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 * A return code of zero implies that the subpool will be under its
1272 * minimum size if the reservation is not restored after page is free.
1273 * Therefore, force restore_reserve operation.
1275 if (hugepage_subpool_put_pages(spool, 1) == 0)
1276 restore_reserve = true;
1278 spin_lock(&hugetlb_lock);
1279 clear_page_huge_active(page);
1280 hugetlb_cgroup_uncharge_page(hstate_index(h),
1281 pages_per_huge_page(h), page);
1282 if (restore_reserve)
1283 h->resv_huge_pages++;
1285 if (PageHugeTemporary(page)) {
1286 list_del(&page->lru);
1287 ClearPageHugeTemporary(page);
1288 update_and_free_page(h, page);
1289 } else if (h->surplus_huge_pages_node[nid]) {
1290 /* remove the page from active list */
1291 list_del(&page->lru);
1292 update_and_free_page(h, page);
1293 h->surplus_huge_pages--;
1294 h->surplus_huge_pages_node[nid]--;
1296 arch_clear_hugepage_flags(page);
1297 enqueue_huge_page(h, page);
1299 spin_unlock(&hugetlb_lock);
1302 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1304 INIT_LIST_HEAD(&page->lru);
1305 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1306 spin_lock(&hugetlb_lock);
1307 set_hugetlb_cgroup(page, NULL);
1309 h->nr_huge_pages_node[nid]++;
1310 spin_unlock(&hugetlb_lock);
1313 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1316 int nr_pages = 1 << order;
1317 struct page *p = page + 1;
1319 /* we rely on prep_new_huge_page to set the destructor */
1320 set_compound_order(page, order);
1321 __ClearPageReserved(page);
1322 __SetPageHead(page);
1323 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1325 * For gigantic hugepages allocated through bootmem at
1326 * boot, it's safer to be consistent with the not-gigantic
1327 * hugepages and clear the PG_reserved bit from all tail pages
1328 * too. Otherwse drivers using get_user_pages() to access tail
1329 * pages may get the reference counting wrong if they see
1330 * PG_reserved set on a tail page (despite the head page not
1331 * having PG_reserved set). Enforcing this consistency between
1332 * head and tail pages allows drivers to optimize away a check
1333 * on the head page when they need know if put_page() is needed
1334 * after get_user_pages().
1336 __ClearPageReserved(p);
1337 set_page_count(p, 0);
1338 set_compound_head(p, page);
1340 atomic_set(compound_mapcount_ptr(page), -1);
1344 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1345 * transparent huge pages. See the PageTransHuge() documentation for more
1348 int PageHuge(struct page *page)
1350 if (!PageCompound(page))
1353 page = compound_head(page);
1354 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1356 EXPORT_SYMBOL_GPL(PageHuge);
1359 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1360 * normal or transparent huge pages.
1362 int PageHeadHuge(struct page *page_head)
1364 if (!PageHead(page_head))
1367 return get_compound_page_dtor(page_head) == free_huge_page;
1370 pgoff_t __basepage_index(struct page *page)
1372 struct page *page_head = compound_head(page);
1373 pgoff_t index = page_index(page_head);
1374 unsigned long compound_idx;
1376 if (!PageHuge(page_head))
1377 return page_index(page);
1379 if (compound_order(page_head) >= MAX_ORDER)
1380 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1382 compound_idx = page - page_head;
1384 return (index << compound_order(page_head)) + compound_idx;
1387 static struct page *alloc_buddy_huge_page(struct hstate *h,
1388 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1390 int order = huge_page_order(h);
1393 gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
1394 if (nid == NUMA_NO_NODE)
1395 nid = numa_mem_id();
1396 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1398 __count_vm_event(HTLB_BUDDY_PGALLOC);
1400 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1406 * Common helper to allocate a fresh hugetlb page. All specific allocators
1407 * should use this function to get new hugetlb pages
1409 static struct page *alloc_fresh_huge_page(struct hstate *h,
1410 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1414 if (hstate_is_gigantic(h))
1415 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1417 page = alloc_buddy_huge_page(h, gfp_mask,
1422 if (hstate_is_gigantic(h))
1423 prep_compound_gigantic_page(page, huge_page_order(h));
1424 prep_new_huge_page(h, page, page_to_nid(page));
1430 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1433 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1437 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1439 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1440 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed);
1448 put_page(page); /* free it into the hugepage allocator */
1454 * Free huge page from pool from next node to free.
1455 * Attempt to keep persistent huge pages more or less
1456 * balanced over allowed nodes.
1457 * Called with hugetlb_lock locked.
1459 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1465 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1467 * If we're returning unused surplus pages, only examine
1468 * nodes with surplus pages.
1470 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1471 !list_empty(&h->hugepage_freelists[node])) {
1473 list_entry(h->hugepage_freelists[node].next,
1475 list_del(&page->lru);
1476 h->free_huge_pages--;
1477 h->free_huge_pages_node[node]--;
1479 h->surplus_huge_pages--;
1480 h->surplus_huge_pages_node[node]--;
1482 update_and_free_page(h, page);
1492 * Dissolve a given free hugepage into free buddy pages. This function does
1493 * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1494 * dissolution fails because a give page is not a free hugepage, or because
1495 * free hugepages are fully reserved.
1497 int dissolve_free_huge_page(struct page *page)
1501 spin_lock(&hugetlb_lock);
1502 if (PageHuge(page) && !page_count(page)) {
1503 struct page *head = compound_head(page);
1504 struct hstate *h = page_hstate(head);
1505 int nid = page_to_nid(head);
1506 if (h->free_huge_pages - h->resv_huge_pages == 0)
1509 * Move PageHWPoison flag from head page to the raw error page,
1510 * which makes any subpages rather than the error page reusable.
1512 if (PageHWPoison(head) && page != head) {
1513 SetPageHWPoison(page);
1514 ClearPageHWPoison(head);
1516 list_del(&head->lru);
1517 h->free_huge_pages--;
1518 h->free_huge_pages_node[nid]--;
1519 h->max_huge_pages--;
1520 update_and_free_page(h, head);
1524 spin_unlock(&hugetlb_lock);
1529 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1530 * make specified memory blocks removable from the system.
1531 * Note that this will dissolve a free gigantic hugepage completely, if any
1532 * part of it lies within the given range.
1533 * Also note that if dissolve_free_huge_page() returns with an error, all
1534 * free hugepages that were dissolved before that error are lost.
1536 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1542 if (!hugepages_supported())
1545 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1546 page = pfn_to_page(pfn);
1547 if (PageHuge(page) && !page_count(page)) {
1548 rc = dissolve_free_huge_page(page);
1558 * Allocates a fresh surplus page from the page allocator.
1560 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1561 int nid, nodemask_t *nmask)
1563 struct page *page = NULL;
1565 if (hstate_is_gigantic(h))
1568 spin_lock(&hugetlb_lock);
1569 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1571 spin_unlock(&hugetlb_lock);
1573 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1577 spin_lock(&hugetlb_lock);
1579 * We could have raced with the pool size change.
1580 * Double check that and simply deallocate the new page
1581 * if we would end up overcommiting the surpluses. Abuse
1582 * temporary page to workaround the nasty free_huge_page
1585 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1586 SetPageHugeTemporary(page);
1587 spin_unlock(&hugetlb_lock);
1591 h->surplus_huge_pages++;
1592 h->surplus_huge_pages_node[page_to_nid(page)]++;
1596 spin_unlock(&hugetlb_lock);
1601 struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1602 int nid, nodemask_t *nmask)
1606 if (hstate_is_gigantic(h))
1609 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1614 * We do not account these pages as surplus because they are only
1615 * temporary and will be released properly on the last reference
1617 SetPageHugeTemporary(page);
1623 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1626 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1627 struct vm_area_struct *vma, unsigned long addr)
1630 struct mempolicy *mpol;
1631 gfp_t gfp_mask = htlb_alloc_mask(h);
1633 nodemask_t *nodemask;
1635 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1636 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1637 mpol_cond_put(mpol);
1642 /* page migration callback function */
1643 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1645 gfp_t gfp_mask = htlb_alloc_mask(h);
1646 struct page *page = NULL;
1648 if (nid != NUMA_NO_NODE)
1649 gfp_mask |= __GFP_THISNODE;
1651 spin_lock(&hugetlb_lock);
1652 if (h->free_huge_pages - h->resv_huge_pages > 0)
1653 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1654 spin_unlock(&hugetlb_lock);
1657 page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1662 /* page migration callback function */
1663 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1666 gfp_t gfp_mask = htlb_alloc_mask(h);
1668 spin_lock(&hugetlb_lock);
1669 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1672 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1674 spin_unlock(&hugetlb_lock);
1678 spin_unlock(&hugetlb_lock);
1680 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1683 /* mempolicy aware migration callback */
1684 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1685 unsigned long address)
1687 struct mempolicy *mpol;
1688 nodemask_t *nodemask;
1693 gfp_mask = htlb_alloc_mask(h);
1694 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1695 page = alloc_huge_page_nodemask(h, node, nodemask);
1696 mpol_cond_put(mpol);
1702 * Increase the hugetlb pool such that it can accommodate a reservation
1705 static int gather_surplus_pages(struct hstate *h, int delta)
1707 struct list_head surplus_list;
1708 struct page *page, *tmp;
1710 int needed, allocated;
1711 bool alloc_ok = true;
1713 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1715 h->resv_huge_pages += delta;
1720 INIT_LIST_HEAD(&surplus_list);
1724 spin_unlock(&hugetlb_lock);
1725 for (i = 0; i < needed; i++) {
1726 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1727 NUMA_NO_NODE, NULL);
1732 list_add(&page->lru, &surplus_list);
1738 * After retaking hugetlb_lock, we need to recalculate 'needed'
1739 * because either resv_huge_pages or free_huge_pages may have changed.
1741 spin_lock(&hugetlb_lock);
1742 needed = (h->resv_huge_pages + delta) -
1743 (h->free_huge_pages + allocated);
1748 * We were not able to allocate enough pages to
1749 * satisfy the entire reservation so we free what
1750 * we've allocated so far.
1755 * The surplus_list now contains _at_least_ the number of extra pages
1756 * needed to accommodate the reservation. Add the appropriate number
1757 * of pages to the hugetlb pool and free the extras back to the buddy
1758 * allocator. Commit the entire reservation here to prevent another
1759 * process from stealing the pages as they are added to the pool but
1760 * before they are reserved.
1762 needed += allocated;
1763 h->resv_huge_pages += delta;
1766 /* Free the needed pages to the hugetlb pool */
1767 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1771 * This page is now managed by the hugetlb allocator and has
1772 * no users -- drop the buddy allocator's reference.
1774 put_page_testzero(page);
1775 VM_BUG_ON_PAGE(page_count(page), page);
1776 enqueue_huge_page(h, page);
1779 spin_unlock(&hugetlb_lock);
1781 /* Free unnecessary surplus pages to the buddy allocator */
1782 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1784 spin_lock(&hugetlb_lock);
1790 * This routine has two main purposes:
1791 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1792 * in unused_resv_pages. This corresponds to the prior adjustments made
1793 * to the associated reservation map.
1794 * 2) Free any unused surplus pages that may have been allocated to satisfy
1795 * the reservation. As many as unused_resv_pages may be freed.
1797 * Called with hugetlb_lock held. However, the lock could be dropped (and
1798 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1799 * we must make sure nobody else can claim pages we are in the process of
1800 * freeing. Do this by ensuring resv_huge_page always is greater than the
1801 * number of huge pages we plan to free when dropping the lock.
1803 static void return_unused_surplus_pages(struct hstate *h,
1804 unsigned long unused_resv_pages)
1806 unsigned long nr_pages;
1808 /* Cannot return gigantic pages currently */
1809 if (hstate_is_gigantic(h))
1813 * Part (or even all) of the reservation could have been backed
1814 * by pre-allocated pages. Only free surplus pages.
1816 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1819 * We want to release as many surplus pages as possible, spread
1820 * evenly across all nodes with memory. Iterate across these nodes
1821 * until we can no longer free unreserved surplus pages. This occurs
1822 * when the nodes with surplus pages have no free pages.
1823 * free_pool_huge_page() will balance the the freed pages across the
1824 * on-line nodes with memory and will handle the hstate accounting.
1826 * Note that we decrement resv_huge_pages as we free the pages. If
1827 * we drop the lock, resv_huge_pages will still be sufficiently large
1828 * to cover subsequent pages we may free.
1830 while (nr_pages--) {
1831 h->resv_huge_pages--;
1832 unused_resv_pages--;
1833 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1835 cond_resched_lock(&hugetlb_lock);
1839 /* Fully uncommit the reservation */
1840 h->resv_huge_pages -= unused_resv_pages;
1845 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1846 * are used by the huge page allocation routines to manage reservations.
1848 * vma_needs_reservation is called to determine if the huge page at addr
1849 * within the vma has an associated reservation. If a reservation is
1850 * needed, the value 1 is returned. The caller is then responsible for
1851 * managing the global reservation and subpool usage counts. After
1852 * the huge page has been allocated, vma_commit_reservation is called
1853 * to add the page to the reservation map. If the page allocation fails,
1854 * the reservation must be ended instead of committed. vma_end_reservation
1855 * is called in such cases.
1857 * In the normal case, vma_commit_reservation returns the same value
1858 * as the preceding vma_needs_reservation call. The only time this
1859 * is not the case is if a reserve map was changed between calls. It
1860 * is the responsibility of the caller to notice the difference and
1861 * take appropriate action.
1863 * vma_add_reservation is used in error paths where a reservation must
1864 * be restored when a newly allocated huge page must be freed. It is
1865 * to be called after calling vma_needs_reservation to determine if a
1866 * reservation exists.
1868 enum vma_resv_mode {
1874 static long __vma_reservation_common(struct hstate *h,
1875 struct vm_area_struct *vma, unsigned long addr,
1876 enum vma_resv_mode mode)
1878 struct resv_map *resv;
1882 resv = vma_resv_map(vma);
1886 idx = vma_hugecache_offset(h, vma, addr);
1888 case VMA_NEEDS_RESV:
1889 ret = region_chg(resv, idx, idx + 1);
1891 case VMA_COMMIT_RESV:
1892 ret = region_add(resv, idx, idx + 1);
1895 region_abort(resv, idx, idx + 1);
1899 if (vma->vm_flags & VM_MAYSHARE)
1900 ret = region_add(resv, idx, idx + 1);
1902 region_abort(resv, idx, idx + 1);
1903 ret = region_del(resv, idx, idx + 1);
1910 if (vma->vm_flags & VM_MAYSHARE)
1912 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1914 * In most cases, reserves always exist for private mappings.
1915 * However, a file associated with mapping could have been
1916 * hole punched or truncated after reserves were consumed.
1917 * As subsequent fault on such a range will not use reserves.
1918 * Subtle - The reserve map for private mappings has the
1919 * opposite meaning than that of shared mappings. If NO
1920 * entry is in the reserve map, it means a reservation exists.
1921 * If an entry exists in the reserve map, it means the
1922 * reservation has already been consumed. As a result, the
1923 * return value of this routine is the opposite of the
1924 * value returned from reserve map manipulation routines above.
1932 return ret < 0 ? ret : 0;
1935 static long vma_needs_reservation(struct hstate *h,
1936 struct vm_area_struct *vma, unsigned long addr)
1938 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1941 static long vma_commit_reservation(struct hstate *h,
1942 struct vm_area_struct *vma, unsigned long addr)
1944 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1947 static void vma_end_reservation(struct hstate *h,
1948 struct vm_area_struct *vma, unsigned long addr)
1950 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1953 static long vma_add_reservation(struct hstate *h,
1954 struct vm_area_struct *vma, unsigned long addr)
1956 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1960 * This routine is called to restore a reservation on error paths. In the
1961 * specific error paths, a huge page was allocated (via alloc_huge_page)
1962 * and is about to be freed. If a reservation for the page existed,
1963 * alloc_huge_page would have consumed the reservation and set PagePrivate
1964 * in the newly allocated page. When the page is freed via free_huge_page,
1965 * the global reservation count will be incremented if PagePrivate is set.
1966 * However, free_huge_page can not adjust the reserve map. Adjust the
1967 * reserve map here to be consistent with global reserve count adjustments
1968 * to be made by free_huge_page.
1970 static void restore_reserve_on_error(struct hstate *h,
1971 struct vm_area_struct *vma, unsigned long address,
1974 if (unlikely(PagePrivate(page))) {
1975 long rc = vma_needs_reservation(h, vma, address);
1977 if (unlikely(rc < 0)) {
1979 * Rare out of memory condition in reserve map
1980 * manipulation. Clear PagePrivate so that
1981 * global reserve count will not be incremented
1982 * by free_huge_page. This will make it appear
1983 * as though the reservation for this page was
1984 * consumed. This may prevent the task from
1985 * faulting in the page at a later time. This
1986 * is better than inconsistent global huge page
1987 * accounting of reserve counts.
1989 ClearPagePrivate(page);
1991 rc = vma_add_reservation(h, vma, address);
1992 if (unlikely(rc < 0))
1994 * See above comment about rare out of
1997 ClearPagePrivate(page);
1999 vma_end_reservation(h, vma, address);
2003 struct page *alloc_huge_page(struct vm_area_struct *vma,
2004 unsigned long addr, int avoid_reserve)
2006 struct hugepage_subpool *spool = subpool_vma(vma);
2007 struct hstate *h = hstate_vma(vma);
2009 long map_chg, map_commit;
2012 struct hugetlb_cgroup *h_cg;
2014 idx = hstate_index(h);
2016 * Examine the region/reserve map to determine if the process
2017 * has a reservation for the page to be allocated. A return
2018 * code of zero indicates a reservation exists (no change).
2020 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2022 return ERR_PTR(-ENOMEM);
2025 * Processes that did not create the mapping will have no
2026 * reserves as indicated by the region/reserve map. Check
2027 * that the allocation will not exceed the subpool limit.
2028 * Allocations for MAP_NORESERVE mappings also need to be
2029 * checked against any subpool limit.
2031 if (map_chg || avoid_reserve) {
2032 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2034 vma_end_reservation(h, vma, addr);
2035 return ERR_PTR(-ENOSPC);
2039 * Even though there was no reservation in the region/reserve
2040 * map, there could be reservations associated with the
2041 * subpool that can be used. This would be indicated if the
2042 * return value of hugepage_subpool_get_pages() is zero.
2043 * However, if avoid_reserve is specified we still avoid even
2044 * the subpool reservations.
2050 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2052 goto out_subpool_put;
2054 spin_lock(&hugetlb_lock);
2056 * glb_chg is passed to indicate whether or not a page must be taken
2057 * from the global free pool (global change). gbl_chg == 0 indicates
2058 * a reservation exists for the allocation.
2060 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2062 spin_unlock(&hugetlb_lock);
2063 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2065 goto out_uncharge_cgroup;
2066 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2067 SetPagePrivate(page);
2068 h->resv_huge_pages--;
2070 spin_lock(&hugetlb_lock);
2071 list_move(&page->lru, &h->hugepage_activelist);
2074 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2075 spin_unlock(&hugetlb_lock);
2077 set_page_private(page, (unsigned long)spool);
2079 map_commit = vma_commit_reservation(h, vma, addr);
2080 if (unlikely(map_chg > map_commit)) {
2082 * The page was added to the reservation map between
2083 * vma_needs_reservation and vma_commit_reservation.
2084 * This indicates a race with hugetlb_reserve_pages.
2085 * Adjust for the subpool count incremented above AND
2086 * in hugetlb_reserve_pages for the same page. Also,
2087 * the reservation count added in hugetlb_reserve_pages
2088 * no longer applies.
2092 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2093 hugetlb_acct_memory(h, -rsv_adjust);
2097 out_uncharge_cgroup:
2098 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2100 if (map_chg || avoid_reserve)
2101 hugepage_subpool_put_pages(spool, 1);
2102 vma_end_reservation(h, vma, addr);
2103 return ERR_PTR(-ENOSPC);
2106 int alloc_bootmem_huge_page(struct hstate *h)
2107 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2108 int __alloc_bootmem_huge_page(struct hstate *h)
2110 struct huge_bootmem_page *m;
2113 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2116 addr = memblock_alloc_try_nid_raw(
2117 huge_page_size(h), huge_page_size(h),
2118 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2121 * Use the beginning of the huge page to store the
2122 * huge_bootmem_page struct (until gather_bootmem
2123 * puts them into the mem_map).
2132 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2133 /* Put them into a private list first because mem_map is not up yet */
2134 INIT_LIST_HEAD(&m->list);
2135 list_add(&m->list, &huge_boot_pages);
2140 static void __init prep_compound_huge_page(struct page *page,
2143 if (unlikely(order > (MAX_ORDER - 1)))
2144 prep_compound_gigantic_page(page, order);
2146 prep_compound_page(page, order);
2149 /* Put bootmem huge pages into the standard lists after mem_map is up */
2150 static void __init gather_bootmem_prealloc(void)
2152 struct huge_bootmem_page *m;
2154 list_for_each_entry(m, &huge_boot_pages, list) {
2155 struct page *page = virt_to_page(m);
2156 struct hstate *h = m->hstate;
2158 WARN_ON(page_count(page) != 1);
2159 prep_compound_huge_page(page, h->order);
2160 WARN_ON(PageReserved(page));
2161 prep_new_huge_page(h, page, page_to_nid(page));
2162 put_page(page); /* free it into the hugepage allocator */
2165 * If we had gigantic hugepages allocated at boot time, we need
2166 * to restore the 'stolen' pages to totalram_pages in order to
2167 * fix confusing memory reports from free(1) and another
2168 * side-effects, like CommitLimit going negative.
2170 if (hstate_is_gigantic(h))
2171 adjust_managed_page_count(page, 1 << h->order);
2176 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2180 for (i = 0; i < h->max_huge_pages; ++i) {
2181 if (hstate_is_gigantic(h)) {
2182 if (!alloc_bootmem_huge_page(h))
2184 } else if (!alloc_pool_huge_page(h,
2185 &node_states[N_MEMORY]))
2189 if (i < h->max_huge_pages) {
2192 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2193 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2194 h->max_huge_pages, buf, i);
2195 h->max_huge_pages = i;
2199 static void __init hugetlb_init_hstates(void)
2203 for_each_hstate(h) {
2204 if (minimum_order > huge_page_order(h))
2205 minimum_order = huge_page_order(h);
2207 /* oversize hugepages were init'ed in early boot */
2208 if (!hstate_is_gigantic(h))
2209 hugetlb_hstate_alloc_pages(h);
2211 VM_BUG_ON(minimum_order == UINT_MAX);
2214 static void __init report_hugepages(void)
2218 for_each_hstate(h) {
2221 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2222 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2223 buf, h->free_huge_pages);
2227 #ifdef CONFIG_HIGHMEM
2228 static void try_to_free_low(struct hstate *h, unsigned long count,
2229 nodemask_t *nodes_allowed)
2233 if (hstate_is_gigantic(h))
2236 for_each_node_mask(i, *nodes_allowed) {
2237 struct page *page, *next;
2238 struct list_head *freel = &h->hugepage_freelists[i];
2239 list_for_each_entry_safe(page, next, freel, lru) {
2240 if (count >= h->nr_huge_pages)
2242 if (PageHighMem(page))
2244 list_del(&page->lru);
2245 update_and_free_page(h, page);
2246 h->free_huge_pages--;
2247 h->free_huge_pages_node[page_to_nid(page)]--;
2252 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2253 nodemask_t *nodes_allowed)
2259 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2260 * balanced by operating on them in a round-robin fashion.
2261 * Returns 1 if an adjustment was made.
2263 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2268 VM_BUG_ON(delta != -1 && delta != 1);
2271 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2272 if (h->surplus_huge_pages_node[node])
2276 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2277 if (h->surplus_huge_pages_node[node] <
2278 h->nr_huge_pages_node[node])
2285 h->surplus_huge_pages += delta;
2286 h->surplus_huge_pages_node[node] += delta;
2290 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2291 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2292 nodemask_t *nodes_allowed)
2294 unsigned long min_count, ret;
2296 spin_lock(&hugetlb_lock);
2299 * Check for a node specific request.
2300 * Changing node specific huge page count may require a corresponding
2301 * change to the global count. In any case, the passed node mask
2302 * (nodes_allowed) will restrict alloc/free to the specified node.
2304 if (nid != NUMA_NO_NODE) {
2305 unsigned long old_count = count;
2307 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2309 * User may have specified a large count value which caused the
2310 * above calculation to overflow. In this case, they wanted
2311 * to allocate as many huge pages as possible. Set count to
2312 * largest possible value to align with their intention.
2314 if (count < old_count)
2319 * Gigantic pages runtime allocation depend on the capability for large
2320 * page range allocation.
2321 * If the system does not provide this feature, return an error when
2322 * the user tries to allocate gigantic pages but let the user free the
2323 * boottime allocated gigantic pages.
2325 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2326 if (count > persistent_huge_pages(h)) {
2327 spin_unlock(&hugetlb_lock);
2330 /* Fall through to decrease pool */
2334 * Increase the pool size
2335 * First take pages out of surplus state. Then make up the
2336 * remaining difference by allocating fresh huge pages.
2338 * We might race with alloc_surplus_huge_page() here and be unable
2339 * to convert a surplus huge page to a normal huge page. That is
2340 * not critical, though, it just means the overall size of the
2341 * pool might be one hugepage larger than it needs to be, but
2342 * within all the constraints specified by the sysctls.
2344 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2345 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2349 while (count > persistent_huge_pages(h)) {
2351 * If this allocation races such that we no longer need the
2352 * page, free_huge_page will handle it by freeing the page
2353 * and reducing the surplus.
2355 spin_unlock(&hugetlb_lock);
2357 /* yield cpu to avoid soft lockup */
2360 ret = alloc_pool_huge_page(h, nodes_allowed);
2361 spin_lock(&hugetlb_lock);
2365 /* Bail for signals. Probably ctrl-c from user */
2366 if (signal_pending(current))
2371 * Decrease the pool size
2372 * First return free pages to the buddy allocator (being careful
2373 * to keep enough around to satisfy reservations). Then place
2374 * pages into surplus state as needed so the pool will shrink
2375 * to the desired size as pages become free.
2377 * By placing pages into the surplus state independent of the
2378 * overcommit value, we are allowing the surplus pool size to
2379 * exceed overcommit. There are few sane options here. Since
2380 * alloc_surplus_huge_page() is checking the global counter,
2381 * though, we'll note that we're not allowed to exceed surplus
2382 * and won't grow the pool anywhere else. Not until one of the
2383 * sysctls are changed, or the surplus pages go out of use.
2385 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2386 min_count = max(count, min_count);
2387 try_to_free_low(h, min_count, nodes_allowed);
2388 while (min_count < persistent_huge_pages(h)) {
2389 if (!free_pool_huge_page(h, nodes_allowed, 0))
2391 cond_resched_lock(&hugetlb_lock);
2393 while (count < persistent_huge_pages(h)) {
2394 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2398 h->max_huge_pages = persistent_huge_pages(h);
2399 spin_unlock(&hugetlb_lock);
2404 #define HSTATE_ATTR_RO(_name) \
2405 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2407 #define HSTATE_ATTR(_name) \
2408 static struct kobj_attribute _name##_attr = \
2409 __ATTR(_name, 0644, _name##_show, _name##_store)
2411 static struct kobject *hugepages_kobj;
2412 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2414 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2416 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2420 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2421 if (hstate_kobjs[i] == kobj) {
2423 *nidp = NUMA_NO_NODE;
2427 return kobj_to_node_hstate(kobj, nidp);
2430 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2431 struct kobj_attribute *attr, char *buf)
2434 unsigned long nr_huge_pages;
2437 h = kobj_to_hstate(kobj, &nid);
2438 if (nid == NUMA_NO_NODE)
2439 nr_huge_pages = h->nr_huge_pages;
2441 nr_huge_pages = h->nr_huge_pages_node[nid];
2443 return sprintf(buf, "%lu\n", nr_huge_pages);
2446 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2447 struct hstate *h, int nid,
2448 unsigned long count, size_t len)
2451 nodemask_t nodes_allowed, *n_mask;
2453 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2456 if (nid == NUMA_NO_NODE) {
2458 * global hstate attribute
2460 if (!(obey_mempolicy &&
2461 init_nodemask_of_mempolicy(&nodes_allowed)))
2462 n_mask = &node_states[N_MEMORY];
2464 n_mask = &nodes_allowed;
2467 * Node specific request. count adjustment happens in
2468 * set_max_huge_pages() after acquiring hugetlb_lock.
2470 init_nodemask_of_node(&nodes_allowed, nid);
2471 n_mask = &nodes_allowed;
2474 err = set_max_huge_pages(h, count, nid, n_mask);
2476 return err ? err : len;
2479 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2480 struct kobject *kobj, const char *buf,
2484 unsigned long count;
2488 err = kstrtoul(buf, 10, &count);
2492 h = kobj_to_hstate(kobj, &nid);
2493 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2496 static ssize_t nr_hugepages_show(struct kobject *kobj,
2497 struct kobj_attribute *attr, char *buf)
2499 return nr_hugepages_show_common(kobj, attr, buf);
2502 static ssize_t nr_hugepages_store(struct kobject *kobj,
2503 struct kobj_attribute *attr, const char *buf, size_t len)
2505 return nr_hugepages_store_common(false, kobj, buf, len);
2507 HSTATE_ATTR(nr_hugepages);
2512 * hstate attribute for optionally mempolicy-based constraint on persistent
2513 * huge page alloc/free.
2515 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2516 struct kobj_attribute *attr, char *buf)
2518 return nr_hugepages_show_common(kobj, attr, buf);
2521 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2522 struct kobj_attribute *attr, const char *buf, size_t len)
2524 return nr_hugepages_store_common(true, kobj, buf, len);
2526 HSTATE_ATTR(nr_hugepages_mempolicy);
2530 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2531 struct kobj_attribute *attr, char *buf)
2533 struct hstate *h = kobj_to_hstate(kobj, NULL);
2534 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2537 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2538 struct kobj_attribute *attr, const char *buf, size_t count)
2541 unsigned long input;
2542 struct hstate *h = kobj_to_hstate(kobj, NULL);
2544 if (hstate_is_gigantic(h))
2547 err = kstrtoul(buf, 10, &input);
2551 spin_lock(&hugetlb_lock);
2552 h->nr_overcommit_huge_pages = input;
2553 spin_unlock(&hugetlb_lock);
2557 HSTATE_ATTR(nr_overcommit_hugepages);
2559 static ssize_t free_hugepages_show(struct kobject *kobj,
2560 struct kobj_attribute *attr, char *buf)
2563 unsigned long free_huge_pages;
2566 h = kobj_to_hstate(kobj, &nid);
2567 if (nid == NUMA_NO_NODE)
2568 free_huge_pages = h->free_huge_pages;
2570 free_huge_pages = h->free_huge_pages_node[nid];
2572 return sprintf(buf, "%lu\n", free_huge_pages);
2574 HSTATE_ATTR_RO(free_hugepages);
2576 static ssize_t resv_hugepages_show(struct kobject *kobj,
2577 struct kobj_attribute *attr, char *buf)
2579 struct hstate *h = kobj_to_hstate(kobj, NULL);
2580 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2582 HSTATE_ATTR_RO(resv_hugepages);
2584 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2585 struct kobj_attribute *attr, char *buf)
2588 unsigned long surplus_huge_pages;
2591 h = kobj_to_hstate(kobj, &nid);
2592 if (nid == NUMA_NO_NODE)
2593 surplus_huge_pages = h->surplus_huge_pages;
2595 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2597 return sprintf(buf, "%lu\n", surplus_huge_pages);
2599 HSTATE_ATTR_RO(surplus_hugepages);
2601 static struct attribute *hstate_attrs[] = {
2602 &nr_hugepages_attr.attr,
2603 &nr_overcommit_hugepages_attr.attr,
2604 &free_hugepages_attr.attr,
2605 &resv_hugepages_attr.attr,
2606 &surplus_hugepages_attr.attr,
2608 &nr_hugepages_mempolicy_attr.attr,
2613 static const struct attribute_group hstate_attr_group = {
2614 .attrs = hstate_attrs,
2617 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2618 struct kobject **hstate_kobjs,
2619 const struct attribute_group *hstate_attr_group)
2622 int hi = hstate_index(h);
2624 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2625 if (!hstate_kobjs[hi])
2628 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2630 kobject_put(hstate_kobjs[hi]);
2635 static void __init hugetlb_sysfs_init(void)
2640 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2641 if (!hugepages_kobj)
2644 for_each_hstate(h) {
2645 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2646 hstate_kobjs, &hstate_attr_group);
2648 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2655 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2656 * with node devices in node_devices[] using a parallel array. The array
2657 * index of a node device or _hstate == node id.
2658 * This is here to avoid any static dependency of the node device driver, in
2659 * the base kernel, on the hugetlb module.
2661 struct node_hstate {
2662 struct kobject *hugepages_kobj;
2663 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2665 static struct node_hstate node_hstates[MAX_NUMNODES];
2668 * A subset of global hstate attributes for node devices
2670 static struct attribute *per_node_hstate_attrs[] = {
2671 &nr_hugepages_attr.attr,
2672 &free_hugepages_attr.attr,
2673 &surplus_hugepages_attr.attr,
2677 static const struct attribute_group per_node_hstate_attr_group = {
2678 .attrs = per_node_hstate_attrs,
2682 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2683 * Returns node id via non-NULL nidp.
2685 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2689 for (nid = 0; nid < nr_node_ids; nid++) {
2690 struct node_hstate *nhs = &node_hstates[nid];
2692 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2693 if (nhs->hstate_kobjs[i] == kobj) {
2705 * Unregister hstate attributes from a single node device.
2706 * No-op if no hstate attributes attached.
2708 static void hugetlb_unregister_node(struct node *node)
2711 struct node_hstate *nhs = &node_hstates[node->dev.id];
2713 if (!nhs->hugepages_kobj)
2714 return; /* no hstate attributes */
2716 for_each_hstate(h) {
2717 int idx = hstate_index(h);
2718 if (nhs->hstate_kobjs[idx]) {
2719 kobject_put(nhs->hstate_kobjs[idx]);
2720 nhs->hstate_kobjs[idx] = NULL;
2724 kobject_put(nhs->hugepages_kobj);
2725 nhs->hugepages_kobj = NULL;
2730 * Register hstate attributes for a single node device.
2731 * No-op if attributes already registered.
2733 static void hugetlb_register_node(struct node *node)
2736 struct node_hstate *nhs = &node_hstates[node->dev.id];
2739 if (nhs->hugepages_kobj)
2740 return; /* already allocated */
2742 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2744 if (!nhs->hugepages_kobj)
2747 for_each_hstate(h) {
2748 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2750 &per_node_hstate_attr_group);
2752 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2753 h->name, node->dev.id);
2754 hugetlb_unregister_node(node);
2761 * hugetlb init time: register hstate attributes for all registered node
2762 * devices of nodes that have memory. All on-line nodes should have
2763 * registered their associated device by this time.
2765 static void __init hugetlb_register_all_nodes(void)
2769 for_each_node_state(nid, N_MEMORY) {
2770 struct node *node = node_devices[nid];
2771 if (node->dev.id == nid)
2772 hugetlb_register_node(node);
2776 * Let the node device driver know we're here so it can
2777 * [un]register hstate attributes on node hotplug.
2779 register_hugetlbfs_with_node(hugetlb_register_node,
2780 hugetlb_unregister_node);
2782 #else /* !CONFIG_NUMA */
2784 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2792 static void hugetlb_register_all_nodes(void) { }
2796 static int __init hugetlb_init(void)
2800 if (!hugepages_supported())
2803 if (!size_to_hstate(default_hstate_size)) {
2804 if (default_hstate_size != 0) {
2805 pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2806 default_hstate_size, HPAGE_SIZE);
2809 default_hstate_size = HPAGE_SIZE;
2810 if (!size_to_hstate(default_hstate_size))
2811 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2813 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2814 if (default_hstate_max_huge_pages) {
2815 if (!default_hstate.max_huge_pages)
2816 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2819 hugetlb_init_hstates();
2820 gather_bootmem_prealloc();
2823 hugetlb_sysfs_init();
2824 hugetlb_register_all_nodes();
2825 hugetlb_cgroup_file_init();
2828 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2830 num_fault_mutexes = 1;
2832 hugetlb_fault_mutex_table =
2833 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
2835 BUG_ON(!hugetlb_fault_mutex_table);
2837 for (i = 0; i < num_fault_mutexes; i++)
2838 mutex_init(&hugetlb_fault_mutex_table[i]);
2841 subsys_initcall(hugetlb_init);
2843 /* Should be called on processing a hugepagesz=... option */
2844 void __init hugetlb_bad_size(void)
2846 parsed_valid_hugepagesz = false;
2849 void __init hugetlb_add_hstate(unsigned int order)
2854 if (size_to_hstate(PAGE_SIZE << order)) {
2855 pr_warn("hugepagesz= specified twice, ignoring\n");
2858 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2860 h = &hstates[hugetlb_max_hstate++];
2862 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2863 h->nr_huge_pages = 0;
2864 h->free_huge_pages = 0;
2865 for (i = 0; i < MAX_NUMNODES; ++i)
2866 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2867 INIT_LIST_HEAD(&h->hugepage_activelist);
2868 h->next_nid_to_alloc = first_memory_node;
2869 h->next_nid_to_free = first_memory_node;
2870 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2871 huge_page_size(h)/1024);
2876 static int __init hugetlb_nrpages_setup(char *s)
2879 static unsigned long *last_mhp;
2881 if (!parsed_valid_hugepagesz) {
2882 pr_warn("hugepages = %s preceded by "
2883 "an unsupported hugepagesz, ignoring\n", s);
2884 parsed_valid_hugepagesz = true;
2888 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2889 * so this hugepages= parameter goes to the "default hstate".
2891 else if (!hugetlb_max_hstate)
2892 mhp = &default_hstate_max_huge_pages;
2894 mhp = &parsed_hstate->max_huge_pages;
2896 if (mhp == last_mhp) {
2897 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2901 if (sscanf(s, "%lu", mhp) <= 0)
2905 * Global state is always initialized later in hugetlb_init.
2906 * But we need to allocate >= MAX_ORDER hstates here early to still
2907 * use the bootmem allocator.
2909 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2910 hugetlb_hstate_alloc_pages(parsed_hstate);
2916 __setup("hugepages=", hugetlb_nrpages_setup);
2918 static int __init hugetlb_default_setup(char *s)
2920 default_hstate_size = memparse(s, &s);
2923 __setup("default_hugepagesz=", hugetlb_default_setup);
2925 static unsigned int cpuset_mems_nr(unsigned int *array)
2928 unsigned int nr = 0;
2930 for_each_node_mask(node, cpuset_current_mems_allowed)
2936 #ifdef CONFIG_SYSCTL
2937 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2938 struct ctl_table *table, int write,
2939 void __user *buffer, size_t *length, loff_t *ppos)
2941 struct hstate *h = &default_hstate;
2942 unsigned long tmp = h->max_huge_pages;
2945 if (!hugepages_supported())
2949 table->maxlen = sizeof(unsigned long);
2950 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2955 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2956 NUMA_NO_NODE, tmp, *length);
2961 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2962 void __user *buffer, size_t *length, loff_t *ppos)
2965 return hugetlb_sysctl_handler_common(false, table, write,
2966 buffer, length, ppos);
2970 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2971 void __user *buffer, size_t *length, loff_t *ppos)
2973 return hugetlb_sysctl_handler_common(true, table, write,
2974 buffer, length, ppos);
2976 #endif /* CONFIG_NUMA */
2978 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2979 void __user *buffer,
2980 size_t *length, loff_t *ppos)
2982 struct hstate *h = &default_hstate;
2986 if (!hugepages_supported())
2989 tmp = h->nr_overcommit_huge_pages;
2991 if (write && hstate_is_gigantic(h))
2995 table->maxlen = sizeof(unsigned long);
2996 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3001 spin_lock(&hugetlb_lock);
3002 h->nr_overcommit_huge_pages = tmp;
3003 spin_unlock(&hugetlb_lock);
3009 #endif /* CONFIG_SYSCTL */
3011 void hugetlb_report_meminfo(struct seq_file *m)
3014 unsigned long total = 0;
3016 if (!hugepages_supported())
3019 for_each_hstate(h) {
3020 unsigned long count = h->nr_huge_pages;
3022 total += (PAGE_SIZE << huge_page_order(h)) * count;
3024 if (h == &default_hstate)
3026 "HugePages_Total: %5lu\n"
3027 "HugePages_Free: %5lu\n"
3028 "HugePages_Rsvd: %5lu\n"
3029 "HugePages_Surp: %5lu\n"
3030 "Hugepagesize: %8lu kB\n",
3034 h->surplus_huge_pages,
3035 (PAGE_SIZE << huge_page_order(h)) / 1024);
3038 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3041 int hugetlb_report_node_meminfo(int nid, char *buf)
3043 struct hstate *h = &default_hstate;
3044 if (!hugepages_supported())
3047 "Node %d HugePages_Total: %5u\n"
3048 "Node %d HugePages_Free: %5u\n"
3049 "Node %d HugePages_Surp: %5u\n",
3050 nid, h->nr_huge_pages_node[nid],
3051 nid, h->free_huge_pages_node[nid],
3052 nid, h->surplus_huge_pages_node[nid]);
3055 void hugetlb_show_meminfo(void)
3060 if (!hugepages_supported())
3063 for_each_node_state(nid, N_MEMORY)
3065 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3067 h->nr_huge_pages_node[nid],
3068 h->free_huge_pages_node[nid],
3069 h->surplus_huge_pages_node[nid],
3070 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3073 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3075 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3076 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3079 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3080 unsigned long hugetlb_total_pages(void)
3083 unsigned long nr_total_pages = 0;
3086 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3087 return nr_total_pages;
3090 static int hugetlb_acct_memory(struct hstate *h, long delta)
3094 spin_lock(&hugetlb_lock);
3096 * When cpuset is configured, it breaks the strict hugetlb page
3097 * reservation as the accounting is done on a global variable. Such
3098 * reservation is completely rubbish in the presence of cpuset because
3099 * the reservation is not checked against page availability for the
3100 * current cpuset. Application can still potentially OOM'ed by kernel
3101 * with lack of free htlb page in cpuset that the task is in.
3102 * Attempt to enforce strict accounting with cpuset is almost
3103 * impossible (or too ugly) because cpuset is too fluid that
3104 * task or memory node can be dynamically moved between cpusets.
3106 * The change of semantics for shared hugetlb mapping with cpuset is
3107 * undesirable. However, in order to preserve some of the semantics,
3108 * we fall back to check against current free page availability as
3109 * a best attempt and hopefully to minimize the impact of changing
3110 * semantics that cpuset has.
3113 if (gather_surplus_pages(h, delta) < 0)
3116 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3117 return_unused_surplus_pages(h, delta);
3124 return_unused_surplus_pages(h, (unsigned long) -delta);
3127 spin_unlock(&hugetlb_lock);
3131 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3133 struct resv_map *resv = vma_resv_map(vma);
3136 * This new VMA should share its siblings reservation map if present.
3137 * The VMA will only ever have a valid reservation map pointer where
3138 * it is being copied for another still existing VMA. As that VMA
3139 * has a reference to the reservation map it cannot disappear until
3140 * after this open call completes. It is therefore safe to take a
3141 * new reference here without additional locking.
3143 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3144 kref_get(&resv->refs);
3147 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3149 struct hstate *h = hstate_vma(vma);
3150 struct resv_map *resv = vma_resv_map(vma);
3151 struct hugepage_subpool *spool = subpool_vma(vma);
3152 unsigned long reserve, start, end;
3155 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3158 start = vma_hugecache_offset(h, vma, vma->vm_start);
3159 end = vma_hugecache_offset(h, vma, vma->vm_end);
3161 reserve = (end - start) - region_count(resv, start, end);
3163 kref_put(&resv->refs, resv_map_release);
3167 * Decrement reserve counts. The global reserve count may be
3168 * adjusted if the subpool has a minimum size.
3170 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3171 hugetlb_acct_memory(h, -gbl_reserve);
3175 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3177 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3182 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3184 struct hstate *hstate = hstate_vma(vma);
3186 return 1UL << huge_page_shift(hstate);
3190 * We cannot handle pagefaults against hugetlb pages at all. They cause
3191 * handle_mm_fault() to try to instantiate regular-sized pages in the
3192 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3195 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3202 * When a new function is introduced to vm_operations_struct and added
3203 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3204 * This is because under System V memory model, mappings created via
3205 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3206 * their original vm_ops are overwritten with shm_vm_ops.
3208 const struct vm_operations_struct hugetlb_vm_ops = {
3209 .fault = hugetlb_vm_op_fault,
3210 .open = hugetlb_vm_op_open,
3211 .close = hugetlb_vm_op_close,
3212 .split = hugetlb_vm_op_split,
3213 .pagesize = hugetlb_vm_op_pagesize,
3216 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3222 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3223 vma->vm_page_prot)));
3225 entry = huge_pte_wrprotect(mk_huge_pte(page,
3226 vma->vm_page_prot));
3228 entry = pte_mkyoung(entry);
3229 entry = pte_mkhuge(entry);
3230 entry = arch_make_huge_pte(entry, vma, page, writable);
3235 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3236 unsigned long address, pte_t *ptep)
3240 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3241 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3242 update_mmu_cache(vma, address, ptep);
3245 bool is_hugetlb_entry_migration(pte_t pte)
3249 if (huge_pte_none(pte) || pte_present(pte))
3251 swp = pte_to_swp_entry(pte);
3252 if (non_swap_entry(swp) && is_migration_entry(swp))
3258 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3262 if (huge_pte_none(pte) || pte_present(pte))
3264 swp = pte_to_swp_entry(pte);
3265 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3271 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3272 struct vm_area_struct *vma)
3274 pte_t *src_pte, *dst_pte, entry, dst_entry;
3275 struct page *ptepage;
3278 struct hstate *h = hstate_vma(vma);
3279 unsigned long sz = huge_page_size(h);
3280 struct mmu_notifier_range range;
3283 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3286 mmu_notifier_range_init(&range, src, vma->vm_start,
3288 mmu_notifier_invalidate_range_start(&range);
3291 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3292 spinlock_t *src_ptl, *dst_ptl;
3293 src_pte = huge_pte_offset(src, addr, sz);
3296 dst_pte = huge_pte_alloc(dst, addr, sz);
3303 * If the pagetables are shared don't copy or take references.
3304 * dst_pte == src_pte is the common case of src/dest sharing.
3306 * However, src could have 'unshared' and dst shares with
3307 * another vma. If dst_pte !none, this implies sharing.
3308 * Check here before taking page table lock, and once again
3309 * after taking the lock below.
3311 dst_entry = huge_ptep_get(dst_pte);
3312 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3315 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3316 src_ptl = huge_pte_lockptr(h, src, src_pte);
3317 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3318 entry = huge_ptep_get(src_pte);
3319 dst_entry = huge_ptep_get(dst_pte);
3320 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3322 * Skip if src entry none. Also, skip in the
3323 * unlikely case dst entry !none as this implies
3324 * sharing with another vma.
3327 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3328 is_hugetlb_entry_hwpoisoned(entry))) {
3329 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3331 if (is_write_migration_entry(swp_entry) && cow) {
3333 * COW mappings require pages in both
3334 * parent and child to be set to read.
3336 make_migration_entry_read(&swp_entry);
3337 entry = swp_entry_to_pte(swp_entry);
3338 set_huge_swap_pte_at(src, addr, src_pte,
3341 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3345 * No need to notify as we are downgrading page
3346 * table protection not changing it to point
3349 * See Documentation/vm/mmu_notifier.rst
3351 huge_ptep_set_wrprotect(src, addr, src_pte);
3353 entry = huge_ptep_get(src_pte);
3354 ptepage = pte_page(entry);
3356 page_dup_rmap(ptepage, true);
3357 set_huge_pte_at(dst, addr, dst_pte, entry);
3358 hugetlb_count_add(pages_per_huge_page(h), dst);
3360 spin_unlock(src_ptl);
3361 spin_unlock(dst_ptl);
3365 mmu_notifier_invalidate_range_end(&range);
3370 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3371 unsigned long start, unsigned long end,
3372 struct page *ref_page)
3374 struct mm_struct *mm = vma->vm_mm;
3375 unsigned long address;
3380 struct hstate *h = hstate_vma(vma);
3381 unsigned long sz = huge_page_size(h);
3382 struct mmu_notifier_range range;
3384 WARN_ON(!is_vm_hugetlb_page(vma));
3385 BUG_ON(start & ~huge_page_mask(h));
3386 BUG_ON(end & ~huge_page_mask(h));
3389 * This is a hugetlb vma, all the pte entries should point
3392 tlb_change_page_size(tlb, sz);
3393 tlb_start_vma(tlb, vma);
3396 * If sharing possible, alert mmu notifiers of worst case.
3398 mmu_notifier_range_init(&range, mm, start, end);
3399 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3400 mmu_notifier_invalidate_range_start(&range);
3402 for (; address < end; address += sz) {
3403 ptep = huge_pte_offset(mm, address, sz);
3407 ptl = huge_pte_lock(h, mm, ptep);
3408 if (huge_pmd_unshare(mm, &address, ptep)) {
3411 * We just unmapped a page of PMDs by clearing a PUD.
3412 * The caller's TLB flush range should cover this area.
3417 pte = huge_ptep_get(ptep);
3418 if (huge_pte_none(pte)) {
3424 * Migrating hugepage or HWPoisoned hugepage is already
3425 * unmapped and its refcount is dropped, so just clear pte here.
3427 if (unlikely(!pte_present(pte))) {
3428 huge_pte_clear(mm, address, ptep, sz);
3433 page = pte_page(pte);
3435 * If a reference page is supplied, it is because a specific
3436 * page is being unmapped, not a range. Ensure the page we
3437 * are about to unmap is the actual page of interest.
3440 if (page != ref_page) {
3445 * Mark the VMA as having unmapped its page so that
3446 * future faults in this VMA will fail rather than
3447 * looking like data was lost
3449 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3452 pte = huge_ptep_get_and_clear(mm, address, ptep);
3453 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3454 if (huge_pte_dirty(pte))
3455 set_page_dirty(page);
3457 hugetlb_count_sub(pages_per_huge_page(h), mm);
3458 page_remove_rmap(page, true);
3461 tlb_remove_page_size(tlb, page, huge_page_size(h));
3463 * Bail out after unmapping reference page if supplied
3468 mmu_notifier_invalidate_range_end(&range);
3469 tlb_end_vma(tlb, vma);
3472 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3473 struct vm_area_struct *vma, unsigned long start,
3474 unsigned long end, struct page *ref_page)
3476 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3479 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3480 * test will fail on a vma being torn down, and not grab a page table
3481 * on its way out. We're lucky that the flag has such an appropriate
3482 * name, and can in fact be safely cleared here. We could clear it
3483 * before the __unmap_hugepage_range above, but all that's necessary
3484 * is to clear it before releasing the i_mmap_rwsem. This works
3485 * because in the context this is called, the VMA is about to be
3486 * destroyed and the i_mmap_rwsem is held.
3488 vma->vm_flags &= ~VM_MAYSHARE;
3491 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3492 unsigned long end, struct page *ref_page)
3494 struct mm_struct *mm;
3495 struct mmu_gather tlb;
3496 unsigned long tlb_start = start;
3497 unsigned long tlb_end = end;
3500 * If shared PMDs were possibly used within this vma range, adjust
3501 * start/end for worst case tlb flushing.
3502 * Note that we can not be sure if PMDs are shared until we try to
3503 * unmap pages. However, we want to make sure TLB flushing covers
3504 * the largest possible range.
3506 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
3510 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
3511 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3512 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
3516 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3517 * mappping it owns the reserve page for. The intention is to unmap the page
3518 * from other VMAs and let the children be SIGKILLed if they are faulting the
3521 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3522 struct page *page, unsigned long address)
3524 struct hstate *h = hstate_vma(vma);
3525 struct vm_area_struct *iter_vma;
3526 struct address_space *mapping;
3530 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3531 * from page cache lookup which is in HPAGE_SIZE units.
3533 address = address & huge_page_mask(h);
3534 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3536 mapping = vma->vm_file->f_mapping;
3539 * Take the mapping lock for the duration of the table walk. As
3540 * this mapping should be shared between all the VMAs,
3541 * __unmap_hugepage_range() is called as the lock is already held
3543 i_mmap_lock_write(mapping);
3544 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3545 /* Do not unmap the current VMA */
3546 if (iter_vma == vma)
3550 * Shared VMAs have their own reserves and do not affect
3551 * MAP_PRIVATE accounting but it is possible that a shared
3552 * VMA is using the same page so check and skip such VMAs.
3554 if (iter_vma->vm_flags & VM_MAYSHARE)
3558 * Unmap the page from other VMAs without their own reserves.
3559 * They get marked to be SIGKILLed if they fault in these
3560 * areas. This is because a future no-page fault on this VMA
3561 * could insert a zeroed page instead of the data existing
3562 * from the time of fork. This would look like data corruption
3564 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3565 unmap_hugepage_range(iter_vma, address,
3566 address + huge_page_size(h), page);
3568 i_mmap_unlock_write(mapping);
3572 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3573 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3574 * cannot race with other handlers or page migration.
3575 * Keep the pte_same checks anyway to make transition from the mutex easier.
3577 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3578 unsigned long address, pte_t *ptep,
3579 struct page *pagecache_page, spinlock_t *ptl)
3582 struct hstate *h = hstate_vma(vma);
3583 struct page *old_page, *new_page;
3584 int outside_reserve = 0;
3586 unsigned long haddr = address & huge_page_mask(h);
3587 struct mmu_notifier_range range;
3589 pte = huge_ptep_get(ptep);
3590 old_page = pte_page(pte);
3593 /* If no-one else is actually using this page, avoid the copy
3594 * and just make the page writable */
3595 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3596 page_move_anon_rmap(old_page, vma);
3597 set_huge_ptep_writable(vma, haddr, ptep);
3602 * If the process that created a MAP_PRIVATE mapping is about to
3603 * perform a COW due to a shared page count, attempt to satisfy
3604 * the allocation without using the existing reserves. The pagecache
3605 * page is used to determine if the reserve at this address was
3606 * consumed or not. If reserves were used, a partial faulted mapping
3607 * at the time of fork() could consume its reserves on COW instead
3608 * of the full address range.
3610 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3611 old_page != pagecache_page)
3612 outside_reserve = 1;
3617 * Drop page table lock as buddy allocator may be called. It will
3618 * be acquired again before returning to the caller, as expected.
3621 new_page = alloc_huge_page(vma, haddr, outside_reserve);
3623 if (IS_ERR(new_page)) {
3625 * If a process owning a MAP_PRIVATE mapping fails to COW,
3626 * it is due to references held by a child and an insufficient
3627 * huge page pool. To guarantee the original mappers
3628 * reliability, unmap the page from child processes. The child
3629 * may get SIGKILLed if it later faults.
3631 if (outside_reserve) {
3633 BUG_ON(huge_pte_none(pte));
3634 unmap_ref_private(mm, vma, old_page, haddr);
3635 BUG_ON(huge_pte_none(pte));
3637 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3639 pte_same(huge_ptep_get(ptep), pte)))
3640 goto retry_avoidcopy;
3642 * race occurs while re-acquiring page table
3643 * lock, and our job is done.
3648 ret = vmf_error(PTR_ERR(new_page));
3649 goto out_release_old;
3653 * When the original hugepage is shared one, it does not have
3654 * anon_vma prepared.
3656 if (unlikely(anon_vma_prepare(vma))) {
3658 goto out_release_all;
3661 copy_user_huge_page(new_page, old_page, address, vma,
3662 pages_per_huge_page(h));
3663 __SetPageUptodate(new_page);
3665 mmu_notifier_range_init(&range, mm, haddr, haddr + huge_page_size(h));
3666 mmu_notifier_invalidate_range_start(&range);
3669 * Retake the page table lock to check for racing updates
3670 * before the page tables are altered
3673 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3674 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3675 ClearPagePrivate(new_page);
3678 huge_ptep_clear_flush(vma, haddr, ptep);
3679 mmu_notifier_invalidate_range(mm, range.start, range.end);
3680 set_huge_pte_at(mm, haddr, ptep,
3681 make_huge_pte(vma, new_page, 1));
3682 page_remove_rmap(old_page, true);
3683 hugepage_add_new_anon_rmap(new_page, vma, haddr);
3684 set_page_huge_active(new_page);
3685 /* Make the old page be freed below */
3686 new_page = old_page;
3689 mmu_notifier_invalidate_range_end(&range);
3691 restore_reserve_on_error(h, vma, haddr, new_page);
3696 spin_lock(ptl); /* Caller expects lock to be held */
3700 /* Return the pagecache page at a given address within a VMA */
3701 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3702 struct vm_area_struct *vma, unsigned long address)
3704 struct address_space *mapping;
3707 mapping = vma->vm_file->f_mapping;
3708 idx = vma_hugecache_offset(h, vma, address);
3710 return find_lock_page(mapping, idx);
3714 * Return whether there is a pagecache page to back given address within VMA.
3715 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3717 static bool hugetlbfs_pagecache_present(struct hstate *h,
3718 struct vm_area_struct *vma, unsigned long address)
3720 struct address_space *mapping;
3724 mapping = vma->vm_file->f_mapping;
3725 idx = vma_hugecache_offset(h, vma, address);
3727 page = find_get_page(mapping, idx);
3730 return page != NULL;
3733 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3736 struct inode *inode = mapping->host;
3737 struct hstate *h = hstate_inode(inode);
3738 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3742 ClearPagePrivate(page);
3745 * set page dirty so that it will not be removed from cache/file
3746 * by non-hugetlbfs specific code paths.
3748 set_page_dirty(page);
3750 spin_lock(&inode->i_lock);
3751 inode->i_blocks += blocks_per_huge_page(h);
3752 spin_unlock(&inode->i_lock);
3756 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
3757 struct vm_area_struct *vma,
3758 struct address_space *mapping, pgoff_t idx,
3759 unsigned long address, pte_t *ptep, unsigned int flags)
3761 struct hstate *h = hstate_vma(vma);
3762 vm_fault_t ret = VM_FAULT_SIGBUS;
3768 unsigned long haddr = address & huge_page_mask(h);
3769 bool new_page = false;
3772 * Currently, we are forced to kill the process in the event the
3773 * original mapper has unmapped pages from the child due to a failed
3774 * COW. Warn that such a situation has occurred as it may not be obvious
3776 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3777 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3783 * Use page lock to guard against racing truncation
3784 * before we get page_table_lock.
3787 page = find_lock_page(mapping, idx);
3789 size = i_size_read(mapping->host) >> huge_page_shift(h);
3794 * Check for page in userfault range
3796 if (userfaultfd_missing(vma)) {
3798 struct vm_fault vmf = {
3803 * Hard to debug if it ends up being
3804 * used by a callee that assumes
3805 * something about the other
3806 * uninitialized fields... same as in
3812 * hugetlb_fault_mutex must be dropped before
3813 * handling userfault. Reacquire after handling
3814 * fault to make calling code simpler.
3816 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping,
3818 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3819 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3820 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3824 page = alloc_huge_page(vma, haddr, 0);
3826 ret = vmf_error(PTR_ERR(page));
3829 clear_huge_page(page, address, pages_per_huge_page(h));
3830 __SetPageUptodate(page);
3833 if (vma->vm_flags & VM_MAYSHARE) {
3834 int err = huge_add_to_page_cache(page, mapping, idx);
3843 if (unlikely(anon_vma_prepare(vma))) {
3845 goto backout_unlocked;
3851 * If memory error occurs between mmap() and fault, some process
3852 * don't have hwpoisoned swap entry for errored virtual address.
3853 * So we need to block hugepage fault by PG_hwpoison bit check.
3855 if (unlikely(PageHWPoison(page))) {
3856 ret = VM_FAULT_HWPOISON |
3857 VM_FAULT_SET_HINDEX(hstate_index(h));
3858 goto backout_unlocked;
3863 * If we are going to COW a private mapping later, we examine the
3864 * pending reservations for this page now. This will ensure that
3865 * any allocations necessary to record that reservation occur outside
3868 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3869 if (vma_needs_reservation(h, vma, haddr) < 0) {
3871 goto backout_unlocked;
3873 /* Just decrements count, does not deallocate */
3874 vma_end_reservation(h, vma, haddr);
3877 ptl = huge_pte_lock(h, mm, ptep);
3878 size = i_size_read(mapping->host) >> huge_page_shift(h);
3883 if (!huge_pte_none(huge_ptep_get(ptep)))
3887 ClearPagePrivate(page);
3888 hugepage_add_new_anon_rmap(page, vma, haddr);
3890 page_dup_rmap(page, true);
3891 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3892 && (vma->vm_flags & VM_SHARED)));
3893 set_huge_pte_at(mm, haddr, ptep, new_pte);
3895 hugetlb_count_add(pages_per_huge_page(h), mm);
3896 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3897 /* Optimization, do the COW without a second fault */
3898 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
3904 * Only make newly allocated pages active. Existing pages found
3905 * in the pagecache could be !page_huge_active() if they have been
3906 * isolated for migration.
3909 set_page_huge_active(page);
3919 restore_reserve_on_error(h, vma, haddr, page);
3925 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3926 struct vm_area_struct *vma,
3927 struct address_space *mapping,
3928 pgoff_t idx, unsigned long address)
3930 unsigned long key[2];
3933 if (vma->vm_flags & VM_SHARED) {
3934 key[0] = (unsigned long) mapping;
3937 key[0] = (unsigned long) mm;
3938 key[1] = address >> huge_page_shift(h);
3941 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3943 return hash & (num_fault_mutexes - 1);
3947 * For uniprocesor systems we always use a single mutex, so just
3948 * return 0 and avoid the hashing overhead.
3950 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3951 struct vm_area_struct *vma,
3952 struct address_space *mapping,
3953 pgoff_t idx, unsigned long address)
3959 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3960 unsigned long address, unsigned int flags)
3967 struct page *page = NULL;
3968 struct page *pagecache_page = NULL;
3969 struct hstate *h = hstate_vma(vma);
3970 struct address_space *mapping;
3971 int need_wait_lock = 0;
3972 unsigned long haddr = address & huge_page_mask(h);
3974 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3976 entry = huge_ptep_get(ptep);
3977 if (unlikely(is_hugetlb_entry_migration(entry))) {
3978 migration_entry_wait_huge(vma, mm, ptep);
3980 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3981 return VM_FAULT_HWPOISON_LARGE |
3982 VM_FAULT_SET_HINDEX(hstate_index(h));
3984 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
3986 return VM_FAULT_OOM;
3989 mapping = vma->vm_file->f_mapping;
3990 idx = vma_hugecache_offset(h, vma, haddr);
3993 * Serialize hugepage allocation and instantiation, so that we don't
3994 * get spurious allocation failures if two CPUs race to instantiate
3995 * the same page in the page cache.
3997 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, haddr);
3998 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4000 entry = huge_ptep_get(ptep);
4001 if (huge_pte_none(entry)) {
4002 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4009 * entry could be a migration/hwpoison entry at this point, so this
4010 * check prevents the kernel from going below assuming that we have
4011 * a active hugepage in pagecache. This goto expects the 2nd page fault,
4012 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
4015 if (!pte_present(entry))
4019 * If we are going to COW the mapping later, we examine the pending
4020 * reservations for this page now. This will ensure that any
4021 * allocations necessary to record that reservation occur outside the
4022 * spinlock. For private mappings, we also lookup the pagecache
4023 * page now as it is used to determine if a reservation has been
4026 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4027 if (vma_needs_reservation(h, vma, haddr) < 0) {
4031 /* Just decrements count, does not deallocate */
4032 vma_end_reservation(h, vma, haddr);
4034 if (!(vma->vm_flags & VM_MAYSHARE))
4035 pagecache_page = hugetlbfs_pagecache_page(h,
4039 ptl = huge_pte_lock(h, mm, ptep);
4041 /* Check for a racing update before calling hugetlb_cow */
4042 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4046 * hugetlb_cow() requires page locks of pte_page(entry) and
4047 * pagecache_page, so here we need take the former one
4048 * when page != pagecache_page or !pagecache_page.
4050 page = pte_page(entry);
4051 if (page != pagecache_page)
4052 if (!trylock_page(page)) {
4059 if (flags & FAULT_FLAG_WRITE) {
4060 if (!huge_pte_write(entry)) {
4061 ret = hugetlb_cow(mm, vma, address, ptep,
4062 pagecache_page, ptl);
4065 entry = huge_pte_mkdirty(entry);
4067 entry = pte_mkyoung(entry);
4068 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4069 flags & FAULT_FLAG_WRITE))
4070 update_mmu_cache(vma, haddr, ptep);
4072 if (page != pagecache_page)
4078 if (pagecache_page) {
4079 unlock_page(pagecache_page);
4080 put_page(pagecache_page);
4083 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4085 * Generally it's safe to hold refcount during waiting page lock. But
4086 * here we just wait to defer the next page fault to avoid busy loop and
4087 * the page is not used after unlocked before returning from the current
4088 * page fault. So we are safe from accessing freed page, even if we wait
4089 * here without taking refcount.
4092 wait_on_page_locked(page);
4097 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4098 * modifications for huge pages.
4100 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4102 struct vm_area_struct *dst_vma,
4103 unsigned long dst_addr,
4104 unsigned long src_addr,
4105 struct page **pagep)
4107 struct address_space *mapping;
4110 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4111 struct hstate *h = hstate_vma(dst_vma);
4119 page = alloc_huge_page(dst_vma, dst_addr, 0);
4123 ret = copy_huge_page_from_user(page,
4124 (const void __user *) src_addr,
4125 pages_per_huge_page(h), false);
4127 /* fallback to copy_from_user outside mmap_sem */
4128 if (unlikely(ret)) {
4131 /* don't free the page */
4140 * The memory barrier inside __SetPageUptodate makes sure that
4141 * preceding stores to the page contents become visible before
4142 * the set_pte_at() write.
4144 __SetPageUptodate(page);
4146 mapping = dst_vma->vm_file->f_mapping;
4147 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4150 * If shared, add to page cache
4153 size = i_size_read(mapping->host) >> huge_page_shift(h);
4156 goto out_release_nounlock;
4159 * Serialization between remove_inode_hugepages() and
4160 * huge_add_to_page_cache() below happens through the
4161 * hugetlb_fault_mutex_table that here must be hold by
4164 ret = huge_add_to_page_cache(page, mapping, idx);
4166 goto out_release_nounlock;
4169 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4173 * Recheck the i_size after holding PT lock to make sure not
4174 * to leave any page mapped (as page_mapped()) beyond the end
4175 * of the i_size (remove_inode_hugepages() is strict about
4176 * enforcing that). If we bail out here, we'll also leave a
4177 * page in the radix tree in the vm_shared case beyond the end
4178 * of the i_size, but remove_inode_hugepages() will take care
4179 * of it as soon as we drop the hugetlb_fault_mutex_table.
4181 size = i_size_read(mapping->host) >> huge_page_shift(h);
4184 goto out_release_unlock;
4187 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4188 goto out_release_unlock;
4191 page_dup_rmap(page, true);
4193 ClearPagePrivate(page);
4194 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4197 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4198 if (dst_vma->vm_flags & VM_WRITE)
4199 _dst_pte = huge_pte_mkdirty(_dst_pte);
4200 _dst_pte = pte_mkyoung(_dst_pte);
4202 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4204 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4205 dst_vma->vm_flags & VM_WRITE);
4206 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4208 /* No need to invalidate - it was non-present before */
4209 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4212 set_page_huge_active(page);
4222 out_release_nounlock:
4227 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4228 struct page **pages, struct vm_area_struct **vmas,
4229 unsigned long *position, unsigned long *nr_pages,
4230 long i, unsigned int flags, int *nonblocking)
4232 unsigned long pfn_offset;
4233 unsigned long vaddr = *position;
4234 unsigned long remainder = *nr_pages;
4235 struct hstate *h = hstate_vma(vma);
4238 while (vaddr < vma->vm_end && remainder) {
4240 spinlock_t *ptl = NULL;
4245 * If we have a pending SIGKILL, don't keep faulting pages and
4246 * potentially allocating memory.
4248 if (fatal_signal_pending(current)) {
4254 * Some archs (sparc64, sh*) have multiple pte_ts to
4255 * each hugepage. We have to make sure we get the
4256 * first, for the page indexing below to work.
4258 * Note that page table lock is not held when pte is null.
4260 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4263 ptl = huge_pte_lock(h, mm, pte);
4264 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4267 * When coredumping, it suits get_dump_page if we just return
4268 * an error where there's an empty slot with no huge pagecache
4269 * to back it. This way, we avoid allocating a hugepage, and
4270 * the sparse dumpfile avoids allocating disk blocks, but its
4271 * huge holes still show up with zeroes where they need to be.
4273 if (absent && (flags & FOLL_DUMP) &&
4274 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4282 * We need call hugetlb_fault for both hugepages under migration
4283 * (in which case hugetlb_fault waits for the migration,) and
4284 * hwpoisoned hugepages (in which case we need to prevent the
4285 * caller from accessing to them.) In order to do this, we use
4286 * here is_swap_pte instead of is_hugetlb_entry_migration and
4287 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4288 * both cases, and because we can't follow correct pages
4289 * directly from any kind of swap entries.
4291 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4292 ((flags & FOLL_WRITE) &&
4293 !huge_pte_write(huge_ptep_get(pte)))) {
4295 unsigned int fault_flags = 0;
4299 if (flags & FOLL_WRITE)
4300 fault_flags |= FAULT_FLAG_WRITE;
4302 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4303 if (flags & FOLL_NOWAIT)
4304 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4305 FAULT_FLAG_RETRY_NOWAIT;
4306 if (flags & FOLL_TRIED) {
4307 VM_WARN_ON_ONCE(fault_flags &
4308 FAULT_FLAG_ALLOW_RETRY);
4309 fault_flags |= FAULT_FLAG_TRIED;
4311 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4312 if (ret & VM_FAULT_ERROR) {
4313 err = vm_fault_to_errno(ret, flags);
4317 if (ret & VM_FAULT_RETRY) {
4319 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4323 * VM_FAULT_RETRY must not return an
4324 * error, it will return zero
4327 * No need to update "position" as the
4328 * caller will not check it after
4329 * *nr_pages is set to 0.
4336 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4337 page = pte_page(huge_ptep_get(pte));
4340 * Instead of doing 'try_get_page()' below in the same_page
4341 * loop, just check the count once here.
4343 if (unlikely(page_count(page) <= 0)) {
4353 pages[i] = mem_map_offset(page, pfn_offset);
4364 if (vaddr < vma->vm_end && remainder &&
4365 pfn_offset < pages_per_huge_page(h)) {
4367 * We use pfn_offset to avoid touching the pageframes
4368 * of this compound page.
4374 *nr_pages = remainder;
4376 * setting position is actually required only if remainder is
4377 * not zero but it's faster not to add a "if (remainder)"
4385 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4387 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4390 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4393 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4394 unsigned long address, unsigned long end, pgprot_t newprot)
4396 struct mm_struct *mm = vma->vm_mm;
4397 unsigned long start = address;
4400 struct hstate *h = hstate_vma(vma);
4401 unsigned long pages = 0;
4402 bool shared_pmd = false;
4403 struct mmu_notifier_range range;
4406 * In the case of shared PMDs, the area to flush could be beyond
4407 * start/end. Set range.start/range.end to cover the maximum possible
4408 * range if PMD sharing is possible.
4410 mmu_notifier_range_init(&range, mm, start, end);
4411 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4413 BUG_ON(address >= end);
4414 flush_cache_range(vma, range.start, range.end);
4416 mmu_notifier_invalidate_range_start(&range);
4417 i_mmap_lock_write(vma->vm_file->f_mapping);
4418 for (; address < end; address += huge_page_size(h)) {
4420 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4423 ptl = huge_pte_lock(h, mm, ptep);
4424 if (huge_pmd_unshare(mm, &address, ptep)) {
4430 pte = huge_ptep_get(ptep);
4431 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4435 if (unlikely(is_hugetlb_entry_migration(pte))) {
4436 swp_entry_t entry = pte_to_swp_entry(pte);
4438 if (is_write_migration_entry(entry)) {
4441 make_migration_entry_read(&entry);
4442 newpte = swp_entry_to_pte(entry);
4443 set_huge_swap_pte_at(mm, address, ptep,
4444 newpte, huge_page_size(h));
4450 if (!huge_pte_none(pte)) {
4453 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
4454 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
4455 pte = arch_make_huge_pte(pte, vma, NULL, 0);
4456 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
4462 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4463 * may have cleared our pud entry and done put_page on the page table:
4464 * once we release i_mmap_rwsem, another task can do the final put_page
4465 * and that page table be reused and filled with junk. If we actually
4466 * did unshare a page of pmds, flush the range corresponding to the pud.
4469 flush_hugetlb_tlb_range(vma, range.start, range.end);
4471 flush_hugetlb_tlb_range(vma, start, end);
4473 * No need to call mmu_notifier_invalidate_range() we are downgrading
4474 * page table protection not changing it to point to a new page.
4476 * See Documentation/vm/mmu_notifier.rst
4478 i_mmap_unlock_write(vma->vm_file->f_mapping);
4479 mmu_notifier_invalidate_range_end(&range);
4481 return pages << h->order;
4484 int hugetlb_reserve_pages(struct inode *inode,
4486 struct vm_area_struct *vma,
4487 vm_flags_t vm_flags)
4490 struct hstate *h = hstate_inode(inode);
4491 struct hugepage_subpool *spool = subpool_inode(inode);
4492 struct resv_map *resv_map;
4495 /* This should never happen */
4497 VM_WARN(1, "%s called with a negative range\n", __func__);
4502 * Only apply hugepage reservation if asked. At fault time, an
4503 * attempt will be made for VM_NORESERVE to allocate a page
4504 * without using reserves
4506 if (vm_flags & VM_NORESERVE)
4510 * Shared mappings base their reservation on the number of pages that
4511 * are already allocated on behalf of the file. Private mappings need
4512 * to reserve the full area even if read-only as mprotect() may be
4513 * called to make the mapping read-write. Assume !vma is a shm mapping
4515 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4516 resv_map = inode_resv_map(inode);
4518 chg = region_chg(resv_map, from, to);
4521 resv_map = resv_map_alloc();
4527 set_vma_resv_map(vma, resv_map);
4528 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4537 * There must be enough pages in the subpool for the mapping. If
4538 * the subpool has a minimum size, there may be some global
4539 * reservations already in place (gbl_reserve).
4541 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4542 if (gbl_reserve < 0) {
4548 * Check enough hugepages are available for the reservation.
4549 * Hand the pages back to the subpool if there are not
4551 ret = hugetlb_acct_memory(h, gbl_reserve);
4553 /* put back original number of pages, chg */
4554 (void)hugepage_subpool_put_pages(spool, chg);
4559 * Account for the reservations made. Shared mappings record regions
4560 * that have reservations as they are shared by multiple VMAs.
4561 * When the last VMA disappears, the region map says how much
4562 * the reservation was and the page cache tells how much of
4563 * the reservation was consumed. Private mappings are per-VMA and
4564 * only the consumed reservations are tracked. When the VMA
4565 * disappears, the original reservation is the VMA size and the
4566 * consumed reservations are stored in the map. Hence, nothing
4567 * else has to be done for private mappings here
4569 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4570 long add = region_add(resv_map, from, to);
4572 if (unlikely(chg > add)) {
4574 * pages in this range were added to the reserve
4575 * map between region_chg and region_add. This
4576 * indicates a race with alloc_huge_page. Adjust
4577 * the subpool and reserve counts modified above
4578 * based on the difference.
4582 rsv_adjust = hugepage_subpool_put_pages(spool,
4584 hugetlb_acct_memory(h, -rsv_adjust);
4589 if (!vma || vma->vm_flags & VM_MAYSHARE)
4590 /* Don't call region_abort if region_chg failed */
4592 region_abort(resv_map, from, to);
4593 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4594 kref_put(&resv_map->refs, resv_map_release);
4598 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4601 struct hstate *h = hstate_inode(inode);
4602 struct resv_map *resv_map = inode_resv_map(inode);
4604 struct hugepage_subpool *spool = subpool_inode(inode);
4608 chg = region_del(resv_map, start, end);
4610 * region_del() can fail in the rare case where a region
4611 * must be split and another region descriptor can not be
4612 * allocated. If end == LONG_MAX, it will not fail.
4618 spin_lock(&inode->i_lock);
4619 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4620 spin_unlock(&inode->i_lock);
4623 * If the subpool has a minimum size, the number of global
4624 * reservations to be released may be adjusted.
4626 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4627 hugetlb_acct_memory(h, -gbl_reserve);
4632 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4633 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4634 struct vm_area_struct *vma,
4635 unsigned long addr, pgoff_t idx)
4637 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4639 unsigned long sbase = saddr & PUD_MASK;
4640 unsigned long s_end = sbase + PUD_SIZE;
4642 /* Allow segments to share if only one is marked locked */
4643 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4644 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4647 * match the virtual addresses, permission and the alignment of the
4650 if (pmd_index(addr) != pmd_index(saddr) ||
4651 vm_flags != svm_flags ||
4652 sbase < svma->vm_start || svma->vm_end < s_end)
4658 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4660 unsigned long base = addr & PUD_MASK;
4661 unsigned long end = base + PUD_SIZE;
4664 * check on proper vm_flags and page table alignment
4666 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
4672 * Determine if start,end range within vma could be mapped by shared pmd.
4673 * If yes, adjust start and end to cover range associated with possible
4674 * shared pmd mappings.
4676 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4677 unsigned long *start, unsigned long *end)
4679 unsigned long check_addr = *start;
4681 if (!(vma->vm_flags & VM_MAYSHARE))
4684 for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
4685 unsigned long a_start = check_addr & PUD_MASK;
4686 unsigned long a_end = a_start + PUD_SIZE;
4689 * If sharing is possible, adjust start/end if necessary.
4691 if (range_in_vma(vma, a_start, a_end)) {
4692 if (a_start < *start)
4701 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4702 * and returns the corresponding pte. While this is not necessary for the
4703 * !shared pmd case because we can allocate the pmd later as well, it makes the
4704 * code much cleaner. pmd allocation is essential for the shared case because
4705 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4706 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4707 * bad pmd for sharing.
4709 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4711 struct vm_area_struct *vma = find_vma(mm, addr);
4712 struct address_space *mapping = vma->vm_file->f_mapping;
4713 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4715 struct vm_area_struct *svma;
4716 unsigned long saddr;
4721 if (!vma_shareable(vma, addr))
4722 return (pte_t *)pmd_alloc(mm, pud, addr);
4724 i_mmap_lock_write(mapping);
4725 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4729 saddr = page_table_shareable(svma, vma, addr, idx);
4731 spte = huge_pte_offset(svma->vm_mm, saddr,
4732 vma_mmu_pagesize(svma));
4734 get_page(virt_to_page(spte));
4743 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4744 if (pud_none(*pud)) {
4745 pud_populate(mm, pud,
4746 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4749 put_page(virt_to_page(spte));
4753 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4754 i_mmap_unlock_write(mapping);
4759 * unmap huge page backed by shared pte.
4761 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4762 * indicated by page_count > 1, unmap is achieved by clearing pud and
4763 * decrementing the ref count. If count == 1, the pte page is not shared.
4765 * called with page table lock held.
4767 * returns: 1 successfully unmapped a shared pte page
4768 * 0 the underlying pte page is not shared, or it is the last user
4770 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4772 pgd_t *pgd = pgd_offset(mm, *addr);
4773 p4d_t *p4d = p4d_offset(pgd, *addr);
4774 pud_t *pud = pud_offset(p4d, *addr);
4776 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4777 if (page_count(virt_to_page(ptep)) == 1)
4781 put_page(virt_to_page(ptep));
4783 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4786 #define want_pmd_share() (1)
4787 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4788 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4793 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4798 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4799 unsigned long *start, unsigned long *end)
4802 #define want_pmd_share() (0)
4803 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4805 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4806 pte_t *huge_pte_alloc(struct mm_struct *mm,
4807 unsigned long addr, unsigned long sz)
4814 pgd = pgd_offset(mm, addr);
4815 p4d = p4d_alloc(mm, pgd, addr);
4818 pud = pud_alloc(mm, p4d, addr);
4820 if (sz == PUD_SIZE) {
4823 BUG_ON(sz != PMD_SIZE);
4824 if (want_pmd_share() && pud_none(*pud))
4825 pte = huge_pmd_share(mm, addr, pud);
4827 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4830 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4836 * huge_pte_offset() - Walk the page table to resolve the hugepage
4837 * entry at address @addr
4839 * Return: Pointer to page table or swap entry (PUD or PMD) for
4840 * address @addr, or NULL if a p*d_none() entry is encountered and the
4841 * size @sz doesn't match the hugepage size at this level of the page
4844 pte_t *huge_pte_offset(struct mm_struct *mm,
4845 unsigned long addr, unsigned long sz)
4852 pgd = pgd_offset(mm, addr);
4853 if (!pgd_present(*pgd))
4855 p4d = p4d_offset(pgd, addr);
4856 if (!p4d_present(*p4d))
4859 pud = pud_offset(p4d, addr);
4860 if (sz != PUD_SIZE && pud_none(*pud))
4862 /* hugepage or swap? */
4863 if (pud_huge(*pud) || !pud_present(*pud))
4864 return (pte_t *)pud;
4866 pmd = pmd_offset(pud, addr);
4867 if (sz != PMD_SIZE && pmd_none(*pmd))
4869 /* hugepage or swap? */
4870 if (pmd_huge(*pmd) || !pmd_present(*pmd))
4871 return (pte_t *)pmd;
4876 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4879 * These functions are overwritable if your architecture needs its own
4882 struct page * __weak
4883 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4886 return ERR_PTR(-EINVAL);
4889 struct page * __weak
4890 follow_huge_pd(struct vm_area_struct *vma,
4891 unsigned long address, hugepd_t hpd, int flags, int pdshift)
4893 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4897 struct page * __weak
4898 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4899 pmd_t *pmd, int flags)
4901 struct page *page = NULL;
4905 ptl = pmd_lockptr(mm, pmd);
4908 * make sure that the address range covered by this pmd is not
4909 * unmapped from other threads.
4911 if (!pmd_huge(*pmd))
4913 pte = huge_ptep_get((pte_t *)pmd);
4914 if (pte_present(pte)) {
4915 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4916 if (flags & FOLL_GET)
4919 if (is_hugetlb_entry_migration(pte)) {
4921 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4925 * hwpoisoned entry is treated as no_page_table in
4926 * follow_page_mask().
4934 struct page * __weak
4935 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4936 pud_t *pud, int flags)
4938 if (flags & FOLL_GET)
4941 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4944 struct page * __weak
4945 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
4947 if (flags & FOLL_GET)
4950 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
4953 bool isolate_huge_page(struct page *page, struct list_head *list)
4957 VM_BUG_ON_PAGE(!PageHead(page), page);
4958 spin_lock(&hugetlb_lock);
4959 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4963 clear_page_huge_active(page);
4964 list_move_tail(&page->lru, list);
4966 spin_unlock(&hugetlb_lock);
4970 void putback_active_hugepage(struct page *page)
4972 VM_BUG_ON_PAGE(!PageHead(page), page);
4973 spin_lock(&hugetlb_lock);
4974 set_page_huge_active(page);
4975 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4976 spin_unlock(&hugetlb_lock);
4980 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
4982 struct hstate *h = page_hstate(oldpage);
4984 hugetlb_cgroup_migrate(oldpage, newpage);
4985 set_page_owner_migrate_reason(newpage, reason);
4988 * transfer temporary state of the new huge page. This is
4989 * reverse to other transitions because the newpage is going to
4990 * be final while the old one will be freed so it takes over
4991 * the temporary status.
4993 * Also note that we have to transfer the per-node surplus state
4994 * here as well otherwise the global surplus count will not match
4997 if (PageHugeTemporary(newpage)) {
4998 int old_nid = page_to_nid(oldpage);
4999 int new_nid = page_to_nid(newpage);
5001 SetPageHugeTemporary(oldpage);
5002 ClearPageHugeTemporary(newpage);
5004 spin_lock(&hugetlb_lock);
5005 if (h->surplus_huge_pages_node[old_nid]) {
5006 h->surplus_huge_pages_node[old_nid]--;
5007 h->surplus_huge_pages_node[new_nid]++;
5009 spin_unlock(&hugetlb_lock);