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/sched/signal.h>
22 #include <linux/rmap.h>
23 #include <linux/string_helpers.h>
24 #include <linux/swap.h>
25 #include <linux/swapops.h>
26 #include <linux/jhash.h>
29 #include <asm/pgtable.h>
33 #include <linux/hugetlb.h>
34 #include <linux/hugetlb_cgroup.h>
35 #include <linux/node.h>
36 #include <linux/userfaultfd_k.h>
39 int hugetlb_max_hstate __read_mostly;
40 unsigned int default_hstate_idx;
41 struct hstate hstates[HUGE_MAX_HSTATE];
43 * Minimum page order among possible hugepage sizes, set to a proper value
46 static unsigned int minimum_order __read_mostly = UINT_MAX;
48 __initdata LIST_HEAD(huge_boot_pages);
50 /* for command line parsing */
51 static struct hstate * __initdata parsed_hstate;
52 static unsigned long __initdata default_hstate_max_huge_pages;
53 static unsigned long __initdata default_hstate_size;
54 static bool __initdata parsed_valid_hugepagesz = true;
57 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
58 * free_huge_pages, and surplus_huge_pages.
60 DEFINE_SPINLOCK(hugetlb_lock);
63 * Serializes faults on the same logical page. This is used to
64 * prevent spurious OOMs when the hugepage pool is fully utilized.
66 static int num_fault_mutexes;
67 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
69 /* Forward declaration */
70 static int hugetlb_acct_memory(struct hstate *h, long delta);
72 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
74 bool free = (spool->count == 0) && (spool->used_hpages == 0);
76 spin_unlock(&spool->lock);
78 /* If no pages are used, and no other handles to the subpool
79 * remain, give up any reservations mased on minimum size and
82 if (spool->min_hpages != -1)
83 hugetlb_acct_memory(spool->hstate,
89 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
92 struct hugepage_subpool *spool;
94 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
98 spin_lock_init(&spool->lock);
100 spool->max_hpages = max_hpages;
102 spool->min_hpages = min_hpages;
104 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
108 spool->rsv_hpages = min_hpages;
113 void hugepage_put_subpool(struct hugepage_subpool *spool)
115 spin_lock(&spool->lock);
116 BUG_ON(!spool->count);
118 unlock_or_release_subpool(spool);
122 * Subpool accounting for allocating and reserving pages.
123 * Return -ENOMEM if there are not enough resources to satisfy the
124 * the request. Otherwise, return the number of pages by which the
125 * global pools must be adjusted (upward). The returned value may
126 * only be different than the passed value (delta) in the case where
127 * a subpool minimum size must be manitained.
129 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
137 spin_lock(&spool->lock);
139 if (spool->max_hpages != -1) { /* maximum size accounting */
140 if ((spool->used_hpages + delta) <= spool->max_hpages)
141 spool->used_hpages += delta;
148 /* minimum size accounting */
149 if (spool->min_hpages != -1 && spool->rsv_hpages) {
150 if (delta > spool->rsv_hpages) {
152 * Asking for more reserves than those already taken on
153 * behalf of subpool. Return difference.
155 ret = delta - spool->rsv_hpages;
156 spool->rsv_hpages = 0;
158 ret = 0; /* reserves already accounted for */
159 spool->rsv_hpages -= delta;
164 spin_unlock(&spool->lock);
169 * Subpool accounting for freeing and unreserving pages.
170 * Return the number of global page reservations that must be dropped.
171 * The return value may only be different than the passed value (delta)
172 * in the case where a subpool minimum size must be maintained.
174 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
182 spin_lock(&spool->lock);
184 if (spool->max_hpages != -1) /* maximum size accounting */
185 spool->used_hpages -= delta;
187 /* minimum size accounting */
188 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
189 if (spool->rsv_hpages + delta <= spool->min_hpages)
192 ret = spool->rsv_hpages + delta - spool->min_hpages;
194 spool->rsv_hpages += delta;
195 if (spool->rsv_hpages > spool->min_hpages)
196 spool->rsv_hpages = spool->min_hpages;
200 * If hugetlbfs_put_super couldn't free spool due to an outstanding
201 * quota reference, free it now.
203 unlock_or_release_subpool(spool);
208 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
210 return HUGETLBFS_SB(inode->i_sb)->spool;
213 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
215 return subpool_inode(file_inode(vma->vm_file));
219 * Region tracking -- allows tracking of reservations and instantiated pages
220 * across the pages in a mapping.
222 * The region data structures are embedded into a resv_map and protected
223 * by a resv_map's lock. The set of regions within the resv_map represent
224 * reservations for huge pages, or huge pages that have already been
225 * instantiated within the map. The from and to elements are huge page
226 * indicies into the associated mapping. from indicates the starting index
227 * of the region. to represents the first index past the end of the region.
229 * For example, a file region structure with from == 0 and to == 4 represents
230 * four huge pages in a mapping. It is important to note that the to element
231 * represents the first element past the end of the region. This is used in
232 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
234 * Interval notation of the form [from, to) will be used to indicate that
235 * the endpoint from is inclusive and to is exclusive.
238 struct list_head link;
244 * Add the huge page range represented by [f, t) to the reserve
245 * map. In the normal case, existing regions will be expanded
246 * to accommodate the specified range. Sufficient regions should
247 * exist for expansion due to the previous call to region_chg
248 * with the same range. However, it is possible that region_del
249 * could have been called after region_chg and modifed the map
250 * in such a way that no region exists to be expanded. In this
251 * case, pull a region descriptor from the cache associated with
252 * the map and use that for the new range.
254 * Return the number of new huge pages added to the map. This
255 * number is greater than or equal to zero.
257 static long region_add(struct resv_map *resv, long f, long t)
259 struct list_head *head = &resv->regions;
260 struct file_region *rg, *nrg, *trg;
263 spin_lock(&resv->lock);
264 /* Locate the region we are either in or before. */
265 list_for_each_entry(rg, head, link)
270 * If no region exists which can be expanded to include the
271 * specified range, the list must have been modified by an
272 * interleving call to region_del(). Pull a region descriptor
273 * from the cache and use it for this range.
275 if (&rg->link == head || t < rg->from) {
276 VM_BUG_ON(resv->region_cache_count <= 0);
278 resv->region_cache_count--;
279 nrg = list_first_entry(&resv->region_cache, struct file_region,
281 list_del(&nrg->link);
285 list_add(&nrg->link, rg->link.prev);
291 /* Round our left edge to the current segment if it encloses us. */
295 /* Check for and consume any regions we now overlap with. */
297 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
298 if (&rg->link == head)
303 /* If this area reaches higher then extend our area to
304 * include it completely. If this is not the first area
305 * which we intend to reuse, free it. */
309 /* Decrement return value by the deleted range.
310 * Another range will span this area so that by
311 * end of routine add will be >= zero
313 add -= (rg->to - rg->from);
319 add += (nrg->from - f); /* Added to beginning of region */
321 add += t - nrg->to; /* Added to end of region */
325 resv->adds_in_progress--;
326 spin_unlock(&resv->lock);
332 * Examine the existing reserve map and determine how many
333 * huge pages in the specified range [f, t) are NOT currently
334 * represented. This routine is called before a subsequent
335 * call to region_add that will actually modify the reserve
336 * map to add the specified range [f, t). region_chg does
337 * not change the number of huge pages represented by the
338 * map. However, if the existing regions in the map can not
339 * be expanded to represent the new range, a new file_region
340 * structure is added to the map as a placeholder. This is
341 * so that the subsequent region_add call will have all the
342 * regions it needs and will not fail.
344 * Upon entry, region_chg will also examine the cache of region descriptors
345 * associated with the map. If there are not enough descriptors cached, one
346 * will be allocated for the in progress add operation.
348 * Returns the number of huge pages that need to be added to the existing
349 * reservation map for the range [f, t). This number is greater or equal to
350 * zero. -ENOMEM is returned if a new file_region structure or cache entry
351 * is needed and can not be allocated.
353 static long region_chg(struct resv_map *resv, long f, long t)
355 struct list_head *head = &resv->regions;
356 struct file_region *rg, *nrg = NULL;
360 spin_lock(&resv->lock);
362 resv->adds_in_progress++;
365 * Check for sufficient descriptors in the cache to accommodate
366 * the number of in progress add operations.
368 if (resv->adds_in_progress > resv->region_cache_count) {
369 struct file_region *trg;
371 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
372 /* Must drop lock to allocate a new descriptor. */
373 resv->adds_in_progress--;
374 spin_unlock(&resv->lock);
376 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
382 spin_lock(&resv->lock);
383 list_add(&trg->link, &resv->region_cache);
384 resv->region_cache_count++;
388 /* Locate the region we are before or in. */
389 list_for_each_entry(rg, head, link)
393 /* If we are below the current region then a new region is required.
394 * Subtle, allocate a new region at the position but make it zero
395 * size such that we can guarantee to record the reservation. */
396 if (&rg->link == head || t < rg->from) {
398 resv->adds_in_progress--;
399 spin_unlock(&resv->lock);
400 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
406 INIT_LIST_HEAD(&nrg->link);
410 list_add(&nrg->link, rg->link.prev);
415 /* Round our left edge to the current segment if it encloses us. */
420 /* Check for and consume any regions we now overlap with. */
421 list_for_each_entry(rg, rg->link.prev, link) {
422 if (&rg->link == head)
427 /* We overlap with this area, if it extends further than
428 * us then we must extend ourselves. Account for its
429 * existing reservation. */
434 chg -= rg->to - rg->from;
438 spin_unlock(&resv->lock);
439 /* We already know we raced and no longer need the new region */
443 spin_unlock(&resv->lock);
448 * Abort the in progress add operation. The adds_in_progress field
449 * of the resv_map keeps track of the operations in progress between
450 * calls to region_chg and region_add. Operations are sometimes
451 * aborted after the call to region_chg. In such cases, region_abort
452 * is called to decrement the adds_in_progress counter.
454 * NOTE: The range arguments [f, t) are not needed or used in this
455 * routine. They are kept to make reading the calling code easier as
456 * arguments will match the associated region_chg call.
458 static void region_abort(struct resv_map *resv, long f, long t)
460 spin_lock(&resv->lock);
461 VM_BUG_ON(!resv->region_cache_count);
462 resv->adds_in_progress--;
463 spin_unlock(&resv->lock);
467 * Delete the specified range [f, t) from the reserve map. If the
468 * t parameter is LONG_MAX, this indicates that ALL regions after f
469 * should be deleted. Locate the regions which intersect [f, t)
470 * and either trim, delete or split the existing regions.
472 * Returns the number of huge pages deleted from the reserve map.
473 * In the normal case, the return value is zero or more. In the
474 * case where a region must be split, a new region descriptor must
475 * be allocated. If the allocation fails, -ENOMEM will be returned.
476 * NOTE: If the parameter t == LONG_MAX, then we will never split
477 * a region and possibly return -ENOMEM. Callers specifying
478 * t == LONG_MAX do not need to check for -ENOMEM error.
480 static long region_del(struct resv_map *resv, long f, long t)
482 struct list_head *head = &resv->regions;
483 struct file_region *rg, *trg;
484 struct file_region *nrg = NULL;
488 spin_lock(&resv->lock);
489 list_for_each_entry_safe(rg, trg, head, link) {
491 * Skip regions before the range to be deleted. file_region
492 * ranges are normally of the form [from, to). However, there
493 * may be a "placeholder" entry in the map which is of the form
494 * (from, to) with from == to. Check for placeholder entries
495 * at the beginning of the range to be deleted.
497 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
503 if (f > rg->from && t < rg->to) { /* Must split region */
505 * Check for an entry in the cache before dropping
506 * lock and attempting allocation.
509 resv->region_cache_count > resv->adds_in_progress) {
510 nrg = list_first_entry(&resv->region_cache,
513 list_del(&nrg->link);
514 resv->region_cache_count--;
518 spin_unlock(&resv->lock);
519 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
527 /* New entry for end of split region */
530 INIT_LIST_HEAD(&nrg->link);
532 /* Original entry is trimmed */
535 list_add(&nrg->link, &rg->link);
540 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
541 del += rg->to - rg->from;
547 if (f <= rg->from) { /* Trim beginning of region */
550 } else { /* Trim end of region */
556 spin_unlock(&resv->lock);
562 * A rare out of memory error was encountered which prevented removal of
563 * the reserve map region for a page. The huge page itself was free'ed
564 * and removed from the page cache. This routine will adjust the subpool
565 * usage count, and the global reserve count if needed. By incrementing
566 * these counts, the reserve map entry which could not be deleted will
567 * appear as a "reserved" entry instead of simply dangling with incorrect
570 void hugetlb_fix_reserve_counts(struct inode *inode)
572 struct hugepage_subpool *spool = subpool_inode(inode);
575 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
577 struct hstate *h = hstate_inode(inode);
579 hugetlb_acct_memory(h, 1);
584 * Count and return the number of huge pages in the reserve map
585 * that intersect with the range [f, t).
587 static long region_count(struct resv_map *resv, long f, long t)
589 struct list_head *head = &resv->regions;
590 struct file_region *rg;
593 spin_lock(&resv->lock);
594 /* Locate each segment we overlap with, and count that overlap. */
595 list_for_each_entry(rg, head, link) {
604 seg_from = max(rg->from, f);
605 seg_to = min(rg->to, t);
607 chg += seg_to - seg_from;
609 spin_unlock(&resv->lock);
615 * Convert the address within this vma to the page offset within
616 * the mapping, in pagecache page units; huge pages here.
618 static pgoff_t vma_hugecache_offset(struct hstate *h,
619 struct vm_area_struct *vma, unsigned long address)
621 return ((address - vma->vm_start) >> huge_page_shift(h)) +
622 (vma->vm_pgoff >> huge_page_order(h));
625 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
626 unsigned long address)
628 return vma_hugecache_offset(hstate_vma(vma), vma, address);
630 EXPORT_SYMBOL_GPL(linear_hugepage_index);
633 * Return the size of the pages allocated when backing a VMA. In the majority
634 * cases this will be same size as used by the page table entries.
636 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
638 struct hstate *hstate;
640 if (!is_vm_hugetlb_page(vma))
643 hstate = hstate_vma(vma);
645 return 1UL << huge_page_shift(hstate);
647 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
650 * Return the page size being used by the MMU to back a VMA. In the majority
651 * of cases, the page size used by the kernel matches the MMU size. On
652 * architectures where it differs, an architecture-specific version of this
653 * function is required.
655 #ifndef vma_mmu_pagesize
656 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
658 return vma_kernel_pagesize(vma);
663 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
664 * bits of the reservation map pointer, which are always clear due to
667 #define HPAGE_RESV_OWNER (1UL << 0)
668 #define HPAGE_RESV_UNMAPPED (1UL << 1)
669 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
672 * These helpers are used to track how many pages are reserved for
673 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
674 * is guaranteed to have their future faults succeed.
676 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
677 * the reserve counters are updated with the hugetlb_lock held. It is safe
678 * to reset the VMA at fork() time as it is not in use yet and there is no
679 * chance of the global counters getting corrupted as a result of the values.
681 * The private mapping reservation is represented in a subtly different
682 * manner to a shared mapping. A shared mapping has a region map associated
683 * with the underlying file, this region map represents the backing file
684 * pages which have ever had a reservation assigned which this persists even
685 * after the page is instantiated. A private mapping has a region map
686 * associated with the original mmap which is attached to all VMAs which
687 * reference it, this region map represents those offsets which have consumed
688 * reservation ie. where pages have been instantiated.
690 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
692 return (unsigned long)vma->vm_private_data;
695 static void set_vma_private_data(struct vm_area_struct *vma,
698 vma->vm_private_data = (void *)value;
701 struct resv_map *resv_map_alloc(void)
703 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
704 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
706 if (!resv_map || !rg) {
712 kref_init(&resv_map->refs);
713 spin_lock_init(&resv_map->lock);
714 INIT_LIST_HEAD(&resv_map->regions);
716 resv_map->adds_in_progress = 0;
718 INIT_LIST_HEAD(&resv_map->region_cache);
719 list_add(&rg->link, &resv_map->region_cache);
720 resv_map->region_cache_count = 1;
725 void resv_map_release(struct kref *ref)
727 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
728 struct list_head *head = &resv_map->region_cache;
729 struct file_region *rg, *trg;
731 /* Clear out any active regions before we release the map. */
732 region_del(resv_map, 0, LONG_MAX);
734 /* ... and any entries left in the cache */
735 list_for_each_entry_safe(rg, trg, head, link) {
740 VM_BUG_ON(resv_map->adds_in_progress);
745 static inline struct resv_map *inode_resv_map(struct inode *inode)
747 return inode->i_mapping->private_data;
750 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
752 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
753 if (vma->vm_flags & VM_MAYSHARE) {
754 struct address_space *mapping = vma->vm_file->f_mapping;
755 struct inode *inode = mapping->host;
757 return inode_resv_map(inode);
760 return (struct resv_map *)(get_vma_private_data(vma) &
765 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
767 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
768 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
770 set_vma_private_data(vma, (get_vma_private_data(vma) &
771 HPAGE_RESV_MASK) | (unsigned long)map);
774 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
776 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
777 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
779 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
782 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
784 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
786 return (get_vma_private_data(vma) & flag) != 0;
789 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
790 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
792 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
793 if (!(vma->vm_flags & VM_MAYSHARE))
794 vma->vm_private_data = (void *)0;
797 /* Returns true if the VMA has associated reserve pages */
798 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
800 if (vma->vm_flags & VM_NORESERVE) {
802 * This address is already reserved by other process(chg == 0),
803 * so, we should decrement reserved count. Without decrementing,
804 * reserve count remains after releasing inode, because this
805 * allocated page will go into page cache and is regarded as
806 * coming from reserved pool in releasing step. Currently, we
807 * don't have any other solution to deal with this situation
808 * properly, so add work-around here.
810 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
816 /* Shared mappings always use reserves */
817 if (vma->vm_flags & VM_MAYSHARE) {
819 * We know VM_NORESERVE is not set. Therefore, there SHOULD
820 * be a region map for all pages. The only situation where
821 * there is no region map is if a hole was punched via
822 * fallocate. In this case, there really are no reverves to
823 * use. This situation is indicated if chg != 0.
832 * Only the process that called mmap() has reserves for
835 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
837 * Like the shared case above, a hole punch or truncate
838 * could have been performed on the private mapping.
839 * Examine the value of chg to determine if reserves
840 * actually exist or were previously consumed.
841 * Very Subtle - The value of chg comes from a previous
842 * call to vma_needs_reserves(). The reserve map for
843 * private mappings has different (opposite) semantics
844 * than that of shared mappings. vma_needs_reserves()
845 * has already taken this difference in semantics into
846 * account. Therefore, the meaning of chg is the same
847 * as in the shared case above. Code could easily be
848 * combined, but keeping it separate draws attention to
849 * subtle differences.
860 static void enqueue_huge_page(struct hstate *h, struct page *page)
862 int nid = page_to_nid(page);
863 list_move(&page->lru, &h->hugepage_freelists[nid]);
864 h->free_huge_pages++;
865 h->free_huge_pages_node[nid]++;
868 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
872 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
873 if (!PageHWPoison(page))
876 * if 'non-isolated free hugepage' not found on the list,
877 * the allocation fails.
879 if (&h->hugepage_freelists[nid] == &page->lru)
881 list_move(&page->lru, &h->hugepage_activelist);
882 set_page_refcounted(page);
883 h->free_huge_pages--;
884 h->free_huge_pages_node[nid]--;
888 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
891 unsigned int cpuset_mems_cookie;
892 struct zonelist *zonelist;
897 zonelist = node_zonelist(nid, gfp_mask);
900 cpuset_mems_cookie = read_mems_allowed_begin();
901 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
904 if (!cpuset_zone_allowed(zone, gfp_mask))
907 * no need to ask again on the same node. Pool is node rather than
910 if (zone_to_nid(zone) == node)
912 node = zone_to_nid(zone);
914 page = dequeue_huge_page_node_exact(h, node);
918 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
924 /* Movability of hugepages depends on migration support. */
925 static inline gfp_t htlb_alloc_mask(struct hstate *h)
927 if (hugepage_migration_supported(h))
928 return GFP_HIGHUSER_MOVABLE;
933 static struct page *dequeue_huge_page_vma(struct hstate *h,
934 struct vm_area_struct *vma,
935 unsigned long address, int avoid_reserve,
939 struct mempolicy *mpol;
941 nodemask_t *nodemask;
945 * A child process with MAP_PRIVATE mappings created by their parent
946 * have no page reserves. This check ensures that reservations are
947 * not "stolen". The child may still get SIGKILLed
949 if (!vma_has_reserves(vma, chg) &&
950 h->free_huge_pages - h->resv_huge_pages == 0)
953 /* If reserves cannot be used, ensure enough pages are in the pool */
954 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
957 gfp_mask = htlb_alloc_mask(h);
958 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
959 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
960 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
961 SetPagePrivate(page);
962 h->resv_huge_pages--;
973 * common helper functions for hstate_next_node_to_{alloc|free}.
974 * We may have allocated or freed a huge page based on a different
975 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
976 * be outside of *nodes_allowed. Ensure that we use an allowed
977 * node for alloc or free.
979 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
981 nid = next_node_in(nid, *nodes_allowed);
982 VM_BUG_ON(nid >= MAX_NUMNODES);
987 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
989 if (!node_isset(nid, *nodes_allowed))
990 nid = next_node_allowed(nid, nodes_allowed);
995 * returns the previously saved node ["this node"] from which to
996 * allocate a persistent huge page for the pool and advance the
997 * next node from which to allocate, handling wrap at end of node
1000 static int hstate_next_node_to_alloc(struct hstate *h,
1001 nodemask_t *nodes_allowed)
1005 VM_BUG_ON(!nodes_allowed);
1007 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1008 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1014 * helper for free_pool_huge_page() - return the previously saved
1015 * node ["this node"] from which to free a huge page. Advance the
1016 * next node id whether or not we find a free huge page to free so
1017 * that the next attempt to free addresses the next node.
1019 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1023 VM_BUG_ON(!nodes_allowed);
1025 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1026 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1031 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1032 for (nr_nodes = nodes_weight(*mask); \
1034 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1037 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1038 for (nr_nodes = nodes_weight(*mask); \
1040 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1043 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1044 static void destroy_compound_gigantic_page(struct page *page,
1048 int nr_pages = 1 << order;
1049 struct page *p = page + 1;
1051 atomic_set(compound_mapcount_ptr(page), 0);
1052 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1053 clear_compound_head(p);
1054 set_page_refcounted(p);
1057 set_compound_order(page, 0);
1058 __ClearPageHead(page);
1061 static void free_gigantic_page(struct page *page, unsigned int order)
1063 free_contig_range(page_to_pfn(page), 1 << order);
1066 static int __alloc_gigantic_page(unsigned long start_pfn,
1067 unsigned long nr_pages, gfp_t gfp_mask)
1069 unsigned long end_pfn = start_pfn + nr_pages;
1070 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1074 static bool pfn_range_valid_gigantic(struct zone *z,
1075 unsigned long start_pfn, unsigned long nr_pages)
1077 unsigned long i, end_pfn = start_pfn + nr_pages;
1080 for (i = start_pfn; i < end_pfn; i++) {
1084 page = pfn_to_page(i);
1086 if (page_zone(page) != z)
1089 if (PageReserved(page))
1092 if (page_count(page) > 0)
1102 static bool zone_spans_last_pfn(const struct zone *zone,
1103 unsigned long start_pfn, unsigned long nr_pages)
1105 unsigned long last_pfn = start_pfn + nr_pages - 1;
1106 return zone_spans_pfn(zone, last_pfn);
1109 static struct page *alloc_gigantic_page(int nid, struct hstate *h)
1111 unsigned int order = huge_page_order(h);
1112 unsigned long nr_pages = 1 << order;
1113 unsigned long ret, pfn, flags;
1114 struct zonelist *zonelist;
1119 gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1120 zonelist = node_zonelist(nid, gfp_mask);
1121 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), NULL) {
1122 spin_lock_irqsave(&zone->lock, flags);
1124 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
1125 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
1126 if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) {
1128 * We release the zone lock here because
1129 * alloc_contig_range() will also lock the zone
1130 * at some point. If there's an allocation
1131 * spinning on this lock, it may win the race
1132 * and cause alloc_contig_range() to fail...
1134 spin_unlock_irqrestore(&zone->lock, flags);
1135 ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask);
1137 return pfn_to_page(pfn);
1138 spin_lock_irqsave(&zone->lock, flags);
1143 spin_unlock_irqrestore(&zone->lock, flags);
1149 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1150 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1152 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1156 page = alloc_gigantic_page(nid, h);
1158 prep_compound_gigantic_page(page, huge_page_order(h));
1159 prep_new_huge_page(h, page, nid);
1165 static int alloc_fresh_gigantic_page(struct hstate *h,
1166 nodemask_t *nodes_allowed)
1168 struct page *page = NULL;
1171 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1172 page = alloc_fresh_gigantic_page_node(h, node);
1180 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1181 static inline bool gigantic_page_supported(void) { return false; }
1182 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1183 static inline void destroy_compound_gigantic_page(struct page *page,
1184 unsigned int order) { }
1185 static inline int alloc_fresh_gigantic_page(struct hstate *h,
1186 nodemask_t *nodes_allowed) { return 0; }
1189 static void update_and_free_page(struct hstate *h, struct page *page)
1193 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1197 h->nr_huge_pages_node[page_to_nid(page)]--;
1198 for (i = 0; i < pages_per_huge_page(h); i++) {
1199 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1200 1 << PG_referenced | 1 << PG_dirty |
1201 1 << PG_active | 1 << PG_private |
1204 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1205 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1206 set_page_refcounted(page);
1207 if (hstate_is_gigantic(h)) {
1208 destroy_compound_gigantic_page(page, huge_page_order(h));
1209 free_gigantic_page(page, huge_page_order(h));
1211 __free_pages(page, huge_page_order(h));
1215 struct hstate *size_to_hstate(unsigned long size)
1219 for_each_hstate(h) {
1220 if (huge_page_size(h) == size)
1227 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1228 * to hstate->hugepage_activelist.)
1230 * This function can be called for tail pages, but never returns true for them.
1232 bool page_huge_active(struct page *page)
1234 VM_BUG_ON_PAGE(!PageHuge(page), page);
1235 return PageHead(page) && PagePrivate(&page[1]);
1238 /* never called for tail page */
1239 static void set_page_huge_active(struct page *page)
1241 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1242 SetPagePrivate(&page[1]);
1245 static void clear_page_huge_active(struct page *page)
1247 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1248 ClearPagePrivate(&page[1]);
1251 void free_huge_page(struct page *page)
1254 * Can't pass hstate in here because it is called from the
1255 * compound page destructor.
1257 struct hstate *h = page_hstate(page);
1258 int nid = page_to_nid(page);
1259 struct hugepage_subpool *spool =
1260 (struct hugepage_subpool *)page_private(page);
1261 bool restore_reserve;
1263 set_page_private(page, 0);
1264 page->mapping = NULL;
1265 VM_BUG_ON_PAGE(page_count(page), page);
1266 VM_BUG_ON_PAGE(page_mapcount(page), page);
1267 restore_reserve = PagePrivate(page);
1268 ClearPagePrivate(page);
1271 * A return code of zero implies that the subpool will be under its
1272 * minimum size if the reservation is not restored after page is free.
1273 * Therefore, force restore_reserve operation.
1275 if (hugepage_subpool_put_pages(spool, 1) == 0)
1276 restore_reserve = true;
1278 spin_lock(&hugetlb_lock);
1279 clear_page_huge_active(page);
1280 hugetlb_cgroup_uncharge_page(hstate_index(h),
1281 pages_per_huge_page(h), page);
1282 if (restore_reserve)
1283 h->resv_huge_pages++;
1285 if (h->surplus_huge_pages_node[nid]) {
1286 /* remove the page from active list */
1287 list_del(&page->lru);
1288 update_and_free_page(h, page);
1289 h->surplus_huge_pages--;
1290 h->surplus_huge_pages_node[nid]--;
1292 arch_clear_hugepage_flags(page);
1293 enqueue_huge_page(h, page);
1295 spin_unlock(&hugetlb_lock);
1298 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1300 INIT_LIST_HEAD(&page->lru);
1301 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1302 spin_lock(&hugetlb_lock);
1303 set_hugetlb_cgroup(page, NULL);
1305 h->nr_huge_pages_node[nid]++;
1306 spin_unlock(&hugetlb_lock);
1307 put_page(page); /* free it into the hugepage allocator */
1310 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1313 int nr_pages = 1 << order;
1314 struct page *p = page + 1;
1316 /* we rely on prep_new_huge_page to set the destructor */
1317 set_compound_order(page, order);
1318 __ClearPageReserved(page);
1319 __SetPageHead(page);
1320 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1322 * For gigantic hugepages allocated through bootmem at
1323 * boot, it's safer to be consistent with the not-gigantic
1324 * hugepages and clear the PG_reserved bit from all tail pages
1325 * too. Otherwse drivers using get_user_pages() to access tail
1326 * pages may get the reference counting wrong if they see
1327 * PG_reserved set on a tail page (despite the head page not
1328 * having PG_reserved set). Enforcing this consistency between
1329 * head and tail pages allows drivers to optimize away a check
1330 * on the head page when they need know if put_page() is needed
1331 * after get_user_pages().
1333 __ClearPageReserved(p);
1334 set_page_count(p, 0);
1335 set_compound_head(p, page);
1337 atomic_set(compound_mapcount_ptr(page), -1);
1341 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1342 * transparent huge pages. See the PageTransHuge() documentation for more
1345 int PageHuge(struct page *page)
1347 if (!PageCompound(page))
1350 page = compound_head(page);
1351 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1353 EXPORT_SYMBOL_GPL(PageHuge);
1356 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1357 * normal or transparent huge pages.
1359 int PageHeadHuge(struct page *page_head)
1361 if (!PageHead(page_head))
1364 return get_compound_page_dtor(page_head) == free_huge_page;
1367 pgoff_t __basepage_index(struct page *page)
1369 struct page *page_head = compound_head(page);
1370 pgoff_t index = page_index(page_head);
1371 unsigned long compound_idx;
1373 if (!PageHuge(page_head))
1374 return page_index(page);
1376 if (compound_order(page_head) >= MAX_ORDER)
1377 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1379 compound_idx = page - page_head;
1381 return (index << compound_order(page_head)) + compound_idx;
1384 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1388 page = __alloc_pages_node(nid,
1389 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1390 __GFP_RETRY_MAYFAIL|__GFP_NOWARN,
1391 huge_page_order(h));
1393 prep_new_huge_page(h, page, nid);
1399 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1405 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1406 page = alloc_fresh_huge_page_node(h, node);
1414 count_vm_event(HTLB_BUDDY_PGALLOC);
1416 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1422 * Free huge page from pool from next node to free.
1423 * Attempt to keep persistent huge pages more or less
1424 * balanced over allowed nodes.
1425 * Called with hugetlb_lock locked.
1427 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1433 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1435 * If we're returning unused surplus pages, only examine
1436 * nodes with surplus pages.
1438 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1439 !list_empty(&h->hugepage_freelists[node])) {
1441 list_entry(h->hugepage_freelists[node].next,
1443 list_del(&page->lru);
1444 h->free_huge_pages--;
1445 h->free_huge_pages_node[node]--;
1447 h->surplus_huge_pages--;
1448 h->surplus_huge_pages_node[node]--;
1450 update_and_free_page(h, page);
1460 * Dissolve a given free hugepage into free buddy pages. This function does
1461 * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1462 * number of free hugepages would be reduced below the number of reserved
1465 int dissolve_free_huge_page(struct page *page)
1469 spin_lock(&hugetlb_lock);
1470 if (PageHuge(page) && !page_count(page)) {
1471 struct page *head = compound_head(page);
1472 struct hstate *h = page_hstate(head);
1473 int nid = page_to_nid(head);
1474 if (h->free_huge_pages - h->resv_huge_pages == 0) {
1479 * Move PageHWPoison flag from head page to the raw error page,
1480 * which makes any subpages rather than the error page reusable.
1482 if (PageHWPoison(head) && page != head) {
1483 SetPageHWPoison(page);
1484 ClearPageHWPoison(head);
1486 list_del(&head->lru);
1487 h->free_huge_pages--;
1488 h->free_huge_pages_node[nid]--;
1489 h->max_huge_pages--;
1490 update_and_free_page(h, head);
1493 spin_unlock(&hugetlb_lock);
1498 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1499 * make specified memory blocks removable from the system.
1500 * Note that this will dissolve a free gigantic hugepage completely, if any
1501 * part of it lies within the given range.
1502 * Also note that if dissolve_free_huge_page() returns with an error, all
1503 * free hugepages that were dissolved before that error are lost.
1505 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1511 if (!hugepages_supported())
1514 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1515 page = pfn_to_page(pfn);
1516 if (PageHuge(page) && !page_count(page)) {
1517 rc = dissolve_free_huge_page(page);
1526 static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1527 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1529 int order = huge_page_order(h);
1531 gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
1532 if (nid == NUMA_NO_NODE)
1533 nid = numa_mem_id();
1534 return __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1537 static struct page *__alloc_buddy_huge_page(struct hstate *h, gfp_t gfp_mask,
1538 int nid, nodemask_t *nmask)
1543 if (hstate_is_gigantic(h))
1547 * Assume we will successfully allocate the surplus page to
1548 * prevent racing processes from causing the surplus to exceed
1551 * This however introduces a different race, where a process B
1552 * tries to grow the static hugepage pool while alloc_pages() is
1553 * called by process A. B will only examine the per-node
1554 * counters in determining if surplus huge pages can be
1555 * converted to normal huge pages in adjust_pool_surplus(). A
1556 * won't be able to increment the per-node counter, until the
1557 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1558 * no more huge pages can be converted from surplus to normal
1559 * state (and doesn't try to convert again). Thus, we have a
1560 * case where a surplus huge page exists, the pool is grown, and
1561 * the surplus huge page still exists after, even though it
1562 * should just have been converted to a normal huge page. This
1563 * does not leak memory, though, as the hugepage will be freed
1564 * once it is out of use. It also does not allow the counters to
1565 * go out of whack in adjust_pool_surplus() as we don't modify
1566 * the node values until we've gotten the hugepage and only the
1567 * per-node value is checked there.
1569 spin_lock(&hugetlb_lock);
1570 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1571 spin_unlock(&hugetlb_lock);
1575 h->surplus_huge_pages++;
1577 spin_unlock(&hugetlb_lock);
1579 page = __hugetlb_alloc_buddy_huge_page(h, gfp_mask, nid, nmask);
1581 spin_lock(&hugetlb_lock);
1583 INIT_LIST_HEAD(&page->lru);
1584 r_nid = page_to_nid(page);
1585 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1586 set_hugetlb_cgroup(page, NULL);
1588 * We incremented the global counters already
1590 h->nr_huge_pages_node[r_nid]++;
1591 h->surplus_huge_pages_node[r_nid]++;
1592 __count_vm_event(HTLB_BUDDY_PGALLOC);
1595 h->surplus_huge_pages--;
1596 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1598 spin_unlock(&hugetlb_lock);
1604 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1607 struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1608 struct vm_area_struct *vma, unsigned long addr)
1611 struct mempolicy *mpol;
1612 gfp_t gfp_mask = htlb_alloc_mask(h);
1614 nodemask_t *nodemask;
1616 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1617 page = __alloc_buddy_huge_page(h, gfp_mask, nid, nodemask);
1618 mpol_cond_put(mpol);
1624 * This allocation function is useful in the context where vma is irrelevant.
1625 * E.g. soft-offlining uses this function because it only cares physical
1626 * address of error page.
1628 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1630 gfp_t gfp_mask = htlb_alloc_mask(h);
1631 struct page *page = NULL;
1633 if (nid != NUMA_NO_NODE)
1634 gfp_mask |= __GFP_THISNODE;
1636 spin_lock(&hugetlb_lock);
1637 if (h->free_huge_pages - h->resv_huge_pages > 0)
1638 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1639 spin_unlock(&hugetlb_lock);
1642 page = __alloc_buddy_huge_page(h, gfp_mask, nid, NULL);
1648 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1651 gfp_t gfp_mask = htlb_alloc_mask(h);
1653 spin_lock(&hugetlb_lock);
1654 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1657 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1659 spin_unlock(&hugetlb_lock);
1663 spin_unlock(&hugetlb_lock);
1665 /* No reservations, try to overcommit */
1667 return __alloc_buddy_huge_page(h, gfp_mask, preferred_nid, nmask);
1671 * Increase the hugetlb pool such that it can accommodate a reservation
1674 static int gather_surplus_pages(struct hstate *h, int delta)
1676 struct list_head surplus_list;
1677 struct page *page, *tmp;
1679 int needed, allocated;
1680 bool alloc_ok = true;
1682 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1684 h->resv_huge_pages += delta;
1689 INIT_LIST_HEAD(&surplus_list);
1693 spin_unlock(&hugetlb_lock);
1694 for (i = 0; i < needed; i++) {
1695 page = __alloc_buddy_huge_page(h, htlb_alloc_mask(h),
1696 NUMA_NO_NODE, NULL);
1701 list_add(&page->lru, &surplus_list);
1707 * After retaking hugetlb_lock, we need to recalculate 'needed'
1708 * because either resv_huge_pages or free_huge_pages may have changed.
1710 spin_lock(&hugetlb_lock);
1711 needed = (h->resv_huge_pages + delta) -
1712 (h->free_huge_pages + allocated);
1717 * We were not able to allocate enough pages to
1718 * satisfy the entire reservation so we free what
1719 * we've allocated so far.
1724 * The surplus_list now contains _at_least_ the number of extra pages
1725 * needed to accommodate the reservation. Add the appropriate number
1726 * of pages to the hugetlb pool and free the extras back to the buddy
1727 * allocator. Commit the entire reservation here to prevent another
1728 * process from stealing the pages as they are added to the pool but
1729 * before they are reserved.
1731 needed += allocated;
1732 h->resv_huge_pages += delta;
1735 /* Free the needed pages to the hugetlb pool */
1736 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1740 * This page is now managed by the hugetlb allocator and has
1741 * no users -- drop the buddy allocator's reference.
1743 put_page_testzero(page);
1744 VM_BUG_ON_PAGE(page_count(page), page);
1745 enqueue_huge_page(h, page);
1748 spin_unlock(&hugetlb_lock);
1750 /* Free unnecessary surplus pages to the buddy allocator */
1751 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1753 spin_lock(&hugetlb_lock);
1759 * This routine has two main purposes:
1760 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1761 * in unused_resv_pages. This corresponds to the prior adjustments made
1762 * to the associated reservation map.
1763 * 2) Free any unused surplus pages that may have been allocated to satisfy
1764 * the reservation. As many as unused_resv_pages may be freed.
1766 * Called with hugetlb_lock held. However, the lock could be dropped (and
1767 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1768 * we must make sure nobody else can claim pages we are in the process of
1769 * freeing. Do this by ensuring resv_huge_page always is greater than the
1770 * number of huge pages we plan to free when dropping the lock.
1772 static void return_unused_surplus_pages(struct hstate *h,
1773 unsigned long unused_resv_pages)
1775 unsigned long nr_pages;
1777 /* Cannot return gigantic pages currently */
1778 if (hstate_is_gigantic(h))
1782 * Part (or even all) of the reservation could have been backed
1783 * by pre-allocated pages. Only free surplus pages.
1785 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1788 * We want to release as many surplus pages as possible, spread
1789 * evenly across all nodes with memory. Iterate across these nodes
1790 * until we can no longer free unreserved surplus pages. This occurs
1791 * when the nodes with surplus pages have no free pages.
1792 * free_pool_huge_page() will balance the the freed pages across the
1793 * on-line nodes with memory and will handle the hstate accounting.
1795 * Note that we decrement resv_huge_pages as we free the pages. If
1796 * we drop the lock, resv_huge_pages will still be sufficiently large
1797 * to cover subsequent pages we may free.
1799 while (nr_pages--) {
1800 h->resv_huge_pages--;
1801 unused_resv_pages--;
1802 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1804 cond_resched_lock(&hugetlb_lock);
1808 /* Fully uncommit the reservation */
1809 h->resv_huge_pages -= unused_resv_pages;
1814 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1815 * are used by the huge page allocation routines to manage reservations.
1817 * vma_needs_reservation is called to determine if the huge page at addr
1818 * within the vma has an associated reservation. If a reservation is
1819 * needed, the value 1 is returned. The caller is then responsible for
1820 * managing the global reservation and subpool usage counts. After
1821 * the huge page has been allocated, vma_commit_reservation is called
1822 * to add the page to the reservation map. If the page allocation fails,
1823 * the reservation must be ended instead of committed. vma_end_reservation
1824 * is called in such cases.
1826 * In the normal case, vma_commit_reservation returns the same value
1827 * as the preceding vma_needs_reservation call. The only time this
1828 * is not the case is if a reserve map was changed between calls. It
1829 * is the responsibility of the caller to notice the difference and
1830 * take appropriate action.
1832 * vma_add_reservation is used in error paths where a reservation must
1833 * be restored when a newly allocated huge page must be freed. It is
1834 * to be called after calling vma_needs_reservation to determine if a
1835 * reservation exists.
1837 enum vma_resv_mode {
1843 static long __vma_reservation_common(struct hstate *h,
1844 struct vm_area_struct *vma, unsigned long addr,
1845 enum vma_resv_mode mode)
1847 struct resv_map *resv;
1851 resv = vma_resv_map(vma);
1855 idx = vma_hugecache_offset(h, vma, addr);
1857 case VMA_NEEDS_RESV:
1858 ret = region_chg(resv, idx, idx + 1);
1860 case VMA_COMMIT_RESV:
1861 ret = region_add(resv, idx, idx + 1);
1864 region_abort(resv, idx, idx + 1);
1868 if (vma->vm_flags & VM_MAYSHARE)
1869 ret = region_add(resv, idx, idx + 1);
1871 region_abort(resv, idx, idx + 1);
1872 ret = region_del(resv, idx, idx + 1);
1879 if (vma->vm_flags & VM_MAYSHARE)
1881 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1883 * In most cases, reserves always exist for private mappings.
1884 * However, a file associated with mapping could have been
1885 * hole punched or truncated after reserves were consumed.
1886 * As subsequent fault on such a range will not use reserves.
1887 * Subtle - The reserve map for private mappings has the
1888 * opposite meaning than that of shared mappings. If NO
1889 * entry is in the reserve map, it means a reservation exists.
1890 * If an entry exists in the reserve map, it means the
1891 * reservation has already been consumed. As a result, the
1892 * return value of this routine is the opposite of the
1893 * value returned from reserve map manipulation routines above.
1901 return ret < 0 ? ret : 0;
1904 static long vma_needs_reservation(struct hstate *h,
1905 struct vm_area_struct *vma, unsigned long addr)
1907 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1910 static long vma_commit_reservation(struct hstate *h,
1911 struct vm_area_struct *vma, unsigned long addr)
1913 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1916 static void vma_end_reservation(struct hstate *h,
1917 struct vm_area_struct *vma, unsigned long addr)
1919 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1922 static long vma_add_reservation(struct hstate *h,
1923 struct vm_area_struct *vma, unsigned long addr)
1925 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1929 * This routine is called to restore a reservation on error paths. In the
1930 * specific error paths, a huge page was allocated (via alloc_huge_page)
1931 * and is about to be freed. If a reservation for the page existed,
1932 * alloc_huge_page would have consumed the reservation and set PagePrivate
1933 * in the newly allocated page. When the page is freed via free_huge_page,
1934 * the global reservation count will be incremented if PagePrivate is set.
1935 * However, free_huge_page can not adjust the reserve map. Adjust the
1936 * reserve map here to be consistent with global reserve count adjustments
1937 * to be made by free_huge_page.
1939 static void restore_reserve_on_error(struct hstate *h,
1940 struct vm_area_struct *vma, unsigned long address,
1943 if (unlikely(PagePrivate(page))) {
1944 long rc = vma_needs_reservation(h, vma, address);
1946 if (unlikely(rc < 0)) {
1948 * Rare out of memory condition in reserve map
1949 * manipulation. Clear PagePrivate so that
1950 * global reserve count will not be incremented
1951 * by free_huge_page. This will make it appear
1952 * as though the reservation for this page was
1953 * consumed. This may prevent the task from
1954 * faulting in the page at a later time. This
1955 * is better than inconsistent global huge page
1956 * accounting of reserve counts.
1958 ClearPagePrivate(page);
1960 rc = vma_add_reservation(h, vma, address);
1961 if (unlikely(rc < 0))
1963 * See above comment about rare out of
1966 ClearPagePrivate(page);
1968 vma_end_reservation(h, vma, address);
1972 struct page *alloc_huge_page(struct vm_area_struct *vma,
1973 unsigned long addr, int avoid_reserve)
1975 struct hugepage_subpool *spool = subpool_vma(vma);
1976 struct hstate *h = hstate_vma(vma);
1978 long map_chg, map_commit;
1981 struct hugetlb_cgroup *h_cg;
1983 idx = hstate_index(h);
1985 * Examine the region/reserve map to determine if the process
1986 * has a reservation for the page to be allocated. A return
1987 * code of zero indicates a reservation exists (no change).
1989 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
1991 return ERR_PTR(-ENOMEM);
1994 * Processes that did not create the mapping will have no
1995 * reserves as indicated by the region/reserve map. Check
1996 * that the allocation will not exceed the subpool limit.
1997 * Allocations for MAP_NORESERVE mappings also need to be
1998 * checked against any subpool limit.
2000 if (map_chg || avoid_reserve) {
2001 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2003 vma_end_reservation(h, vma, addr);
2004 return ERR_PTR(-ENOSPC);
2008 * Even though there was no reservation in the region/reserve
2009 * map, there could be reservations associated with the
2010 * subpool that can be used. This would be indicated if the
2011 * return value of hugepage_subpool_get_pages() is zero.
2012 * However, if avoid_reserve is specified we still avoid even
2013 * the subpool reservations.
2019 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2021 goto out_subpool_put;
2023 spin_lock(&hugetlb_lock);
2025 * glb_chg is passed to indicate whether or not a page must be taken
2026 * from the global free pool (global change). gbl_chg == 0 indicates
2027 * a reservation exists for the allocation.
2029 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2031 spin_unlock(&hugetlb_lock);
2032 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
2034 goto out_uncharge_cgroup;
2035 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2036 SetPagePrivate(page);
2037 h->resv_huge_pages--;
2039 spin_lock(&hugetlb_lock);
2040 list_move(&page->lru, &h->hugepage_activelist);
2043 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2044 spin_unlock(&hugetlb_lock);
2046 set_page_private(page, (unsigned long)spool);
2048 map_commit = vma_commit_reservation(h, vma, addr);
2049 if (unlikely(map_chg > map_commit)) {
2051 * The page was added to the reservation map between
2052 * vma_needs_reservation and vma_commit_reservation.
2053 * This indicates a race with hugetlb_reserve_pages.
2054 * Adjust for the subpool count incremented above AND
2055 * in hugetlb_reserve_pages for the same page. Also,
2056 * the reservation count added in hugetlb_reserve_pages
2057 * no longer applies.
2061 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2062 hugetlb_acct_memory(h, -rsv_adjust);
2066 out_uncharge_cgroup:
2067 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2069 if (map_chg || avoid_reserve)
2070 hugepage_subpool_put_pages(spool, 1);
2071 vma_end_reservation(h, vma, addr);
2072 return ERR_PTR(-ENOSPC);
2076 * alloc_huge_page()'s wrapper which simply returns the page if allocation
2077 * succeeds, otherwise NULL. This function is called from new_vma_page(),
2078 * where no ERR_VALUE is expected to be returned.
2080 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
2081 unsigned long addr, int avoid_reserve)
2083 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
2089 int alloc_bootmem_huge_page(struct hstate *h)
2090 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2091 int __alloc_bootmem_huge_page(struct hstate *h)
2093 struct huge_bootmem_page *m;
2096 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2099 addr = memblock_virt_alloc_try_nid_nopanic(
2100 huge_page_size(h), huge_page_size(h),
2101 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
2104 * Use the beginning of the huge page to store the
2105 * huge_bootmem_page struct (until gather_bootmem
2106 * puts them into the mem_map).
2115 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2116 /* Put them into a private list first because mem_map is not up yet */
2117 list_add(&m->list, &huge_boot_pages);
2122 static void __init prep_compound_huge_page(struct page *page,
2125 if (unlikely(order > (MAX_ORDER - 1)))
2126 prep_compound_gigantic_page(page, order);
2128 prep_compound_page(page, order);
2131 /* Put bootmem huge pages into the standard lists after mem_map is up */
2132 static void __init gather_bootmem_prealloc(void)
2134 struct huge_bootmem_page *m;
2136 list_for_each_entry(m, &huge_boot_pages, list) {
2137 struct hstate *h = m->hstate;
2140 #ifdef CONFIG_HIGHMEM
2141 page = pfn_to_page(m->phys >> PAGE_SHIFT);
2142 memblock_free_late(__pa(m),
2143 sizeof(struct huge_bootmem_page));
2145 page = virt_to_page(m);
2147 WARN_ON(page_count(page) != 1);
2148 prep_compound_huge_page(page, h->order);
2149 WARN_ON(PageReserved(page));
2150 prep_new_huge_page(h, page, page_to_nid(page));
2152 * If we had gigantic hugepages allocated at boot time, we need
2153 * to restore the 'stolen' pages to totalram_pages in order to
2154 * fix confusing memory reports from free(1) and another
2155 * side-effects, like CommitLimit going negative.
2157 if (hstate_is_gigantic(h))
2158 adjust_managed_page_count(page, 1 << h->order);
2162 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2166 for (i = 0; i < h->max_huge_pages; ++i) {
2167 if (hstate_is_gigantic(h)) {
2168 if (!alloc_bootmem_huge_page(h))
2170 } else if (!alloc_fresh_huge_page(h,
2171 &node_states[N_MEMORY]))
2175 if (i < h->max_huge_pages) {
2178 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2179 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2180 h->max_huge_pages, buf, i);
2181 h->max_huge_pages = i;
2185 static void __init hugetlb_init_hstates(void)
2189 for_each_hstate(h) {
2190 if (minimum_order > huge_page_order(h))
2191 minimum_order = huge_page_order(h);
2193 /* oversize hugepages were init'ed in early boot */
2194 if (!hstate_is_gigantic(h))
2195 hugetlb_hstate_alloc_pages(h);
2197 VM_BUG_ON(minimum_order == UINT_MAX);
2200 static void __init report_hugepages(void)
2204 for_each_hstate(h) {
2207 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2208 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2209 buf, h->free_huge_pages);
2213 #ifdef CONFIG_HIGHMEM
2214 static void try_to_free_low(struct hstate *h, unsigned long count,
2215 nodemask_t *nodes_allowed)
2219 if (hstate_is_gigantic(h))
2222 for_each_node_mask(i, *nodes_allowed) {
2223 struct page *page, *next;
2224 struct list_head *freel = &h->hugepage_freelists[i];
2225 list_for_each_entry_safe(page, next, freel, lru) {
2226 if (count >= h->nr_huge_pages)
2228 if (PageHighMem(page))
2230 list_del(&page->lru);
2231 update_and_free_page(h, page);
2232 h->free_huge_pages--;
2233 h->free_huge_pages_node[page_to_nid(page)]--;
2238 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2239 nodemask_t *nodes_allowed)
2245 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2246 * balanced by operating on them in a round-robin fashion.
2247 * Returns 1 if an adjustment was made.
2249 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2254 VM_BUG_ON(delta != -1 && delta != 1);
2257 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2258 if (h->surplus_huge_pages_node[node])
2262 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2263 if (h->surplus_huge_pages_node[node] <
2264 h->nr_huge_pages_node[node])
2271 h->surplus_huge_pages += delta;
2272 h->surplus_huge_pages_node[node] += delta;
2276 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2277 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2278 nodemask_t *nodes_allowed)
2280 unsigned long min_count, ret;
2282 if (hstate_is_gigantic(h) && !gigantic_page_supported())
2283 return h->max_huge_pages;
2286 * Increase the pool size
2287 * First take pages out of surplus state. Then make up the
2288 * remaining difference by allocating fresh huge pages.
2290 * We might race with __alloc_buddy_huge_page() here and be unable
2291 * to convert a surplus huge page to a normal huge page. That is
2292 * not critical, though, it just means the overall size of the
2293 * pool might be one hugepage larger than it needs to be, but
2294 * within all the constraints specified by the sysctls.
2296 spin_lock(&hugetlb_lock);
2297 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2298 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2302 while (count > persistent_huge_pages(h)) {
2304 * If this allocation races such that we no longer need the
2305 * page, free_huge_page will handle it by freeing the page
2306 * and reducing the surplus.
2308 spin_unlock(&hugetlb_lock);
2310 /* yield cpu to avoid soft lockup */
2313 if (hstate_is_gigantic(h))
2314 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2316 ret = alloc_fresh_huge_page(h, nodes_allowed);
2317 spin_lock(&hugetlb_lock);
2321 /* Bail for signals. Probably ctrl-c from user */
2322 if (signal_pending(current))
2327 * Decrease the pool size
2328 * First return free pages to the buddy allocator (being careful
2329 * to keep enough around to satisfy reservations). Then place
2330 * pages into surplus state as needed so the pool will shrink
2331 * to the desired size as pages become free.
2333 * By placing pages into the surplus state independent of the
2334 * overcommit value, we are allowing the surplus pool size to
2335 * exceed overcommit. There are few sane options here. Since
2336 * __alloc_buddy_huge_page() is checking the global counter,
2337 * though, we'll note that we're not allowed to exceed surplus
2338 * and won't grow the pool anywhere else. Not until one of the
2339 * sysctls are changed, or the surplus pages go out of use.
2341 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2342 min_count = max(count, min_count);
2343 try_to_free_low(h, min_count, nodes_allowed);
2344 while (min_count < persistent_huge_pages(h)) {
2345 if (!free_pool_huge_page(h, nodes_allowed, 0))
2347 cond_resched_lock(&hugetlb_lock);
2349 while (count < persistent_huge_pages(h)) {
2350 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2354 ret = persistent_huge_pages(h);
2355 spin_unlock(&hugetlb_lock);
2359 #define HSTATE_ATTR_RO(_name) \
2360 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2362 #define HSTATE_ATTR(_name) \
2363 static struct kobj_attribute _name##_attr = \
2364 __ATTR(_name, 0644, _name##_show, _name##_store)
2366 static struct kobject *hugepages_kobj;
2367 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2369 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2371 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2375 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2376 if (hstate_kobjs[i] == kobj) {
2378 *nidp = NUMA_NO_NODE;
2382 return kobj_to_node_hstate(kobj, nidp);
2385 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2386 struct kobj_attribute *attr, char *buf)
2389 unsigned long nr_huge_pages;
2392 h = kobj_to_hstate(kobj, &nid);
2393 if (nid == NUMA_NO_NODE)
2394 nr_huge_pages = h->nr_huge_pages;
2396 nr_huge_pages = h->nr_huge_pages_node[nid];
2398 return sprintf(buf, "%lu\n", nr_huge_pages);
2401 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2402 struct hstate *h, int nid,
2403 unsigned long count, size_t len)
2406 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2408 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2413 if (nid == NUMA_NO_NODE) {
2415 * global hstate attribute
2417 if (!(obey_mempolicy &&
2418 init_nodemask_of_mempolicy(nodes_allowed))) {
2419 NODEMASK_FREE(nodes_allowed);
2420 nodes_allowed = &node_states[N_MEMORY];
2422 } else if (nodes_allowed) {
2424 * per node hstate attribute: adjust count to global,
2425 * but restrict alloc/free to the specified node.
2427 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2428 init_nodemask_of_node(nodes_allowed, nid);
2430 nodes_allowed = &node_states[N_MEMORY];
2432 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2434 if (nodes_allowed != &node_states[N_MEMORY])
2435 NODEMASK_FREE(nodes_allowed);
2439 NODEMASK_FREE(nodes_allowed);
2443 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2444 struct kobject *kobj, const char *buf,
2448 unsigned long count;
2452 err = kstrtoul(buf, 10, &count);
2456 h = kobj_to_hstate(kobj, &nid);
2457 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2460 static ssize_t nr_hugepages_show(struct kobject *kobj,
2461 struct kobj_attribute *attr, char *buf)
2463 return nr_hugepages_show_common(kobj, attr, buf);
2466 static ssize_t nr_hugepages_store(struct kobject *kobj,
2467 struct kobj_attribute *attr, const char *buf, size_t len)
2469 return nr_hugepages_store_common(false, kobj, buf, len);
2471 HSTATE_ATTR(nr_hugepages);
2476 * hstate attribute for optionally mempolicy-based constraint on persistent
2477 * huge page alloc/free.
2479 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2480 struct kobj_attribute *attr, char *buf)
2482 return nr_hugepages_show_common(kobj, attr, buf);
2485 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2486 struct kobj_attribute *attr, const char *buf, size_t len)
2488 return nr_hugepages_store_common(true, kobj, buf, len);
2490 HSTATE_ATTR(nr_hugepages_mempolicy);
2494 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2495 struct kobj_attribute *attr, char *buf)
2497 struct hstate *h = kobj_to_hstate(kobj, NULL);
2498 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2501 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2502 struct kobj_attribute *attr, const char *buf, size_t count)
2505 unsigned long input;
2506 struct hstate *h = kobj_to_hstate(kobj, NULL);
2508 if (hstate_is_gigantic(h))
2511 err = kstrtoul(buf, 10, &input);
2515 spin_lock(&hugetlb_lock);
2516 h->nr_overcommit_huge_pages = input;
2517 spin_unlock(&hugetlb_lock);
2521 HSTATE_ATTR(nr_overcommit_hugepages);
2523 static ssize_t free_hugepages_show(struct kobject *kobj,
2524 struct kobj_attribute *attr, char *buf)
2527 unsigned long free_huge_pages;
2530 h = kobj_to_hstate(kobj, &nid);
2531 if (nid == NUMA_NO_NODE)
2532 free_huge_pages = h->free_huge_pages;
2534 free_huge_pages = h->free_huge_pages_node[nid];
2536 return sprintf(buf, "%lu\n", free_huge_pages);
2538 HSTATE_ATTR_RO(free_hugepages);
2540 static ssize_t resv_hugepages_show(struct kobject *kobj,
2541 struct kobj_attribute *attr, char *buf)
2543 struct hstate *h = kobj_to_hstate(kobj, NULL);
2544 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2546 HSTATE_ATTR_RO(resv_hugepages);
2548 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2549 struct kobj_attribute *attr, char *buf)
2552 unsigned long surplus_huge_pages;
2555 h = kobj_to_hstate(kobj, &nid);
2556 if (nid == NUMA_NO_NODE)
2557 surplus_huge_pages = h->surplus_huge_pages;
2559 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2561 return sprintf(buf, "%lu\n", surplus_huge_pages);
2563 HSTATE_ATTR_RO(surplus_hugepages);
2565 static struct attribute *hstate_attrs[] = {
2566 &nr_hugepages_attr.attr,
2567 &nr_overcommit_hugepages_attr.attr,
2568 &free_hugepages_attr.attr,
2569 &resv_hugepages_attr.attr,
2570 &surplus_hugepages_attr.attr,
2572 &nr_hugepages_mempolicy_attr.attr,
2577 static const struct attribute_group hstate_attr_group = {
2578 .attrs = hstate_attrs,
2581 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2582 struct kobject **hstate_kobjs,
2583 const struct attribute_group *hstate_attr_group)
2586 int hi = hstate_index(h);
2588 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2589 if (!hstate_kobjs[hi])
2592 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2594 kobject_put(hstate_kobjs[hi]);
2599 static void __init hugetlb_sysfs_init(void)
2604 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2605 if (!hugepages_kobj)
2608 for_each_hstate(h) {
2609 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2610 hstate_kobjs, &hstate_attr_group);
2612 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2619 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2620 * with node devices in node_devices[] using a parallel array. The array
2621 * index of a node device or _hstate == node id.
2622 * This is here to avoid any static dependency of the node device driver, in
2623 * the base kernel, on the hugetlb module.
2625 struct node_hstate {
2626 struct kobject *hugepages_kobj;
2627 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2629 static struct node_hstate node_hstates[MAX_NUMNODES];
2632 * A subset of global hstate attributes for node devices
2634 static struct attribute *per_node_hstate_attrs[] = {
2635 &nr_hugepages_attr.attr,
2636 &free_hugepages_attr.attr,
2637 &surplus_hugepages_attr.attr,
2641 static const struct attribute_group per_node_hstate_attr_group = {
2642 .attrs = per_node_hstate_attrs,
2646 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2647 * Returns node id via non-NULL nidp.
2649 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2653 for (nid = 0; nid < nr_node_ids; nid++) {
2654 struct node_hstate *nhs = &node_hstates[nid];
2656 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2657 if (nhs->hstate_kobjs[i] == kobj) {
2669 * Unregister hstate attributes from a single node device.
2670 * No-op if no hstate attributes attached.
2672 static void hugetlb_unregister_node(struct node *node)
2675 struct node_hstate *nhs = &node_hstates[node->dev.id];
2677 if (!nhs->hugepages_kobj)
2678 return; /* no hstate attributes */
2680 for_each_hstate(h) {
2681 int idx = hstate_index(h);
2682 if (nhs->hstate_kobjs[idx]) {
2683 kobject_put(nhs->hstate_kobjs[idx]);
2684 nhs->hstate_kobjs[idx] = NULL;
2688 kobject_put(nhs->hugepages_kobj);
2689 nhs->hugepages_kobj = NULL;
2694 * Register hstate attributes for a single node device.
2695 * No-op if attributes already registered.
2697 static void hugetlb_register_node(struct node *node)
2700 struct node_hstate *nhs = &node_hstates[node->dev.id];
2703 if (nhs->hugepages_kobj)
2704 return; /* already allocated */
2706 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2708 if (!nhs->hugepages_kobj)
2711 for_each_hstate(h) {
2712 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2714 &per_node_hstate_attr_group);
2716 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2717 h->name, node->dev.id);
2718 hugetlb_unregister_node(node);
2725 * hugetlb init time: register hstate attributes for all registered node
2726 * devices of nodes that have memory. All on-line nodes should have
2727 * registered their associated device by this time.
2729 static void __init hugetlb_register_all_nodes(void)
2733 for_each_node_state(nid, N_MEMORY) {
2734 struct node *node = node_devices[nid];
2735 if (node->dev.id == nid)
2736 hugetlb_register_node(node);
2740 * Let the node device driver know we're here so it can
2741 * [un]register hstate attributes on node hotplug.
2743 register_hugetlbfs_with_node(hugetlb_register_node,
2744 hugetlb_unregister_node);
2746 #else /* !CONFIG_NUMA */
2748 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2756 static void hugetlb_register_all_nodes(void) { }
2760 static int __init hugetlb_init(void)
2764 if (!hugepages_supported())
2767 if (!size_to_hstate(default_hstate_size)) {
2768 if (default_hstate_size != 0) {
2769 pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2770 default_hstate_size, HPAGE_SIZE);
2773 default_hstate_size = HPAGE_SIZE;
2774 if (!size_to_hstate(default_hstate_size))
2775 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2777 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2778 if (default_hstate_max_huge_pages) {
2779 if (!default_hstate.max_huge_pages)
2780 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2783 hugetlb_init_hstates();
2784 gather_bootmem_prealloc();
2787 hugetlb_sysfs_init();
2788 hugetlb_register_all_nodes();
2789 hugetlb_cgroup_file_init();
2792 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2794 num_fault_mutexes = 1;
2796 hugetlb_fault_mutex_table =
2797 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2798 BUG_ON(!hugetlb_fault_mutex_table);
2800 for (i = 0; i < num_fault_mutexes; i++)
2801 mutex_init(&hugetlb_fault_mutex_table[i]);
2804 subsys_initcall(hugetlb_init);
2806 /* Should be called on processing a hugepagesz=... option */
2807 void __init hugetlb_bad_size(void)
2809 parsed_valid_hugepagesz = false;
2812 void __init hugetlb_add_hstate(unsigned int order)
2817 if (size_to_hstate(PAGE_SIZE << order)) {
2818 pr_warn("hugepagesz= specified twice, ignoring\n");
2821 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2823 h = &hstates[hugetlb_max_hstate++];
2825 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2826 h->nr_huge_pages = 0;
2827 h->free_huge_pages = 0;
2828 for (i = 0; i < MAX_NUMNODES; ++i)
2829 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2830 INIT_LIST_HEAD(&h->hugepage_activelist);
2831 h->next_nid_to_alloc = first_memory_node;
2832 h->next_nid_to_free = first_memory_node;
2833 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2834 huge_page_size(h)/1024);
2839 static int __init hugetlb_nrpages_setup(char *s)
2842 static unsigned long *last_mhp;
2844 if (!parsed_valid_hugepagesz) {
2845 pr_warn("hugepages = %s preceded by "
2846 "an unsupported hugepagesz, ignoring\n", s);
2847 parsed_valid_hugepagesz = true;
2851 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2852 * so this hugepages= parameter goes to the "default hstate".
2854 else if (!hugetlb_max_hstate)
2855 mhp = &default_hstate_max_huge_pages;
2857 mhp = &parsed_hstate->max_huge_pages;
2859 if (mhp == last_mhp) {
2860 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2864 if (sscanf(s, "%lu", mhp) <= 0)
2868 * Global state is always initialized later in hugetlb_init.
2869 * But we need to allocate >= MAX_ORDER hstates here early to still
2870 * use the bootmem allocator.
2872 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2873 hugetlb_hstate_alloc_pages(parsed_hstate);
2879 __setup("hugepages=", hugetlb_nrpages_setup);
2881 static int __init hugetlb_default_setup(char *s)
2883 default_hstate_size = memparse(s, &s);
2886 __setup("default_hugepagesz=", hugetlb_default_setup);
2888 static unsigned int cpuset_mems_nr(unsigned int *array)
2891 unsigned int nr = 0;
2893 for_each_node_mask(node, cpuset_current_mems_allowed)
2899 #ifdef CONFIG_SYSCTL
2900 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2901 struct ctl_table *table, int write,
2902 void __user *buffer, size_t *length, loff_t *ppos)
2904 struct hstate *h = &default_hstate;
2905 unsigned long tmp = h->max_huge_pages;
2908 if (!hugepages_supported())
2912 table->maxlen = sizeof(unsigned long);
2913 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2918 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2919 NUMA_NO_NODE, tmp, *length);
2924 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2925 void __user *buffer, size_t *length, loff_t *ppos)
2928 return hugetlb_sysctl_handler_common(false, table, write,
2929 buffer, length, ppos);
2933 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2934 void __user *buffer, size_t *length, loff_t *ppos)
2936 return hugetlb_sysctl_handler_common(true, table, write,
2937 buffer, length, ppos);
2939 #endif /* CONFIG_NUMA */
2941 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2942 void __user *buffer,
2943 size_t *length, loff_t *ppos)
2945 struct hstate *h = &default_hstate;
2949 if (!hugepages_supported())
2952 tmp = h->nr_overcommit_huge_pages;
2954 if (write && hstate_is_gigantic(h))
2958 table->maxlen = sizeof(unsigned long);
2959 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2964 spin_lock(&hugetlb_lock);
2965 h->nr_overcommit_huge_pages = tmp;
2966 spin_unlock(&hugetlb_lock);
2972 #endif /* CONFIG_SYSCTL */
2974 void hugetlb_report_meminfo(struct seq_file *m)
2977 unsigned long total = 0;
2979 if (!hugepages_supported())
2982 for_each_hstate(h) {
2983 unsigned long count = h->nr_huge_pages;
2985 total += (PAGE_SIZE << huge_page_order(h)) * count;
2987 if (h == &default_hstate)
2989 "HugePages_Total: %5lu\n"
2990 "HugePages_Free: %5lu\n"
2991 "HugePages_Rsvd: %5lu\n"
2992 "HugePages_Surp: %5lu\n"
2993 "Hugepagesize: %8lu kB\n",
2997 h->surplus_huge_pages,
2998 (PAGE_SIZE << huge_page_order(h)) / 1024);
3001 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3004 int hugetlb_report_node_meminfo(int nid, char *buf)
3006 struct hstate *h = &default_hstate;
3007 if (!hugepages_supported())
3010 "Node %d HugePages_Total: %5u\n"
3011 "Node %d HugePages_Free: %5u\n"
3012 "Node %d HugePages_Surp: %5u\n",
3013 nid, h->nr_huge_pages_node[nid],
3014 nid, h->free_huge_pages_node[nid],
3015 nid, h->surplus_huge_pages_node[nid]);
3018 void hugetlb_show_meminfo(void)
3023 if (!hugepages_supported())
3026 for_each_node_state(nid, N_MEMORY)
3028 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3030 h->nr_huge_pages_node[nid],
3031 h->free_huge_pages_node[nid],
3032 h->surplus_huge_pages_node[nid],
3033 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3036 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3038 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3039 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3042 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3043 unsigned long hugetlb_total_pages(void)
3046 unsigned long nr_total_pages = 0;
3049 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3050 return nr_total_pages;
3053 static int hugetlb_acct_memory(struct hstate *h, long delta)
3057 spin_lock(&hugetlb_lock);
3059 * When cpuset is configured, it breaks the strict hugetlb page
3060 * reservation as the accounting is done on a global variable. Such
3061 * reservation is completely rubbish in the presence of cpuset because
3062 * the reservation is not checked against page availability for the
3063 * current cpuset. Application can still potentially OOM'ed by kernel
3064 * with lack of free htlb page in cpuset that the task is in.
3065 * Attempt to enforce strict accounting with cpuset is almost
3066 * impossible (or too ugly) because cpuset is too fluid that
3067 * task or memory node can be dynamically moved between cpusets.
3069 * The change of semantics for shared hugetlb mapping with cpuset is
3070 * undesirable. However, in order to preserve some of the semantics,
3071 * we fall back to check against current free page availability as
3072 * a best attempt and hopefully to minimize the impact of changing
3073 * semantics that cpuset has.
3076 if (gather_surplus_pages(h, delta) < 0)
3079 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3080 return_unused_surplus_pages(h, delta);
3087 return_unused_surplus_pages(h, (unsigned long) -delta);
3090 spin_unlock(&hugetlb_lock);
3094 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3096 struct resv_map *resv = vma_resv_map(vma);
3099 * This new VMA should share its siblings reservation map if present.
3100 * The VMA will only ever have a valid reservation map pointer where
3101 * it is being copied for another still existing VMA. As that VMA
3102 * has a reference to the reservation map it cannot disappear until
3103 * after this open call completes. It is therefore safe to take a
3104 * new reference here without additional locking.
3106 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3107 kref_get(&resv->refs);
3110 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3112 struct hstate *h = hstate_vma(vma);
3113 struct resv_map *resv = vma_resv_map(vma);
3114 struct hugepage_subpool *spool = subpool_vma(vma);
3115 unsigned long reserve, start, end;
3118 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3121 start = vma_hugecache_offset(h, vma, vma->vm_start);
3122 end = vma_hugecache_offset(h, vma, vma->vm_end);
3124 reserve = (end - start) - region_count(resv, start, end);
3126 kref_put(&resv->refs, resv_map_release);
3130 * Decrement reserve counts. The global reserve count may be
3131 * adjusted if the subpool has a minimum size.
3133 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3134 hugetlb_acct_memory(h, -gbl_reserve);
3138 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3140 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3146 * We cannot handle pagefaults against hugetlb pages at all. They cause
3147 * handle_mm_fault() to try to instantiate regular-sized pages in the
3148 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3151 static int hugetlb_vm_op_fault(struct vm_fault *vmf)
3157 const struct vm_operations_struct hugetlb_vm_ops = {
3158 .fault = hugetlb_vm_op_fault,
3159 .open = hugetlb_vm_op_open,
3160 .close = hugetlb_vm_op_close,
3161 .split = hugetlb_vm_op_split,
3164 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3170 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3171 vma->vm_page_prot)));
3173 entry = huge_pte_wrprotect(mk_huge_pte(page,
3174 vma->vm_page_prot));
3176 entry = pte_mkyoung(entry);
3177 entry = pte_mkhuge(entry);
3178 entry = arch_make_huge_pte(entry, vma, page, writable);
3183 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3184 unsigned long address, pte_t *ptep)
3188 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3189 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3190 update_mmu_cache(vma, address, ptep);
3193 bool is_hugetlb_entry_migration(pte_t pte)
3197 if (huge_pte_none(pte) || pte_present(pte))
3199 swp = pte_to_swp_entry(pte);
3200 if (non_swap_entry(swp) && is_migration_entry(swp))
3206 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3210 if (huge_pte_none(pte) || pte_present(pte))
3212 swp = pte_to_swp_entry(pte);
3213 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3219 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3220 struct vm_area_struct *vma)
3222 pte_t *src_pte, *dst_pte, entry;
3223 struct page *ptepage;
3226 struct hstate *h = hstate_vma(vma);
3227 unsigned long sz = huge_page_size(h);
3228 unsigned long mmun_start; /* For mmu_notifiers */
3229 unsigned long mmun_end; /* For mmu_notifiers */
3232 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3234 mmun_start = vma->vm_start;
3235 mmun_end = vma->vm_end;
3237 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3239 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3240 spinlock_t *src_ptl, *dst_ptl;
3241 src_pte = huge_pte_offset(src, addr, sz);
3244 dst_pte = huge_pte_alloc(dst, addr, sz);
3250 /* If the pagetables are shared don't copy or take references */
3251 if (dst_pte == src_pte)
3254 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3255 src_ptl = huge_pte_lockptr(h, src, src_pte);
3256 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3257 entry = huge_ptep_get(src_pte);
3258 if (huge_pte_none(entry)) { /* skip none entry */
3260 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3261 is_hugetlb_entry_hwpoisoned(entry))) {
3262 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3264 if (is_write_migration_entry(swp_entry) && cow) {
3266 * COW mappings require pages in both
3267 * parent and child to be set to read.
3269 make_migration_entry_read(&swp_entry);
3270 entry = swp_entry_to_pte(swp_entry);
3271 set_huge_swap_pte_at(src, addr, src_pte,
3274 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3278 * No need to notify as we are downgrading page
3279 * table protection not changing it to point
3282 * See Documentation/vm/mmu_notifier.txt
3284 huge_ptep_set_wrprotect(src, addr, src_pte);
3286 entry = huge_ptep_get(src_pte);
3287 ptepage = pte_page(entry);
3289 page_dup_rmap(ptepage, true);
3290 set_huge_pte_at(dst, addr, dst_pte, entry);
3291 hugetlb_count_add(pages_per_huge_page(h), dst);
3293 spin_unlock(src_ptl);
3294 spin_unlock(dst_ptl);
3298 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3303 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3304 unsigned long start, unsigned long end,
3305 struct page *ref_page)
3307 struct mm_struct *mm = vma->vm_mm;
3308 unsigned long address;
3313 struct hstate *h = hstate_vma(vma);
3314 unsigned long sz = huge_page_size(h);
3315 const unsigned long mmun_start = start; /* For mmu_notifiers */
3316 const unsigned long mmun_end = end; /* For mmu_notifiers */
3318 WARN_ON(!is_vm_hugetlb_page(vma));
3319 BUG_ON(start & ~huge_page_mask(h));
3320 BUG_ON(end & ~huge_page_mask(h));
3323 * This is a hugetlb vma, all the pte entries should point
3326 tlb_remove_check_page_size_change(tlb, sz);
3327 tlb_start_vma(tlb, vma);
3328 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3330 for (; address < end; address += sz) {
3331 ptep = huge_pte_offset(mm, address, sz);
3335 ptl = huge_pte_lock(h, mm, ptep);
3336 if (huge_pmd_unshare(mm, &address, ptep)) {
3341 pte = huge_ptep_get(ptep);
3342 if (huge_pte_none(pte)) {
3348 * Migrating hugepage or HWPoisoned hugepage is already
3349 * unmapped and its refcount is dropped, so just clear pte here.
3351 if (unlikely(!pte_present(pte))) {
3352 huge_pte_clear(mm, address, ptep, sz);
3357 page = pte_page(pte);
3359 * If a reference page is supplied, it is because a specific
3360 * page is being unmapped, not a range. Ensure the page we
3361 * are about to unmap is the actual page of interest.
3364 if (page != ref_page) {
3369 * Mark the VMA as having unmapped its page so that
3370 * future faults in this VMA will fail rather than
3371 * looking like data was lost
3373 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3376 pte = huge_ptep_get_and_clear(mm, address, ptep);
3377 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3378 if (huge_pte_dirty(pte))
3379 set_page_dirty(page);
3381 hugetlb_count_sub(pages_per_huge_page(h), mm);
3382 page_remove_rmap(page, true);
3385 tlb_remove_page_size(tlb, page, huge_page_size(h));
3387 * Bail out after unmapping reference page if supplied
3392 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3393 tlb_end_vma(tlb, vma);
3396 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3397 struct vm_area_struct *vma, unsigned long start,
3398 unsigned long end, struct page *ref_page)
3400 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3403 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3404 * test will fail on a vma being torn down, and not grab a page table
3405 * on its way out. We're lucky that the flag has such an appropriate
3406 * name, and can in fact be safely cleared here. We could clear it
3407 * before the __unmap_hugepage_range above, but all that's necessary
3408 * is to clear it before releasing the i_mmap_rwsem. This works
3409 * because in the context this is called, the VMA is about to be
3410 * destroyed and the i_mmap_rwsem is held.
3412 vma->vm_flags &= ~VM_MAYSHARE;
3415 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3416 unsigned long end, struct page *ref_page)
3418 struct mm_struct *mm;
3419 struct mmu_gather tlb;
3423 tlb_gather_mmu(&tlb, mm, start, end);
3424 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3425 tlb_finish_mmu(&tlb, start, end);
3429 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3430 * mappping it owns the reserve page for. The intention is to unmap the page
3431 * from other VMAs and let the children be SIGKILLed if they are faulting the
3434 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3435 struct page *page, unsigned long address)
3437 struct hstate *h = hstate_vma(vma);
3438 struct vm_area_struct *iter_vma;
3439 struct address_space *mapping;
3443 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3444 * from page cache lookup which is in HPAGE_SIZE units.
3446 address = address & huge_page_mask(h);
3447 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3449 mapping = vma->vm_file->f_mapping;
3452 * Take the mapping lock for the duration of the table walk. As
3453 * this mapping should be shared between all the VMAs,
3454 * __unmap_hugepage_range() is called as the lock is already held
3456 i_mmap_lock_write(mapping);
3457 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3458 /* Do not unmap the current VMA */
3459 if (iter_vma == vma)
3463 * Shared VMAs have their own reserves and do not affect
3464 * MAP_PRIVATE accounting but it is possible that a shared
3465 * VMA is using the same page so check and skip such VMAs.
3467 if (iter_vma->vm_flags & VM_MAYSHARE)
3471 * Unmap the page from other VMAs without their own reserves.
3472 * They get marked to be SIGKILLed if they fault in these
3473 * areas. This is because a future no-page fault on this VMA
3474 * could insert a zeroed page instead of the data existing
3475 * from the time of fork. This would look like data corruption
3477 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3478 unmap_hugepage_range(iter_vma, address,
3479 address + huge_page_size(h), page);
3481 i_mmap_unlock_write(mapping);
3485 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3486 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3487 * cannot race with other handlers or page migration.
3488 * Keep the pte_same checks anyway to make transition from the mutex easier.
3490 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3491 unsigned long address, pte_t *ptep,
3492 struct page *pagecache_page, spinlock_t *ptl)
3495 struct hstate *h = hstate_vma(vma);
3496 struct page *old_page, *new_page;
3497 int ret = 0, outside_reserve = 0;
3498 unsigned long mmun_start; /* For mmu_notifiers */
3499 unsigned long mmun_end; /* For mmu_notifiers */
3501 pte = huge_ptep_get(ptep);
3502 old_page = pte_page(pte);
3505 /* If no-one else is actually using this page, avoid the copy
3506 * and just make the page writable */
3507 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3508 page_move_anon_rmap(old_page, vma);
3509 set_huge_ptep_writable(vma, address, ptep);
3514 * If the process that created a MAP_PRIVATE mapping is about to
3515 * perform a COW due to a shared page count, attempt to satisfy
3516 * the allocation without using the existing reserves. The pagecache
3517 * page is used to determine if the reserve at this address was
3518 * consumed or not. If reserves were used, a partial faulted mapping
3519 * at the time of fork() could consume its reserves on COW instead
3520 * of the full address range.
3522 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3523 old_page != pagecache_page)
3524 outside_reserve = 1;
3529 * Drop page table lock as buddy allocator may be called. It will
3530 * be acquired again before returning to the caller, as expected.
3533 new_page = alloc_huge_page(vma, address, outside_reserve);
3535 if (IS_ERR(new_page)) {
3537 * If a process owning a MAP_PRIVATE mapping fails to COW,
3538 * it is due to references held by a child and an insufficient
3539 * huge page pool. To guarantee the original mappers
3540 * reliability, unmap the page from child processes. The child
3541 * may get SIGKILLed if it later faults.
3543 if (outside_reserve) {
3545 BUG_ON(huge_pte_none(pte));
3546 unmap_ref_private(mm, vma, old_page, address);
3547 BUG_ON(huge_pte_none(pte));
3549 ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3552 pte_same(huge_ptep_get(ptep), pte)))
3553 goto retry_avoidcopy;
3555 * race occurs while re-acquiring page table
3556 * lock, and our job is done.
3561 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3562 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3563 goto out_release_old;
3567 * When the original hugepage is shared one, it does not have
3568 * anon_vma prepared.
3570 if (unlikely(anon_vma_prepare(vma))) {
3572 goto out_release_all;
3575 copy_user_huge_page(new_page, old_page, address, vma,
3576 pages_per_huge_page(h));
3577 __SetPageUptodate(new_page);
3578 set_page_huge_active(new_page);
3580 mmun_start = address & huge_page_mask(h);
3581 mmun_end = mmun_start + huge_page_size(h);
3582 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3585 * Retake the page table lock to check for racing updates
3586 * before the page tables are altered
3589 ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3591 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3592 ClearPagePrivate(new_page);
3595 huge_ptep_clear_flush(vma, address, ptep);
3596 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3597 set_huge_pte_at(mm, address, ptep,
3598 make_huge_pte(vma, new_page, 1));
3599 page_remove_rmap(old_page, true);
3600 hugepage_add_new_anon_rmap(new_page, vma, address);
3601 /* Make the old page be freed below */
3602 new_page = old_page;
3605 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3607 restore_reserve_on_error(h, vma, address, new_page);
3612 spin_lock(ptl); /* Caller expects lock to be held */
3616 /* Return the pagecache page at a given address within a VMA */
3617 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3618 struct vm_area_struct *vma, unsigned long address)
3620 struct address_space *mapping;
3623 mapping = vma->vm_file->f_mapping;
3624 idx = vma_hugecache_offset(h, vma, address);
3626 return find_lock_page(mapping, idx);
3630 * Return whether there is a pagecache page to back given address within VMA.
3631 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3633 static bool hugetlbfs_pagecache_present(struct hstate *h,
3634 struct vm_area_struct *vma, unsigned long address)
3636 struct address_space *mapping;
3640 mapping = vma->vm_file->f_mapping;
3641 idx = vma_hugecache_offset(h, vma, address);
3643 page = find_get_page(mapping, idx);
3646 return page != NULL;
3649 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3652 struct inode *inode = mapping->host;
3653 struct hstate *h = hstate_inode(inode);
3654 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3658 ClearPagePrivate(page);
3660 spin_lock(&inode->i_lock);
3661 inode->i_blocks += blocks_per_huge_page(h);
3662 spin_unlock(&inode->i_lock);
3666 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3667 struct address_space *mapping, pgoff_t idx,
3668 unsigned long address, pte_t *ptep, unsigned int flags)
3670 struct hstate *h = hstate_vma(vma);
3671 int ret = VM_FAULT_SIGBUS;
3679 * Currently, we are forced to kill the process in the event the
3680 * original mapper has unmapped pages from the child due to a failed
3681 * COW. Warn that such a situation has occurred as it may not be obvious
3683 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3684 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3690 * Use page lock to guard against racing truncation
3691 * before we get page_table_lock.
3694 page = find_lock_page(mapping, idx);
3696 size = i_size_read(mapping->host) >> huge_page_shift(h);
3701 * Check for page in userfault range
3703 if (userfaultfd_missing(vma)) {
3705 struct vm_fault vmf = {
3710 * Hard to debug if it ends up being
3711 * used by a callee that assumes
3712 * something about the other
3713 * uninitialized fields... same as in
3719 * hugetlb_fault_mutex must be dropped before
3720 * handling userfault. Reacquire after handling
3721 * fault to make calling code simpler.
3723 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping,
3725 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3726 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3727 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3731 page = alloc_huge_page(vma, address, 0);
3733 ret = PTR_ERR(page);
3737 ret = VM_FAULT_SIGBUS;
3740 clear_huge_page(page, address, pages_per_huge_page(h));
3741 __SetPageUptodate(page);
3742 set_page_huge_active(page);
3744 if (vma->vm_flags & VM_MAYSHARE) {
3745 int err = huge_add_to_page_cache(page, mapping, idx);
3754 if (unlikely(anon_vma_prepare(vma))) {
3756 goto backout_unlocked;
3762 * If memory error occurs between mmap() and fault, some process
3763 * don't have hwpoisoned swap entry for errored virtual address.
3764 * So we need to block hugepage fault by PG_hwpoison bit check.
3766 if (unlikely(PageHWPoison(page))) {
3767 ret = VM_FAULT_HWPOISON |
3768 VM_FAULT_SET_HINDEX(hstate_index(h));
3769 goto backout_unlocked;
3774 * If we are going to COW a private mapping later, we examine the
3775 * pending reservations for this page now. This will ensure that
3776 * any allocations necessary to record that reservation occur outside
3779 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3780 if (vma_needs_reservation(h, vma, address) < 0) {
3782 goto backout_unlocked;
3784 /* Just decrements count, does not deallocate */
3785 vma_end_reservation(h, vma, address);
3788 ptl = huge_pte_lock(h, mm, ptep);
3789 size = i_size_read(mapping->host) >> huge_page_shift(h);
3794 if (!huge_pte_none(huge_ptep_get(ptep)))
3798 ClearPagePrivate(page);
3799 hugepage_add_new_anon_rmap(page, vma, address);
3801 page_dup_rmap(page, true);
3802 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3803 && (vma->vm_flags & VM_SHARED)));
3804 set_huge_pte_at(mm, address, ptep, new_pte);
3806 hugetlb_count_add(pages_per_huge_page(h), mm);
3807 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3808 /* Optimization, do the COW without a second fault */
3809 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
3821 restore_reserve_on_error(h, vma, address, page);
3827 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3828 struct vm_area_struct *vma,
3829 struct address_space *mapping,
3830 pgoff_t idx, unsigned long address)
3832 unsigned long key[2];
3835 if (vma->vm_flags & VM_SHARED) {
3836 key[0] = (unsigned long) mapping;
3839 key[0] = (unsigned long) mm;
3840 key[1] = address >> huge_page_shift(h);
3843 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3845 return hash & (num_fault_mutexes - 1);
3849 * For uniprocesor systems we always use a single mutex, so just
3850 * return 0 and avoid the hashing overhead.
3852 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3853 struct vm_area_struct *vma,
3854 struct address_space *mapping,
3855 pgoff_t idx, unsigned long address)
3861 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3862 unsigned long address, unsigned int flags)
3869 struct page *page = NULL;
3870 struct page *pagecache_page = NULL;
3871 struct hstate *h = hstate_vma(vma);
3872 struct address_space *mapping;
3873 int need_wait_lock = 0;
3875 address &= huge_page_mask(h);
3877 ptep = huge_pte_offset(mm, address, huge_page_size(h));
3879 entry = huge_ptep_get(ptep);
3880 if (unlikely(is_hugetlb_entry_migration(entry))) {
3881 migration_entry_wait_huge(vma, mm, ptep);
3883 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3884 return VM_FAULT_HWPOISON_LARGE |
3885 VM_FAULT_SET_HINDEX(hstate_index(h));
3887 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3889 return VM_FAULT_OOM;
3892 mapping = vma->vm_file->f_mapping;
3893 idx = vma_hugecache_offset(h, vma, address);
3896 * Serialize hugepage allocation and instantiation, so that we don't
3897 * get spurious allocation failures if two CPUs race to instantiate
3898 * the same page in the page cache.
3900 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3901 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3903 entry = huge_ptep_get(ptep);
3904 if (huge_pte_none(entry)) {
3905 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3912 * entry could be a migration/hwpoison entry at this point, so this
3913 * check prevents the kernel from going below assuming that we have
3914 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3915 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3918 if (!pte_present(entry))
3922 * If we are going to COW the mapping later, we examine the pending
3923 * reservations for this page now. This will ensure that any
3924 * allocations necessary to record that reservation occur outside the
3925 * spinlock. For private mappings, we also lookup the pagecache
3926 * page now as it is used to determine if a reservation has been
3929 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3930 if (vma_needs_reservation(h, vma, address) < 0) {
3934 /* Just decrements count, does not deallocate */
3935 vma_end_reservation(h, vma, address);
3937 if (!(vma->vm_flags & VM_MAYSHARE))
3938 pagecache_page = hugetlbfs_pagecache_page(h,
3942 ptl = huge_pte_lock(h, mm, ptep);
3944 /* Check for a racing update before calling hugetlb_cow */
3945 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3949 * hugetlb_cow() requires page locks of pte_page(entry) and
3950 * pagecache_page, so here we need take the former one
3951 * when page != pagecache_page or !pagecache_page.
3953 page = pte_page(entry);
3954 if (page != pagecache_page)
3955 if (!trylock_page(page)) {
3962 if (flags & FAULT_FLAG_WRITE) {
3963 if (!huge_pte_write(entry)) {
3964 ret = hugetlb_cow(mm, vma, address, ptep,
3965 pagecache_page, ptl);
3968 entry = huge_pte_mkdirty(entry);
3970 entry = pte_mkyoung(entry);
3971 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3972 flags & FAULT_FLAG_WRITE))
3973 update_mmu_cache(vma, address, ptep);
3975 if (page != pagecache_page)
3981 if (pagecache_page) {
3982 unlock_page(pagecache_page);
3983 put_page(pagecache_page);
3986 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3988 * Generally it's safe to hold refcount during waiting page lock. But
3989 * here we just wait to defer the next page fault to avoid busy loop and
3990 * the page is not used after unlocked before returning from the current
3991 * page fault. So we are safe from accessing freed page, even if we wait
3992 * here without taking refcount.
3995 wait_on_page_locked(page);
4000 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4001 * modifications for huge pages.
4003 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4005 struct vm_area_struct *dst_vma,
4006 unsigned long dst_addr,
4007 unsigned long src_addr,
4008 struct page **pagep)
4010 struct address_space *mapping;
4013 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4014 struct hstate *h = hstate_vma(dst_vma);
4022 page = alloc_huge_page(dst_vma, dst_addr, 0);
4026 ret = copy_huge_page_from_user(page,
4027 (const void __user *) src_addr,
4028 pages_per_huge_page(h), false);
4030 /* fallback to copy_from_user outside mmap_sem */
4031 if (unlikely(ret)) {
4034 /* don't free the page */
4043 * The memory barrier inside __SetPageUptodate makes sure that
4044 * preceding stores to the page contents become visible before
4045 * the set_pte_at() write.
4047 __SetPageUptodate(page);
4048 set_page_huge_active(page);
4050 mapping = dst_vma->vm_file->f_mapping;
4051 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4054 * If shared, add to page cache
4057 size = i_size_read(mapping->host) >> huge_page_shift(h);
4060 goto out_release_nounlock;
4063 * Serialization between remove_inode_hugepages() and
4064 * huge_add_to_page_cache() below happens through the
4065 * hugetlb_fault_mutex_table that here must be hold by
4068 ret = huge_add_to_page_cache(page, mapping, idx);
4070 goto out_release_nounlock;
4073 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4077 * Recheck the i_size after holding PT lock to make sure not
4078 * to leave any page mapped (as page_mapped()) beyond the end
4079 * of the i_size (remove_inode_hugepages() is strict about
4080 * enforcing that). If we bail out here, we'll also leave a
4081 * page in the radix tree in the vm_shared case beyond the end
4082 * of the i_size, but remove_inode_hugepages() will take care
4083 * of it as soon as we drop the hugetlb_fault_mutex_table.
4085 size = i_size_read(mapping->host) >> huge_page_shift(h);
4088 goto out_release_unlock;
4091 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4092 goto out_release_unlock;
4095 page_dup_rmap(page, true);
4097 ClearPagePrivate(page);
4098 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4101 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4102 if (dst_vma->vm_flags & VM_WRITE)
4103 _dst_pte = huge_pte_mkdirty(_dst_pte);
4104 _dst_pte = pte_mkyoung(_dst_pte);
4106 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4108 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4109 dst_vma->vm_flags & VM_WRITE);
4110 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4112 /* No need to invalidate - it was non-present before */
4113 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4125 out_release_nounlock:
4130 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4131 struct page **pages, struct vm_area_struct **vmas,
4132 unsigned long *position, unsigned long *nr_pages,
4133 long i, unsigned int flags, int *nonblocking)
4135 unsigned long pfn_offset;
4136 unsigned long vaddr = *position;
4137 unsigned long remainder = *nr_pages;
4138 struct hstate *h = hstate_vma(vma);
4141 while (vaddr < vma->vm_end && remainder) {
4143 spinlock_t *ptl = NULL;
4148 * If we have a pending SIGKILL, don't keep faulting pages and
4149 * potentially allocating memory.
4151 if (unlikely(fatal_signal_pending(current))) {
4157 * Some archs (sparc64, sh*) have multiple pte_ts to
4158 * each hugepage. We have to make sure we get the
4159 * first, for the page indexing below to work.
4161 * Note that page table lock is not held when pte is null.
4163 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4166 ptl = huge_pte_lock(h, mm, pte);
4167 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4170 * When coredumping, it suits get_dump_page if we just return
4171 * an error where there's an empty slot with no huge pagecache
4172 * to back it. This way, we avoid allocating a hugepage, and
4173 * the sparse dumpfile avoids allocating disk blocks, but its
4174 * huge holes still show up with zeroes where they need to be.
4176 if (absent && (flags & FOLL_DUMP) &&
4177 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4185 * We need call hugetlb_fault for both hugepages under migration
4186 * (in which case hugetlb_fault waits for the migration,) and
4187 * hwpoisoned hugepages (in which case we need to prevent the
4188 * caller from accessing to them.) In order to do this, we use
4189 * here is_swap_pte instead of is_hugetlb_entry_migration and
4190 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4191 * both cases, and because we can't follow correct pages
4192 * directly from any kind of swap entries.
4194 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4195 ((flags & FOLL_WRITE) &&
4196 !huge_pte_write(huge_ptep_get(pte)))) {
4198 unsigned int fault_flags = 0;
4202 if (flags & FOLL_WRITE)
4203 fault_flags |= FAULT_FLAG_WRITE;
4205 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4206 if (flags & FOLL_NOWAIT)
4207 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4208 FAULT_FLAG_RETRY_NOWAIT;
4209 if (flags & FOLL_TRIED) {
4210 VM_WARN_ON_ONCE(fault_flags &
4211 FAULT_FLAG_ALLOW_RETRY);
4212 fault_flags |= FAULT_FLAG_TRIED;
4214 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4215 if (ret & VM_FAULT_ERROR) {
4216 err = vm_fault_to_errno(ret, flags);
4220 if (ret & VM_FAULT_RETRY) {
4225 * VM_FAULT_RETRY must not return an
4226 * error, it will return zero
4229 * No need to update "position" as the
4230 * caller will not check it after
4231 * *nr_pages is set to 0.
4238 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4239 page = pte_page(huge_ptep_get(pte));
4242 pages[i] = mem_map_offset(page, pfn_offset);
4253 if (vaddr < vma->vm_end && remainder &&
4254 pfn_offset < pages_per_huge_page(h)) {
4256 * We use pfn_offset to avoid touching the pageframes
4257 * of this compound page.
4263 *nr_pages = remainder;
4265 * setting position is actually required only if remainder is
4266 * not zero but it's faster not to add a "if (remainder)"
4274 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4276 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4279 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4282 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4283 unsigned long address, unsigned long end, pgprot_t newprot)
4285 struct mm_struct *mm = vma->vm_mm;
4286 unsigned long start = address;
4289 struct hstate *h = hstate_vma(vma);
4290 unsigned long pages = 0;
4292 BUG_ON(address >= end);
4293 flush_cache_range(vma, address, end);
4295 mmu_notifier_invalidate_range_start(mm, start, end);
4296 i_mmap_lock_write(vma->vm_file->f_mapping);
4297 for (; address < end; address += huge_page_size(h)) {
4299 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4302 ptl = huge_pte_lock(h, mm, ptep);
4303 if (huge_pmd_unshare(mm, &address, ptep)) {
4308 pte = huge_ptep_get(ptep);
4309 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4313 if (unlikely(is_hugetlb_entry_migration(pte))) {
4314 swp_entry_t entry = pte_to_swp_entry(pte);
4316 if (is_write_migration_entry(entry)) {
4319 make_migration_entry_read(&entry);
4320 newpte = swp_entry_to_pte(entry);
4321 set_huge_swap_pte_at(mm, address, ptep,
4322 newpte, huge_page_size(h));
4328 if (!huge_pte_none(pte)) {
4329 pte = huge_ptep_get_and_clear(mm, address, ptep);
4330 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
4331 pte = arch_make_huge_pte(pte, vma, NULL, 0);
4332 set_huge_pte_at(mm, address, ptep, pte);
4338 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4339 * may have cleared our pud entry and done put_page on the page table:
4340 * once we release i_mmap_rwsem, another task can do the final put_page
4341 * and that page table be reused and filled with junk.
4343 flush_hugetlb_tlb_range(vma, start, end);
4345 * No need to call mmu_notifier_invalidate_range() we are downgrading
4346 * page table protection not changing it to point to a new page.
4348 * See Documentation/vm/mmu_notifier.txt
4350 i_mmap_unlock_write(vma->vm_file->f_mapping);
4351 mmu_notifier_invalidate_range_end(mm, start, end);
4353 return pages << h->order;
4356 int hugetlb_reserve_pages(struct inode *inode,
4358 struct vm_area_struct *vma,
4359 vm_flags_t vm_flags)
4362 struct hstate *h = hstate_inode(inode);
4363 struct hugepage_subpool *spool = subpool_inode(inode);
4364 struct resv_map *resv_map;
4368 * Only apply hugepage reservation if asked. At fault time, an
4369 * attempt will be made for VM_NORESERVE to allocate a page
4370 * without using reserves
4372 if (vm_flags & VM_NORESERVE)
4376 * Shared mappings base their reservation on the number of pages that
4377 * are already allocated on behalf of the file. Private mappings need
4378 * to reserve the full area even if read-only as mprotect() may be
4379 * called to make the mapping read-write. Assume !vma is a shm mapping
4381 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4382 resv_map = inode_resv_map(inode);
4384 chg = region_chg(resv_map, from, to);
4387 resv_map = resv_map_alloc();
4393 set_vma_resv_map(vma, resv_map);
4394 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4403 * There must be enough pages in the subpool for the mapping. If
4404 * the subpool has a minimum size, there may be some global
4405 * reservations already in place (gbl_reserve).
4407 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4408 if (gbl_reserve < 0) {
4414 * Check enough hugepages are available for the reservation.
4415 * Hand the pages back to the subpool if there are not
4417 ret = hugetlb_acct_memory(h, gbl_reserve);
4419 /* put back original number of pages, chg */
4420 (void)hugepage_subpool_put_pages(spool, chg);
4425 * Account for the reservations made. Shared mappings record regions
4426 * that have reservations as they are shared by multiple VMAs.
4427 * When the last VMA disappears, the region map says how much
4428 * the reservation was and the page cache tells how much of
4429 * the reservation was consumed. Private mappings are per-VMA and
4430 * only the consumed reservations are tracked. When the VMA
4431 * disappears, the original reservation is the VMA size and the
4432 * consumed reservations are stored in the map. Hence, nothing
4433 * else has to be done for private mappings here
4435 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4436 long add = region_add(resv_map, from, to);
4438 if (unlikely(chg > add)) {
4440 * pages in this range were added to the reserve
4441 * map between region_chg and region_add. This
4442 * indicates a race with alloc_huge_page. Adjust
4443 * the subpool and reserve counts modified above
4444 * based on the difference.
4448 rsv_adjust = hugepage_subpool_put_pages(spool,
4450 hugetlb_acct_memory(h, -rsv_adjust);
4455 if (!vma || vma->vm_flags & VM_MAYSHARE)
4456 /* Don't call region_abort if region_chg failed */
4458 region_abort(resv_map, from, to);
4459 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4460 kref_put(&resv_map->refs, resv_map_release);
4464 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4467 struct hstate *h = hstate_inode(inode);
4468 struct resv_map *resv_map = inode_resv_map(inode);
4470 struct hugepage_subpool *spool = subpool_inode(inode);
4474 chg = region_del(resv_map, start, end);
4476 * region_del() can fail in the rare case where a region
4477 * must be split and another region descriptor can not be
4478 * allocated. If end == LONG_MAX, it will not fail.
4484 spin_lock(&inode->i_lock);
4485 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4486 spin_unlock(&inode->i_lock);
4489 * If the subpool has a minimum size, the number of global
4490 * reservations to be released may be adjusted.
4492 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4493 hugetlb_acct_memory(h, -gbl_reserve);
4498 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4499 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4500 struct vm_area_struct *vma,
4501 unsigned long addr, pgoff_t idx)
4503 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4505 unsigned long sbase = saddr & PUD_MASK;
4506 unsigned long s_end = sbase + PUD_SIZE;
4508 /* Allow segments to share if only one is marked locked */
4509 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4510 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4513 * match the virtual addresses, permission and the alignment of the
4516 if (pmd_index(addr) != pmd_index(saddr) ||
4517 vm_flags != svm_flags ||
4518 sbase < svma->vm_start || svma->vm_end < s_end)
4524 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4526 unsigned long base = addr & PUD_MASK;
4527 unsigned long end = base + PUD_SIZE;
4530 * check on proper vm_flags and page table alignment
4532 if (vma->vm_flags & VM_MAYSHARE &&
4533 vma->vm_start <= base && end <= vma->vm_end)
4539 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4540 * and returns the corresponding pte. While this is not necessary for the
4541 * !shared pmd case because we can allocate the pmd later as well, it makes the
4542 * code much cleaner. pmd allocation is essential for the shared case because
4543 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4544 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4545 * bad pmd for sharing.
4547 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4549 struct vm_area_struct *vma = find_vma(mm, addr);
4550 struct address_space *mapping = vma->vm_file->f_mapping;
4551 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4553 struct vm_area_struct *svma;
4554 unsigned long saddr;
4559 if (!vma_shareable(vma, addr))
4560 return (pte_t *)pmd_alloc(mm, pud, addr);
4562 i_mmap_lock_write(mapping);
4563 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4567 saddr = page_table_shareable(svma, vma, addr, idx);
4569 spte = huge_pte_offset(svma->vm_mm, saddr,
4570 vma_mmu_pagesize(svma));
4572 get_page(virt_to_page(spte));
4581 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4582 if (pud_none(*pud)) {
4583 pud_populate(mm, pud,
4584 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4587 put_page(virt_to_page(spte));
4591 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4592 i_mmap_unlock_write(mapping);
4597 * unmap huge page backed by shared pte.
4599 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4600 * indicated by page_count > 1, unmap is achieved by clearing pud and
4601 * decrementing the ref count. If count == 1, the pte page is not shared.
4603 * called with page table lock held.
4605 * returns: 1 successfully unmapped a shared pte page
4606 * 0 the underlying pte page is not shared, or it is the last user
4608 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4610 pgd_t *pgd = pgd_offset(mm, *addr);
4611 p4d_t *p4d = p4d_offset(pgd, *addr);
4612 pud_t *pud = pud_offset(p4d, *addr);
4614 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4615 if (page_count(virt_to_page(ptep)) == 1)
4619 put_page(virt_to_page(ptep));
4621 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4624 #define want_pmd_share() (1)
4625 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4626 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4631 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4635 #define want_pmd_share() (0)
4636 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4638 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4639 pte_t *huge_pte_alloc(struct mm_struct *mm,
4640 unsigned long addr, unsigned long sz)
4647 pgd = pgd_offset(mm, addr);
4648 p4d = p4d_alloc(mm, pgd, addr);
4651 pud = pud_alloc(mm, p4d, addr);
4653 if (sz == PUD_SIZE) {
4656 BUG_ON(sz != PMD_SIZE);
4657 if (want_pmd_share() && pud_none(*pud))
4658 pte = huge_pmd_share(mm, addr, pud);
4660 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4663 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4669 * huge_pte_offset() - Walk the page table to resolve the hugepage
4670 * entry at address @addr
4672 * Return: Pointer to page table or swap entry (PUD or PMD) for
4673 * address @addr, or NULL if a p*d_none() entry is encountered and the
4674 * size @sz doesn't match the hugepage size at this level of the page
4677 pte_t *huge_pte_offset(struct mm_struct *mm,
4678 unsigned long addr, unsigned long sz)
4685 pgd = pgd_offset(mm, addr);
4686 if (!pgd_present(*pgd))
4688 p4d = p4d_offset(pgd, addr);
4689 if (!p4d_present(*p4d))
4692 pud = pud_offset(p4d, addr);
4693 if (sz != PUD_SIZE && pud_none(*pud))
4695 /* hugepage or swap? */
4696 if (pud_huge(*pud) || !pud_present(*pud))
4697 return (pte_t *)pud;
4699 pmd = pmd_offset(pud, addr);
4700 if (sz != PMD_SIZE && pmd_none(*pmd))
4702 /* hugepage or swap? */
4703 if (pmd_huge(*pmd) || !pmd_present(*pmd))
4704 return (pte_t *)pmd;
4709 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4712 * These functions are overwritable if your architecture needs its own
4715 struct page * __weak
4716 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4719 return ERR_PTR(-EINVAL);
4722 struct page * __weak
4723 follow_huge_pd(struct vm_area_struct *vma,
4724 unsigned long address, hugepd_t hpd, int flags, int pdshift)
4726 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4730 struct page * __weak
4731 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4732 pmd_t *pmd, int flags)
4734 struct page *page = NULL;
4738 ptl = pmd_lockptr(mm, pmd);
4741 * make sure that the address range covered by this pmd is not
4742 * unmapped from other threads.
4744 if (!pmd_huge(*pmd))
4746 pte = huge_ptep_get((pte_t *)pmd);
4747 if (pte_present(pte)) {
4748 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4749 if (flags & FOLL_GET)
4752 if (is_hugetlb_entry_migration(pte)) {
4754 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4758 * hwpoisoned entry is treated as no_page_table in
4759 * follow_page_mask().
4767 struct page * __weak
4768 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4769 pud_t *pud, int flags)
4771 if (flags & FOLL_GET)
4774 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4777 struct page * __weak
4778 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
4780 if (flags & FOLL_GET)
4783 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
4786 bool isolate_huge_page(struct page *page, struct list_head *list)
4790 VM_BUG_ON_PAGE(!PageHead(page), page);
4791 spin_lock(&hugetlb_lock);
4792 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4796 clear_page_huge_active(page);
4797 list_move_tail(&page->lru, list);
4799 spin_unlock(&hugetlb_lock);
4803 void putback_active_hugepage(struct page *page)
4805 VM_BUG_ON_PAGE(!PageHead(page), page);
4806 spin_lock(&hugetlb_lock);
4807 set_page_huge_active(page);
4808 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4809 spin_unlock(&hugetlb_lock);