2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
8 #include <linux/seq_file.h>
9 #include <linux/sysctl.h>
10 #include <linux/highmem.h>
11 #include <linux/mmu_notifier.h>
12 #include <linux/nodemask.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/compiler.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/mmdebug.h>
22 #include <linux/sched/signal.h>
23 #include <linux/rmap.h>
24 #include <linux/string_helpers.h>
25 #include <linux/swap.h>
26 #include <linux/swapops.h>
27 #include <linux/jhash.h>
30 #include <asm/pgtable.h>
34 #include <linux/hugetlb.h>
35 #include <linux/hugetlb_cgroup.h>
36 #include <linux/node.h>
37 #include <linux/userfaultfd_k.h>
40 int hugepages_treat_as_movable;
42 int hugetlb_max_hstate __read_mostly;
43 unsigned int default_hstate_idx;
44 struct hstate hstates[HUGE_MAX_HSTATE];
46 * Minimum page order among possible hugepage sizes, set to a proper value
49 static unsigned int minimum_order __read_mostly = UINT_MAX;
51 __initdata LIST_HEAD(huge_boot_pages);
53 /* for command line parsing */
54 static struct hstate * __initdata parsed_hstate;
55 static unsigned long __initdata default_hstate_max_huge_pages;
56 static unsigned long __initdata default_hstate_size;
57 static bool __initdata parsed_valid_hugepagesz = true;
60 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
61 * free_huge_pages, and surplus_huge_pages.
63 DEFINE_SPINLOCK(hugetlb_lock);
66 * Serializes faults on the same logical page. This is used to
67 * prevent spurious OOMs when the hugepage pool is fully utilized.
69 static int num_fault_mutexes;
70 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
72 /* 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 struct hstate *hstate;
643 if (!is_vm_hugetlb_page(vma))
646 hstate = hstate_vma(vma);
648 return 1UL << huge_page_shift(hstate);
650 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
653 * Return the page size being used by the MMU to back a VMA. In the majority
654 * of cases, the page size used by the kernel matches the MMU size. On
655 * architectures where it differs, an architecture-specific version of this
656 * function is required.
658 #ifndef vma_mmu_pagesize
659 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
661 return vma_kernel_pagesize(vma);
666 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
667 * bits of the reservation map pointer, which are always clear due to
670 #define HPAGE_RESV_OWNER (1UL << 0)
671 #define HPAGE_RESV_UNMAPPED (1UL << 1)
672 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
675 * These helpers are used to track how many pages are reserved for
676 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
677 * is guaranteed to have their future faults succeed.
679 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
680 * the reserve counters are updated with the hugetlb_lock held. It is safe
681 * to reset the VMA at fork() time as it is not in use yet and there is no
682 * chance of the global counters getting corrupted as a result of the values.
684 * The private mapping reservation is represented in a subtly different
685 * manner to a shared mapping. A shared mapping has a region map associated
686 * with the underlying file, this region map represents the backing file
687 * pages which have ever had a reservation assigned which this persists even
688 * after the page is instantiated. A private mapping has a region map
689 * associated with the original mmap which is attached to all VMAs which
690 * reference it, this region map represents those offsets which have consumed
691 * reservation ie. where pages have been instantiated.
693 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
695 return (unsigned long)vma->vm_private_data;
698 static void set_vma_private_data(struct vm_area_struct *vma,
701 vma->vm_private_data = (void *)value;
704 struct resv_map *resv_map_alloc(void)
706 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
707 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
709 if (!resv_map || !rg) {
715 kref_init(&resv_map->refs);
716 spin_lock_init(&resv_map->lock);
717 INIT_LIST_HEAD(&resv_map->regions);
719 resv_map->adds_in_progress = 0;
721 INIT_LIST_HEAD(&resv_map->region_cache);
722 list_add(&rg->link, &resv_map->region_cache);
723 resv_map->region_cache_count = 1;
728 void resv_map_release(struct kref *ref)
730 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
731 struct list_head *head = &resv_map->region_cache;
732 struct file_region *rg, *trg;
734 /* Clear out any active regions before we release the map. */
735 region_del(resv_map, 0, LONG_MAX);
737 /* ... and any entries left in the cache */
738 list_for_each_entry_safe(rg, trg, head, link) {
743 VM_BUG_ON(resv_map->adds_in_progress);
748 static inline struct resv_map *inode_resv_map(struct inode *inode)
750 return inode->i_mapping->private_data;
753 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
755 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
756 if (vma->vm_flags & VM_MAYSHARE) {
757 struct address_space *mapping = vma->vm_file->f_mapping;
758 struct inode *inode = mapping->host;
760 return inode_resv_map(inode);
763 return (struct resv_map *)(get_vma_private_data(vma) &
768 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
770 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
771 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
773 set_vma_private_data(vma, (get_vma_private_data(vma) &
774 HPAGE_RESV_MASK) | (unsigned long)map);
777 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
779 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
780 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
782 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
785 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
787 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
789 return (get_vma_private_data(vma) & flag) != 0;
792 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
793 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
795 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
796 if (!(vma->vm_flags & VM_MAYSHARE))
797 vma->vm_private_data = (void *)0;
800 /* Returns true if the VMA has associated reserve pages */
801 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
803 if (vma->vm_flags & VM_NORESERVE) {
805 * This address is already reserved by other process(chg == 0),
806 * so, we should decrement reserved count. Without decrementing,
807 * reserve count remains after releasing inode, because this
808 * allocated page will go into page cache and is regarded as
809 * coming from reserved pool in releasing step. Currently, we
810 * don't have any other solution to deal with this situation
811 * properly, so add work-around here.
813 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
819 /* Shared mappings always use reserves */
820 if (vma->vm_flags & VM_MAYSHARE) {
822 * We know VM_NORESERVE is not set. Therefore, there SHOULD
823 * be a region map for all pages. The only situation where
824 * there is no region map is if a hole was punched via
825 * fallocate. In this case, there really are no reverves to
826 * use. This situation is indicated if chg != 0.
835 * Only the process that called mmap() has reserves for
838 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
840 * Like the shared case above, a hole punch or truncate
841 * could have been performed on the private mapping.
842 * Examine the value of chg to determine if reserves
843 * actually exist or were previously consumed.
844 * Very Subtle - The value of chg comes from a previous
845 * call to vma_needs_reserves(). The reserve map for
846 * private mappings has different (opposite) semantics
847 * than that of shared mappings. vma_needs_reserves()
848 * has already taken this difference in semantics into
849 * account. Therefore, the meaning of chg is the same
850 * as in the shared case above. Code could easily be
851 * combined, but keeping it separate draws attention to
852 * subtle differences.
863 static void enqueue_huge_page(struct hstate *h, struct page *page)
865 int nid = page_to_nid(page);
866 list_move(&page->lru, &h->hugepage_freelists[nid]);
867 h->free_huge_pages++;
868 h->free_huge_pages_node[nid]++;
871 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
875 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
876 if (!PageHWPoison(page))
879 * if 'non-isolated free hugepage' not found on the list,
880 * the allocation fails.
882 if (&h->hugepage_freelists[nid] == &page->lru)
884 list_move(&page->lru, &h->hugepage_activelist);
885 set_page_refcounted(page);
886 h->free_huge_pages--;
887 h->free_huge_pages_node[nid]--;
891 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
894 unsigned int cpuset_mems_cookie;
895 struct zonelist *zonelist;
900 zonelist = node_zonelist(nid, gfp_mask);
903 cpuset_mems_cookie = read_mems_allowed_begin();
904 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
907 if (!cpuset_zone_allowed(zone, gfp_mask))
910 * no need to ask again on the same node. Pool is node rather than
913 if (zone_to_nid(zone) == node)
915 node = zone_to_nid(zone);
917 page = dequeue_huge_page_node_exact(h, node);
921 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
927 /* Movability of hugepages depends on migration support. */
928 static inline gfp_t htlb_alloc_mask(struct hstate *h)
930 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
931 return GFP_HIGHUSER_MOVABLE;
936 static struct page *dequeue_huge_page_vma(struct hstate *h,
937 struct vm_area_struct *vma,
938 unsigned long address, int avoid_reserve,
942 struct mempolicy *mpol;
944 nodemask_t *nodemask;
948 * A child process with MAP_PRIVATE mappings created by their parent
949 * have no page reserves. This check ensures that reservations are
950 * not "stolen". The child may still get SIGKILLed
952 if (!vma_has_reserves(vma, chg) &&
953 h->free_huge_pages - h->resv_huge_pages == 0)
956 /* If reserves cannot be used, ensure enough pages are in the pool */
957 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
960 gfp_mask = htlb_alloc_mask(h);
961 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
962 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
963 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
964 SetPagePrivate(page);
965 h->resv_huge_pages--;
976 * common helper functions for hstate_next_node_to_{alloc|free}.
977 * We may have allocated or freed a huge page based on a different
978 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
979 * be outside of *nodes_allowed. Ensure that we use an allowed
980 * node for alloc or free.
982 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
984 nid = next_node_in(nid, *nodes_allowed);
985 VM_BUG_ON(nid >= MAX_NUMNODES);
990 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
992 if (!node_isset(nid, *nodes_allowed))
993 nid = next_node_allowed(nid, nodes_allowed);
998 * returns the previously saved node ["this node"] from which to
999 * allocate a persistent huge page for the pool and advance the
1000 * next node from which to allocate, handling wrap at end of node
1003 static int hstate_next_node_to_alloc(struct hstate *h,
1004 nodemask_t *nodes_allowed)
1008 VM_BUG_ON(!nodes_allowed);
1010 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1011 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1017 * helper for free_pool_huge_page() - return the previously saved
1018 * node ["this node"] from which to free a huge page. Advance the
1019 * next node id whether or not we find a free huge page to free so
1020 * that the next attempt to free addresses the next node.
1022 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1026 VM_BUG_ON(!nodes_allowed);
1028 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1029 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1034 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1035 for (nr_nodes = nodes_weight(*mask); \
1037 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1040 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1041 for (nr_nodes = nodes_weight(*mask); \
1043 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1046 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1047 static void destroy_compound_gigantic_page(struct page *page,
1051 int nr_pages = 1 << order;
1052 struct page *p = page + 1;
1054 atomic_set(compound_mapcount_ptr(page), 0);
1055 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1056 clear_compound_head(p);
1057 set_page_refcounted(p);
1060 set_compound_order(page, 0);
1061 __ClearPageHead(page);
1064 static void free_gigantic_page(struct page *page, unsigned int order)
1066 free_contig_range(page_to_pfn(page), 1 << order);
1069 static int __alloc_gigantic_page(unsigned long start_pfn,
1070 unsigned long nr_pages, gfp_t gfp_mask)
1072 unsigned long end_pfn = start_pfn + nr_pages;
1073 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1077 static bool pfn_range_valid_gigantic(struct zone *z,
1078 unsigned long start_pfn, unsigned long nr_pages)
1080 unsigned long i, end_pfn = start_pfn + nr_pages;
1083 for (i = start_pfn; i < end_pfn; i++) {
1087 page = pfn_to_page(i);
1089 if (page_zone(page) != z)
1092 if (PageReserved(page))
1095 if (page_count(page) > 0)
1105 static bool zone_spans_last_pfn(const struct zone *zone,
1106 unsigned long start_pfn, unsigned long nr_pages)
1108 unsigned long last_pfn = start_pfn + nr_pages - 1;
1109 return zone_spans_pfn(zone, last_pfn);
1112 static struct page *alloc_gigantic_page(int nid, struct hstate *h)
1114 unsigned int order = huge_page_order(h);
1115 unsigned long nr_pages = 1 << order;
1116 unsigned long ret, pfn, flags;
1117 struct zonelist *zonelist;
1122 gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1123 zonelist = node_zonelist(nid, gfp_mask);
1124 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), NULL) {
1125 spin_lock_irqsave(&zone->lock, flags);
1127 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
1128 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
1129 if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) {
1131 * We release the zone lock here because
1132 * alloc_contig_range() will also lock the zone
1133 * at some point. If there's an allocation
1134 * spinning on this lock, it may win the race
1135 * and cause alloc_contig_range() to fail...
1137 spin_unlock_irqrestore(&zone->lock, flags);
1138 ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask);
1140 return pfn_to_page(pfn);
1141 spin_lock_irqsave(&zone->lock, flags);
1146 spin_unlock_irqrestore(&zone->lock, flags);
1152 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1153 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1155 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1159 page = alloc_gigantic_page(nid, h);
1161 prep_compound_gigantic_page(page, huge_page_order(h));
1162 prep_new_huge_page(h, page, nid);
1168 static int alloc_fresh_gigantic_page(struct hstate *h,
1169 nodemask_t *nodes_allowed)
1171 struct page *page = NULL;
1174 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1175 page = alloc_fresh_gigantic_page_node(h, node);
1183 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1184 static inline bool gigantic_page_supported(void) { return false; }
1185 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1186 static inline void destroy_compound_gigantic_page(struct page *page,
1187 unsigned int order) { }
1188 static inline int alloc_fresh_gigantic_page(struct hstate *h,
1189 nodemask_t *nodes_allowed) { return 0; }
1192 static void update_and_free_page(struct hstate *h, struct page *page)
1196 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1200 h->nr_huge_pages_node[page_to_nid(page)]--;
1201 for (i = 0; i < pages_per_huge_page(h); i++) {
1202 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1203 1 << PG_referenced | 1 << PG_dirty |
1204 1 << PG_active | 1 << PG_private |
1207 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1208 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1209 set_page_refcounted(page);
1210 if (hstate_is_gigantic(h)) {
1211 destroy_compound_gigantic_page(page, huge_page_order(h));
1212 free_gigantic_page(page, huge_page_order(h));
1214 __free_pages(page, huge_page_order(h));
1218 struct hstate *size_to_hstate(unsigned long size)
1222 for_each_hstate(h) {
1223 if (huge_page_size(h) == size)
1230 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1231 * to hstate->hugepage_activelist.)
1233 * This function can be called for tail pages, but never returns true for them.
1235 bool page_huge_active(struct page *page)
1237 VM_BUG_ON_PAGE(!PageHuge(page), page);
1238 return PageHead(page) && PagePrivate(&page[1]);
1241 /* never called for tail page */
1242 static void set_page_huge_active(struct page *page)
1244 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1245 SetPagePrivate(&page[1]);
1248 static void clear_page_huge_active(struct page *page)
1250 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1251 ClearPagePrivate(&page[1]);
1254 void free_huge_page(struct page *page)
1257 * Can't pass hstate in here because it is called from the
1258 * compound page destructor.
1260 struct hstate *h = page_hstate(page);
1261 int nid = page_to_nid(page);
1262 struct hugepage_subpool *spool =
1263 (struct hugepage_subpool *)page_private(page);
1264 bool restore_reserve;
1266 set_page_private(page, 0);
1267 page->mapping = NULL;
1268 VM_BUG_ON_PAGE(page_count(page), page);
1269 VM_BUG_ON_PAGE(page_mapcount(page), page);
1270 restore_reserve = PagePrivate(page);
1271 ClearPagePrivate(page);
1274 * A return code of zero implies that the subpool will be under its
1275 * minimum size if the reservation is not restored after page is free.
1276 * Therefore, force restore_reserve operation.
1278 if (hugepage_subpool_put_pages(spool, 1) == 0)
1279 restore_reserve = true;
1281 spin_lock(&hugetlb_lock);
1282 clear_page_huge_active(page);
1283 hugetlb_cgroup_uncharge_page(hstate_index(h),
1284 pages_per_huge_page(h), page);
1285 if (restore_reserve)
1286 h->resv_huge_pages++;
1288 if (h->surplus_huge_pages_node[nid]) {
1289 /* remove the page from active list */
1290 list_del(&page->lru);
1291 update_and_free_page(h, page);
1292 h->surplus_huge_pages--;
1293 h->surplus_huge_pages_node[nid]--;
1295 arch_clear_hugepage_flags(page);
1296 enqueue_huge_page(h, page);
1298 spin_unlock(&hugetlb_lock);
1301 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1303 INIT_LIST_HEAD(&page->lru);
1304 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1305 spin_lock(&hugetlb_lock);
1306 set_hugetlb_cgroup(page, NULL);
1308 h->nr_huge_pages_node[nid]++;
1309 spin_unlock(&hugetlb_lock);
1310 put_page(page); /* free it into the hugepage allocator */
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_fresh_huge_page_node(struct hstate *h, int nid)
1391 page = __alloc_pages_node(nid,
1392 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1393 __GFP_RETRY_MAYFAIL|__GFP_NOWARN,
1394 huge_page_order(h));
1396 prep_new_huge_page(h, page, nid);
1402 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1408 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1409 page = alloc_fresh_huge_page_node(h, node);
1417 count_vm_event(HTLB_BUDDY_PGALLOC);
1419 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1425 * Free huge page from pool from next node to free.
1426 * Attempt to keep persistent huge pages more or less
1427 * balanced over allowed nodes.
1428 * Called with hugetlb_lock locked.
1430 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1436 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1438 * If we're returning unused surplus pages, only examine
1439 * nodes with surplus pages.
1441 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1442 !list_empty(&h->hugepage_freelists[node])) {
1444 list_entry(h->hugepage_freelists[node].next,
1446 list_del(&page->lru);
1447 h->free_huge_pages--;
1448 h->free_huge_pages_node[node]--;
1450 h->surplus_huge_pages--;
1451 h->surplus_huge_pages_node[node]--;
1453 update_and_free_page(h, page);
1463 * Dissolve a given free hugepage into free buddy pages. This function does
1464 * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1465 * number of free hugepages would be reduced below the number of reserved
1468 int dissolve_free_huge_page(struct page *page)
1472 spin_lock(&hugetlb_lock);
1473 if (PageHuge(page) && !page_count(page)) {
1474 struct page *head = compound_head(page);
1475 struct hstate *h = page_hstate(head);
1476 int nid = page_to_nid(head);
1477 if (h->free_huge_pages - h->resv_huge_pages == 0) {
1482 * Move PageHWPoison flag from head page to the raw error page,
1483 * which makes any subpages rather than the error page reusable.
1485 if (PageHWPoison(head) && page != head) {
1486 SetPageHWPoison(page);
1487 ClearPageHWPoison(head);
1489 list_del(&head->lru);
1490 h->free_huge_pages--;
1491 h->free_huge_pages_node[nid]--;
1492 h->max_huge_pages--;
1493 update_and_free_page(h, head);
1496 spin_unlock(&hugetlb_lock);
1501 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1502 * make specified memory blocks removable from the system.
1503 * Note that this will dissolve a free gigantic hugepage completely, if any
1504 * part of it lies within the given range.
1505 * Also note that if dissolve_free_huge_page() returns with an error, all
1506 * free hugepages that were dissolved before that error are lost.
1508 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1514 if (!hugepages_supported())
1517 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1518 page = pfn_to_page(pfn);
1519 if (PageHuge(page) && !page_count(page)) {
1520 rc = dissolve_free_huge_page(page);
1529 static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1530 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1532 int order = huge_page_order(h);
1534 gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
1535 if (nid == NUMA_NO_NODE)
1536 nid = numa_mem_id();
1537 return __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1540 static struct page *__alloc_buddy_huge_page(struct hstate *h, gfp_t gfp_mask,
1541 int nid, nodemask_t *nmask)
1546 if (hstate_is_gigantic(h))
1550 * Assume we will successfully allocate the surplus page to
1551 * prevent racing processes from causing the surplus to exceed
1554 * This however introduces a different race, where a process B
1555 * tries to grow the static hugepage pool while alloc_pages() is
1556 * called by process A. B will only examine the per-node
1557 * counters in determining if surplus huge pages can be
1558 * converted to normal huge pages in adjust_pool_surplus(). A
1559 * won't be able to increment the per-node counter, until the
1560 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1561 * no more huge pages can be converted from surplus to normal
1562 * state (and doesn't try to convert again). Thus, we have a
1563 * case where a surplus huge page exists, the pool is grown, and
1564 * the surplus huge page still exists after, even though it
1565 * should just have been converted to a normal huge page. This
1566 * does not leak memory, though, as the hugepage will be freed
1567 * once it is out of use. It also does not allow the counters to
1568 * go out of whack in adjust_pool_surplus() as we don't modify
1569 * the node values until we've gotten the hugepage and only the
1570 * per-node value is checked there.
1572 spin_lock(&hugetlb_lock);
1573 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1574 spin_unlock(&hugetlb_lock);
1578 h->surplus_huge_pages++;
1580 spin_unlock(&hugetlb_lock);
1582 page = __hugetlb_alloc_buddy_huge_page(h, gfp_mask, nid, nmask);
1584 spin_lock(&hugetlb_lock);
1586 INIT_LIST_HEAD(&page->lru);
1587 r_nid = page_to_nid(page);
1588 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1589 set_hugetlb_cgroup(page, NULL);
1591 * We incremented the global counters already
1593 h->nr_huge_pages_node[r_nid]++;
1594 h->surplus_huge_pages_node[r_nid]++;
1595 __count_vm_event(HTLB_BUDDY_PGALLOC);
1598 h->surplus_huge_pages--;
1599 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1601 spin_unlock(&hugetlb_lock);
1607 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1610 struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1611 struct vm_area_struct *vma, unsigned long addr)
1614 struct mempolicy *mpol;
1615 gfp_t gfp_mask = htlb_alloc_mask(h);
1617 nodemask_t *nodemask;
1619 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1620 page = __alloc_buddy_huge_page(h, gfp_mask, nid, nodemask);
1621 mpol_cond_put(mpol);
1627 * This allocation function is useful in the context where vma is irrelevant.
1628 * E.g. soft-offlining uses this function because it only cares physical
1629 * address of error page.
1631 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1633 gfp_t gfp_mask = htlb_alloc_mask(h);
1634 struct page *page = NULL;
1636 if (nid != NUMA_NO_NODE)
1637 gfp_mask |= __GFP_THISNODE;
1639 spin_lock(&hugetlb_lock);
1640 if (h->free_huge_pages - h->resv_huge_pages > 0)
1641 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1642 spin_unlock(&hugetlb_lock);
1645 page = __alloc_buddy_huge_page(h, gfp_mask, nid, NULL);
1651 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1654 gfp_t gfp_mask = htlb_alloc_mask(h);
1656 spin_lock(&hugetlb_lock);
1657 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1660 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1662 spin_unlock(&hugetlb_lock);
1666 spin_unlock(&hugetlb_lock);
1668 /* No reservations, try to overcommit */
1670 return __alloc_buddy_huge_page(h, gfp_mask, preferred_nid, nmask);
1674 * Increase the hugetlb pool such that it can accommodate a reservation
1677 static int gather_surplus_pages(struct hstate *h, int delta)
1679 struct list_head surplus_list;
1680 struct page *page, *tmp;
1682 int needed, allocated;
1683 bool alloc_ok = true;
1685 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1687 h->resv_huge_pages += delta;
1692 INIT_LIST_HEAD(&surplus_list);
1696 spin_unlock(&hugetlb_lock);
1697 for (i = 0; i < needed; i++) {
1698 page = __alloc_buddy_huge_page(h, htlb_alloc_mask(h),
1699 NUMA_NO_NODE, NULL);
1704 list_add(&page->lru, &surplus_list);
1710 * After retaking hugetlb_lock, we need to recalculate 'needed'
1711 * because either resv_huge_pages or free_huge_pages may have changed.
1713 spin_lock(&hugetlb_lock);
1714 needed = (h->resv_huge_pages + delta) -
1715 (h->free_huge_pages + allocated);
1720 * We were not able to allocate enough pages to
1721 * satisfy the entire reservation so we free what
1722 * we've allocated so far.
1727 * The surplus_list now contains _at_least_ the number of extra pages
1728 * needed to accommodate the reservation. Add the appropriate number
1729 * of pages to the hugetlb pool and free the extras back to the buddy
1730 * allocator. Commit the entire reservation here to prevent another
1731 * process from stealing the pages as they are added to the pool but
1732 * before they are reserved.
1734 needed += allocated;
1735 h->resv_huge_pages += delta;
1738 /* Free the needed pages to the hugetlb pool */
1739 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1743 * This page is now managed by the hugetlb allocator and has
1744 * no users -- drop the buddy allocator's reference.
1746 put_page_testzero(page);
1747 VM_BUG_ON_PAGE(page_count(page), page);
1748 enqueue_huge_page(h, page);
1751 spin_unlock(&hugetlb_lock);
1753 /* Free unnecessary surplus pages to the buddy allocator */
1754 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1756 spin_lock(&hugetlb_lock);
1762 * This routine has two main purposes:
1763 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1764 * in unused_resv_pages. This corresponds to the prior adjustments made
1765 * to the associated reservation map.
1766 * 2) Free any unused surplus pages that may have been allocated to satisfy
1767 * the reservation. As many as unused_resv_pages may be freed.
1769 * Called with hugetlb_lock held. However, the lock could be dropped (and
1770 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1771 * we must make sure nobody else can claim pages we are in the process of
1772 * freeing. Do this by ensuring resv_huge_page always is greater than the
1773 * number of huge pages we plan to free when dropping the lock.
1775 static void return_unused_surplus_pages(struct hstate *h,
1776 unsigned long unused_resv_pages)
1778 unsigned long nr_pages;
1780 /* Cannot return gigantic pages currently */
1781 if (hstate_is_gigantic(h))
1785 * Part (or even all) of the reservation could have been backed
1786 * by pre-allocated pages. Only free surplus pages.
1788 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1791 * We want to release as many surplus pages as possible, spread
1792 * evenly across all nodes with memory. Iterate across these nodes
1793 * until we can no longer free unreserved surplus pages. This occurs
1794 * when the nodes with surplus pages have no free pages.
1795 * free_pool_huge_page() will balance the the freed pages across the
1796 * on-line nodes with memory and will handle the hstate accounting.
1798 * Note that we decrement resv_huge_pages as we free the pages. If
1799 * we drop the lock, resv_huge_pages will still be sufficiently large
1800 * to cover subsequent pages we may free.
1802 while (nr_pages--) {
1803 h->resv_huge_pages--;
1804 unused_resv_pages--;
1805 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1807 cond_resched_lock(&hugetlb_lock);
1811 /* Fully uncommit the reservation */
1812 h->resv_huge_pages -= unused_resv_pages;
1817 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1818 * are used by the huge page allocation routines to manage reservations.
1820 * vma_needs_reservation is called to determine if the huge page at addr
1821 * within the vma has an associated reservation. If a reservation is
1822 * needed, the value 1 is returned. The caller is then responsible for
1823 * managing the global reservation and subpool usage counts. After
1824 * the huge page has been allocated, vma_commit_reservation is called
1825 * to add the page to the reservation map. If the page allocation fails,
1826 * the reservation must be ended instead of committed. vma_end_reservation
1827 * is called in such cases.
1829 * In the normal case, vma_commit_reservation returns the same value
1830 * as the preceding vma_needs_reservation call. The only time this
1831 * is not the case is if a reserve map was changed between calls. It
1832 * is the responsibility of the caller to notice the difference and
1833 * take appropriate action.
1835 * vma_add_reservation is used in error paths where a reservation must
1836 * be restored when a newly allocated huge page must be freed. It is
1837 * to be called after calling vma_needs_reservation to determine if a
1838 * reservation exists.
1840 enum vma_resv_mode {
1846 static long __vma_reservation_common(struct hstate *h,
1847 struct vm_area_struct *vma, unsigned long addr,
1848 enum vma_resv_mode mode)
1850 struct resv_map *resv;
1854 resv = vma_resv_map(vma);
1858 idx = vma_hugecache_offset(h, vma, addr);
1860 case VMA_NEEDS_RESV:
1861 ret = region_chg(resv, idx, idx + 1);
1863 case VMA_COMMIT_RESV:
1864 ret = region_add(resv, idx, idx + 1);
1867 region_abort(resv, idx, idx + 1);
1871 if (vma->vm_flags & VM_MAYSHARE)
1872 ret = region_add(resv, idx, idx + 1);
1874 region_abort(resv, idx, idx + 1);
1875 ret = region_del(resv, idx, idx + 1);
1882 if (vma->vm_flags & VM_MAYSHARE)
1884 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1886 * In most cases, reserves always exist for private mappings.
1887 * However, a file associated with mapping could have been
1888 * hole punched or truncated after reserves were consumed.
1889 * As subsequent fault on such a range will not use reserves.
1890 * Subtle - The reserve map for private mappings has the
1891 * opposite meaning than that of shared mappings. If NO
1892 * entry is in the reserve map, it means a reservation exists.
1893 * If an entry exists in the reserve map, it means the
1894 * reservation has already been consumed. As a result, the
1895 * return value of this routine is the opposite of the
1896 * value returned from reserve map manipulation routines above.
1904 return ret < 0 ? ret : 0;
1907 static long vma_needs_reservation(struct hstate *h,
1908 struct vm_area_struct *vma, unsigned long addr)
1910 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1913 static long vma_commit_reservation(struct hstate *h,
1914 struct vm_area_struct *vma, unsigned long addr)
1916 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1919 static void vma_end_reservation(struct hstate *h,
1920 struct vm_area_struct *vma, unsigned long addr)
1922 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1925 static long vma_add_reservation(struct hstate *h,
1926 struct vm_area_struct *vma, unsigned long addr)
1928 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1932 * This routine is called to restore a reservation on error paths. In the
1933 * specific error paths, a huge page was allocated (via alloc_huge_page)
1934 * and is about to be freed. If a reservation for the page existed,
1935 * alloc_huge_page would have consumed the reservation and set PagePrivate
1936 * in the newly allocated page. When the page is freed via free_huge_page,
1937 * the global reservation count will be incremented if PagePrivate is set.
1938 * However, free_huge_page can not adjust the reserve map. Adjust the
1939 * reserve map here to be consistent with global reserve count adjustments
1940 * to be made by free_huge_page.
1942 static void restore_reserve_on_error(struct hstate *h,
1943 struct vm_area_struct *vma, unsigned long address,
1946 if (unlikely(PagePrivate(page))) {
1947 long rc = vma_needs_reservation(h, vma, address);
1949 if (unlikely(rc < 0)) {
1951 * Rare out of memory condition in reserve map
1952 * manipulation. Clear PagePrivate so that
1953 * global reserve count will not be incremented
1954 * by free_huge_page. This will make it appear
1955 * as though the reservation for this page was
1956 * consumed. This may prevent the task from
1957 * faulting in the page at a later time. This
1958 * is better than inconsistent global huge page
1959 * accounting of reserve counts.
1961 ClearPagePrivate(page);
1963 rc = vma_add_reservation(h, vma, address);
1964 if (unlikely(rc < 0))
1966 * See above comment about rare out of
1969 ClearPagePrivate(page);
1971 vma_end_reservation(h, vma, address);
1975 struct page *alloc_huge_page(struct vm_area_struct *vma,
1976 unsigned long addr, int avoid_reserve)
1978 struct hugepage_subpool *spool = subpool_vma(vma);
1979 struct hstate *h = hstate_vma(vma);
1981 long map_chg, map_commit;
1984 struct hugetlb_cgroup *h_cg;
1986 idx = hstate_index(h);
1988 * Examine the region/reserve map to determine if the process
1989 * has a reservation for the page to be allocated. A return
1990 * code of zero indicates a reservation exists (no change).
1992 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
1994 return ERR_PTR(-ENOMEM);
1997 * Processes that did not create the mapping will have no
1998 * reserves as indicated by the region/reserve map. Check
1999 * that the allocation will not exceed the subpool limit.
2000 * Allocations for MAP_NORESERVE mappings also need to be
2001 * checked against any subpool limit.
2003 if (map_chg || avoid_reserve) {
2004 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2006 vma_end_reservation(h, vma, addr);
2007 return ERR_PTR(-ENOSPC);
2011 * Even though there was no reservation in the region/reserve
2012 * map, there could be reservations associated with the
2013 * subpool that can be used. This would be indicated if the
2014 * return value of hugepage_subpool_get_pages() is zero.
2015 * However, if avoid_reserve is specified we still avoid even
2016 * the subpool reservations.
2022 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2024 goto out_subpool_put;
2026 spin_lock(&hugetlb_lock);
2028 * glb_chg is passed to indicate whether or not a page must be taken
2029 * from the global free pool (global change). gbl_chg == 0 indicates
2030 * a reservation exists for the allocation.
2032 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2034 spin_unlock(&hugetlb_lock);
2035 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
2037 goto out_uncharge_cgroup;
2038 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2039 SetPagePrivate(page);
2040 h->resv_huge_pages--;
2042 spin_lock(&hugetlb_lock);
2043 list_move(&page->lru, &h->hugepage_activelist);
2046 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2047 spin_unlock(&hugetlb_lock);
2049 set_page_private(page, (unsigned long)spool);
2051 map_commit = vma_commit_reservation(h, vma, addr);
2052 if (unlikely(map_chg > map_commit)) {
2054 * The page was added to the reservation map between
2055 * vma_needs_reservation and vma_commit_reservation.
2056 * This indicates a race with hugetlb_reserve_pages.
2057 * Adjust for the subpool count incremented above AND
2058 * in hugetlb_reserve_pages for the same page. Also,
2059 * the reservation count added in hugetlb_reserve_pages
2060 * no longer applies.
2064 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2065 hugetlb_acct_memory(h, -rsv_adjust);
2069 out_uncharge_cgroup:
2070 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2072 if (map_chg || avoid_reserve)
2073 hugepage_subpool_put_pages(spool, 1);
2074 vma_end_reservation(h, vma, addr);
2075 return ERR_PTR(-ENOSPC);
2079 * alloc_huge_page()'s wrapper which simply returns the page if allocation
2080 * succeeds, otherwise NULL. This function is called from new_vma_page(),
2081 * where no ERR_VALUE is expected to be returned.
2083 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
2084 unsigned long addr, int avoid_reserve)
2086 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
2092 int alloc_bootmem_huge_page(struct hstate *h)
2093 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2094 int __alloc_bootmem_huge_page(struct hstate *h)
2096 struct huge_bootmem_page *m;
2099 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2102 addr = memblock_virt_alloc_try_nid_nopanic(
2103 huge_page_size(h), huge_page_size(h),
2104 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
2107 * Use the beginning of the huge page to store the
2108 * huge_bootmem_page struct (until gather_bootmem
2109 * puts them into the mem_map).
2118 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2119 /* Put them into a private list first because mem_map is not up yet */
2120 list_add(&m->list, &huge_boot_pages);
2125 static void __init prep_compound_huge_page(struct page *page,
2128 if (unlikely(order > (MAX_ORDER - 1)))
2129 prep_compound_gigantic_page(page, order);
2131 prep_compound_page(page, order);
2134 /* Put bootmem huge pages into the standard lists after mem_map is up */
2135 static void __init gather_bootmem_prealloc(void)
2137 struct huge_bootmem_page *m;
2139 list_for_each_entry(m, &huge_boot_pages, list) {
2140 struct hstate *h = m->hstate;
2143 #ifdef CONFIG_HIGHMEM
2144 page = pfn_to_page(m->phys >> PAGE_SHIFT);
2145 memblock_free_late(__pa(m),
2146 sizeof(struct huge_bootmem_page));
2148 page = virt_to_page(m);
2150 WARN_ON(page_count(page) != 1);
2151 prep_compound_huge_page(page, h->order);
2152 WARN_ON(PageReserved(page));
2153 prep_new_huge_page(h, page, page_to_nid(page));
2155 * If we had gigantic hugepages allocated at boot time, we need
2156 * to restore the 'stolen' pages to totalram_pages in order to
2157 * fix confusing memory reports from free(1) and another
2158 * side-effects, like CommitLimit going negative.
2160 if (hstate_is_gigantic(h))
2161 adjust_managed_page_count(page, 1 << h->order);
2165 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2169 for (i = 0; i < h->max_huge_pages; ++i) {
2170 if (hstate_is_gigantic(h)) {
2171 if (!alloc_bootmem_huge_page(h))
2173 } else if (!alloc_fresh_huge_page(h,
2174 &node_states[N_MEMORY]))
2178 if (i < h->max_huge_pages) {
2181 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2182 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2183 h->max_huge_pages, buf, i);
2184 h->max_huge_pages = i;
2188 static void __init hugetlb_init_hstates(void)
2192 for_each_hstate(h) {
2193 if (minimum_order > huge_page_order(h))
2194 minimum_order = huge_page_order(h);
2196 /* oversize hugepages were init'ed in early boot */
2197 if (!hstate_is_gigantic(h))
2198 hugetlb_hstate_alloc_pages(h);
2200 VM_BUG_ON(minimum_order == UINT_MAX);
2203 static void __init report_hugepages(void)
2207 for_each_hstate(h) {
2210 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2211 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2212 buf, h->free_huge_pages);
2216 #ifdef CONFIG_HIGHMEM
2217 static void try_to_free_low(struct hstate *h, unsigned long count,
2218 nodemask_t *nodes_allowed)
2222 if (hstate_is_gigantic(h))
2225 for_each_node_mask(i, *nodes_allowed) {
2226 struct page *page, *next;
2227 struct list_head *freel = &h->hugepage_freelists[i];
2228 list_for_each_entry_safe(page, next, freel, lru) {
2229 if (count >= h->nr_huge_pages)
2231 if (PageHighMem(page))
2233 list_del(&page->lru);
2234 update_and_free_page(h, page);
2235 h->free_huge_pages--;
2236 h->free_huge_pages_node[page_to_nid(page)]--;
2241 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2242 nodemask_t *nodes_allowed)
2248 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2249 * balanced by operating on them in a round-robin fashion.
2250 * Returns 1 if an adjustment was made.
2252 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2257 VM_BUG_ON(delta != -1 && delta != 1);
2260 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2261 if (h->surplus_huge_pages_node[node])
2265 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2266 if (h->surplus_huge_pages_node[node] <
2267 h->nr_huge_pages_node[node])
2274 h->surplus_huge_pages += delta;
2275 h->surplus_huge_pages_node[node] += delta;
2279 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2280 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2281 nodemask_t *nodes_allowed)
2283 unsigned long min_count, ret;
2285 if (hstate_is_gigantic(h) && !gigantic_page_supported())
2286 return h->max_huge_pages;
2289 * Increase the pool size
2290 * First take pages out of surplus state. Then make up the
2291 * remaining difference by allocating fresh huge pages.
2293 * We might race with __alloc_buddy_huge_page() here and be unable
2294 * to convert a surplus huge page to a normal huge page. That is
2295 * not critical, though, it just means the overall size of the
2296 * pool might be one hugepage larger than it needs to be, but
2297 * within all the constraints specified by the sysctls.
2299 spin_lock(&hugetlb_lock);
2300 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2301 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2305 while (count > persistent_huge_pages(h)) {
2307 * If this allocation races such that we no longer need the
2308 * page, free_huge_page will handle it by freeing the page
2309 * and reducing the surplus.
2311 spin_unlock(&hugetlb_lock);
2313 /* yield cpu to avoid soft lockup */
2316 if (hstate_is_gigantic(h))
2317 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2319 ret = alloc_fresh_huge_page(h, nodes_allowed);
2320 spin_lock(&hugetlb_lock);
2324 /* Bail for signals. Probably ctrl-c from user */
2325 if (signal_pending(current))
2330 * Decrease the pool size
2331 * First return free pages to the buddy allocator (being careful
2332 * to keep enough around to satisfy reservations). Then place
2333 * pages into surplus state as needed so the pool will shrink
2334 * to the desired size as pages become free.
2336 * By placing pages into the surplus state independent of the
2337 * overcommit value, we are allowing the surplus pool size to
2338 * exceed overcommit. There are few sane options here. Since
2339 * __alloc_buddy_huge_page() is checking the global counter,
2340 * though, we'll note that we're not allowed to exceed surplus
2341 * and won't grow the pool anywhere else. Not until one of the
2342 * sysctls are changed, or the surplus pages go out of use.
2344 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2345 min_count = max(count, min_count);
2346 try_to_free_low(h, min_count, nodes_allowed);
2347 while (min_count < persistent_huge_pages(h)) {
2348 if (!free_pool_huge_page(h, nodes_allowed, 0))
2350 cond_resched_lock(&hugetlb_lock);
2352 while (count < persistent_huge_pages(h)) {
2353 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2357 ret = persistent_huge_pages(h);
2358 spin_unlock(&hugetlb_lock);
2362 #define HSTATE_ATTR_RO(_name) \
2363 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2365 #define HSTATE_ATTR(_name) \
2366 static struct kobj_attribute _name##_attr = \
2367 __ATTR(_name, 0644, _name##_show, _name##_store)
2369 static struct kobject *hugepages_kobj;
2370 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2372 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2374 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2378 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2379 if (hstate_kobjs[i] == kobj) {
2381 *nidp = NUMA_NO_NODE;
2385 return kobj_to_node_hstate(kobj, nidp);
2388 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2389 struct kobj_attribute *attr, char *buf)
2392 unsigned long nr_huge_pages;
2395 h = kobj_to_hstate(kobj, &nid);
2396 if (nid == NUMA_NO_NODE)
2397 nr_huge_pages = h->nr_huge_pages;
2399 nr_huge_pages = h->nr_huge_pages_node[nid];
2401 return sprintf(buf, "%lu\n", nr_huge_pages);
2404 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2405 struct hstate *h, int nid,
2406 unsigned long count, size_t len)
2409 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2411 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2416 if (nid == NUMA_NO_NODE) {
2418 * global hstate attribute
2420 if (!(obey_mempolicy &&
2421 init_nodemask_of_mempolicy(nodes_allowed))) {
2422 NODEMASK_FREE(nodes_allowed);
2423 nodes_allowed = &node_states[N_MEMORY];
2425 } else if (nodes_allowed) {
2427 * per node hstate attribute: adjust count to global,
2428 * but restrict alloc/free to the specified node.
2430 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2431 init_nodemask_of_node(nodes_allowed, nid);
2433 nodes_allowed = &node_states[N_MEMORY];
2435 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2437 if (nodes_allowed != &node_states[N_MEMORY])
2438 NODEMASK_FREE(nodes_allowed);
2442 NODEMASK_FREE(nodes_allowed);
2446 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2447 struct kobject *kobj, const char *buf,
2451 unsigned long count;
2455 err = kstrtoul(buf, 10, &count);
2459 h = kobj_to_hstate(kobj, &nid);
2460 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2463 static ssize_t nr_hugepages_show(struct kobject *kobj,
2464 struct kobj_attribute *attr, char *buf)
2466 return nr_hugepages_show_common(kobj, attr, buf);
2469 static ssize_t nr_hugepages_store(struct kobject *kobj,
2470 struct kobj_attribute *attr, const char *buf, size_t len)
2472 return nr_hugepages_store_common(false, kobj, buf, len);
2474 HSTATE_ATTR(nr_hugepages);
2479 * hstate attribute for optionally mempolicy-based constraint on persistent
2480 * huge page alloc/free.
2482 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2483 struct kobj_attribute *attr, char *buf)
2485 return nr_hugepages_show_common(kobj, attr, buf);
2488 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2489 struct kobj_attribute *attr, const char *buf, size_t len)
2491 return nr_hugepages_store_common(true, kobj, buf, len);
2493 HSTATE_ATTR(nr_hugepages_mempolicy);
2497 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2498 struct kobj_attribute *attr, char *buf)
2500 struct hstate *h = kobj_to_hstate(kobj, NULL);
2501 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2504 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2505 struct kobj_attribute *attr, const char *buf, size_t count)
2508 unsigned long input;
2509 struct hstate *h = kobj_to_hstate(kobj, NULL);
2511 if (hstate_is_gigantic(h))
2514 err = kstrtoul(buf, 10, &input);
2518 spin_lock(&hugetlb_lock);
2519 h->nr_overcommit_huge_pages = input;
2520 spin_unlock(&hugetlb_lock);
2524 HSTATE_ATTR(nr_overcommit_hugepages);
2526 static ssize_t free_hugepages_show(struct kobject *kobj,
2527 struct kobj_attribute *attr, char *buf)
2530 unsigned long free_huge_pages;
2533 h = kobj_to_hstate(kobj, &nid);
2534 if (nid == NUMA_NO_NODE)
2535 free_huge_pages = h->free_huge_pages;
2537 free_huge_pages = h->free_huge_pages_node[nid];
2539 return sprintf(buf, "%lu\n", free_huge_pages);
2541 HSTATE_ATTR_RO(free_hugepages);
2543 static ssize_t resv_hugepages_show(struct kobject *kobj,
2544 struct kobj_attribute *attr, char *buf)
2546 struct hstate *h = kobj_to_hstate(kobj, NULL);
2547 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2549 HSTATE_ATTR_RO(resv_hugepages);
2551 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2552 struct kobj_attribute *attr, char *buf)
2555 unsigned long surplus_huge_pages;
2558 h = kobj_to_hstate(kobj, &nid);
2559 if (nid == NUMA_NO_NODE)
2560 surplus_huge_pages = h->surplus_huge_pages;
2562 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2564 return sprintf(buf, "%lu\n", surplus_huge_pages);
2566 HSTATE_ATTR_RO(surplus_hugepages);
2568 static struct attribute *hstate_attrs[] = {
2569 &nr_hugepages_attr.attr,
2570 &nr_overcommit_hugepages_attr.attr,
2571 &free_hugepages_attr.attr,
2572 &resv_hugepages_attr.attr,
2573 &surplus_hugepages_attr.attr,
2575 &nr_hugepages_mempolicy_attr.attr,
2580 static const struct attribute_group hstate_attr_group = {
2581 .attrs = hstate_attrs,
2584 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2585 struct kobject **hstate_kobjs,
2586 const struct attribute_group *hstate_attr_group)
2589 int hi = hstate_index(h);
2591 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2592 if (!hstate_kobjs[hi])
2595 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2597 kobject_put(hstate_kobjs[hi]);
2602 static void __init hugetlb_sysfs_init(void)
2607 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2608 if (!hugepages_kobj)
2611 for_each_hstate(h) {
2612 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2613 hstate_kobjs, &hstate_attr_group);
2615 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2622 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2623 * with node devices in node_devices[] using a parallel array. The array
2624 * index of a node device or _hstate == node id.
2625 * This is here to avoid any static dependency of the node device driver, in
2626 * the base kernel, on the hugetlb module.
2628 struct node_hstate {
2629 struct kobject *hugepages_kobj;
2630 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2632 static struct node_hstate node_hstates[MAX_NUMNODES];
2635 * A subset of global hstate attributes for node devices
2637 static struct attribute *per_node_hstate_attrs[] = {
2638 &nr_hugepages_attr.attr,
2639 &free_hugepages_attr.attr,
2640 &surplus_hugepages_attr.attr,
2644 static const struct attribute_group per_node_hstate_attr_group = {
2645 .attrs = per_node_hstate_attrs,
2649 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2650 * Returns node id via non-NULL nidp.
2652 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2656 for (nid = 0; nid < nr_node_ids; nid++) {
2657 struct node_hstate *nhs = &node_hstates[nid];
2659 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2660 if (nhs->hstate_kobjs[i] == kobj) {
2672 * Unregister hstate attributes from a single node device.
2673 * No-op if no hstate attributes attached.
2675 static void hugetlb_unregister_node(struct node *node)
2678 struct node_hstate *nhs = &node_hstates[node->dev.id];
2680 if (!nhs->hugepages_kobj)
2681 return; /* no hstate attributes */
2683 for_each_hstate(h) {
2684 int idx = hstate_index(h);
2685 if (nhs->hstate_kobjs[idx]) {
2686 kobject_put(nhs->hstate_kobjs[idx]);
2687 nhs->hstate_kobjs[idx] = NULL;
2691 kobject_put(nhs->hugepages_kobj);
2692 nhs->hugepages_kobj = NULL;
2697 * Register hstate attributes for a single node device.
2698 * No-op if attributes already registered.
2700 static void hugetlb_register_node(struct node *node)
2703 struct node_hstate *nhs = &node_hstates[node->dev.id];
2706 if (nhs->hugepages_kobj)
2707 return; /* already allocated */
2709 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2711 if (!nhs->hugepages_kobj)
2714 for_each_hstate(h) {
2715 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2717 &per_node_hstate_attr_group);
2719 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2720 h->name, node->dev.id);
2721 hugetlb_unregister_node(node);
2728 * hugetlb init time: register hstate attributes for all registered node
2729 * devices of nodes that have memory. All on-line nodes should have
2730 * registered their associated device by this time.
2732 static void __init hugetlb_register_all_nodes(void)
2736 for_each_node_state(nid, N_MEMORY) {
2737 struct node *node = node_devices[nid];
2738 if (node->dev.id == nid)
2739 hugetlb_register_node(node);
2743 * Let the node device driver know we're here so it can
2744 * [un]register hstate attributes on node hotplug.
2746 register_hugetlbfs_with_node(hugetlb_register_node,
2747 hugetlb_unregister_node);
2749 #else /* !CONFIG_NUMA */
2751 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2759 static void hugetlb_register_all_nodes(void) { }
2763 static int __init hugetlb_init(void)
2767 if (!hugepages_supported())
2770 if (!size_to_hstate(default_hstate_size)) {
2771 if (default_hstate_size != 0) {
2772 pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2773 default_hstate_size, HPAGE_SIZE);
2776 default_hstate_size = HPAGE_SIZE;
2777 if (!size_to_hstate(default_hstate_size))
2778 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2780 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2781 if (default_hstate_max_huge_pages) {
2782 if (!default_hstate.max_huge_pages)
2783 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2786 hugetlb_init_hstates();
2787 gather_bootmem_prealloc();
2790 hugetlb_sysfs_init();
2791 hugetlb_register_all_nodes();
2792 hugetlb_cgroup_file_init();
2795 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2797 num_fault_mutexes = 1;
2799 hugetlb_fault_mutex_table =
2800 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2801 BUG_ON(!hugetlb_fault_mutex_table);
2803 for (i = 0; i < num_fault_mutexes; i++)
2804 mutex_init(&hugetlb_fault_mutex_table[i]);
2807 subsys_initcall(hugetlb_init);
2809 /* Should be called on processing a hugepagesz=... option */
2810 void __init hugetlb_bad_size(void)
2812 parsed_valid_hugepagesz = false;
2815 void __init hugetlb_add_hstate(unsigned int order)
2820 if (size_to_hstate(PAGE_SIZE << order)) {
2821 pr_warn("hugepagesz= specified twice, ignoring\n");
2824 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2826 h = &hstates[hugetlb_max_hstate++];
2828 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2829 h->nr_huge_pages = 0;
2830 h->free_huge_pages = 0;
2831 for (i = 0; i < MAX_NUMNODES; ++i)
2832 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2833 INIT_LIST_HEAD(&h->hugepage_activelist);
2834 h->next_nid_to_alloc = first_memory_node;
2835 h->next_nid_to_free = first_memory_node;
2836 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2837 huge_page_size(h)/1024);
2842 static int __init hugetlb_nrpages_setup(char *s)
2845 static unsigned long *last_mhp;
2847 if (!parsed_valid_hugepagesz) {
2848 pr_warn("hugepages = %s preceded by "
2849 "an unsupported hugepagesz, ignoring\n", s);
2850 parsed_valid_hugepagesz = true;
2854 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2855 * so this hugepages= parameter goes to the "default hstate".
2857 else if (!hugetlb_max_hstate)
2858 mhp = &default_hstate_max_huge_pages;
2860 mhp = &parsed_hstate->max_huge_pages;
2862 if (mhp == last_mhp) {
2863 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2867 if (sscanf(s, "%lu", mhp) <= 0)
2871 * Global state is always initialized later in hugetlb_init.
2872 * But we need to allocate >= MAX_ORDER hstates here early to still
2873 * use the bootmem allocator.
2875 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2876 hugetlb_hstate_alloc_pages(parsed_hstate);
2882 __setup("hugepages=", hugetlb_nrpages_setup);
2884 static int __init hugetlb_default_setup(char *s)
2886 default_hstate_size = memparse(s, &s);
2889 __setup("default_hugepagesz=", hugetlb_default_setup);
2891 static unsigned int cpuset_mems_nr(unsigned int *array)
2894 unsigned int nr = 0;
2896 for_each_node_mask(node, cpuset_current_mems_allowed)
2902 #ifdef CONFIG_SYSCTL
2903 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2904 struct ctl_table *table, int write,
2905 void __user *buffer, size_t *length, loff_t *ppos)
2907 struct hstate *h = &default_hstate;
2908 unsigned long tmp = h->max_huge_pages;
2911 if (!hugepages_supported())
2915 table->maxlen = sizeof(unsigned long);
2916 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2921 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2922 NUMA_NO_NODE, tmp, *length);
2927 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2928 void __user *buffer, size_t *length, loff_t *ppos)
2931 return hugetlb_sysctl_handler_common(false, table, write,
2932 buffer, length, ppos);
2936 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2937 void __user *buffer, size_t *length, loff_t *ppos)
2939 return hugetlb_sysctl_handler_common(true, table, write,
2940 buffer, length, ppos);
2942 #endif /* CONFIG_NUMA */
2944 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2945 void __user *buffer,
2946 size_t *length, loff_t *ppos)
2948 struct hstate *h = &default_hstate;
2952 if (!hugepages_supported())
2955 tmp = h->nr_overcommit_huge_pages;
2957 if (write && hstate_is_gigantic(h))
2961 table->maxlen = sizeof(unsigned long);
2962 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2967 spin_lock(&hugetlb_lock);
2968 h->nr_overcommit_huge_pages = tmp;
2969 spin_unlock(&hugetlb_lock);
2975 #endif /* CONFIG_SYSCTL */
2977 void hugetlb_report_meminfo(struct seq_file *m)
2979 struct hstate *h = &default_hstate;
2980 if (!hugepages_supported())
2983 "HugePages_Total: %5lu\n"
2984 "HugePages_Free: %5lu\n"
2985 "HugePages_Rsvd: %5lu\n"
2986 "HugePages_Surp: %5lu\n"
2987 "Hugepagesize: %8lu kB\n",
2991 h->surplus_huge_pages,
2992 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2995 int hugetlb_report_node_meminfo(int nid, char *buf)
2997 struct hstate *h = &default_hstate;
2998 if (!hugepages_supported())
3001 "Node %d HugePages_Total: %5u\n"
3002 "Node %d HugePages_Free: %5u\n"
3003 "Node %d HugePages_Surp: %5u\n",
3004 nid, h->nr_huge_pages_node[nid],
3005 nid, h->free_huge_pages_node[nid],
3006 nid, h->surplus_huge_pages_node[nid]);
3009 void hugetlb_show_meminfo(void)
3014 if (!hugepages_supported())
3017 for_each_node_state(nid, N_MEMORY)
3019 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3021 h->nr_huge_pages_node[nid],
3022 h->free_huge_pages_node[nid],
3023 h->surplus_huge_pages_node[nid],
3024 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3027 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3029 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3030 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3033 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3034 unsigned long hugetlb_total_pages(void)
3037 unsigned long nr_total_pages = 0;
3040 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3041 return nr_total_pages;
3044 static int hugetlb_acct_memory(struct hstate *h, long delta)
3048 spin_lock(&hugetlb_lock);
3050 * When cpuset is configured, it breaks the strict hugetlb page
3051 * reservation as the accounting is done on a global variable. Such
3052 * reservation is completely rubbish in the presence of cpuset because
3053 * the reservation is not checked against page availability for the
3054 * current cpuset. Application can still potentially OOM'ed by kernel
3055 * with lack of free htlb page in cpuset that the task is in.
3056 * Attempt to enforce strict accounting with cpuset is almost
3057 * impossible (or too ugly) because cpuset is too fluid that
3058 * task or memory node can be dynamically moved between cpusets.
3060 * The change of semantics for shared hugetlb mapping with cpuset is
3061 * undesirable. However, in order to preserve some of the semantics,
3062 * we fall back to check against current free page availability as
3063 * a best attempt and hopefully to minimize the impact of changing
3064 * semantics that cpuset has.
3067 if (gather_surplus_pages(h, delta) < 0)
3070 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3071 return_unused_surplus_pages(h, delta);
3078 return_unused_surplus_pages(h, (unsigned long) -delta);
3081 spin_unlock(&hugetlb_lock);
3085 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3087 struct resv_map *resv = vma_resv_map(vma);
3090 * This new VMA should share its siblings reservation map if present.
3091 * The VMA will only ever have a valid reservation map pointer where
3092 * it is being copied for another still existing VMA. As that VMA
3093 * has a reference to the reservation map it cannot disappear until
3094 * after this open call completes. It is therefore safe to take a
3095 * new reference here without additional locking.
3097 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3098 kref_get(&resv->refs);
3101 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3103 struct hstate *h = hstate_vma(vma);
3104 struct resv_map *resv = vma_resv_map(vma);
3105 struct hugepage_subpool *spool = subpool_vma(vma);
3106 unsigned long reserve, start, end;
3109 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3112 start = vma_hugecache_offset(h, vma, vma->vm_start);
3113 end = vma_hugecache_offset(h, vma, vma->vm_end);
3115 reserve = (end - start) - region_count(resv, start, end);
3117 kref_put(&resv->refs, resv_map_release);
3121 * Decrement reserve counts. The global reserve count may be
3122 * adjusted if the subpool has a minimum size.
3124 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3125 hugetlb_acct_memory(h, -gbl_reserve);
3129 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3131 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3137 * We cannot handle pagefaults against hugetlb pages at all. They cause
3138 * handle_mm_fault() to try to instantiate regular-sized pages in the
3139 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3142 static int hugetlb_vm_op_fault(struct vm_fault *vmf)
3148 const struct vm_operations_struct hugetlb_vm_ops = {
3149 .fault = hugetlb_vm_op_fault,
3150 .open = hugetlb_vm_op_open,
3151 .close = hugetlb_vm_op_close,
3152 .split = hugetlb_vm_op_split,
3155 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3161 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3162 vma->vm_page_prot)));
3164 entry = huge_pte_wrprotect(mk_huge_pte(page,
3165 vma->vm_page_prot));
3167 entry = pte_mkyoung(entry);
3168 entry = pte_mkhuge(entry);
3169 entry = arch_make_huge_pte(entry, vma, page, writable);
3174 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3175 unsigned long address, pte_t *ptep)
3179 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3180 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3181 update_mmu_cache(vma, address, ptep);
3184 bool is_hugetlb_entry_migration(pte_t pte)
3188 if (huge_pte_none(pte) || pte_present(pte))
3190 swp = pte_to_swp_entry(pte);
3191 if (non_swap_entry(swp) && is_migration_entry(swp))
3197 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3201 if (huge_pte_none(pte) || pte_present(pte))
3203 swp = pte_to_swp_entry(pte);
3204 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3210 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3211 struct vm_area_struct *vma)
3213 pte_t *src_pte, *dst_pte, entry;
3214 struct page *ptepage;
3217 struct hstate *h = hstate_vma(vma);
3218 unsigned long sz = huge_page_size(h);
3219 unsigned long mmun_start; /* For mmu_notifiers */
3220 unsigned long mmun_end; /* For mmu_notifiers */
3223 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3225 mmun_start = vma->vm_start;
3226 mmun_end = vma->vm_end;
3228 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3230 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3231 spinlock_t *src_ptl, *dst_ptl;
3232 src_pte = huge_pte_offset(src, addr, sz);
3235 dst_pte = huge_pte_alloc(dst, addr, sz);
3241 /* If the pagetables are shared don't copy or take references */
3242 if (dst_pte == src_pte)
3245 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3246 src_ptl = huge_pte_lockptr(h, src, src_pte);
3247 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3248 entry = huge_ptep_get(src_pte);
3249 if (huge_pte_none(entry)) { /* skip none entry */
3251 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3252 is_hugetlb_entry_hwpoisoned(entry))) {
3253 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3255 if (is_write_migration_entry(swp_entry) && cow) {
3257 * COW mappings require pages in both
3258 * parent and child to be set to read.
3260 make_migration_entry_read(&swp_entry);
3261 entry = swp_entry_to_pte(swp_entry);
3262 set_huge_swap_pte_at(src, addr, src_pte,
3265 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3268 huge_ptep_set_wrprotect(src, addr, src_pte);
3269 mmu_notifier_invalidate_range(src, mmun_start,
3272 entry = huge_ptep_get(src_pte);
3273 ptepage = pte_page(entry);
3275 page_dup_rmap(ptepage, true);
3276 set_huge_pte_at(dst, addr, dst_pte, entry);
3277 hugetlb_count_add(pages_per_huge_page(h), dst);
3279 spin_unlock(src_ptl);
3280 spin_unlock(dst_ptl);
3284 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3289 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3290 unsigned long start, unsigned long end,
3291 struct page *ref_page)
3293 struct mm_struct *mm = vma->vm_mm;
3294 unsigned long address;
3299 struct hstate *h = hstate_vma(vma);
3300 unsigned long sz = huge_page_size(h);
3301 const unsigned long mmun_start = start; /* For mmu_notifiers */
3302 const unsigned long mmun_end = end; /* For mmu_notifiers */
3304 WARN_ON(!is_vm_hugetlb_page(vma));
3305 BUG_ON(start & ~huge_page_mask(h));
3306 BUG_ON(end & ~huge_page_mask(h));
3309 * This is a hugetlb vma, all the pte entries should point
3312 tlb_remove_check_page_size_change(tlb, sz);
3313 tlb_start_vma(tlb, vma);
3314 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3316 for (; address < end; address += sz) {
3317 ptep = huge_pte_offset(mm, address, sz);
3321 ptl = huge_pte_lock(h, mm, ptep);
3322 if (huge_pmd_unshare(mm, &address, ptep)) {
3327 pte = huge_ptep_get(ptep);
3328 if (huge_pte_none(pte)) {
3334 * Migrating hugepage or HWPoisoned hugepage is already
3335 * unmapped and its refcount is dropped, so just clear pte here.
3337 if (unlikely(!pte_present(pte))) {
3338 huge_pte_clear(mm, address, ptep, sz);
3343 page = pte_page(pte);
3345 * If a reference page is supplied, it is because a specific
3346 * page is being unmapped, not a range. Ensure the page we
3347 * are about to unmap is the actual page of interest.
3350 if (page != ref_page) {
3355 * Mark the VMA as having unmapped its page so that
3356 * future faults in this VMA will fail rather than
3357 * looking like data was lost
3359 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3362 pte = huge_ptep_get_and_clear(mm, address, ptep);
3363 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3364 if (huge_pte_dirty(pte))
3365 set_page_dirty(page);
3367 hugetlb_count_sub(pages_per_huge_page(h), mm);
3368 page_remove_rmap(page, true);
3371 tlb_remove_page_size(tlb, page, huge_page_size(h));
3373 * Bail out after unmapping reference page if supplied
3378 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3379 tlb_end_vma(tlb, vma);
3382 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3383 struct vm_area_struct *vma, unsigned long start,
3384 unsigned long end, struct page *ref_page)
3386 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3389 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3390 * test will fail on a vma being torn down, and not grab a page table
3391 * on its way out. We're lucky that the flag has such an appropriate
3392 * name, and can in fact be safely cleared here. We could clear it
3393 * before the __unmap_hugepage_range above, but all that's necessary
3394 * is to clear it before releasing the i_mmap_rwsem. This works
3395 * because in the context this is called, the VMA is about to be
3396 * destroyed and the i_mmap_rwsem is held.
3398 vma->vm_flags &= ~VM_MAYSHARE;
3401 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3402 unsigned long end, struct page *ref_page)
3404 struct mm_struct *mm;
3405 struct mmu_gather tlb;
3409 tlb_gather_mmu(&tlb, mm, start, end);
3410 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3411 tlb_finish_mmu(&tlb, start, end);
3415 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3416 * mappping it owns the reserve page for. The intention is to unmap the page
3417 * from other VMAs and let the children be SIGKILLed if they are faulting the
3420 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3421 struct page *page, unsigned long address)
3423 struct hstate *h = hstate_vma(vma);
3424 struct vm_area_struct *iter_vma;
3425 struct address_space *mapping;
3429 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3430 * from page cache lookup which is in HPAGE_SIZE units.
3432 address = address & huge_page_mask(h);
3433 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3435 mapping = vma->vm_file->f_mapping;
3438 * Take the mapping lock for the duration of the table walk. As
3439 * this mapping should be shared between all the VMAs,
3440 * __unmap_hugepage_range() is called as the lock is already held
3442 i_mmap_lock_write(mapping);
3443 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3444 /* Do not unmap the current VMA */
3445 if (iter_vma == vma)
3449 * Shared VMAs have their own reserves and do not affect
3450 * MAP_PRIVATE accounting but it is possible that a shared
3451 * VMA is using the same page so check and skip such VMAs.
3453 if (iter_vma->vm_flags & VM_MAYSHARE)
3457 * Unmap the page from other VMAs without their own reserves.
3458 * They get marked to be SIGKILLed if they fault in these
3459 * areas. This is because a future no-page fault on this VMA
3460 * could insert a zeroed page instead of the data existing
3461 * from the time of fork. This would look like data corruption
3463 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3464 unmap_hugepage_range(iter_vma, address,
3465 address + huge_page_size(h), page);
3467 i_mmap_unlock_write(mapping);
3471 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3472 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3473 * cannot race with other handlers or page migration.
3474 * Keep the pte_same checks anyway to make transition from the mutex easier.
3476 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3477 unsigned long address, pte_t *ptep,
3478 struct page *pagecache_page, spinlock_t *ptl)
3481 struct hstate *h = hstate_vma(vma);
3482 struct page *old_page, *new_page;
3483 int ret = 0, outside_reserve = 0;
3484 unsigned long mmun_start; /* For mmu_notifiers */
3485 unsigned long mmun_end; /* For mmu_notifiers */
3487 pte = huge_ptep_get(ptep);
3488 old_page = pte_page(pte);
3491 /* If no-one else is actually using this page, avoid the copy
3492 * and just make the page writable */
3493 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3494 page_move_anon_rmap(old_page, vma);
3495 set_huge_ptep_writable(vma, address, ptep);
3500 * If the process that created a MAP_PRIVATE mapping is about to
3501 * perform a COW due to a shared page count, attempt to satisfy
3502 * the allocation without using the existing reserves. The pagecache
3503 * page is used to determine if the reserve at this address was
3504 * consumed or not. If reserves were used, a partial faulted mapping
3505 * at the time of fork() could consume its reserves on COW instead
3506 * of the full address range.
3508 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3509 old_page != pagecache_page)
3510 outside_reserve = 1;
3515 * Drop page table lock as buddy allocator may be called. It will
3516 * be acquired again before returning to the caller, as expected.
3519 new_page = alloc_huge_page(vma, address, outside_reserve);
3521 if (IS_ERR(new_page)) {
3523 * If a process owning a MAP_PRIVATE mapping fails to COW,
3524 * it is due to references held by a child and an insufficient
3525 * huge page pool. To guarantee the original mappers
3526 * reliability, unmap the page from child processes. The child
3527 * may get SIGKILLed if it later faults.
3529 if (outside_reserve) {
3531 BUG_ON(huge_pte_none(pte));
3532 unmap_ref_private(mm, vma, old_page, address);
3533 BUG_ON(huge_pte_none(pte));
3535 ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3538 pte_same(huge_ptep_get(ptep), pte)))
3539 goto retry_avoidcopy;
3541 * race occurs while re-acquiring page table
3542 * lock, and our job is done.
3547 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3548 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3549 goto out_release_old;
3553 * When the original hugepage is shared one, it does not have
3554 * anon_vma prepared.
3556 if (unlikely(anon_vma_prepare(vma))) {
3558 goto out_release_all;
3561 copy_user_huge_page(new_page, old_page, address, vma,
3562 pages_per_huge_page(h));
3563 __SetPageUptodate(new_page);
3564 set_page_huge_active(new_page);
3566 mmun_start = address & huge_page_mask(h);
3567 mmun_end = mmun_start + huge_page_size(h);
3568 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3571 * Retake the page table lock to check for racing updates
3572 * before the page tables are altered
3575 ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3577 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3578 ClearPagePrivate(new_page);
3581 huge_ptep_clear_flush(vma, address, ptep);
3582 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3583 set_huge_pte_at(mm, address, ptep,
3584 make_huge_pte(vma, new_page, 1));
3585 page_remove_rmap(old_page, true);
3586 hugepage_add_new_anon_rmap(new_page, vma, address);
3587 /* Make the old page be freed below */
3588 new_page = old_page;
3591 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3593 restore_reserve_on_error(h, vma, address, new_page);
3598 spin_lock(ptl); /* Caller expects lock to be held */
3602 /* Return the pagecache page at a given address within a VMA */
3603 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3604 struct vm_area_struct *vma, unsigned long address)
3606 struct address_space *mapping;
3609 mapping = vma->vm_file->f_mapping;
3610 idx = vma_hugecache_offset(h, vma, address);
3612 return find_lock_page(mapping, idx);
3616 * Return whether there is a pagecache page to back given address within VMA.
3617 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3619 static bool hugetlbfs_pagecache_present(struct hstate *h,
3620 struct vm_area_struct *vma, unsigned long address)
3622 struct address_space *mapping;
3626 mapping = vma->vm_file->f_mapping;
3627 idx = vma_hugecache_offset(h, vma, address);
3629 page = find_get_page(mapping, idx);
3632 return page != NULL;
3635 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3638 struct inode *inode = mapping->host;
3639 struct hstate *h = hstate_inode(inode);
3640 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3644 ClearPagePrivate(page);
3646 spin_lock(&inode->i_lock);
3647 inode->i_blocks += blocks_per_huge_page(h);
3648 spin_unlock(&inode->i_lock);
3652 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3653 struct address_space *mapping, pgoff_t idx,
3654 unsigned long address, pte_t *ptep, unsigned int flags)
3656 struct hstate *h = hstate_vma(vma);
3657 int ret = VM_FAULT_SIGBUS;
3665 * Currently, we are forced to kill the process in the event the
3666 * original mapper has unmapped pages from the child due to a failed
3667 * COW. Warn that such a situation has occurred as it may not be obvious
3669 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3670 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3676 * Use page lock to guard against racing truncation
3677 * before we get page_table_lock.
3680 page = find_lock_page(mapping, idx);
3682 size = i_size_read(mapping->host) >> huge_page_shift(h);
3687 * Check for page in userfault range
3689 if (userfaultfd_missing(vma)) {
3691 struct vm_fault vmf = {
3696 * Hard to debug if it ends up being
3697 * used by a callee that assumes
3698 * something about the other
3699 * uninitialized fields... same as in
3705 * hugetlb_fault_mutex must be dropped before
3706 * handling userfault. Reacquire after handling
3707 * fault to make calling code simpler.
3709 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping,
3711 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3712 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3713 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3717 page = alloc_huge_page(vma, address, 0);
3719 ret = PTR_ERR(page);
3723 ret = VM_FAULT_SIGBUS;
3726 clear_huge_page(page, address, pages_per_huge_page(h));
3727 __SetPageUptodate(page);
3728 set_page_huge_active(page);
3730 if (vma->vm_flags & VM_MAYSHARE) {
3731 int err = huge_add_to_page_cache(page, mapping, idx);
3740 if (unlikely(anon_vma_prepare(vma))) {
3742 goto backout_unlocked;
3748 * If memory error occurs between mmap() and fault, some process
3749 * don't have hwpoisoned swap entry for errored virtual address.
3750 * So we need to block hugepage fault by PG_hwpoison bit check.
3752 if (unlikely(PageHWPoison(page))) {
3753 ret = VM_FAULT_HWPOISON |
3754 VM_FAULT_SET_HINDEX(hstate_index(h));
3755 goto backout_unlocked;
3760 * If we are going to COW a private mapping later, we examine the
3761 * pending reservations for this page now. This will ensure that
3762 * any allocations necessary to record that reservation occur outside
3765 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3766 if (vma_needs_reservation(h, vma, address) < 0) {
3768 goto backout_unlocked;
3770 /* Just decrements count, does not deallocate */
3771 vma_end_reservation(h, vma, address);
3774 ptl = huge_pte_lock(h, mm, ptep);
3775 size = i_size_read(mapping->host) >> huge_page_shift(h);
3780 if (!huge_pte_none(huge_ptep_get(ptep)))
3784 ClearPagePrivate(page);
3785 hugepage_add_new_anon_rmap(page, vma, address);
3787 page_dup_rmap(page, true);
3788 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3789 && (vma->vm_flags & VM_SHARED)));
3790 set_huge_pte_at(mm, address, ptep, new_pte);
3792 hugetlb_count_add(pages_per_huge_page(h), mm);
3793 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3794 /* Optimization, do the COW without a second fault */
3795 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
3807 restore_reserve_on_error(h, vma, address, page);
3813 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3814 struct vm_area_struct *vma,
3815 struct address_space *mapping,
3816 pgoff_t idx, unsigned long address)
3818 unsigned long key[2];
3821 if (vma->vm_flags & VM_SHARED) {
3822 key[0] = (unsigned long) mapping;
3825 key[0] = (unsigned long) mm;
3826 key[1] = address >> huge_page_shift(h);
3829 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3831 return hash & (num_fault_mutexes - 1);
3835 * For uniprocesor systems we always use a single mutex, so just
3836 * return 0 and avoid the hashing overhead.
3838 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3839 struct vm_area_struct *vma,
3840 struct address_space *mapping,
3841 pgoff_t idx, unsigned long address)
3847 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3848 unsigned long address, unsigned int flags)
3855 struct page *page = NULL;
3856 struct page *pagecache_page = NULL;
3857 struct hstate *h = hstate_vma(vma);
3858 struct address_space *mapping;
3859 int need_wait_lock = 0;
3861 address &= huge_page_mask(h);
3863 ptep = huge_pte_offset(mm, address, huge_page_size(h));
3865 entry = huge_ptep_get(ptep);
3866 if (unlikely(is_hugetlb_entry_migration(entry))) {
3867 migration_entry_wait_huge(vma, mm, ptep);
3869 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3870 return VM_FAULT_HWPOISON_LARGE |
3871 VM_FAULT_SET_HINDEX(hstate_index(h));
3873 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3875 return VM_FAULT_OOM;
3878 mapping = vma->vm_file->f_mapping;
3879 idx = vma_hugecache_offset(h, vma, address);
3882 * Serialize hugepage allocation and instantiation, so that we don't
3883 * get spurious allocation failures if two CPUs race to instantiate
3884 * the same page in the page cache.
3886 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3887 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3889 entry = huge_ptep_get(ptep);
3890 if (huge_pte_none(entry)) {
3891 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3898 * entry could be a migration/hwpoison entry at this point, so this
3899 * check prevents the kernel from going below assuming that we have
3900 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3901 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3904 if (!pte_present(entry))
3908 * If we are going to COW the mapping later, we examine the pending
3909 * reservations for this page now. This will ensure that any
3910 * allocations necessary to record that reservation occur outside the
3911 * spinlock. For private mappings, we also lookup the pagecache
3912 * page now as it is used to determine if a reservation has been
3915 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3916 if (vma_needs_reservation(h, vma, address) < 0) {
3920 /* Just decrements count, does not deallocate */
3921 vma_end_reservation(h, vma, address);
3923 if (!(vma->vm_flags & VM_MAYSHARE))
3924 pagecache_page = hugetlbfs_pagecache_page(h,
3928 ptl = huge_pte_lock(h, mm, ptep);
3930 /* Check for a racing update before calling hugetlb_cow */
3931 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3935 * hugetlb_cow() requires page locks of pte_page(entry) and
3936 * pagecache_page, so here we need take the former one
3937 * when page != pagecache_page or !pagecache_page.
3939 page = pte_page(entry);
3940 if (page != pagecache_page)
3941 if (!trylock_page(page)) {
3948 if (flags & FAULT_FLAG_WRITE) {
3949 if (!huge_pte_write(entry)) {
3950 ret = hugetlb_cow(mm, vma, address, ptep,
3951 pagecache_page, ptl);
3954 entry = huge_pte_mkdirty(entry);
3956 entry = pte_mkyoung(entry);
3957 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3958 flags & FAULT_FLAG_WRITE))
3959 update_mmu_cache(vma, address, ptep);
3961 if (page != pagecache_page)
3967 if (pagecache_page) {
3968 unlock_page(pagecache_page);
3969 put_page(pagecache_page);
3972 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3974 * Generally it's safe to hold refcount during waiting page lock. But
3975 * here we just wait to defer the next page fault to avoid busy loop and
3976 * the page is not used after unlocked before returning from the current
3977 * page fault. So we are safe from accessing freed page, even if we wait
3978 * here without taking refcount.
3981 wait_on_page_locked(page);
3986 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
3987 * modifications for huge pages.
3989 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
3991 struct vm_area_struct *dst_vma,
3992 unsigned long dst_addr,
3993 unsigned long src_addr,
3994 struct page **pagep)
3996 struct address_space *mapping;
3999 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4000 struct hstate *h = hstate_vma(dst_vma);
4008 page = alloc_huge_page(dst_vma, dst_addr, 0);
4012 ret = copy_huge_page_from_user(page,
4013 (const void __user *) src_addr,
4014 pages_per_huge_page(h), false);
4016 /* fallback to copy_from_user outside mmap_sem */
4017 if (unlikely(ret)) {
4020 /* don't free the page */
4029 * The memory barrier inside __SetPageUptodate makes sure that
4030 * preceding stores to the page contents become visible before
4031 * the set_pte_at() write.
4033 __SetPageUptodate(page);
4034 set_page_huge_active(page);
4036 mapping = dst_vma->vm_file->f_mapping;
4037 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4040 * If shared, add to page cache
4043 size = i_size_read(mapping->host) >> huge_page_shift(h);
4046 goto out_release_nounlock;
4049 * Serialization between remove_inode_hugepages() and
4050 * huge_add_to_page_cache() below happens through the
4051 * hugetlb_fault_mutex_table that here must be hold by
4054 ret = huge_add_to_page_cache(page, mapping, idx);
4056 goto out_release_nounlock;
4059 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4063 * Recheck the i_size after holding PT lock to make sure not
4064 * to leave any page mapped (as page_mapped()) beyond the end
4065 * of the i_size (remove_inode_hugepages() is strict about
4066 * enforcing that). If we bail out here, we'll also leave a
4067 * page in the radix tree in the vm_shared case beyond the end
4068 * of the i_size, but remove_inode_hugepages() will take care
4069 * of it as soon as we drop the hugetlb_fault_mutex_table.
4071 size = i_size_read(mapping->host) >> huge_page_shift(h);
4074 goto out_release_unlock;
4077 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4078 goto out_release_unlock;
4081 page_dup_rmap(page, true);
4083 ClearPagePrivate(page);
4084 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4087 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4088 if (dst_vma->vm_flags & VM_WRITE)
4089 _dst_pte = huge_pte_mkdirty(_dst_pte);
4090 _dst_pte = pte_mkyoung(_dst_pte);
4092 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4094 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4095 dst_vma->vm_flags & VM_WRITE);
4096 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4098 /* No need to invalidate - it was non-present before */
4099 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4111 out_release_nounlock:
4116 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4117 struct page **pages, struct vm_area_struct **vmas,
4118 unsigned long *position, unsigned long *nr_pages,
4119 long i, unsigned int flags, int *nonblocking)
4121 unsigned long pfn_offset;
4122 unsigned long vaddr = *position;
4123 unsigned long remainder = *nr_pages;
4124 struct hstate *h = hstate_vma(vma);
4127 while (vaddr < vma->vm_end && remainder) {
4129 spinlock_t *ptl = NULL;
4134 * If we have a pending SIGKILL, don't keep faulting pages and
4135 * potentially allocating memory.
4137 if (unlikely(fatal_signal_pending(current))) {
4143 * Some archs (sparc64, sh*) have multiple pte_ts to
4144 * each hugepage. We have to make sure we get the
4145 * first, for the page indexing below to work.
4147 * Note that page table lock is not held when pte is null.
4149 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4152 ptl = huge_pte_lock(h, mm, pte);
4153 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4156 * When coredumping, it suits get_dump_page if we just return
4157 * an error where there's an empty slot with no huge pagecache
4158 * to back it. This way, we avoid allocating a hugepage, and
4159 * the sparse dumpfile avoids allocating disk blocks, but its
4160 * huge holes still show up with zeroes where they need to be.
4162 if (absent && (flags & FOLL_DUMP) &&
4163 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4171 * We need call hugetlb_fault for both hugepages under migration
4172 * (in which case hugetlb_fault waits for the migration,) and
4173 * hwpoisoned hugepages (in which case we need to prevent the
4174 * caller from accessing to them.) In order to do this, we use
4175 * here is_swap_pte instead of is_hugetlb_entry_migration and
4176 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4177 * both cases, and because we can't follow correct pages
4178 * directly from any kind of swap entries.
4180 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4181 ((flags & FOLL_WRITE) &&
4182 !huge_pte_write(huge_ptep_get(pte)))) {
4184 unsigned int fault_flags = 0;
4188 if (flags & FOLL_WRITE)
4189 fault_flags |= FAULT_FLAG_WRITE;
4191 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4192 if (flags & FOLL_NOWAIT)
4193 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4194 FAULT_FLAG_RETRY_NOWAIT;
4195 if (flags & FOLL_TRIED) {
4196 VM_WARN_ON_ONCE(fault_flags &
4197 FAULT_FLAG_ALLOW_RETRY);
4198 fault_flags |= FAULT_FLAG_TRIED;
4200 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4201 if (ret & VM_FAULT_ERROR) {
4202 err = vm_fault_to_errno(ret, flags);
4206 if (ret & VM_FAULT_RETRY) {
4211 * VM_FAULT_RETRY must not return an
4212 * error, it will return zero
4215 * No need to update "position" as the
4216 * caller will not check it after
4217 * *nr_pages is set to 0.
4224 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4225 page = pte_page(huge_ptep_get(pte));
4228 pages[i] = mem_map_offset(page, pfn_offset);
4239 if (vaddr < vma->vm_end && remainder &&
4240 pfn_offset < pages_per_huge_page(h)) {
4242 * We use pfn_offset to avoid touching the pageframes
4243 * of this compound page.
4249 *nr_pages = remainder;
4251 * setting position is actually required only if remainder is
4252 * not zero but it's faster not to add a "if (remainder)"
4260 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4262 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4265 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4268 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4269 unsigned long address, unsigned long end, pgprot_t newprot)
4271 struct mm_struct *mm = vma->vm_mm;
4272 unsigned long start = address;
4275 struct hstate *h = hstate_vma(vma);
4276 unsigned long pages = 0;
4278 BUG_ON(address >= end);
4279 flush_cache_range(vma, address, end);
4281 mmu_notifier_invalidate_range_start(mm, start, end);
4282 i_mmap_lock_write(vma->vm_file->f_mapping);
4283 for (; address < end; address += huge_page_size(h)) {
4285 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4288 ptl = huge_pte_lock(h, mm, ptep);
4289 if (huge_pmd_unshare(mm, &address, ptep)) {
4294 pte = huge_ptep_get(ptep);
4295 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4299 if (unlikely(is_hugetlb_entry_migration(pte))) {
4300 swp_entry_t entry = pte_to_swp_entry(pte);
4302 if (is_write_migration_entry(entry)) {
4305 make_migration_entry_read(&entry);
4306 newpte = swp_entry_to_pte(entry);
4307 set_huge_swap_pte_at(mm, address, ptep,
4308 newpte, huge_page_size(h));
4314 if (!huge_pte_none(pte)) {
4315 pte = huge_ptep_get_and_clear(mm, address, ptep);
4316 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
4317 pte = arch_make_huge_pte(pte, vma, NULL, 0);
4318 set_huge_pte_at(mm, address, ptep, pte);
4324 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4325 * may have cleared our pud entry and done put_page on the page table:
4326 * once we release i_mmap_rwsem, another task can do the final put_page
4327 * and that page table be reused and filled with junk.
4329 flush_hugetlb_tlb_range(vma, start, end);
4330 mmu_notifier_invalidate_range(mm, start, end);
4331 i_mmap_unlock_write(vma->vm_file->f_mapping);
4332 mmu_notifier_invalidate_range_end(mm, start, end);
4334 return pages << h->order;
4337 int hugetlb_reserve_pages(struct inode *inode,
4339 struct vm_area_struct *vma,
4340 vm_flags_t vm_flags)
4343 struct hstate *h = hstate_inode(inode);
4344 struct hugepage_subpool *spool = subpool_inode(inode);
4345 struct resv_map *resv_map;
4348 /* This should never happen */
4350 VM_WARN(1, "%s called with a negative range\n", __func__);
4355 * Only apply hugepage reservation if asked. At fault time, an
4356 * attempt will be made for VM_NORESERVE to allocate a page
4357 * without using reserves
4359 if (vm_flags & VM_NORESERVE)
4363 * Shared mappings base their reservation on the number of pages that
4364 * are already allocated on behalf of the file. Private mappings need
4365 * to reserve the full area even if read-only as mprotect() may be
4366 * called to make the mapping read-write. Assume !vma is a shm mapping
4368 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4369 resv_map = inode_resv_map(inode);
4371 chg = region_chg(resv_map, from, to);
4374 resv_map = resv_map_alloc();
4380 set_vma_resv_map(vma, resv_map);
4381 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4390 * There must be enough pages in the subpool for the mapping. If
4391 * the subpool has a minimum size, there may be some global
4392 * reservations already in place (gbl_reserve).
4394 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4395 if (gbl_reserve < 0) {
4401 * Check enough hugepages are available for the reservation.
4402 * Hand the pages back to the subpool if there are not
4404 ret = hugetlb_acct_memory(h, gbl_reserve);
4406 /* put back original number of pages, chg */
4407 (void)hugepage_subpool_put_pages(spool, chg);
4412 * Account for the reservations made. Shared mappings record regions
4413 * that have reservations as they are shared by multiple VMAs.
4414 * When the last VMA disappears, the region map says how much
4415 * the reservation was and the page cache tells how much of
4416 * the reservation was consumed. Private mappings are per-VMA and
4417 * only the consumed reservations are tracked. When the VMA
4418 * disappears, the original reservation is the VMA size and the
4419 * consumed reservations are stored in the map. Hence, nothing
4420 * else has to be done for private mappings here
4422 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4423 long add = region_add(resv_map, from, to);
4425 if (unlikely(chg > add)) {
4427 * pages in this range were added to the reserve
4428 * map between region_chg and region_add. This
4429 * indicates a race with alloc_huge_page. Adjust
4430 * the subpool and reserve counts modified above
4431 * based on the difference.
4435 rsv_adjust = hugepage_subpool_put_pages(spool,
4437 hugetlb_acct_memory(h, -rsv_adjust);
4442 if (!vma || vma->vm_flags & VM_MAYSHARE)
4443 /* Don't call region_abort if region_chg failed */
4445 region_abort(resv_map, from, to);
4446 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4447 kref_put(&resv_map->refs, resv_map_release);
4451 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4454 struct hstate *h = hstate_inode(inode);
4455 struct resv_map *resv_map = inode_resv_map(inode);
4457 struct hugepage_subpool *spool = subpool_inode(inode);
4461 chg = region_del(resv_map, start, end);
4463 * region_del() can fail in the rare case where a region
4464 * must be split and another region descriptor can not be
4465 * allocated. If end == LONG_MAX, it will not fail.
4471 spin_lock(&inode->i_lock);
4472 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4473 spin_unlock(&inode->i_lock);
4476 * If the subpool has a minimum size, the number of global
4477 * reservations to be released may be adjusted.
4479 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4480 hugetlb_acct_memory(h, -gbl_reserve);
4485 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4486 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4487 struct vm_area_struct *vma,
4488 unsigned long addr, pgoff_t idx)
4490 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4492 unsigned long sbase = saddr & PUD_MASK;
4493 unsigned long s_end = sbase + PUD_SIZE;
4495 /* Allow segments to share if only one is marked locked */
4496 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4497 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4500 * match the virtual addresses, permission and the alignment of the
4503 if (pmd_index(addr) != pmd_index(saddr) ||
4504 vm_flags != svm_flags ||
4505 sbase < svma->vm_start || svma->vm_end < s_end)
4511 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4513 unsigned long base = addr & PUD_MASK;
4514 unsigned long end = base + PUD_SIZE;
4517 * check on proper vm_flags and page table alignment
4519 if (vma->vm_flags & VM_MAYSHARE &&
4520 vma->vm_start <= base && end <= vma->vm_end)
4526 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4527 * and returns the corresponding pte. While this is not necessary for the
4528 * !shared pmd case because we can allocate the pmd later as well, it makes the
4529 * code much cleaner. pmd allocation is essential for the shared case because
4530 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4531 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4532 * bad pmd for sharing.
4534 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4536 struct vm_area_struct *vma = find_vma(mm, addr);
4537 struct address_space *mapping = vma->vm_file->f_mapping;
4538 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4540 struct vm_area_struct *svma;
4541 unsigned long saddr;
4546 if (!vma_shareable(vma, addr))
4547 return (pte_t *)pmd_alloc(mm, pud, addr);
4549 i_mmap_lock_write(mapping);
4550 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4554 saddr = page_table_shareable(svma, vma, addr, idx);
4556 spte = huge_pte_offset(svma->vm_mm, saddr,
4557 vma_mmu_pagesize(svma));
4559 get_page(virt_to_page(spte));
4568 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4569 if (pud_none(*pud)) {
4570 pud_populate(mm, pud,
4571 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4574 put_page(virt_to_page(spte));
4578 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4579 i_mmap_unlock_write(mapping);
4584 * unmap huge page backed by shared pte.
4586 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4587 * indicated by page_count > 1, unmap is achieved by clearing pud and
4588 * decrementing the ref count. If count == 1, the pte page is not shared.
4590 * called with page table lock held.
4592 * returns: 1 successfully unmapped a shared pte page
4593 * 0 the underlying pte page is not shared, or it is the last user
4595 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4597 pgd_t *pgd = pgd_offset(mm, *addr);
4598 p4d_t *p4d = p4d_offset(pgd, *addr);
4599 pud_t *pud = pud_offset(p4d, *addr);
4601 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4602 if (page_count(virt_to_page(ptep)) == 1)
4606 put_page(virt_to_page(ptep));
4608 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4611 #define want_pmd_share() (1)
4612 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4613 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4618 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4622 #define want_pmd_share() (0)
4623 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4625 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4626 pte_t *huge_pte_alloc(struct mm_struct *mm,
4627 unsigned long addr, unsigned long sz)
4634 pgd = pgd_offset(mm, addr);
4635 p4d = p4d_alloc(mm, pgd, addr);
4638 pud = pud_alloc(mm, p4d, addr);
4640 if (sz == PUD_SIZE) {
4643 BUG_ON(sz != PMD_SIZE);
4644 if (want_pmd_share() && pud_none(*pud))
4645 pte = huge_pmd_share(mm, addr, pud);
4647 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4650 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4656 * huge_pte_offset() - Walk the page table to resolve the hugepage
4657 * entry at address @addr
4659 * Return: Pointer to page table or swap entry (PUD or PMD) for
4660 * address @addr, or NULL if a p*d_none() entry is encountered and the
4661 * size @sz doesn't match the hugepage size at this level of the page
4664 pte_t *huge_pte_offset(struct mm_struct *mm,
4665 unsigned long addr, unsigned long sz)
4672 pgd = pgd_offset(mm, addr);
4673 if (!pgd_present(*pgd))
4675 p4d = p4d_offset(pgd, addr);
4676 if (!p4d_present(*p4d))
4679 pud = pud_offset(p4d, addr);
4680 if (sz != PUD_SIZE && pud_none(*pud))
4682 /* hugepage or swap? */
4683 if (pud_huge(*pud) || !pud_present(*pud))
4684 return (pte_t *)pud;
4686 pmd = pmd_offset(pud, addr);
4687 if (sz != PMD_SIZE && pmd_none(*pmd))
4689 /* hugepage or swap? */
4690 if (pmd_huge(*pmd) || !pmd_present(*pmd))
4691 return (pte_t *)pmd;
4696 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4699 * These functions are overwritable if your architecture needs its own
4702 struct page * __weak
4703 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4706 return ERR_PTR(-EINVAL);
4709 struct page * __weak
4710 follow_huge_pd(struct vm_area_struct *vma,
4711 unsigned long address, hugepd_t hpd, int flags, int pdshift)
4713 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4717 struct page * __weak
4718 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4719 pmd_t *pmd, int flags)
4721 struct page *page = NULL;
4725 ptl = pmd_lockptr(mm, pmd);
4728 * make sure that the address range covered by this pmd is not
4729 * unmapped from other threads.
4731 if (!pmd_huge(*pmd))
4733 pte = huge_ptep_get((pte_t *)pmd);
4734 if (pte_present(pte)) {
4735 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4736 if (flags & FOLL_GET)
4739 if (is_hugetlb_entry_migration(pte)) {
4741 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4745 * hwpoisoned entry is treated as no_page_table in
4746 * follow_page_mask().
4754 struct page * __weak
4755 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4756 pud_t *pud, int flags)
4758 if (flags & FOLL_GET)
4761 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4764 struct page * __weak
4765 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
4767 if (flags & FOLL_GET)
4770 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
4773 bool isolate_huge_page(struct page *page, struct list_head *list)
4777 VM_BUG_ON_PAGE(!PageHead(page), page);
4778 spin_lock(&hugetlb_lock);
4779 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4783 clear_page_huge_active(page);
4784 list_move_tail(&page->lru, list);
4786 spin_unlock(&hugetlb_lock);
4790 void putback_active_hugepage(struct page *page)
4792 VM_BUG_ON_PAGE(!PageHead(page), page);
4793 spin_lock(&hugetlb_lock);
4794 set_page_huge_active(page);
4795 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4796 spin_unlock(&hugetlb_lock);