = hstate_index(h);
2635 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2636 if (!hstate_kobjs[hi])
2639 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2641 kobject_put(hstate_kobjs[hi]);
2646 static void __init hugetlb_sysfs_init(void)
2651 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2652 if (!hugepages_kobj)
2655 for_each_hstate(h) {
2656 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2657 hstate_kobjs, &hstate_attr_group);
2659 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2666 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2667 * with node devices in node_devices[] using a parallel array. The array
2668 * index of a node device or _hstate == node id.
2669 * This is here to avoid any static dependency of the node device driver, in
2670 * the base kernel, on the hugetlb module.
2672 struct node_hstate {
2673 struct kobject *hugepages_kobj;
2674 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2676 static struct node_hstate node_hstates[MAX_NUMNODES];
2679 * A subset of global hstate attributes for node devices
2681 static struct attribute *per_node_hstate_attrs[] = {
2682 &nr_hugepages_attr.attr,
2683 &free_hugepages_attr.attr,
2684 &surplus_hugepages_attr.attr,
2688 static const struct attribute_group per_node_hstate_attr_group = {
2689 .attrs = per_node_hstate_attrs,
2693 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2694 * Returns node id via non-NULL nidp.
2696 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2700 for (nid = 0; nid < nr_node_ids; nid++) {
2701 struct node_hstate *nhs = &node_hstates[nid];
2703 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2704 if (nhs->hstate_kobjs[i] == kobj) {
2716 * Unregister hstate attributes from a single node device.
2717 * No-op if no hstate attributes attached.
2719 static void hugetlb_unregister_node(struct node *node)
2722 struct node_hstate *nhs = &node_hstates[node->dev.id];
2724 if (!nhs->hugepages_kobj)
2725 return; /* no hstate attributes */
2727 for_each_hstate(h) {
2728 int idx = hstate_index(h);
2729 if (nhs->hstate_kobjs[idx]) {
2730 kobject_put(nhs->hstate_kobjs[idx]);
2731 nhs->hstate_kobjs[idx] = NULL;
2735 kobject_put(nhs->hugepages_kobj);
2736 nhs->hugepages_kobj = NULL;
2741 * Register hstate attributes for a single node device.
2742 * No-op if attributes already registered.
2744 static void hugetlb_register_node(struct node *node)
2747 struct node_hstate *nhs = &node_hstates[node->dev.id];
2750 if (nhs->hugepages_kobj)
2751 return; /* already allocated */
2753 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2755 if (!nhs->hugepages_kobj)
2758 for_each_hstate(h) {
2759 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2761 &per_node_hstate_attr_group);
2763 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2764 h->name, node->dev.id);
2765 hugetlb_unregister_node(node);
2772 * hugetlb init time: register hstate attributes for all registered node
2773 * devices of nodes that have memory. All on-line nodes should have
2774 * registered their associated device by this time.
2776 static void __init hugetlb_register_all_nodes(void)
2780 for_each_node_state(nid, N_MEMORY) {
2781 struct node *node = node_devices[nid];
2782 if (node->dev.id == nid)
2783 hugetlb_register_node(node);
2787 * Let the node device driver know we're here so it can
2788 * [un]register hstate attributes on node hotplug.
2790 register_hugetlbfs_with_node(hugetlb_register_node,
2791 hugetlb_unregister_node);
2793 #else /* !CONFIG_NUMA */
2795 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2803 static void hugetlb_register_all_nodes(void) { }
2807 static int __init hugetlb_init(void)
2811 if (!hugepages_supported())
2814 if (!size_to_hstate(default_hstate_size)) {
2815 if (default_hstate_size != 0) {
2816 pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2817 default_hstate_size, HPAGE_SIZE);
2820 default_hstate_size = HPAGE_SIZE;
2821 if (!size_to_hstate(default_hstate_size))
2822 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2824 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2825 if (default_hstate_max_huge_pages) {
2826 if (!default_hstate.max_huge_pages)
2827 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2830 hugetlb_init_hstates();
2831 gather_bootmem_prealloc();
2834 hugetlb_sysfs_init();
2835 hugetlb_register_all_nodes();
2836 hugetlb_cgroup_file_init();
2839 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2841 num_fault_mutexes = 1;
2843 hugetlb_fault_mutex_table =
2844 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
2846 BUG_ON(!hugetlb_fault_mutex_table);
2848 for (i = 0; i < num_fault_mutexes; i++)
2849 mutex_init(&hugetlb_fault_mutex_table[i]);
2852 subsys_initcall(hugetlb_init);
2854 /* Should be called on processing a hugepagesz=... option */
2855 void __init hugetlb_bad_size(void)
2857 parsed_valid_hugepagesz = false;
2860 void __init hugetlb_add_hstate(unsigned int order)
2865 if (size_to_hstate(PAGE_SIZE << order)) {
2866 pr_warn("hugepagesz= specified twice, ignoring\n");
2869 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2871 h = &hstates[hugetlb_max_hstate++];
2873 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2874 h->nr_huge_pages = 0;
2875 h->free_huge_pages = 0;
2876 for (i = 0; i < MAX_NUMNODES; ++i)
2877 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2878 INIT_LIST_HEAD(&h->hugepage_activelist);
2879 h->next_nid_to_alloc = first_memory_node;
2880 h->next_nid_to_free = first_memory_node;
2881 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2882 huge_page_size(h)/1024);
2887 static int __init hugetlb_nrpages_setup(char *s)
2890 static unsigned long *last_mhp;
2892 if (!parsed_valid_hugepagesz) {
2893 pr_warn("hugepages = %s preceded by "
2894 "an unsupported hugepagesz, ignoring\n", s);
2895 parsed_valid_hugepagesz = true;
2899 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2900 * so this hugepages= parameter goes to the "default hstate".
2902 else if (!hugetlb_max_hstate)
2903 mhp = &default_hstate_max_huge_pages;
2905 mhp = &parsed_hstate->max_huge_pages;
2907 if (mhp == last_mhp) {
2908 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2912 if (sscanf(s, "%lu", mhp) <= 0)
2916 * Global state is always initialized later in hugetlb_init.
2917 * But we need to allocate >= MAX_ORDER hstates here early to still
2918 * use the bootmem allocator.
2920 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2921 hugetlb_hstate_alloc_pages(parsed_hstate);
2927 __setup("hugepages=", hugetlb_nrpages_setup);
2929 static int __init hugetlb_default_setup(char *s)
2931 default_hstate_size = memparse(s, &s);
2934 __setup("default_hugepagesz=", hugetlb_default_setup);
2936 static unsigned int cpuset_mems_nr(unsigned int *array)
2939 unsigned int nr = 0;
2941 for_each_node_mask(node, cpuset_current_mems_allowed)
2947 #ifdef CONFIG_SYSCTL
2948 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2949 struct ctl_table *table, int write,
2950 void __user *buffer, size_t *length, loff_t *ppos)
2952 struct hstate *h = &default_hstate;
2953 unsigned long tmp = h->max_huge_pages;
2956 if (!hugepages_supported())
2960 table->maxlen = sizeof(unsigned long);
2961 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2966 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2967 NUMA_NO_NODE, tmp, *length);
2972 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2973 void __user *buffer, size_t *length, loff_t *ppos)
2976 return hugetlb_sysctl_handler_common(false, table, write,
2977 buffer, length, ppos);
2981 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2982 void __user *buffer, size_t *length, loff_t *ppos)
2984 return hugetlb_sysctl_handler_common(true, table, write,
2985 buffer, length, ppos);
2987 #endif /* CONFIG_NUMA */
2989 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2990 void __user *buffer,
2991 size_t *length, loff_t *ppos)
2993 struct hstate *h = &default_hstate;
2997 if (!hugepages_supported())
3000 tmp = h->nr_overcommit_huge_pages;
3002 if (write && hstate_is_gigantic(h))
3006 table->maxlen = sizeof(unsigned long);
3007 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3012 spin_lock(&hugetlb_lock);
3013 h->nr_overcommit_huge_pages = tmp;
3014 spin_unlock(&hugetlb_lock);
3020 #endif /* CONFIG_SYSCTL */
3022 void hugetlb_report_meminfo(struct seq_file *m)
3025 unsigned long total = 0;
3027 if (!hugepages_supported())
3030 for_each_hstate(h) {
3031 unsigned long count = h->nr_huge_pages;
3033 total += (PAGE_SIZE << huge_page_order(h)) * count;
3035 if (h == &default_hstate)
3037 "HugePages_Total: %5lu\n"
3038 "HugePages_Free: %5lu\n"
3039 "HugePages_Rsvd: %5lu\n"
3040 "HugePages_Surp: %5lu\n"
3041 "Hugepagesize: %8lu kB\n",
3045 h->surplus_huge_pages,
3046 (PAGE_SIZE << huge_page_order(h)) / 1024);
3049 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3052 int hugetlb_report_node_meminfo(int nid, char *buf)
3054 struct hstate *h = &default_hstate;
3055 if (!hugepages_supported())
3058 "Node %d HugePages_Total: %5u\n"
3059 "Node %d HugePages_Free: %5u\n"
3060 "Node %d HugePages_Surp: %5u\n",
3061 nid, h->nr_huge_pages_node[nid],
3062 nid, h->free_huge_pages_node[nid],
3063 nid, h->surplus_huge_pages_node[nid]);
3066 void hugetlb_show_meminfo(void)
3071 if (!hugepages_supported())
3074 for_each_node_state(nid, N_MEMORY)
3076 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3078 h->nr_huge_pages_node[nid],
3079 h->free_huge_pages_node[nid],
3080 h->surplus_huge_pages_node[nid],
3081 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3084 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3086 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3087 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3090 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3091 unsigned long hugetlb_total_pages(void)
3094 unsigned long nr_total_pages = 0;
3097 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3098 return nr_total_pages;
3101 static int hugetlb_acct_memory(struct hstate *h, long delta)
3105 spin_lock(&hugetlb_lock);
3107 * When cpuset is configured, it breaks the strict hugetlb page
3108 * reservation as the accounting is done on a global variable. Such
3109 * reservation is completely rubbish in the presence of cpuset because
3110 * the reservation is not checked against page availability for the
3111 * current cpuset. Application can still potentially OOM'ed by kernel
3112 * with lack of free htlb page in cpuset that the task is in.
3113 * Attempt to enforce strict accounting with cpuset is almost
3114 * impossible (or too ugly) because cpuset is too fluid that
3115 * task or memory node can be dynamically moved between cpusets.
3117 * The change of semantics for shared hugetlb mapping with cpuset is
3118 * undesirable. However, in order to preserve some of the semantics,
3119 * we fall back to check against current free page availability as
3120 * a best attempt and hopefully to minimize the impact of changing
3121 * semantics that cpuset has.
3124 if (gather_surplus_pages(h, delta) < 0)
3127 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3128 return_unused_surplus_pages(h, delta);
3135 return_unused_surplus_pages(h, (unsigned long) -delta);
3138 spin_unlock(&hugetlb_lock);
3142 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3144 struct resv_map *resv = vma_resv_map(vma);
3147 * This new VMA should share its siblings reservation map if present.
3148 * The VMA will only ever have a valid reservation map pointer where
3149 * it is being copied for another still existing VMA. As that VMA
3150 * has a reference to the reservation map it cannot disappear until
3151 * after this open call completes. It is therefore safe to take a
3152 * new reference here without additional locking.
3154 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3155 kref_get(&resv->refs);
3158 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3160 struct hstate *h = hstate_vma(vma);
3161 struct resv_map *resv = vma_resv_map(vma);
3162 struct hugepage_subpool *spool = subpool_vma(vma);
3163 unsigned long reserve, start, end;
3166 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3169 start = vma_hugecache_offset(h, vma, vma->vm_start);
3170 end = vma_hugecache_offset(h, vma, vma->vm_end);
3172 reserve = (end - start) - region_count(resv, start, end);
3174 kref_put(&resv->refs, resv_map_release);
3178 * Decrement reserve counts. The global reserve count may be
3179 * adjusted if the subpool has a minimum size.
3181 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3182 hugetlb_acct_memory(h, -gbl_reserve);
3186 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3188 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3193 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3195 struct hstate *hstate = hstate_vma(vma);
3197 return 1UL << huge_page_shift(hstate);
3201 * We cannot handle pagefaults against hugetlb pages at all. They cause
3202 * handle_mm_fault() to try to instantiate regular-sized pages in the
3203 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3206 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3213 * When a new function is introduced to vm_operations_struct and added
3214 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3215 * This is because under System V memory model, mappings created via
3216 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3217 * their original vm_ops are overwritten with shm_vm_ops.
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,
3224 .pagesize = hugetlb_vm_op_pagesize,
3227 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3233 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3234 vma->vm_page_prot)));
3236 entry = huge_pte_wrprotect(mk_huge_pte(page,
3237 vma->vm_page_prot));
3239 entry = pte_mkyoung(entry);
3240 entry = pte_mkhuge(entry);
3241 entry = arch_make_huge_pte(entry, vma, page, writable);
3246 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3247 unsigned long address, pte_t *ptep)
3251 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3252 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3253 update_mmu_cache(vma, address, ptep);
3256 bool is_hugetlb_entry_migration(pte_t pte)
3260 if (huge_pte_none(pte) || pte_present(pte))
3262 swp = pte_to_swp_entry(pte);
3263 if (non_swap_entry(swp) && is_migration_entry(swp))
3269 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3273 if (huge_pte_none(pte) || pte_present(pte))
3275 swp = pte_to_swp_entry(pte);
3276 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3282 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3283 struct vm_area_struct *vma)
3285 pte_t *src_pte, *dst_pte, entry, dst_entry;
3286 struct page *ptepage;
3289 struct hstate *h = hstate_vma(vma);
3290 unsigned long sz = huge_page_size(h);
3291 struct mmu_notifier_range range;
3294 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3297 mmu_notifier_range_init(&range, src, vma->vm_start,
3299 mmu_notifier_invalidate_range_start(&range);
3302 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3303 spinlock_t *src_ptl, *dst_ptl;
3304 src_pte = huge_pte_offset(src, addr, sz);
3307 dst_pte = huge_pte_alloc(dst, addr, sz);
3314 * If the pagetables are shared don't copy or take references.
3315 * dst_pte == src_pte is the common case of src/dest sharing.
3317 * However, src could have 'unshared' and dst shares with
3318 * another vma. If dst_pte !none, this implies sharing.
3319 * Check here before taking page table lock, and once again
3320 * after taking the lock below.
3322 dst_entry = huge_ptep_get(dst_pte);
3323 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3326 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3327 src_ptl = huge_pte_lockptr(h, src, src_pte);
3328 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3329 entry = huge_ptep_get(src_pte);
3330 dst_entry = huge_ptep_get(dst_pte);
3331 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3333 * Skip if src entry none. Also, skip in the
3334 * unlikely case dst entry !none as this implies
3335 * sharing with another vma.
3338 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3339 is_hugetlb_entry_hwpoisoned(entry))) {
3340 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3342 if (is_write_migration_entry(swp_entry) && cow) {
3344 * COW mappings require pages in both
3345 * parent and child to be set to read.
3347 make_migration_entry_read(&swp_entry);
3348 entry = swp_entry_to_pte(swp_entry);
3349 set_huge_swap_pte_at(src, addr, src_pte,
3352 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3356 * No need to notify as we are downgrading page
3357 * table protection not changing it to point
3360 * See Documentation/vm/mmu_notifier.rst
3362 huge_ptep_set_wrprotect(src, addr, src_pte);
3364 entry = huge_ptep_get(src_pte);
3365 ptepage = pte_page(entry);
3367 page_dup_rmap(ptepage, true);
3368 set_huge_pte_at(dst, addr, dst_pte, entry);
3369 hugetlb_count_add(pages_per_huge_page(h), dst);
3371 spin_unlock(src_ptl);
3372 spin_unlock(dst_ptl);
3376 mmu_notifier_invalidate_range_end(&range);
3381 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3382 unsigned long start, unsigned long end,
3383 struct page *ref_page)
3385 struct mm_struct *mm = vma->vm_mm;
3386 unsigned long address;
3391 struct hstate *h = hstate_vma(vma);
3392 unsigned long sz = huge_page_size(h);
3393 struct mmu_notifier_range range;
3395 WARN_ON(!is_vm_hugetlb_page(vma));
3396 BUG_ON(start & ~huge_page_mask(h));
3397 BUG_ON(end & ~huge_page_mask(h));
3400 * This is a hugetlb vma, all the pte entries should point
3403 tlb_change_page_size(tlb, sz);
3404 tlb_start_vma(tlb, vma);
3407 * If sharing possible, alert mmu notifiers of worst case.
3409 mmu_notifier_range_init(&range, mm, start, end);
3410 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3411 mmu_notifier_invalidate_range_start(&range);
3413 for (; address < end; address += sz) {
3414 ptep = huge_pte_offset(mm, address, sz);
3418 ptl = huge_pte_lock(h, mm, ptep);
3419 if (huge_pmd_unshare(mm, &address, ptep)) {
3422 * We just unmapped a page of PMDs by clearing a PUD.
3423 * The caller's TLB flush range should cover this area.
3428 pte = huge_ptep_get(ptep);
3429 if (huge_pte_none(pte)) {
3435 * Migrating hugepage or HWPoisoned hugepage is already
3436 * unmapped and its refcount is dropped, so just clear pte here.
3438 if (unlikely(!pte_present(pte))) {
3439 huge_pte_clear(mm, address, ptep, sz);
3444 page = pte_page(pte);
3446 * If a reference page is supplied, it is because a specific
3447 * page is being unmapped, not a range. Ensure the page we
3448 * are about to unmap is the actual page of interest.
3451 if (page != ref_page) {
3456 * Mark the VMA as having unmapped its page so that
3457 * future faults in this VMA will fail rather than
3458 * looking like data was lost
3460 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3463 pte = huge_ptep_get_and_clear(mm, address, ptep);
3464 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3465 if (huge_pte_dirty(pte))
3466 set_page_dirty(page);
3468 hugetlb_count_sub(pages_per_huge_page(h), mm);
3469 page_remove_rmap(page, true);
3472 tlb_remove_page_size(tlb, page, huge_page_size(h));
3474 * Bail out after unmapping reference page if supplied
3479 mmu_notifier_invalidate_range_end(&range);
3480 tlb_end_vma(tlb, vma);
3483 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3484 struct vm_area_struct *vma, unsigned long start,
3485 unsigned long end, struct page *ref_page)
3487 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3490 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3491 * test will fail on a vma being torn down, and not grab a page table
3492 * on its way out. We're lucky that the flag has such an appropriate
3493 * name, and can in fact be safely cleared here. We could clear it
3494 * before the __unmap_hugepage_range above, but all that's necessary
3495 * is to clear it before releasing the i_mmap_rwsem. This works
3496 * because in the context this is called, the VMA is about to be
3497 * destroyed and the i_mmap_rwsem is held.
3499 vma->vm_flags &= ~VM_MAYSHARE;
3502 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3503 unsigned long end, struct page *ref_page)
3505 struct mm_struct *mm;
3506 struct mmu_gather tlb;
3507 unsigned long tlb_start = start;
3508 unsigned long tlb_end = end;
3511 * If shared PMDs were possibly used within this vma range, adjust
3512 * start/end for worst case tlb flushing.
3513 * Note that we can not be sure if PMDs are shared until we try to
3514 * unmap pages. However, we want to make sure TLB flushing covers
3515 * the largest possible range.
3517 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
3521 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
3522 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3523 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
3527 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3528 * mappping it owns the reserve page for. The intention is to unmap the page
3529 * from other VMAs and let the children be SIGKILLed if they are faulting the
3532 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3533 struct page *page, unsigned long address)
3535 struct hstate *h = hstate_vma(vma);
3536 struct vm_area_struct *iter_vma;
3537 struct address_space *mapping;
3541 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3542 * from page cache lookup which is in HPAGE_SIZE units.
3544 address = address & huge_page_mask(h);
3545 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3547 mapping = vma->vm_file->f_mapping;
3550 * Take the mapping lock for the duration of the table walk. As
3551 * this mapping should be shared between all the VMAs,
3552 * __unmap_hugepage_range() is called as the lock is already held
3554 i_mmap_lock_write(mapping);
3555 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3556 /* Do not unmap the current VMA */
3557 if (iter_vma == vma)
3561 * Shared VMAs have their own reserves and do not affect
3562 * MAP_PRIVATE accounting but it is possible that a shared
3563 * VMA is using the same page so check and skip such VMAs.
3565 if (iter_vma->vm_flags & VM_MAYSHARE)
3569 * Unmap the page from other VMAs without their own reserves.
3570 * They get marked to be SIGKILLed if they fault in these
3571 * areas. This is because a future no-page fault on this VMA
3572 * could insert a zeroed page instead of the data existing
3573 * from the time of fork. This would look like data corruption
3575 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3576 unmap_hugepage_range(iter_vma, address,
3577 address + huge_page_size(h), page);
3579 i_mmap_unlock_write(mapping);
3583 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3584 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3585 * cannot race with other handlers or page migration.
3586 * Keep the pte_same checks anyway to make transition from the mutex easier.
3588 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3589 unsigned long address, pte_t *ptep,
3590 struct page *pagecache_page, spinlock_t *ptl)
3593 struct hstate *h = hstate_vma(vma);
3594 struct page *old_page, *new_page;
3595 int outside_reserve = 0;
3597 unsigned long haddr = address & huge_page_mask(h);
3598 struct mmu_notifier_range range;
3600 pte = huge_ptep_get(ptep);
3601 old_page = pte_page(pte);
3604 /* If no-one else is actually using this page, avoid the copy
3605 * and just make the page writable */
3606 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3607 page_move_anon_rmap(old_page, vma);
3608 set_huge_ptep_writable(vma, haddr, ptep);
3613 * If the process that created a MAP_PRIVATE mapping is about to
3614 * perform a COW due to a shared page count, attempt to satisfy
3615 * the allocation without using the existing reserves. The pagecache
3616 * page is used to determine if the reserve at this address was
3617 * consumed or not. If reserves were used, a partial faulted mapping
3618 * at the time of fork() could consume its reserves on COW instead
3619 * of the full address range.
3621 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3622 old_page != pagecache_page)
3623 outside_reserve = 1;
3628 * Drop page table lock as buddy allocator may be called. It will
3629 * be acquired again before returning to the caller, as expected.
3632 new_page = alloc_huge_page(vma, haddr, outside_reserve);
3634 if (IS_ERR(new_page)) {
3636 * If a process owning a MAP_PRIVATE mapping fails to COW,
3637 * it is due to references held by a child and an insufficient
3638 * huge page pool. To guarantee the original mappers
3639 * reliability, unmap the page from child processes. The child
3640 * may get SIGKILLed if it later faults.
3642 if (outside_reserve) {
3644 BUG_ON(huge_pte_none(pte));
3645 unmap_ref_private(mm, vma, old_page, haddr);
3646 BUG_ON(huge_pte_none(pte));
3648 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3650 pte_same(huge_ptep_get(ptep), pte)))
3651 goto retry_avoidcopy;
3653 * race occurs while re-acquiring page table
3654 * lock, and our job is done.
3659 ret = vmf_error(PTR_ERR(new_page));
3660 goto out_release_old;
3664 * When the original hugepage is shared one, it does not have
3665 * anon_vma prepared.
3667 if (unlikely(anon_vma_prepare(vma))) {
3669 goto out_release_all;
3672 copy_user_huge_page(new_page, old_page, address, vma,
3673 pages_per_huge_page(h));
3674 __SetPageUptodate(new_page);
3676 mmu_notifier_range_init(&range, mm, haddr, haddr + huge_page_size(h));
3677 mmu_notifier_invalidate_range_start(&range);
3680 * Retake the page table lock to check for racing updates
3681 * before the page tables are altered
3684 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3685 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3686 ClearPagePrivate(new_page);
3689 huge_ptep_clear_flush(vma, haddr, ptep);
3690 mmu_notifier_invalidate_range(mm, range.start, range.end);
3691 set_huge_pte_at(mm, haddr, ptep,
3692 make_huge_pte(vma, new_page, 1));
3693 page_remove_rmap(old_page, true);
3694 hugepage_add_new_anon_rmap(new_page, vma, haddr);
3695 set_page_huge_active(new_page);
3696 /* Make the old page be freed below */
3697 new_page = old_page;
3700 mmu_notifier_invalidate_range_end(&range);
3702 restore_reserve_on_error(h, vma, haddr, new_page);
3707 spin_lock(ptl); /* Caller expects lock to be held */
3711 /* Return the pagecache page at a given address within a VMA */
3712 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3713 struct vm_area_struct *vma, unsigned long address)
3715 struct address_space *mapping;
3718 mapping = vma->vm_file->f_mapping;
3719 idx = vma_hugecache_offset(h, vma, address);
3721 return find_lock_page(mapping, idx);
3725 * Return whether there is a pagecache page to back given address within VMA.
3726 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3728 static bool hugetlbfs_pagecache_present(struct hstate *h,
3729 struct vm_area_struct *vma, unsigned long address)
3731 struct address_space *mapping;
3735 mapping = vma->vm_file->f_mapping;
3736 idx = vma_hugecache_offset(h, vma, address);
3738 page = find_get_page(mapping, idx);
3741 return page != NULL;
3744 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3747 struct inode *inode = mapping->host;
3748 struct hstate *h = hstate_inode(inode);
3749 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3753 ClearPagePrivate(page);
3756 * set page dirty so that it will not be removed from cache/file
3757 * by non-hugetlbfs specific code paths.
3759 set_page_dirty(page);
3761 spin_lock(&inode->i_lock);
3762 inode->i_blocks += blocks_per_huge_page(h);
3763 spin_unlock(&inode->i_lock);
3767 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
3768 struct vm_area_struct *vma,
3769 struct address_space *mapping, pgoff_t idx,
3770 unsigned long address, pte_t *ptep, unsigned int flags)
3772 struct hstate *h = hstate_vma(vma);
3773 vm_fault_t ret = VM_FAULT_SIGBUS;
3779 unsigned long haddr = address & huge_page_mask(h);
3780 bool new_page = false;
3783 * Currently, we are forced to kill the process in the event the
3784 * original mapper has unmapped pages from the child due to a failed
3785 * COW. Warn that such a situation has occurred as it may not be obvious
3787 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3788 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3794 * Use page lock to guard against racing truncation
3795 * before we get page_table_lock.
3798 page = find_lock_page(mapping, idx);
3800 size = i_size_read(mapping->host) >> huge_page_shift(h);
3805 * Check for page in userfault range
3807 if (userfaultfd_missing(vma)) {
3809 struct vm_fault vmf = {
3814 * Hard to debug if it ends up being
3815 * used by a callee that assumes
3816 * something about the other
3817 * uninitialized fields... same as in
3823 * hugetlb_fault_mutex must be dropped before
3824 * handling userfault. Reacquire after handling
3825 * fault to make calling code simpler.
3827 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping,
3829 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3830 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3831 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3835 page = alloc_huge_page(vma, haddr, 0);
3837 ret = vmf_error(PTR_ERR(page));
3840 clear_huge_page(page, address, pages_per_huge_page(h));
3841 __SetPageUptodate(page);
3844 if (vma->vm_flags & VM_MAYSHARE) {
3845 int err = huge_add_to_page_cache(page, mapping, idx);
3854 if (unlikely(anon_vma_prepare(vma))) {
3856 goto backout_unlocked;
3862 * If memory error occurs between mmap() and fault, some process
3863 * don't have hwpoisoned swap entry for errored virtual address.
3864 * So we need to block hugepage fault by PG_hwpoison bit check.
3866 if (unlikely(PageHWPoison(page))) {
3867 ret = VM_FAULT_HWPOISON |
3868 VM_FAULT_SET_HINDEX(hstate_index(h));
3869 goto backout_unlocked;
3874 * If we are going to COW a private mapping later, we examine the
3875 * pending reservations for this page now. This will ensure that
3876 * any allocations necessary to record that reservation occur outside
3879 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3880 if (vma_needs_reservation(h, vma, haddr) < 0) {
3882 goto backout_unlocked;
3884 /* Just decrements count, does not deallocate */
3885 vma_end_reservation(h, vma, haddr);
3888 ptl = huge_pte_lock(h, mm, ptep);
3889 size = i_size_read(mapping->host) >> huge_page_shift(h);
3894 if (!huge_pte_none(huge_ptep_get(ptep)))
3898 ClearPagePrivate(page);
3899 hugepage_add_new_anon_rmap(page, vma, haddr);
3901 page_dup_rmap(page, true);
3902 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3903 && (vma->vm_flags & VM_SHARED)));
3904 set_huge_pte_at(mm, haddr, ptep, new_pte);
3906 hugetlb_count_add(pages_per_huge_page(h), mm);
3907 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3908 /* Optimization, do the COW without a second fault */
3909 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
3915 * Only make newly allocated pages active. Existing pages found
3916 * in the pagecache could be !page_huge_active() if they have been
3917 * isolated for migration.
3920 set_page_huge_active(page);
3930 restore_reserve_on_error(h, vma, haddr, page);
3936 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3937 struct vm_area_struct *vma,
3938 struct address_space *mapping,
3939 pgoff_t idx, unsigned long address)
3941 unsigned long key[2];
3944 if (vma->vm_flags & VM_SHARED) {
3945 key[0] = (unsigned long) mapping;
3948 key[0] = (unsigned long) mm;
3949 key[1] = address >> huge_page_shift(h);
3952 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3954 return hash & (num_fault_mutexes - 1);
3958 * For uniprocesor systems we always use a single mutex, so just
3959 * return 0 and avoid the hashing overhead.
3961 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3962 struct vm_area_struct *vma,
3963 struct address_space *mapping,
3964 pgoff_t idx, unsigned long address)
3970 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3971 unsigned long address, unsigned int flags)
3978 struct page *page = NULL;
3979 struct page *pagecache_page = NULL;
3980 struct hstate *h = hstate_vma(vma);
3981 struct address_space *mapping;
3982 int need_wait_lock = 0;
3983 unsigned long haddr = address & huge_page_mask(h);
3985 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3987 entry = huge_ptep_get(ptep);
3988 if (unlikely(is_hugetlb_entry_migration(entry))) {
3989 migration_entry_wait_huge(vma, mm, ptep);
3991 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3992 return VM_FAULT_HWPOISON_LARGE |
3993 VM_FAULT_SET_HINDEX(hstate_index(h));
3995 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
3997 return VM_FAULT_OOM;
4000 mapping = vma->vm_file->f_mapping;
4001 idx = vma_hugecache_offset(h, vma, haddr);
4004 * Serialize hugepage allocation and instantiation, so that we don't
4005 * get spurious allocation failures if two CPUs race to instantiate
4006 * the same page in the page cache.
4008 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, haddr);
4009 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4011 entry = huge_ptep_get(ptep);
4012 if (huge_pte_none(entry)) {
4013 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4020 * entry could be a migration/hwpoison entry at this point, so this
4021 * check prevents the kernel from going below assuming that we have
4022 * a active hugepage in pagecache. This goto expects the 2nd page fault,
4023 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
4026 if (!pte_present(entry))
4030 * If we are going to COW the mapping later, we examine the pending
4031 * reservations for this page now. This will ensure that any
4032 * allocations necessary to record that reservation occur outside the
4033 * spinlock. For private mappings, we also lookup the pagecache
4034 * page now as it is used to determine if a reservation has been
4037 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4038 if (vma_needs_reservation(h, vma, haddr) < 0) {
4042 /* Just decrements count, does not deallocate */
4043 vma_end_reservation(h, vma, haddr);
4045 if (!(vma->vm_flags & VM_MAYSHARE))
4046 pagecache_page = hugetlbfs_pagecache_page(h,
4050 ptl = huge_pte_lock(h, mm, ptep);
4052 /* Check for a racing update before calling hugetlb_cow */
4053 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4057 * hugetlb_cow() requires page locks of pte_page(entry) and
4058 * pagecache_page, so here we need take the former one
4059 * when page != pagecache_page or !pagecache_page.
4061 page = pte_page(entry);
4062 if (page != pagecache_page)
4063 if (!trylock_page(page)) {
4070 if (flags & FAULT_FLAG_WRITE) {
4071 if (!huge_pte_write(entry)) {
4072 ret = hugetlb_cow(mm, vma, address, ptep,
4073 pagecache_page, ptl);
4076 entry = huge_pte_mkdirty(entry);
4078 entry = pte_mkyoung(entry);
4079 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4080 flags & FAULT_FLAG_WRITE))
4081 update_mmu_cache(vma, haddr, ptep);
4083 if (page != pagecache_page)
4089 if (pagecache_page) {
4090 unlock_page(pagecache_page);
4091 put_page(pagecache_page);
4094 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4096 * Generally it's safe to hold refcount during waiting page lock. But
4097 * here we just wait to defer the next page fault to avoid busy loop and
4098 * the page is not used after unlocked before returning from the current
4099 * page fault. So we are safe from accessing freed page, even if we wait
4100 * here without taking refcount.
4103 wait_on_page_locked(page);
4108 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4109 * modifications for huge pages.
4111 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4113 struct vm_area_struct *dst_vma,
4114 unsigned long dst_addr,
4115 unsigned long src_addr,
4116 struct page **pagep)
4118 struct address_space *mapping;
4121 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4122 struct hstate *h = hstate_vma(dst_vma);
4130 page = alloc_huge_page(dst_vma, dst_addr, 0);
4134 ret = copy_huge_page_from_user(page,
4135 (const void __user *) src_addr,
4136 pages_per_huge_page(h), false);
4138 /* fallback to copy_from_user outside mmap_sem */
4139 if (unlikely(ret)) {
4142 /* don't free the page */
4151 * The memory barrier inside __SetPageUptodate makes sure that
4152 * preceding stores to the page contents become visible before
4153 * the set_pte_at() write.
4155 __SetPageUptodate(page);
4157 mapping = dst_vma->vm_file->f_mapping;
4158 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4161 * If shared, add to page cache
4164 size = i_size_read(mapping->host) >> huge_page_shift(h);
4167 goto out_release_nounlock;
4170 * Serialization between remove_inode_hugepages() and
4171 * huge_add_to_page_cache() below happens through the
4172 * hugetlb_fault_mutex_table that here must be hold by
4175 ret = huge_add_to_page_cache(page, mapping, idx);
4177 goto out_release_nounlock;
4180 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4184 * Recheck the i_size after holding PT lock to make sure not
4185 * to leave any page mapped (as page_mapped()) beyond the end
4186 * of the i_size (remove_inode_hugepages() is strict about
4187 * enforcing that). If we bail out here, we'll also leave a
4188 * page in the radix tree in the vm_shared case beyond the end
4189 * of the i_size, but remove_inode_hugepages() will take care
4190 * of it as soon as we drop the hugetlb_fault_mutex_table.
4192 size = i_size_read(mapping->host) >> huge_page_shift(h);
4195 goto out_release_unlock;
4198 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4199 goto out_release_unlock;
4202 page_dup_rmap(page, true);
4204 ClearPagePrivate(page);
4205 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4208 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4209 if (dst_vma->vm_flags & VM_WRITE)
4210 _dst_pte = huge_pte_mkdirty(_dst_pte);
4211 _dst_pte = pte_mkyoung(_dst_pte);
4213 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4215 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4216 dst_vma->vm_flags & VM_WRITE);
4217 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4219 /* No need to invalidate - it was non-present before */
4220 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4223 set_page_huge_active(page);
4233 out_release_nounlock:
4238 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4239 struct page **pages, struct vm_area_struct **vmas,
4240 unsigned long *position, unsigned long *nr_pages,
4241 long i, unsigned int flags, int *nonblocking)
4243 unsigned long pfn_offset;
4244 unsigned long vaddr = *position;
4245 unsigned long remainder = *nr_pages;
4246 struct hstate *h = hstate_vma(vma);
4249 while (vaddr < vma->vm_end && remainder) {
4251 spinlock_t *ptl = NULL;
4256 * If we have a pending SIGKILL, don't keep faulting pages and
4257 * potentially allocating memory.
4259 if (fatal_signal_pending(current)) {
4265 * Some archs (sparc64, sh*) have multiple pte_ts to
4266 * each hugepage. We have to make sure we get the
4267 * first, for the page indexing below to work.
4269 * Note that page table lock is not held when pte is null.
4271 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4274 ptl = huge_pte_lock(h, mm, pte);
4275 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4278 * When coredumping, it suits get_dump_page if we just return
4279 * an error where there's an empty slot with no huge pagecache
4280 * to back it. This way, we avoid allocating a hugepage, and
4281 * the sparse dumpfile avoids allocating disk blocks, but its
4282 * huge holes still show up with zeroes where they need to be.
4284 if (absent && (flags & FOLL_DUMP) &&
4285 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4293 * We need call hugetlb_fault for both hugepages under migration
4294 * (in which case hugetlb_fault waits for the migration,) and
4295 * hwpoisoned hugepages (in which case we need to prevent the
4296 * caller from accessing to them.) In order to do this, we use
4297 * here is_swap_pte instead of is_hugetlb_entry_migration and
4298 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4299 * both cases, and because we can't follow correct pages
4300 * directly from any kind of swap entries.
4302 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4303 ((flags & FOLL_WRITE) &&
4304 !huge_pte_write(huge_ptep_get(pte)))) {
4306 unsigned int fault_flags = 0;
4310 if (flags & FOLL_WRITE)
4311 fault_flags |= FAULT_FLAG_WRITE;
4313 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4314 if (flags & FOLL_NOWAIT)
4315 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4316 FAULT_FLAG_RETRY_NOWAIT;
4317 if (flags & FOLL_TRIED) {
4318 VM_WARN_ON_ONCE(fault_flags &
4319 FAULT_FLAG_ALLOW_RETRY);
4320 fault_flags |= FAULT_FLAG_TRIED;
4322 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4323 if (ret & VM_FAULT_ERROR) {
4324 err = vm_fault_to_errno(ret, flags);
4328 if (ret & VM_FAULT_RETRY) {
4330 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4334 * VM_FAULT_RETRY must not return an
4335 * error, it will return zero
4338 * No need to update "position" as the
4339 * caller will not check it after
4340 * *nr_pages is set to 0.
4347 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4348 page = pte_page(huge_ptep_get(pte));
4351 * Instead of doing 'try_get_page()' below in the same_page
4352 * loop, just check the count once here.
4354 if (unlikely(page_count(page) <= 0)) {
4364 pages[i] = mem_map_offset(page, pfn_offset);
4375 if (vaddr < vma->vm_end && remainder &&
4376 pfn_offset < pages_per_huge_page(h)) {
4378 * We use pfn_offset to avoid touching the pageframes
4379 * of this compound page.
4385 *nr_pages = remainder;
4387 * setting position is actually required only if remainder is
4388 * not zero but it's faster not to add a "if (remainder)"
4396 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4398 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4401 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4404 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4405 unsigned long address, unsigned long end, pgprot_t newprot)
4407 struct mm_struct *mm = vma->vm_mm;
4408 unsigned long start = address;
4411 struct hstate *h = hstate_vma(vma);
4412 unsigned long pages = 0;
4413 bool shared_pmd = false;
4414 struct mmu_notifier_range range;
4417 * In the case of shared PMDs, the area to flush could be beyond
4418 * start/end. Set range.start/range.end to cover the maximum possible
4419 * range if PMD sharing is possible.
4421 mmu_notifier_range_init(&range, mm, start, end);
4422 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4424 BUG_ON(address >= end);
4425 flush_cache_range(vma, range.start, range.end);
4427 mmu_notifier_invalidate_range_start(&range);
4428 i_mmap_lock_write(vma->vm_file->f_mapping);
4429 for (; address < end; address += huge_page_size(h)) {
4431 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4434 ptl = huge_pte_lock(h, mm, ptep);
4435 if (huge_pmd_unshare(mm, &address, ptep)) {
4441 pte = huge_ptep_get(ptep);
4442 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4446 if (unlikely(is_hugetlb_entry_migration(pte))) {
4447 swp_entry_t entry = pte_to_swp_entry(pte);
4449 if (is_write_migration_entry(entry)) {
4452 make_migration_entry_read(&entry);
4453 newpte = swp_entry_to_pte(entry);
4454 set_huge_swap_pte_at(mm, address, ptep,
4455 newpte, huge_page_size(h));
4461 if (!huge_pte_none(pte)) {
4464 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
4465 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
4466 pte = arch_make_huge_pte(pte, vma, NULL, 0);
4467 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
4473 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4474 * may have cleared our pud entry and done put_page on the page table:
4475 * once we release i_mmap_rwsem, another task can do the final put_page
4476 * and that page table be reused and filled with junk. If we actually
4477 * did unshare a page of pmds, flush the range corresponding to the pud.
4480 flush_hugetlb_tlb_range(vma, range.start, range.end);
4482 flush_hugetlb_tlb_range(vma, start, end);
4484 * No need to call mmu_notifier_invalidate_range() we are downgrading
4485 * page table protection not changing it to point to a new page.
4487 * See Documentation/vm/mmu_notifier.rst
4489 i_mmap_unlock_write(vma->vm_file->f_mapping);
4490 mmu_notifier_invalidate_range_end(&range);
4492 return pages << h->order;
4495 int hugetlb_reserve_pages(struct inode *inode,
4497 struct vm_area_struct *vma,
4498 vm_flags_t vm_flags)
4501 struct hstate *h = hstate_inode(inode);
4502 struct hugepage_subpool *spool = subpool_inode(inode);
4503 struct resv_map *resv_map;
4506 /* This should never happen */
4508 VM_WARN(1, "%s called with a negative range\n", __func__);
4513 * Only apply hugepage reservation if asked. At fault time, an
4514 * attempt will be made for VM_NORESERVE to allocate a page
4515 * without using reserves
4517 if (vm_flags & VM_NORESERVE)
4521 * Shared mappings base their reservation on the number of pages that
4522 * are already allocated on behalf of the file. Private mappings need
4523 * to reserve the full area even if read-only as mprotect() may be
4524 * called to make the mapping read-write. Assume !vma is a shm mapping
4526 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4527 resv_map = inode_resv_map(inode);
4529 chg = region_chg(resv_map, from, to);
4532 resv_map = resv_map_alloc();
4538 set_vma_resv_map(vma, resv_map);
4539 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4548 * There must be enough pages in the subpool for the mapping. If
4549 * the subpool has a minimum size, there may be some global
4550 * reservations already in place (gbl_reserve).
4552 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4553 if (gbl_reserve < 0) {
4559 * Check enough hugepages are available for the reservation.
4560 * Hand the pages back to the subpool if there are not
4562 ret = hugetlb_acct_memory(h, gbl_reserve);
4564 /* put back original number of pages, chg */
4565 (void)hugepage_subpool_put_pages(spool, chg);
4570 * Account for the reservations made. Shared mappings record regions
4571 * that have reservations as they are shared by multiple VMAs.
4572 * When the last VMA disappears, the region map says how much
4573 * the reservation was and the page cache tells how much of
4574 * the reservation was consumed. Private mappings are per-VMA and
4575 * only the consumed reservations are tracked. When the VMA
4576 * disappears, the original reservation is the VMA size and the
4577 * consumed reservations are stored in the map. Hence, nothing
4578 * else has to be done for private mappings here
4580 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4581 long add = region_add(resv_map, from, to);
4583 if (unlikely(chg > add)) {
4585 * pages in this range were added to the reserve
4586 * map between region_chg and region_add. This
4587 * indicates a race with alloc_huge_page. Adjust
4588 * the subpool and reserve counts modified above
4589 * based on the difference.
4593 rsv_adjust = hugepage_subpool_put_pages(spool,
4595 hugetlb_acct_memory(h, -rsv_adjust);
4600 if (!vma || vma->vm_flags & VM_MAYSHARE)
4601 /* Don't call region_abort if region_chg failed */
4603 region_abort(resv_map, from, to);
4604 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4605 kref_put(&resv_map->refs, resv_map_release);
4609 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4612 struct hstate *h = hstate_inode(inode);
4613 struct resv_map *resv_map = inode_resv_map(inode);
4615 struct hugepage_subpool *spool = subpool_inode(inode);
4619 chg = region_del(resv_map, start, end);
4621 * region_del() can fail in the rare case where a region
4622 * must be split and another region descriptor can not be
4623 * allocated. If end == LONG_MAX, it will not fail.
4629 spin_lock(&inode->i_lock);
4630 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4631 spin_unlock(&inode->i_lock);
4634 * If the subpool has a minimum size, the number of global
4635 * reservations to be released may be adjusted.
4637 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4638 hugetlb_acct_memory(h, -gbl_reserve);
4643 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4644 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4645 struct vm_area_struct *vma,
4646 unsigned long addr, pgoff_t idx)
4648 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4650 unsigned long sbase = saddr & PUD_MASK;
4651 unsigned long s_end = sbase + PUD_SIZE;
4653 /* Allow segments to share if only one is marked locked */
4654 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4655 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4658 * match the virtual addresses, permission and the alignment of the
4661 if (pmd_index(addr) != pmd_index(saddr) ||
4662 vm_flags != svm_flags ||
4663 sbase < svma->vm_start || svma->vm_end < s_end)
4669 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4671 unsigned long base = addr & PUD_MASK;
4672 unsigned long end = base + PUD_SIZE;
4675 * check on proper vm_flags and page table alignment
4677 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
4683 * Determine if start,end range within vma could be mapped by shared pmd.
4684 * If yes, adjust start and end to cover range associated with possible
4685 * shared pmd mappings.
4687 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4688 unsigned long *start, unsigned long *end)
4690 unsigned long check_addr = *start;
4692 if (!(vma->vm_flags & VM_MAYSHARE))
4695 for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
4696 unsigned long a_start = check_addr & PUD_MASK;
4697 unsigned long a_end = a_start + PUD_SIZE;
4700 * If sharing is possible, adjust start/end if necessary.
4702 if (range_in_vma(vma, a_start, a_end)) {
4703 if (a_start < *start)
4712 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4713 * and returns the corresponding pte. While this is not necessary for the
4714 * !shared pmd case because we can allocate the pmd later as well, it makes the
4715 * code much cleaner. pmd allocation is essential for the shared case because
4716 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4717 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4718 * bad pmd for sharing.
4720 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4722 struct vm_area_struct *vma = find_vma(mm, addr);
4723 struct address_space *mapping = vma->vm_file->f_mapping;
4724 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4726 struct vm_area_struct *svma;
4727 unsigned long saddr;
4732 if (!vma_shareable(vma, addr))
4733 return (pte_t *)pmd_alloc(mm, pud, addr);
4735 i_mmap_lock_write(mapping);
4736 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4740 saddr = page_table_shareable(svma, vma, addr, idx);
4742 spte = huge_pte_offset(svma->vm_mm, saddr,
4743 vma_mmu_pagesize(svma));
4745 get_page(virt_to_page(spte));
4754 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4755 if (pud_none(*pud)) {
4756 pud_populate(mm, pud,
4757 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4760 put_page(virt_to_page(spte));
4764 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4765 i_mmap_unlock_write(mapping);
4770 * unmap huge page backed by shared pte.
4772 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4773 * indicated by page_count > 1, unmap is achieved by clearing pud and
4774 * decrementing the ref count. If count == 1, the pte page is not shared.
4776 * called with page table lock held.
4778 * returns: 1 successfully unmapped a shared pte page
4779 * 0 the underlying pte page is not shared, or it is the last user
4781 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4783 pgd_t *pgd = pgd_offset(mm, *addr);
4784 p4d_t *p4d = p4d_offset(pgd, *addr);
4785 pud_t *pud = pud_offset(p4d, *addr);
4787 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4788 if (page_count(virt_to_page(ptep)) == 1)
4792 put_page(virt_to_page(ptep));
4794 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4797 #define want_pmd_share() (1)
4798 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4799 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4804 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4809 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4810 unsigned long *start, unsigned long *end)
4813 #define want_pmd_share() (0)
4814 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4816 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4817 pte_t *huge_pte_alloc(struct mm_struct *mm,
4818 unsigned long addr, unsigned long sz)
4825 pgd = pgd_offset(mm, addr);
4826 p4d = p4d_alloc(mm, pgd, addr);
4829 pud = pud_alloc(mm, p4d, addr);
4831 if (sz == PUD_SIZE) {
4834 BUG_ON(sz != PMD_SIZE);
4835 if (want_pmd_share() && pud_none(*pud))
4836 pte = huge_pmd_share(mm, addr, pud);
4838 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4841 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4847 * huge_pte_offset() - Walk the page table to resolve the hugepage
4848 * entry at address @addr
4850 * Return: Pointer to page table or swap entry (PUD or PMD) for
4851 * address @addr, or NULL if a p*d_none() entry is encountered and the
4852 * size @sz doesn't match the hugepage size at this level of the page
4855 pte_t *huge_pte_offset(struct mm_struct *mm,
4856 unsigned long addr, unsigned long sz)
4863 pgd = pgd_offset(mm, addr);
4864 if (!pgd_present(*pgd))
4866 p4d = p4d_offset(pgd, addr);
4867 if (!p4d_present(*p4d))
4870 pud = pud_offset(p4d, addr);
4871 if (sz != PUD_SIZE && pud_none(*pud))
4873 /* hugepage or swap? */
4874 if (pud_huge(*pud) || !pud_present(*pud))
4875 return (pte_t *)pud;
4877 pmd = pmd_offset(pud, addr);
4878 if (sz != PMD_SIZE && pmd_none(*pmd))
4880 /* hugepage or swap? */
4881 if (pmd_huge(*pmd) || !pmd_present(*pmd))
4882 return (pte_t *)pmd;
4887 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4890 * These functions are overwritable if your architecture needs its own
4893 struct page * __weak
4894 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4897 return ERR_PTR(-EINVAL);
4900 struct page * __weak
4901 follow_huge_pd(struct vm_area_struct *vma,
4902 unsigned long address, hugepd_t hpd, int flags, int pdshift)
4904 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4908 struct page * __weak
4909 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4910 pmd_t *pmd, int flags)
4912 struct page *page = NULL;
4916 ptl = pmd_lockptr(mm, pmd);
4919 * make sure that the address range covered by this pmd is not
4920 * unmapped from other threads.
4922 if (!pmd_huge(*pmd))
4924 pte = huge_ptep_get((pte_t *)pmd);
4925 if (pte_present(pte)) {
4926 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4927 if (flags & FOLL_GET)
4930 if (is_hugetlb_entry_migration(pte)) {
4932 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4936 * hwpoisoned entry is treated as no_page_table in
4937 * follow_page_mask().
4945 struct page * __weak
4946 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4947 pud_t *pud, int flags)
4949 if (flags & FOLL_GET)
4952 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4955 struct page * __weak
4956 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
4958 if (flags & FOLL_GET)
4961 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
4964 bool isolate_huge_page(struct page *page, struct list_head *list)
4968 VM_BUG_ON_PAGE(!PageHead(page), page);
4969 spin_lock(&hugetlb_lock);
4970 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4974 clear_page_huge_active(page);
4975 list_move_tail(&page->lru, list);
4977 spin_unlock(&hugetlb_lock);
4981 void putback_active_hugepage(struct page *page)
4983 VM_BUG_ON_PAGE(!PageHead(page), page);
4984 spin_lock(&hugetlb_lock);
4985 set_page_huge_active(page);
4986 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4987 spin_unlock(&hugetlb_lock);
4991 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
4993 struct hstate *h = page_hstate(oldpage);
4995 hugetlb_cgroup_migrate(oldpage, newpage);
4996 set_page_owner_migrate_reason(newpage, reason);
4999 * transfer temporary state of the new huge page. This is
5000 * reverse to other transitions because the newpage is going to
5001 * be final while the old one will be freed so it takes over
5002 * the temporary status.
5004 * Also note that we have to transfer the per-node surplus state
5005 * here as well otherwise the global surplus count will not match
5008 if (PageHugeTemporary(newpage)) {
5009 int old_nid = page_to_nid(oldpage);
5010 int new_nid = page_to_nid(newpage);
5012 SetPageHugeTemporary(oldpage);
5013 ClearPageHugeTemporary(newpage);
5015 spin_lock(&hugetlb_lock);
5016 if (h->surplus_huge_pages_node[old_nid]) {
5017 h->surplus_huge_pages_node[old_nid]--;
5018 h->surplus_huge_pages_node[new_nid]++;
5020 spin_unlock(&hugetlb_lock);