1 // SPDX-License-Identifier: GPL-2.0-only
3 * Generic hugetlb support.
4 * (C) Nadia Yvette Chambers, April 2004
6 #include <linux/list.h>
7 #include <linux/init.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/sched/mm.h>
23 #include <linux/mmdebug.h>
24 #include <linux/sched/signal.h>
25 #include <linux/rmap.h>
26 #include <linux/string_helpers.h>
27 #include <linux/swap.h>
28 #include <linux/swapops.h>
29 #include <linux/jhash.h>
30 #include <linux/numa.h>
31 #include <linux/llist.h>
32 #include <linux/cma.h>
35 #include <asm/pgalloc.h>
39 #include <linux/hugetlb.h>
40 #include <linux/hugetlb_cgroup.h>
41 #include <linux/node.h>
42 #include <linux/userfaultfd_k.h>
43 #include <linux/page_owner.h>
46 int hugetlb_max_hstate __read_mostly;
47 unsigned int default_hstate_idx;
48 struct hstate hstates[HUGE_MAX_HSTATE];
51 static struct cma *hugetlb_cma[MAX_NUMNODES];
53 static unsigned long hugetlb_cma_size __initdata;
56 * Minimum page order among possible hugepage sizes, set to a proper value
59 static unsigned int minimum_order __read_mostly = UINT_MAX;
61 __initdata LIST_HEAD(huge_boot_pages);
63 /* for command line parsing */
64 static struct hstate * __initdata parsed_hstate;
65 static unsigned long __initdata default_hstate_max_huge_pages;
66 static bool __initdata parsed_valid_hugepagesz = true;
67 static bool __initdata parsed_default_hugepagesz;
70 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
71 * free_huge_pages, and surplus_huge_pages.
73 DEFINE_SPINLOCK(hugetlb_lock);
76 * Serializes faults on the same logical page. This is used to
77 * prevent spurious OOMs when the hugepage pool is fully utilized.
79 static int num_fault_mutexes;
80 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
82 static inline bool PageHugeFreed(struct page *head)
84 return page_private(head + 4) == -1UL;
87 static inline void SetPageHugeFreed(struct page *head)
89 set_page_private(head + 4, -1UL);
92 static inline void ClearPageHugeFreed(struct page *head)
94 set_page_private(head + 4, 0);
97 /* Forward declaration */
98 static int hugetlb_acct_memory(struct hstate *h, long delta);
100 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
102 bool free = (spool->count == 0) && (spool->used_hpages == 0);
104 spin_unlock(&spool->lock);
106 /* If no pages are used, and no other handles to the subpool
107 * remain, give up any reservations based on minimum size and
108 * free the subpool */
110 if (spool->min_hpages != -1)
111 hugetlb_acct_memory(spool->hstate,
117 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
120 struct hugepage_subpool *spool;
122 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
126 spin_lock_init(&spool->lock);
128 spool->max_hpages = max_hpages;
130 spool->min_hpages = min_hpages;
132 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
136 spool->rsv_hpages = min_hpages;
141 void hugepage_put_subpool(struct hugepage_subpool *spool)
143 spin_lock(&spool->lock);
144 BUG_ON(!spool->count);
146 unlock_or_release_subpool(spool);
150 * Subpool accounting for allocating and reserving pages.
151 * Return -ENOMEM if there are not enough resources to satisfy the
152 * request. Otherwise, return the number of pages by which the
153 * global pools must be adjusted (upward). The returned value may
154 * only be different than the passed value (delta) in the case where
155 * a subpool minimum size must be maintained.
157 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
165 spin_lock(&spool->lock);
167 if (spool->max_hpages != -1) { /* maximum size accounting */
168 if ((spool->used_hpages + delta) <= spool->max_hpages)
169 spool->used_hpages += delta;
176 /* minimum size accounting */
177 if (spool->min_hpages != -1 && spool->rsv_hpages) {
178 if (delta > spool->rsv_hpages) {
180 * Asking for more reserves than those already taken on
181 * behalf of subpool. Return difference.
183 ret = delta - spool->rsv_hpages;
184 spool->rsv_hpages = 0;
186 ret = 0; /* reserves already accounted for */
187 spool->rsv_hpages -= delta;
192 spin_unlock(&spool->lock);
197 * Subpool accounting for freeing and unreserving pages.
198 * Return the number of global page reservations that must be dropped.
199 * The return value may only be different than the passed value (delta)
200 * in the case where a subpool minimum size must be maintained.
202 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
210 spin_lock(&spool->lock);
212 if (spool->max_hpages != -1) /* maximum size accounting */
213 spool->used_hpages -= delta;
215 /* minimum size accounting */
216 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
217 if (spool->rsv_hpages + delta <= spool->min_hpages)
220 ret = spool->rsv_hpages + delta - spool->min_hpages;
222 spool->rsv_hpages += delta;
223 if (spool->rsv_hpages > spool->min_hpages)
224 spool->rsv_hpages = spool->min_hpages;
228 * If hugetlbfs_put_super couldn't free spool due to an outstanding
229 * quota reference, free it now.
231 unlock_or_release_subpool(spool);
236 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
238 return HUGETLBFS_SB(inode->i_sb)->spool;
241 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
243 return subpool_inode(file_inode(vma->vm_file));
246 /* Helper that removes a struct file_region from the resv_map cache and returns
249 static struct file_region *
250 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
252 struct file_region *nrg = NULL;
254 VM_BUG_ON(resv->region_cache_count <= 0);
256 resv->region_cache_count--;
257 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
258 list_del(&nrg->link);
266 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
267 struct file_region *rg)
269 #ifdef CONFIG_CGROUP_HUGETLB
270 nrg->reservation_counter = rg->reservation_counter;
277 /* Helper that records hugetlb_cgroup uncharge info. */
278 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
280 struct resv_map *resv,
281 struct file_region *nrg)
283 #ifdef CONFIG_CGROUP_HUGETLB
285 nrg->reservation_counter =
286 &h_cg->rsvd_hugepage[hstate_index(h)];
287 nrg->css = &h_cg->css;
288 if (!resv->pages_per_hpage)
289 resv->pages_per_hpage = pages_per_huge_page(h);
290 /* pages_per_hpage should be the same for all entries in
293 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
295 nrg->reservation_counter = NULL;
301 static bool has_same_uncharge_info(struct file_region *rg,
302 struct file_region *org)
304 #ifdef CONFIG_CGROUP_HUGETLB
306 rg->reservation_counter == org->reservation_counter &&
314 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
316 struct file_region *nrg = NULL, *prg = NULL;
318 prg = list_prev_entry(rg, link);
319 if (&prg->link != &resv->regions && prg->to == rg->from &&
320 has_same_uncharge_info(prg, rg)) {
329 nrg = list_next_entry(rg, link);
330 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
331 has_same_uncharge_info(nrg, rg)) {
332 nrg->from = rg->from;
340 * Must be called with resv->lock held.
342 * Calling this with regions_needed != NULL will count the number of pages
343 * to be added but will not modify the linked list. And regions_needed will
344 * indicate the number of file_regions needed in the cache to carry out to add
345 * the regions for this range.
347 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
348 struct hugetlb_cgroup *h_cg,
349 struct hstate *h, long *regions_needed)
352 struct list_head *head = &resv->regions;
353 long last_accounted_offset = f;
354 struct file_region *rg = NULL, *trg = NULL, *nrg = NULL;
359 /* In this loop, we essentially handle an entry for the range
360 * [last_accounted_offset, rg->from), at every iteration, with some
363 list_for_each_entry_safe(rg, trg, head, link) {
364 /* Skip irrelevant regions that start before our range. */
366 /* If this region ends after the last accounted offset,
367 * then we need to update last_accounted_offset.
369 if (rg->to > last_accounted_offset)
370 last_accounted_offset = rg->to;
374 /* When we find a region that starts beyond our range, we've
380 /* Add an entry for last_accounted_offset -> rg->from, and
381 * update last_accounted_offset.
383 if (rg->from > last_accounted_offset) {
384 add += rg->from - last_accounted_offset;
385 if (!regions_needed) {
386 nrg = get_file_region_entry_from_cache(
387 resv, last_accounted_offset, rg->from);
388 record_hugetlb_cgroup_uncharge_info(h_cg, h,
390 list_add(&nrg->link, rg->link.prev);
391 coalesce_file_region(resv, nrg);
393 *regions_needed += 1;
396 last_accounted_offset = rg->to;
399 /* Handle the case where our range extends beyond
400 * last_accounted_offset.
402 if (last_accounted_offset < t) {
403 add += t - last_accounted_offset;
404 if (!regions_needed) {
405 nrg = get_file_region_entry_from_cache(
406 resv, last_accounted_offset, t);
407 record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg);
408 list_add(&nrg->link, rg->link.prev);
409 coalesce_file_region(resv, nrg);
411 *regions_needed += 1;
418 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
420 static int allocate_file_region_entries(struct resv_map *resv,
422 __must_hold(&resv->lock)
424 struct list_head allocated_regions;
425 int to_allocate = 0, i = 0;
426 struct file_region *trg = NULL, *rg = NULL;
428 VM_BUG_ON(regions_needed < 0);
430 INIT_LIST_HEAD(&allocated_regions);
433 * Check for sufficient descriptors in the cache to accommodate
434 * the number of in progress add operations plus regions_needed.
436 * This is a while loop because when we drop the lock, some other call
437 * to region_add or region_del may have consumed some region_entries,
438 * so we keep looping here until we finally have enough entries for
439 * (adds_in_progress + regions_needed).
441 while (resv->region_cache_count <
442 (resv->adds_in_progress + regions_needed)) {
443 to_allocate = resv->adds_in_progress + regions_needed -
444 resv->region_cache_count;
446 /* At this point, we should have enough entries in the cache
447 * for all the existings adds_in_progress. We should only be
448 * needing to allocate for regions_needed.
450 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
452 spin_unlock(&resv->lock);
453 for (i = 0; i < to_allocate; i++) {
454 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
457 list_add(&trg->link, &allocated_regions);
460 spin_lock(&resv->lock);
462 list_splice(&allocated_regions, &resv->region_cache);
463 resv->region_cache_count += to_allocate;
469 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
477 * Add the huge page range represented by [f, t) to the reserve
478 * map. Regions will be taken from the cache to fill in this range.
479 * Sufficient regions should exist in the cache due to the previous
480 * call to region_chg with the same range, but in some cases the cache will not
481 * have sufficient entries due to races with other code doing region_add or
482 * region_del. The extra needed entries will be allocated.
484 * regions_needed is the out value provided by a previous call to region_chg.
486 * Return the number of new huge pages added to the map. This number is greater
487 * than or equal to zero. If file_region entries needed to be allocated for
488 * this operation and we were not able to allocate, it returns -ENOMEM.
489 * region_add of regions of length 1 never allocate file_regions and cannot
490 * fail; region_chg will always allocate at least 1 entry and a region_add for
491 * 1 page will only require at most 1 entry.
493 static long region_add(struct resv_map *resv, long f, long t,
494 long in_regions_needed, struct hstate *h,
495 struct hugetlb_cgroup *h_cg)
497 long add = 0, actual_regions_needed = 0;
499 spin_lock(&resv->lock);
502 /* Count how many regions are actually needed to execute this add. */
503 add_reservation_in_range(resv, f, t, NULL, NULL,
504 &actual_regions_needed);
507 * Check for sufficient descriptors in the cache to accommodate
508 * this add operation. Note that actual_regions_needed may be greater
509 * than in_regions_needed, as the resv_map may have been modified since
510 * the region_chg call. In this case, we need to make sure that we
511 * allocate extra entries, such that we have enough for all the
512 * existing adds_in_progress, plus the excess needed for this
515 if (actual_regions_needed > in_regions_needed &&
516 resv->region_cache_count <
517 resv->adds_in_progress +
518 (actual_regions_needed - in_regions_needed)) {
519 /* region_add operation of range 1 should never need to
520 * allocate file_region entries.
522 VM_BUG_ON(t - f <= 1);
524 if (allocate_file_region_entries(
525 resv, actual_regions_needed - in_regions_needed)) {
532 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
534 resv->adds_in_progress -= in_regions_needed;
536 spin_unlock(&resv->lock);
542 * Examine the existing reserve map and determine how many
543 * huge pages in the specified range [f, t) are NOT currently
544 * represented. This routine is called before a subsequent
545 * call to region_add that will actually modify the reserve
546 * map to add the specified range [f, t). region_chg does
547 * not change the number of huge pages represented by the
548 * map. A number of new file_region structures is added to the cache as a
549 * placeholder, for the subsequent region_add call to use. At least 1
550 * file_region structure is added.
552 * out_regions_needed is the number of regions added to the
553 * resv->adds_in_progress. This value needs to be provided to a follow up call
554 * to region_add or region_abort for proper accounting.
556 * Returns the number of huge pages that need to be added to the existing
557 * reservation map for the range [f, t). This number is greater or equal to
558 * zero. -ENOMEM is returned if a new file_region structure or cache entry
559 * is needed and can not be allocated.
561 static long region_chg(struct resv_map *resv, long f, long t,
562 long *out_regions_needed)
566 spin_lock(&resv->lock);
568 /* Count how many hugepages in this range are NOT represented. */
569 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
572 if (*out_regions_needed == 0)
573 *out_regions_needed = 1;
575 if (allocate_file_region_entries(resv, *out_regions_needed))
578 resv->adds_in_progress += *out_regions_needed;
580 spin_unlock(&resv->lock);
585 * Abort the in progress add operation. The adds_in_progress field
586 * of the resv_map keeps track of the operations in progress between
587 * calls to region_chg and region_add. Operations are sometimes
588 * aborted after the call to region_chg. In such cases, region_abort
589 * is called to decrement the adds_in_progress counter. regions_needed
590 * is the value returned by the region_chg call, it is used to decrement
591 * the adds_in_progress counter.
593 * NOTE: The range arguments [f, t) are not needed or used in this
594 * routine. They are kept to make reading the calling code easier as
595 * arguments will match the associated region_chg call.
597 static void region_abort(struct resv_map *resv, long f, long t,
600 spin_lock(&resv->lock);
601 VM_BUG_ON(!resv->region_cache_count);
602 resv->adds_in_progress -= regions_needed;
603 spin_unlock(&resv->lock);
607 * Delete the specified range [f, t) from the reserve map. If the
608 * t parameter is LONG_MAX, this indicates that ALL regions after f
609 * should be deleted. Locate the regions which intersect [f, t)
610 * and either trim, delete or split the existing regions.
612 * Returns the number of huge pages deleted from the reserve map.
613 * In the normal case, the return value is zero or more. In the
614 * case where a region must be split, a new region descriptor must
615 * be allocated. If the allocation fails, -ENOMEM will be returned.
616 * NOTE: If the parameter t == LONG_MAX, then we will never split
617 * a region and possibly return -ENOMEM. Callers specifying
618 * t == LONG_MAX do not need to check for -ENOMEM error.
620 static long region_del(struct resv_map *resv, long f, long t)
622 struct list_head *head = &resv->regions;
623 struct file_region *rg, *trg;
624 struct file_region *nrg = NULL;
628 spin_lock(&resv->lock);
629 list_for_each_entry_safe(rg, trg, head, link) {
631 * Skip regions before the range to be deleted. file_region
632 * ranges are normally of the form [from, to). However, there
633 * may be a "placeholder" entry in the map which is of the form
634 * (from, to) with from == to. Check for placeholder entries
635 * at the beginning of the range to be deleted.
637 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
643 if (f > rg->from && t < rg->to) { /* Must split region */
645 * Check for an entry in the cache before dropping
646 * lock and attempting allocation.
649 resv->region_cache_count > resv->adds_in_progress) {
650 nrg = list_first_entry(&resv->region_cache,
653 list_del(&nrg->link);
654 resv->region_cache_count--;
658 spin_unlock(&resv->lock);
659 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
666 hugetlb_cgroup_uncharge_file_region(
669 /* New entry for end of split region */
673 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
675 INIT_LIST_HEAD(&nrg->link);
677 /* Original entry is trimmed */
680 list_add(&nrg->link, &rg->link);
685 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
686 del += rg->to - rg->from;
687 hugetlb_cgroup_uncharge_file_region(resv, rg,
694 if (f <= rg->from) { /* Trim beginning of region */
695 hugetlb_cgroup_uncharge_file_region(resv, rg,
700 } else { /* Trim end of region */
701 hugetlb_cgroup_uncharge_file_region(resv, rg,
709 spin_unlock(&resv->lock);
715 * A rare out of memory error was encountered which prevented removal of
716 * the reserve map region for a page. The huge page itself was free'ed
717 * and removed from the page cache. This routine will adjust the subpool
718 * usage count, and the global reserve count if needed. By incrementing
719 * these counts, the reserve map entry which could not be deleted will
720 * appear as a "reserved" entry instead of simply dangling with incorrect
723 void hugetlb_fix_reserve_counts(struct inode *inode)
725 struct hugepage_subpool *spool = subpool_inode(inode);
728 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
730 struct hstate *h = hstate_inode(inode);
732 hugetlb_acct_memory(h, 1);
737 * Count and return the number of huge pages in the reserve map
738 * that intersect with the range [f, t).
740 static long region_count(struct resv_map *resv, long f, long t)
742 struct list_head *head = &resv->regions;
743 struct file_region *rg;
746 spin_lock(&resv->lock);
747 /* Locate each segment we overlap with, and count that overlap. */
748 list_for_each_entry(rg, head, link) {
757 seg_from = max(rg->from, f);
758 seg_to = min(rg->to, t);
760 chg += seg_to - seg_from;
762 spin_unlock(&resv->lock);
768 * Convert the address within this vma to the page offset within
769 * the mapping, in pagecache page units; huge pages here.
771 static pgoff_t vma_hugecache_offset(struct hstate *h,
772 struct vm_area_struct *vma, unsigned long address)
774 return ((address - vma->vm_start) >> huge_page_shift(h)) +
775 (vma->vm_pgoff >> huge_page_order(h));
778 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
779 unsigned long address)
781 return vma_hugecache_offset(hstate_vma(vma), vma, address);
783 EXPORT_SYMBOL_GPL(linear_hugepage_index);
786 * Return the size of the pages allocated when backing a VMA. In the majority
787 * cases this will be same size as used by the page table entries.
789 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
791 if (vma->vm_ops && vma->vm_ops->pagesize)
792 return vma->vm_ops->pagesize(vma);
795 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
798 * Return the page size being used by the MMU to back a VMA. In the majority
799 * of cases, the page size used by the kernel matches the MMU size. On
800 * architectures where it differs, an architecture-specific 'strong'
801 * version of this symbol is required.
803 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
805 return vma_kernel_pagesize(vma);
809 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
810 * bits of the reservation map pointer, which are always clear due to
813 #define HPAGE_RESV_OWNER (1UL << 0)
814 #define HPAGE_RESV_UNMAPPED (1UL << 1)
815 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
818 * These helpers are used to track how many pages are reserved for
819 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
820 * is guaranteed to have their future faults succeed.
822 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
823 * the reserve counters are updated with the hugetlb_lock held. It is safe
824 * to reset the VMA at fork() time as it is not in use yet and there is no
825 * chance of the global counters getting corrupted as a result of the values.
827 * The private mapping reservation is represented in a subtly different
828 * manner to a shared mapping. A shared mapping has a region map associated
829 * with the underlying file, this region map represents the backing file
830 * pages which have ever had a reservation assigned which this persists even
831 * after the page is instantiated. A private mapping has a region map
832 * associated with the original mmap which is attached to all VMAs which
833 * reference it, this region map represents those offsets which have consumed
834 * reservation ie. where pages have been instantiated.
836 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
838 return (unsigned long)vma->vm_private_data;
841 static void set_vma_private_data(struct vm_area_struct *vma,
844 vma->vm_private_data = (void *)value;
848 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
849 struct hugetlb_cgroup *h_cg,
852 #ifdef CONFIG_CGROUP_HUGETLB
854 resv_map->reservation_counter = NULL;
855 resv_map->pages_per_hpage = 0;
856 resv_map->css = NULL;
858 resv_map->reservation_counter =
859 &h_cg->rsvd_hugepage[hstate_index(h)];
860 resv_map->pages_per_hpage = pages_per_huge_page(h);
861 resv_map->css = &h_cg->css;
866 struct resv_map *resv_map_alloc(void)
868 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
869 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
871 if (!resv_map || !rg) {
877 kref_init(&resv_map->refs);
878 spin_lock_init(&resv_map->lock);
879 INIT_LIST_HEAD(&resv_map->regions);
881 resv_map->adds_in_progress = 0;
883 * Initialize these to 0. On shared mappings, 0's here indicate these
884 * fields don't do cgroup accounting. On private mappings, these will be
885 * re-initialized to the proper values, to indicate that hugetlb cgroup
886 * reservations are to be un-charged from here.
888 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
890 INIT_LIST_HEAD(&resv_map->region_cache);
891 list_add(&rg->link, &resv_map->region_cache);
892 resv_map->region_cache_count = 1;
897 void resv_map_release(struct kref *ref)
899 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
900 struct list_head *head = &resv_map->region_cache;
901 struct file_region *rg, *trg;
903 /* Clear out any active regions before we release the map. */
904 region_del(resv_map, 0, LONG_MAX);
906 /* ... and any entries left in the cache */
907 list_for_each_entry_safe(rg, trg, head, link) {
912 VM_BUG_ON(resv_map->adds_in_progress);
917 static inline struct resv_map *inode_resv_map(struct inode *inode)
920 * At inode evict time, i_mapping may not point to the original
921 * address space within the inode. This original address space
922 * contains the pointer to the resv_map. So, always use the
923 * address space embedded within the inode.
924 * The VERY common case is inode->mapping == &inode->i_data but,
925 * this may not be true for device special inodes.
927 return (struct resv_map *)(&inode->i_data)->private_data;
930 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
932 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
933 if (vma->vm_flags & VM_MAYSHARE) {
934 struct address_space *mapping = vma->vm_file->f_mapping;
935 struct inode *inode = mapping->host;
937 return inode_resv_map(inode);
940 return (struct resv_map *)(get_vma_private_data(vma) &
945 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
947 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
948 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
950 set_vma_private_data(vma, (get_vma_private_data(vma) &
951 HPAGE_RESV_MASK) | (unsigned long)map);
954 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
956 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
957 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
959 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
962 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
964 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
966 return (get_vma_private_data(vma) & flag) != 0;
969 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
970 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
972 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
973 if (!(vma->vm_flags & VM_MAYSHARE))
974 vma->vm_private_data = (void *)0;
977 /* Returns true if the VMA has associated reserve pages */
978 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
980 if (vma->vm_flags & VM_NORESERVE) {
982 * This address is already reserved by other process(chg == 0),
983 * so, we should decrement reserved count. Without decrementing,
984 * reserve count remains after releasing inode, because this
985 * allocated page will go into page cache and is regarded as
986 * coming from reserved pool in releasing step. Currently, we
987 * don't have any other solution to deal with this situation
988 * properly, so add work-around here.
990 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
996 /* Shared mappings always use reserves */
997 if (vma->vm_flags & VM_MAYSHARE) {
999 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1000 * be a region map for all pages. The only situation where
1001 * there is no region map is if a hole was punched via
1002 * fallocate. In this case, there really are no reserves to
1003 * use. This situation is indicated if chg != 0.
1012 * Only the process that called mmap() has reserves for
1015 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1017 * Like the shared case above, a hole punch or truncate
1018 * could have been performed on the private mapping.
1019 * Examine the value of chg to determine if reserves
1020 * actually exist or were previously consumed.
1021 * Very Subtle - The value of chg comes from a previous
1022 * call to vma_needs_reserves(). The reserve map for
1023 * private mappings has different (opposite) semantics
1024 * than that of shared mappings. vma_needs_reserves()
1025 * has already taken this difference in semantics into
1026 * account. Therefore, the meaning of chg is the same
1027 * as in the shared case above. Code could easily be
1028 * combined, but keeping it separate draws attention to
1029 * subtle differences.
1040 static void enqueue_huge_page(struct hstate *h, struct page *page)
1042 int nid = page_to_nid(page);
1043 list_move(&page->lru, &h->hugepage_freelists[nid]);
1044 h->free_huge_pages++;
1045 h->free_huge_pages_node[nid]++;
1046 SetPageHugeFreed(page);
1049 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1052 bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA);
1054 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1055 if (nocma && is_migrate_cma_page(page))
1058 if (PageHWPoison(page))
1061 list_move(&page->lru, &h->hugepage_activelist);
1062 set_page_refcounted(page);
1063 ClearPageHugeFreed(page);
1064 h->free_huge_pages--;
1065 h->free_huge_pages_node[nid]--;
1072 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1075 unsigned int cpuset_mems_cookie;
1076 struct zonelist *zonelist;
1079 int node = NUMA_NO_NODE;
1081 zonelist = node_zonelist(nid, gfp_mask);
1084 cpuset_mems_cookie = read_mems_allowed_begin();
1085 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1088 if (!cpuset_zone_allowed(zone, gfp_mask))
1091 * no need to ask again on the same node. Pool is node rather than
1094 if (zone_to_nid(zone) == node)
1096 node = zone_to_nid(zone);
1098 page = dequeue_huge_page_node_exact(h, node);
1102 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1108 static struct page *dequeue_huge_page_vma(struct hstate *h,
1109 struct vm_area_struct *vma,
1110 unsigned long address, int avoid_reserve,
1114 struct mempolicy *mpol;
1116 nodemask_t *nodemask;
1120 * A child process with MAP_PRIVATE mappings created by their parent
1121 * have no page reserves. This check ensures that reservations are
1122 * not "stolen". The child may still get SIGKILLed
1124 if (!vma_has_reserves(vma, chg) &&
1125 h->free_huge_pages - h->resv_huge_pages == 0)
1128 /* If reserves cannot be used, ensure enough pages are in the pool */
1129 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1132 gfp_mask = htlb_alloc_mask(h);
1133 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1134 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1135 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1136 SetPagePrivate(page);
1137 h->resv_huge_pages--;
1140 mpol_cond_put(mpol);
1148 * common helper functions for hstate_next_node_to_{alloc|free}.
1149 * We may have allocated or freed a huge page based on a different
1150 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1151 * be outside of *nodes_allowed. Ensure that we use an allowed
1152 * node for alloc or free.
1154 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1156 nid = next_node_in(nid, *nodes_allowed);
1157 VM_BUG_ON(nid >= MAX_NUMNODES);
1162 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1164 if (!node_isset(nid, *nodes_allowed))
1165 nid = next_node_allowed(nid, nodes_allowed);
1170 * returns the previously saved node ["this node"] from which to
1171 * allocate a persistent huge page for the pool and advance the
1172 * next node from which to allocate, handling wrap at end of node
1175 static int hstate_next_node_to_alloc(struct hstate *h,
1176 nodemask_t *nodes_allowed)
1180 VM_BUG_ON(!nodes_allowed);
1182 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1183 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1189 * helper for free_pool_huge_page() - return the previously saved
1190 * node ["this node"] from which to free a huge page. Advance the
1191 * next node id whether or not we find a free huge page to free so
1192 * that the next attempt to free addresses the next node.
1194 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1198 VM_BUG_ON(!nodes_allowed);
1200 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1201 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1206 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1207 for (nr_nodes = nodes_weight(*mask); \
1209 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1212 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1213 for (nr_nodes = nodes_weight(*mask); \
1215 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1218 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1219 static void destroy_compound_gigantic_page(struct page *page,
1223 int nr_pages = 1 << order;
1224 struct page *p = page + 1;
1226 atomic_set(compound_mapcount_ptr(page), 0);
1227 if (hpage_pincount_available(page))
1228 atomic_set(compound_pincount_ptr(page), 0);
1230 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1231 clear_compound_head(p);
1232 set_page_refcounted(p);
1235 set_compound_order(page, 0);
1236 page[1].compound_nr = 0;
1237 __ClearPageHead(page);
1240 static void free_gigantic_page(struct page *page, unsigned int order)
1243 * If the page isn't allocated using the cma allocator,
1244 * cma_release() returns false.
1247 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1251 free_contig_range(page_to_pfn(page), 1 << order);
1254 #ifdef CONFIG_CONTIG_ALLOC
1255 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1256 int nid, nodemask_t *nodemask)
1258 unsigned long nr_pages = 1UL << huge_page_order(h);
1259 if (nid == NUMA_NO_NODE)
1260 nid = numa_mem_id();
1267 if (hugetlb_cma[nid]) {
1268 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1269 huge_page_order(h), true);
1274 if (!(gfp_mask & __GFP_THISNODE)) {
1275 for_each_node_mask(node, *nodemask) {
1276 if (node == nid || !hugetlb_cma[node])
1279 page = cma_alloc(hugetlb_cma[node], nr_pages,
1280 huge_page_order(h), true);
1288 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1291 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1292 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1293 #else /* !CONFIG_CONTIG_ALLOC */
1294 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1295 int nid, nodemask_t *nodemask)
1299 #endif /* CONFIG_CONTIG_ALLOC */
1301 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1302 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1303 int nid, nodemask_t *nodemask)
1307 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1308 static inline void destroy_compound_gigantic_page(struct page *page,
1309 unsigned int order) { }
1312 static void update_and_free_page(struct hstate *h, struct page *page)
1315 struct page *subpage = page;
1317 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1321 h->nr_huge_pages_node[page_to_nid(page)]--;
1322 for (i = 0; i < pages_per_huge_page(h);
1323 i++, subpage = mem_map_next(subpage, page, i)) {
1324 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1325 1 << PG_referenced | 1 << PG_dirty |
1326 1 << PG_active | 1 << PG_private |
1329 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1330 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1331 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1332 set_page_refcounted(page);
1333 if (hstate_is_gigantic(h)) {
1335 * Temporarily drop the hugetlb_lock, because
1336 * we might block in free_gigantic_page().
1338 spin_unlock(&hugetlb_lock);
1339 destroy_compound_gigantic_page(page, huge_page_order(h));
1340 free_gigantic_page(page, huge_page_order(h));
1341 spin_lock(&hugetlb_lock);
1343 __free_pages(page, huge_page_order(h));
1347 struct hstate *size_to_hstate(unsigned long size)
1351 for_each_hstate(h) {
1352 if (huge_page_size(h) == size)
1359 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1360 * to hstate->hugepage_activelist.)
1362 * This function can be called for tail pages, but never returns true for them.
1364 bool page_huge_active(struct page *page)
1366 return PageHeadHuge(page) && PagePrivate(&page[1]);
1369 /* never called for tail page */
1370 void set_page_huge_active(struct page *page)
1372 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1373 SetPagePrivate(&page[1]);
1376 static void clear_page_huge_active(struct page *page)
1378 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1379 ClearPagePrivate(&page[1]);
1383 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1386 static inline bool PageHugeTemporary(struct page *page)
1388 if (!PageHuge(page))
1391 return (unsigned long)page[2].mapping == -1U;
1394 static inline void SetPageHugeTemporary(struct page *page)
1396 page[2].mapping = (void *)-1U;
1399 static inline void ClearPageHugeTemporary(struct page *page)
1401 page[2].mapping = NULL;
1404 static void __free_huge_page(struct page *page)
1407 * Can't pass hstate in here because it is called from the
1408 * compound page destructor.
1410 struct hstate *h = page_hstate(page);
1411 int nid = page_to_nid(page);
1412 struct hugepage_subpool *spool =
1413 (struct hugepage_subpool *)page_private(page);
1414 bool restore_reserve;
1416 VM_BUG_ON_PAGE(page_count(page), page);
1417 VM_BUG_ON_PAGE(page_mapcount(page), page);
1419 set_page_private(page, 0);
1420 page->mapping = NULL;
1421 restore_reserve = PagePrivate(page);
1422 ClearPagePrivate(page);
1425 * If PagePrivate() was set on page, page allocation consumed a
1426 * reservation. If the page was associated with a subpool, there
1427 * would have been a page reserved in the subpool before allocation
1428 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1429 * reservtion, do not call hugepage_subpool_put_pages() as this will
1430 * remove the reserved page from the subpool.
1432 if (!restore_reserve) {
1434 * A return code of zero implies that the subpool will be
1435 * under its minimum size if the reservation is not restored
1436 * after page is free. Therefore, force restore_reserve
1439 if (hugepage_subpool_put_pages(spool, 1) == 0)
1440 restore_reserve = true;
1443 spin_lock(&hugetlb_lock);
1444 clear_page_huge_active(page);
1445 hugetlb_cgroup_uncharge_page(hstate_index(h),
1446 pages_per_huge_page(h), page);
1447 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1448 pages_per_huge_page(h), page);
1449 if (restore_reserve)
1450 h->resv_huge_pages++;
1452 if (PageHugeTemporary(page)) {
1453 list_del(&page->lru);
1454 ClearPageHugeTemporary(page);
1455 update_and_free_page(h, page);
1456 } else if (h->surplus_huge_pages_node[nid]) {
1457 /* remove the page from active list */
1458 list_del(&page->lru);
1459 update_and_free_page(h, page);
1460 h->surplus_huge_pages--;
1461 h->surplus_huge_pages_node[nid]--;
1463 arch_clear_hugepage_flags(page);
1464 enqueue_huge_page(h, page);
1466 spin_unlock(&hugetlb_lock);
1470 * As free_huge_page() can be called from a non-task context, we have
1471 * to defer the actual freeing in a workqueue to prevent potential
1472 * hugetlb_lock deadlock.
1474 * free_hpage_workfn() locklessly retrieves the linked list of pages to
1475 * be freed and frees them one-by-one. As the page->mapping pointer is
1476 * going to be cleared in __free_huge_page() anyway, it is reused as the
1477 * llist_node structure of a lockless linked list of huge pages to be freed.
1479 static LLIST_HEAD(hpage_freelist);
1481 static void free_hpage_workfn(struct work_struct *work)
1483 struct llist_node *node;
1486 node = llist_del_all(&hpage_freelist);
1489 page = container_of((struct address_space **)node,
1490 struct page, mapping);
1492 __free_huge_page(page);
1495 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1497 void free_huge_page(struct page *page)
1500 * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1504 * Only call schedule_work() if hpage_freelist is previously
1505 * empty. Otherwise, schedule_work() had been called but the
1506 * workfn hasn't retrieved the list yet.
1508 if (llist_add((struct llist_node *)&page->mapping,
1510 schedule_work(&free_hpage_work);
1514 __free_huge_page(page);
1517 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1519 INIT_LIST_HEAD(&page->lru);
1520 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1521 set_hugetlb_cgroup(page, NULL);
1522 set_hugetlb_cgroup_rsvd(page, NULL);
1523 spin_lock(&hugetlb_lock);
1525 h->nr_huge_pages_node[nid]++;
1526 ClearPageHugeFreed(page);
1527 spin_unlock(&hugetlb_lock);
1530 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1533 int nr_pages = 1 << order;
1534 struct page *p = page + 1;
1536 /* we rely on prep_new_huge_page to set the destructor */
1537 set_compound_order(page, order);
1538 __ClearPageReserved(page);
1539 __SetPageHead(page);
1540 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1542 * For gigantic hugepages allocated through bootmem at
1543 * boot, it's safer to be consistent with the not-gigantic
1544 * hugepages and clear the PG_reserved bit from all tail pages
1545 * too. Otherwise drivers using get_user_pages() to access tail
1546 * pages may get the reference counting wrong if they see
1547 * PG_reserved set on a tail page (despite the head page not
1548 * having PG_reserved set). Enforcing this consistency between
1549 * head and tail pages allows drivers to optimize away a check
1550 * on the head page when they need know if put_page() is needed
1551 * after get_user_pages().
1553 __ClearPageReserved(p);
1554 set_page_count(p, 0);
1555 set_compound_head(p, page);
1557 atomic_set(compound_mapcount_ptr(page), -1);
1559 if (hpage_pincount_available(page))
1560 atomic_set(compound_pincount_ptr(page), 0);
1564 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1565 * transparent huge pages. See the PageTransHuge() documentation for more
1568 int PageHuge(struct page *page)
1570 if (!PageCompound(page))
1573 page = compound_head(page);
1574 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1576 EXPORT_SYMBOL_GPL(PageHuge);
1579 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1580 * normal or transparent huge pages.
1582 int PageHeadHuge(struct page *page_head)
1584 if (!PageHead(page_head))
1587 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1591 * Find and lock address space (mapping) in write mode.
1593 * Upon entry, the page is locked which means that page_mapping() is
1594 * stable. Due to locking order, we can only trylock_write. If we can
1595 * not get the lock, simply return NULL to caller.
1597 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1599 struct address_space *mapping = page_mapping(hpage);
1604 if (i_mmap_trylock_write(mapping))
1610 pgoff_t __basepage_index(struct page *page)
1612 struct page *page_head = compound_head(page);
1613 pgoff_t index = page_index(page_head);
1614 unsigned long compound_idx;
1616 if (!PageHuge(page_head))
1617 return page_index(page);
1619 if (compound_order(page_head) >= MAX_ORDER)
1620 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1622 compound_idx = page - page_head;
1624 return (index << compound_order(page_head)) + compound_idx;
1627 static struct page *alloc_buddy_huge_page(struct hstate *h,
1628 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1629 nodemask_t *node_alloc_noretry)
1631 int order = huge_page_order(h);
1633 bool alloc_try_hard = true;
1636 * By default we always try hard to allocate the page with
1637 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1638 * a loop (to adjust global huge page counts) and previous allocation
1639 * failed, do not continue to try hard on the same node. Use the
1640 * node_alloc_noretry bitmap to manage this state information.
1642 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1643 alloc_try_hard = false;
1644 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1646 gfp_mask |= __GFP_RETRY_MAYFAIL;
1647 if (nid == NUMA_NO_NODE)
1648 nid = numa_mem_id();
1649 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1651 __count_vm_event(HTLB_BUDDY_PGALLOC);
1653 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1656 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1657 * indicates an overall state change. Clear bit so that we resume
1658 * normal 'try hard' allocations.
1660 if (node_alloc_noretry && page && !alloc_try_hard)
1661 node_clear(nid, *node_alloc_noretry);
1664 * If we tried hard to get a page but failed, set bit so that
1665 * subsequent attempts will not try as hard until there is an
1666 * overall state change.
1668 if (node_alloc_noretry && !page && alloc_try_hard)
1669 node_set(nid, *node_alloc_noretry);
1675 * Common helper to allocate a fresh hugetlb page. All specific allocators
1676 * should use this function to get new hugetlb pages
1678 static struct page *alloc_fresh_huge_page(struct hstate *h,
1679 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1680 nodemask_t *node_alloc_noretry)
1684 if (hstate_is_gigantic(h))
1685 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1687 page = alloc_buddy_huge_page(h, gfp_mask,
1688 nid, nmask, node_alloc_noretry);
1692 if (hstate_is_gigantic(h))
1693 prep_compound_gigantic_page(page, huge_page_order(h));
1694 prep_new_huge_page(h, page, page_to_nid(page));
1700 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1703 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1704 nodemask_t *node_alloc_noretry)
1708 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1710 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1711 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1712 node_alloc_noretry);
1720 put_page(page); /* free it into the hugepage allocator */
1726 * Free huge page from pool from next node to free.
1727 * Attempt to keep persistent huge pages more or less
1728 * balanced over allowed nodes.
1729 * Called with hugetlb_lock locked.
1731 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1737 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1739 * If we're returning unused surplus pages, only examine
1740 * nodes with surplus pages.
1742 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1743 !list_empty(&h->hugepage_freelists[node])) {
1745 list_entry(h->hugepage_freelists[node].next,
1747 list_del(&page->lru);
1748 h->free_huge_pages--;
1749 h->free_huge_pages_node[node]--;
1751 h->surplus_huge_pages--;
1752 h->surplus_huge_pages_node[node]--;
1754 update_and_free_page(h, page);
1764 * Dissolve a given free hugepage into free buddy pages. This function does
1765 * nothing for in-use hugepages and non-hugepages.
1766 * This function returns values like below:
1768 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1769 * (allocated or reserved.)
1770 * 0: successfully dissolved free hugepages or the page is not a
1771 * hugepage (considered as already dissolved)
1773 int dissolve_free_huge_page(struct page *page)
1778 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1779 if (!PageHuge(page))
1782 spin_lock(&hugetlb_lock);
1783 if (!PageHuge(page)) {
1788 if (!page_count(page)) {
1789 struct page *head = compound_head(page);
1790 struct hstate *h = page_hstate(head);
1791 int nid = page_to_nid(head);
1792 if (h->free_huge_pages - h->resv_huge_pages == 0)
1796 * We should make sure that the page is already on the free list
1797 * when it is dissolved.
1799 if (unlikely(!PageHugeFreed(head))) {
1800 spin_unlock(&hugetlb_lock);
1804 * Theoretically, we should return -EBUSY when we
1805 * encounter this race. In fact, we have a chance
1806 * to successfully dissolve the page if we do a
1807 * retry. Because the race window is quite small.
1808 * If we seize this opportunity, it is an optimization
1809 * for increasing the success rate of dissolving page.
1815 * Move PageHWPoison flag from head page to the raw error page,
1816 * which makes any subpages rather than the error page reusable.
1818 if (PageHWPoison(head) && page != head) {
1819 SetPageHWPoison(page);
1820 ClearPageHWPoison(head);
1822 list_del(&head->lru);
1823 h->free_huge_pages--;
1824 h->free_huge_pages_node[nid]--;
1825 h->max_huge_pages--;
1826 update_and_free_page(h, head);
1830 spin_unlock(&hugetlb_lock);
1835 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1836 * make specified memory blocks removable from the system.
1837 * Note that this will dissolve a free gigantic hugepage completely, if any
1838 * part of it lies within the given range.
1839 * Also note that if dissolve_free_huge_page() returns with an error, all
1840 * free hugepages that were dissolved before that error are lost.
1842 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1848 if (!hugepages_supported())
1851 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1852 page = pfn_to_page(pfn);
1853 rc = dissolve_free_huge_page(page);
1862 * Allocates a fresh surplus page from the page allocator.
1864 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1865 int nid, nodemask_t *nmask)
1867 struct page *page = NULL;
1869 if (hstate_is_gigantic(h))
1872 spin_lock(&hugetlb_lock);
1873 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1875 spin_unlock(&hugetlb_lock);
1877 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1881 spin_lock(&hugetlb_lock);
1883 * We could have raced with the pool size change.
1884 * Double check that and simply deallocate the new page
1885 * if we would end up overcommiting the surpluses. Abuse
1886 * temporary page to workaround the nasty free_huge_page
1889 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1890 SetPageHugeTemporary(page);
1891 spin_unlock(&hugetlb_lock);
1895 h->surplus_huge_pages++;
1896 h->surplus_huge_pages_node[page_to_nid(page)]++;
1900 spin_unlock(&hugetlb_lock);
1905 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1906 int nid, nodemask_t *nmask)
1910 if (hstate_is_gigantic(h))
1913 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1918 * We do not account these pages as surplus because they are only
1919 * temporary and will be released properly on the last reference
1921 SetPageHugeTemporary(page);
1927 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1930 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1931 struct vm_area_struct *vma, unsigned long addr)
1934 struct mempolicy *mpol;
1935 gfp_t gfp_mask = htlb_alloc_mask(h);
1937 nodemask_t *nodemask;
1939 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1940 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1941 mpol_cond_put(mpol);
1946 /* page migration callback function */
1947 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1948 nodemask_t *nmask, gfp_t gfp_mask)
1950 spin_lock(&hugetlb_lock);
1951 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1954 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1956 spin_unlock(&hugetlb_lock);
1960 spin_unlock(&hugetlb_lock);
1962 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1965 /* mempolicy aware migration callback */
1966 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1967 unsigned long address)
1969 struct mempolicy *mpol;
1970 nodemask_t *nodemask;
1975 gfp_mask = htlb_alloc_mask(h);
1976 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1977 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
1978 mpol_cond_put(mpol);
1984 * Increase the hugetlb pool such that it can accommodate a reservation
1987 static int gather_surplus_pages(struct hstate *h, int delta)
1988 __must_hold(&hugetlb_lock)
1990 struct list_head surplus_list;
1991 struct page *page, *tmp;
1993 int needed, allocated;
1994 bool alloc_ok = true;
1996 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1998 h->resv_huge_pages += delta;
2003 INIT_LIST_HEAD(&surplus_list);
2007 spin_unlock(&hugetlb_lock);
2008 for (i = 0; i < needed; i++) {
2009 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2010 NUMA_NO_NODE, NULL);
2015 list_add(&page->lru, &surplus_list);
2021 * After retaking hugetlb_lock, we need to recalculate 'needed'
2022 * because either resv_huge_pages or free_huge_pages may have changed.
2024 spin_lock(&hugetlb_lock);
2025 needed = (h->resv_huge_pages + delta) -
2026 (h->free_huge_pages + allocated);
2031 * We were not able to allocate enough pages to
2032 * satisfy the entire reservation so we free what
2033 * we've allocated so far.
2038 * The surplus_list now contains _at_least_ the number of extra pages
2039 * needed to accommodate the reservation. Add the appropriate number
2040 * of pages to the hugetlb pool and free the extras back to the buddy
2041 * allocator. Commit the entire reservation here to prevent another
2042 * process from stealing the pages as they are added to the pool but
2043 * before they are reserved.
2045 needed += allocated;
2046 h->resv_huge_pages += delta;
2049 /* Free the needed pages to the hugetlb pool */
2050 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2054 * This page is now managed by the hugetlb allocator and has
2055 * no users -- drop the buddy allocator's reference.
2057 put_page_testzero(page);
2058 VM_BUG_ON_PAGE(page_count(page), page);
2059 enqueue_huge_page(h, page);
2062 spin_unlock(&hugetlb_lock);
2064 /* Free unnecessary surplus pages to the buddy allocator */
2065 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2067 spin_lock(&hugetlb_lock);
2073 * This routine has two main purposes:
2074 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2075 * in unused_resv_pages. This corresponds to the prior adjustments made
2076 * to the associated reservation map.
2077 * 2) Free any unused surplus pages that may have been allocated to satisfy
2078 * the reservation. As many as unused_resv_pages may be freed.
2080 * Called with hugetlb_lock held. However, the lock could be dropped (and
2081 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
2082 * we must make sure nobody else can claim pages we are in the process of
2083 * freeing. Do this by ensuring resv_huge_page always is greater than the
2084 * number of huge pages we plan to free when dropping the lock.
2086 static void return_unused_surplus_pages(struct hstate *h,
2087 unsigned long unused_resv_pages)
2089 unsigned long nr_pages;
2091 /* Cannot return gigantic pages currently */
2092 if (hstate_is_gigantic(h))
2096 * Part (or even all) of the reservation could have been backed
2097 * by pre-allocated pages. Only free surplus pages.
2099 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2102 * We want to release as many surplus pages as possible, spread
2103 * evenly across all nodes with memory. Iterate across these nodes
2104 * until we can no longer free unreserved surplus pages. This occurs
2105 * when the nodes with surplus pages have no free pages.
2106 * free_pool_huge_page() will balance the freed pages across the
2107 * on-line nodes with memory and will handle the hstate accounting.
2109 * Note that we decrement resv_huge_pages as we free the pages. If
2110 * we drop the lock, resv_huge_pages will still be sufficiently large
2111 * to cover subsequent pages we may free.
2113 while (nr_pages--) {
2114 h->resv_huge_pages--;
2115 unused_resv_pages--;
2116 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
2118 cond_resched_lock(&hugetlb_lock);
2122 /* Fully uncommit the reservation */
2123 h->resv_huge_pages -= unused_resv_pages;
2128 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2129 * are used by the huge page allocation routines to manage reservations.
2131 * vma_needs_reservation is called to determine if the huge page at addr
2132 * within the vma has an associated reservation. If a reservation is
2133 * needed, the value 1 is returned. The caller is then responsible for
2134 * managing the global reservation and subpool usage counts. After
2135 * the huge page has been allocated, vma_commit_reservation is called
2136 * to add the page to the reservation map. If the page allocation fails,
2137 * the reservation must be ended instead of committed. vma_end_reservation
2138 * is called in such cases.
2140 * In the normal case, vma_commit_reservation returns the same value
2141 * as the preceding vma_needs_reservation call. The only time this
2142 * is not the case is if a reserve map was changed between calls. It
2143 * is the responsibility of the caller to notice the difference and
2144 * take appropriate action.
2146 * vma_add_reservation is used in error paths where a reservation must
2147 * be restored when a newly allocated huge page must be freed. It is
2148 * to be called after calling vma_needs_reservation to determine if a
2149 * reservation exists.
2151 enum vma_resv_mode {
2157 static long __vma_reservation_common(struct hstate *h,
2158 struct vm_area_struct *vma, unsigned long addr,
2159 enum vma_resv_mode mode)
2161 struct resv_map *resv;
2164 long dummy_out_regions_needed;
2166 resv = vma_resv_map(vma);
2170 idx = vma_hugecache_offset(h, vma, addr);
2172 case VMA_NEEDS_RESV:
2173 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2174 /* We assume that vma_reservation_* routines always operate on
2175 * 1 page, and that adding to resv map a 1 page entry can only
2176 * ever require 1 region.
2178 VM_BUG_ON(dummy_out_regions_needed != 1);
2180 case VMA_COMMIT_RESV:
2181 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2182 /* region_add calls of range 1 should never fail. */
2186 region_abort(resv, idx, idx + 1, 1);
2190 if (vma->vm_flags & VM_MAYSHARE) {
2191 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2192 /* region_add calls of range 1 should never fail. */
2195 region_abort(resv, idx, idx + 1, 1);
2196 ret = region_del(resv, idx, idx + 1);
2203 if (vma->vm_flags & VM_MAYSHARE)
2205 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
2207 * In most cases, reserves always exist for private mappings.
2208 * However, a file associated with mapping could have been
2209 * hole punched or truncated after reserves were consumed.
2210 * As subsequent fault on such a range will not use reserves.
2211 * Subtle - The reserve map for private mappings has the
2212 * opposite meaning than that of shared mappings. If NO
2213 * entry is in the reserve map, it means a reservation exists.
2214 * If an entry exists in the reserve map, it means the
2215 * reservation has already been consumed. As a result, the
2216 * return value of this routine is the opposite of the
2217 * value returned from reserve map manipulation routines above.
2225 return ret < 0 ? ret : 0;
2228 static long vma_needs_reservation(struct hstate *h,
2229 struct vm_area_struct *vma, unsigned long addr)
2231 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2234 static long vma_commit_reservation(struct hstate *h,
2235 struct vm_area_struct *vma, unsigned long addr)
2237 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2240 static void vma_end_reservation(struct hstate *h,
2241 struct vm_area_struct *vma, unsigned long addr)
2243 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2246 static long vma_add_reservation(struct hstate *h,
2247 struct vm_area_struct *vma, unsigned long addr)
2249 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2253 * This routine is called to restore a reservation on error paths. In the
2254 * specific error paths, a huge page was allocated (via alloc_huge_page)
2255 * and is about to be freed. If a reservation for the page existed,
2256 * alloc_huge_page would have consumed the reservation and set PagePrivate
2257 * in the newly allocated page. When the page is freed via free_huge_page,
2258 * the global reservation count will be incremented if PagePrivate is set.
2259 * However, free_huge_page can not adjust the reserve map. Adjust the
2260 * reserve map here to be consistent with global reserve count adjustments
2261 * to be made by free_huge_page.
2263 static void restore_reserve_on_error(struct hstate *h,
2264 struct vm_area_struct *vma, unsigned long address,
2267 if (unlikely(PagePrivate(page))) {
2268 long rc = vma_needs_reservation(h, vma, address);
2270 if (unlikely(rc < 0)) {
2272 * Rare out of memory condition in reserve map
2273 * manipulation. Clear PagePrivate so that
2274 * global reserve count will not be incremented
2275 * by free_huge_page. This will make it appear
2276 * as though the reservation for this page was
2277 * consumed. This may prevent the task from
2278 * faulting in the page at a later time. This
2279 * is better than inconsistent global huge page
2280 * accounting of reserve counts.
2282 ClearPagePrivate(page);
2284 rc = vma_add_reservation(h, vma, address);
2285 if (unlikely(rc < 0))
2287 * See above comment about rare out of
2290 ClearPagePrivate(page);
2292 vma_end_reservation(h, vma, address);
2296 struct page *alloc_huge_page(struct vm_area_struct *vma,
2297 unsigned long addr, int avoid_reserve)
2299 struct hugepage_subpool *spool = subpool_vma(vma);
2300 struct hstate *h = hstate_vma(vma);
2302 long map_chg, map_commit;
2305 struct hugetlb_cgroup *h_cg;
2306 bool deferred_reserve;
2308 idx = hstate_index(h);
2310 * Examine the region/reserve map to determine if the process
2311 * has a reservation for the page to be allocated. A return
2312 * code of zero indicates a reservation exists (no change).
2314 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2316 return ERR_PTR(-ENOMEM);
2319 * Processes that did not create the mapping will have no
2320 * reserves as indicated by the region/reserve map. Check
2321 * that the allocation will not exceed the subpool limit.
2322 * Allocations for MAP_NORESERVE mappings also need to be
2323 * checked against any subpool limit.
2325 if (map_chg || avoid_reserve) {
2326 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2328 vma_end_reservation(h, vma, addr);
2329 return ERR_PTR(-ENOSPC);
2333 * Even though there was no reservation in the region/reserve
2334 * map, there could be reservations associated with the
2335 * subpool that can be used. This would be indicated if the
2336 * return value of hugepage_subpool_get_pages() is zero.
2337 * However, if avoid_reserve is specified we still avoid even
2338 * the subpool reservations.
2344 /* If this allocation is not consuming a reservation, charge it now.
2346 deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma);
2347 if (deferred_reserve) {
2348 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2349 idx, pages_per_huge_page(h), &h_cg);
2351 goto out_subpool_put;
2354 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2356 goto out_uncharge_cgroup_reservation;
2358 spin_lock(&hugetlb_lock);
2360 * glb_chg is passed to indicate whether or not a page must be taken
2361 * from the global free pool (global change). gbl_chg == 0 indicates
2362 * a reservation exists for the allocation.
2364 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2366 spin_unlock(&hugetlb_lock);
2367 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2369 goto out_uncharge_cgroup;
2370 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2371 SetPagePrivate(page);
2372 h->resv_huge_pages--;
2374 spin_lock(&hugetlb_lock);
2375 list_add(&page->lru, &h->hugepage_activelist);
2378 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2379 /* If allocation is not consuming a reservation, also store the
2380 * hugetlb_cgroup pointer on the page.
2382 if (deferred_reserve) {
2383 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2387 spin_unlock(&hugetlb_lock);
2389 set_page_private(page, (unsigned long)spool);
2391 map_commit = vma_commit_reservation(h, vma, addr);
2392 if (unlikely(map_chg > map_commit)) {
2394 * The page was added to the reservation map between
2395 * vma_needs_reservation and vma_commit_reservation.
2396 * This indicates a race with hugetlb_reserve_pages.
2397 * Adjust for the subpool count incremented above AND
2398 * in hugetlb_reserve_pages for the same page. Also,
2399 * the reservation count added in hugetlb_reserve_pages
2400 * no longer applies.
2404 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2405 hugetlb_acct_memory(h, -rsv_adjust);
2406 if (deferred_reserve)
2407 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2408 pages_per_huge_page(h), page);
2412 out_uncharge_cgroup:
2413 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2414 out_uncharge_cgroup_reservation:
2415 if (deferred_reserve)
2416 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2419 if (map_chg || avoid_reserve)
2420 hugepage_subpool_put_pages(spool, 1);
2421 vma_end_reservation(h, vma, addr);
2422 return ERR_PTR(-ENOSPC);
2425 int alloc_bootmem_huge_page(struct hstate *h)
2426 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2427 int __alloc_bootmem_huge_page(struct hstate *h)
2429 struct huge_bootmem_page *m;
2432 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2435 addr = memblock_alloc_try_nid_raw(
2436 huge_page_size(h), huge_page_size(h),
2437 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2440 * Use the beginning of the huge page to store the
2441 * huge_bootmem_page struct (until gather_bootmem
2442 * puts them into the mem_map).
2451 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2452 /* Put them into a private list first because mem_map is not up yet */
2453 INIT_LIST_HEAD(&m->list);
2454 list_add(&m->list, &huge_boot_pages);
2459 static void __init prep_compound_huge_page(struct page *page,
2462 if (unlikely(order > (MAX_ORDER - 1)))
2463 prep_compound_gigantic_page(page, order);
2465 prep_compound_page(page, order);
2468 /* Put bootmem huge pages into the standard lists after mem_map is up */
2469 static void __init gather_bootmem_prealloc(void)
2471 struct huge_bootmem_page *m;
2473 list_for_each_entry(m, &huge_boot_pages, list) {
2474 struct page *page = virt_to_page(m);
2475 struct hstate *h = m->hstate;
2477 WARN_ON(page_count(page) != 1);
2478 prep_compound_huge_page(page, h->order);
2479 WARN_ON(PageReserved(page));
2480 prep_new_huge_page(h, page, page_to_nid(page));
2481 put_page(page); /* free it into the hugepage allocator */
2484 * If we had gigantic hugepages allocated at boot time, we need
2485 * to restore the 'stolen' pages to totalram_pages in order to
2486 * fix confusing memory reports from free(1) and another
2487 * side-effects, like CommitLimit going negative.
2489 if (hstate_is_gigantic(h))
2490 adjust_managed_page_count(page, 1 << h->order);
2495 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2498 nodemask_t *node_alloc_noretry;
2500 if (!hstate_is_gigantic(h)) {
2502 * Bit mask controlling how hard we retry per-node allocations.
2503 * Ignore errors as lower level routines can deal with
2504 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2505 * time, we are likely in bigger trouble.
2507 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2510 /* allocations done at boot time */
2511 node_alloc_noretry = NULL;
2514 /* bit mask controlling how hard we retry per-node allocations */
2515 if (node_alloc_noretry)
2516 nodes_clear(*node_alloc_noretry);
2518 for (i = 0; i < h->max_huge_pages; ++i) {
2519 if (hstate_is_gigantic(h)) {
2520 if (hugetlb_cma_size) {
2521 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2524 if (!alloc_bootmem_huge_page(h))
2526 } else if (!alloc_pool_huge_page(h,
2527 &node_states[N_MEMORY],
2528 node_alloc_noretry))
2532 if (i < h->max_huge_pages) {
2535 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2536 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2537 h->max_huge_pages, buf, i);
2538 h->max_huge_pages = i;
2541 kfree(node_alloc_noretry);
2544 static void __init hugetlb_init_hstates(void)
2548 for_each_hstate(h) {
2549 if (minimum_order > huge_page_order(h))
2550 minimum_order = huge_page_order(h);
2552 /* oversize hugepages were init'ed in early boot */
2553 if (!hstate_is_gigantic(h))
2554 hugetlb_hstate_alloc_pages(h);
2556 VM_BUG_ON(minimum_order == UINT_MAX);
2559 static void __init report_hugepages(void)
2563 for_each_hstate(h) {
2566 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2567 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2568 buf, h->free_huge_pages);
2572 #ifdef CONFIG_HIGHMEM
2573 static void try_to_free_low(struct hstate *h, unsigned long count,
2574 nodemask_t *nodes_allowed)
2578 if (hstate_is_gigantic(h))
2581 for_each_node_mask(i, *nodes_allowed) {
2582 struct page *page, *next;
2583 struct list_head *freel = &h->hugepage_freelists[i];
2584 list_for_each_entry_safe(page, next, freel, lru) {
2585 if (count >= h->nr_huge_pages)
2587 if (PageHighMem(page))
2589 list_del(&page->lru);
2590 update_and_free_page(h, page);
2591 h->free_huge_pages--;
2592 h->free_huge_pages_node[page_to_nid(page)]--;
2597 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2598 nodemask_t *nodes_allowed)
2604 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2605 * balanced by operating on them in a round-robin fashion.
2606 * Returns 1 if an adjustment was made.
2608 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2613 VM_BUG_ON(delta != -1 && delta != 1);
2616 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2617 if (h->surplus_huge_pages_node[node])
2621 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2622 if (h->surplus_huge_pages_node[node] <
2623 h->nr_huge_pages_node[node])
2630 h->surplus_huge_pages += delta;
2631 h->surplus_huge_pages_node[node] += delta;
2635 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2636 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2637 nodemask_t *nodes_allowed)
2639 unsigned long min_count, ret;
2640 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2643 * Bit mask controlling how hard we retry per-node allocations.
2644 * If we can not allocate the bit mask, do not attempt to allocate
2645 * the requested huge pages.
2647 if (node_alloc_noretry)
2648 nodes_clear(*node_alloc_noretry);
2652 spin_lock(&hugetlb_lock);
2655 * Check for a node specific request.
2656 * Changing node specific huge page count may require a corresponding
2657 * change to the global count. In any case, the passed node mask
2658 * (nodes_allowed) will restrict alloc/free to the specified node.
2660 if (nid != NUMA_NO_NODE) {
2661 unsigned long old_count = count;
2663 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2665 * User may have specified a large count value which caused the
2666 * above calculation to overflow. In this case, they wanted
2667 * to allocate as many huge pages as possible. Set count to
2668 * largest possible value to align with their intention.
2670 if (count < old_count)
2675 * Gigantic pages runtime allocation depend on the capability for large
2676 * page range allocation.
2677 * If the system does not provide this feature, return an error when
2678 * the user tries to allocate gigantic pages but let the user free the
2679 * boottime allocated gigantic pages.
2681 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2682 if (count > persistent_huge_pages(h)) {
2683 spin_unlock(&hugetlb_lock);
2684 NODEMASK_FREE(node_alloc_noretry);
2687 /* Fall through to decrease pool */
2691 * Increase the pool size
2692 * First take pages out of surplus state. Then make up the
2693 * remaining difference by allocating fresh huge pages.
2695 * We might race with alloc_surplus_huge_page() here and be unable
2696 * to convert a surplus huge page to a normal huge page. That is
2697 * not critical, though, it just means the overall size of the
2698 * pool might be one hugepage larger than it needs to be, but
2699 * within all the constraints specified by the sysctls.
2701 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2702 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2706 while (count > persistent_huge_pages(h)) {
2708 * If this allocation races such that we no longer need the
2709 * page, free_huge_page will handle it by freeing the page
2710 * and reducing the surplus.
2712 spin_unlock(&hugetlb_lock);
2714 /* yield cpu to avoid soft lockup */
2717 ret = alloc_pool_huge_page(h, nodes_allowed,
2718 node_alloc_noretry);
2719 spin_lock(&hugetlb_lock);
2723 /* Bail for signals. Probably ctrl-c from user */
2724 if (signal_pending(current))
2729 * Decrease the pool size
2730 * First return free pages to the buddy allocator (being careful
2731 * to keep enough around to satisfy reservations). Then place
2732 * pages into surplus state as needed so the pool will shrink
2733 * to the desired size as pages become free.
2735 * By placing pages into the surplus state independent of the
2736 * overcommit value, we are allowing the surplus pool size to
2737 * exceed overcommit. There are few sane options here. Since
2738 * alloc_surplus_huge_page() is checking the global counter,
2739 * though, we'll note that we're not allowed to exceed surplus
2740 * and won't grow the pool anywhere else. Not until one of the
2741 * sysctls are changed, or the surplus pages go out of use.
2743 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2744 min_count = max(count, min_count);
2745 try_to_free_low(h, min_count, nodes_allowed);
2746 while (min_count < persistent_huge_pages(h)) {
2747 if (!free_pool_huge_page(h, nodes_allowed, 0))
2749 cond_resched_lock(&hugetlb_lock);
2751 while (count < persistent_huge_pages(h)) {
2752 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2756 h->max_huge_pages = persistent_huge_pages(h);
2757 spin_unlock(&hugetlb_lock);
2759 NODEMASK_FREE(node_alloc_noretry);
2764 #define HSTATE_ATTR_RO(_name) \
2765 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2767 #define HSTATE_ATTR(_name) \
2768 static struct kobj_attribute _name##_attr = \
2769 __ATTR(_name, 0644, _name##_show, _name##_store)
2771 static struct kobject *hugepages_kobj;
2772 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2774 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2776 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2780 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2781 if (hstate_kobjs[i] == kobj) {
2783 *nidp = NUMA_NO_NODE;
2787 return kobj_to_node_hstate(kobj, nidp);
2790 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2791 struct kobj_attribute *attr, char *buf)
2794 unsigned long nr_huge_pages;
2797 h = kobj_to_hstate(kobj, &nid);
2798 if (nid == NUMA_NO_NODE)
2799 nr_huge_pages = h->nr_huge_pages;
2801 nr_huge_pages = h->nr_huge_pages_node[nid];
2803 return sprintf(buf, "%lu\n", nr_huge_pages);
2806 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2807 struct hstate *h, int nid,
2808 unsigned long count, size_t len)
2811 nodemask_t nodes_allowed, *n_mask;
2813 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2816 if (nid == NUMA_NO_NODE) {
2818 * global hstate attribute
2820 if (!(obey_mempolicy &&
2821 init_nodemask_of_mempolicy(&nodes_allowed)))
2822 n_mask = &node_states[N_MEMORY];
2824 n_mask = &nodes_allowed;
2827 * Node specific request. count adjustment happens in
2828 * set_max_huge_pages() after acquiring hugetlb_lock.
2830 init_nodemask_of_node(&nodes_allowed, nid);
2831 n_mask = &nodes_allowed;
2834 err = set_max_huge_pages(h, count, nid, n_mask);
2836 return err ? err : len;
2839 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2840 struct kobject *kobj, const char *buf,
2844 unsigned long count;
2848 err = kstrtoul(buf, 10, &count);
2852 h = kobj_to_hstate(kobj, &nid);
2853 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2856 static ssize_t nr_hugepages_show(struct kobject *kobj,
2857 struct kobj_attribute *attr, char *buf)
2859 return nr_hugepages_show_common(kobj, attr, buf);
2862 static ssize_t nr_hugepages_store(struct kobject *kobj,
2863 struct kobj_attribute *attr, const char *buf, size_t len)
2865 return nr_hugepages_store_common(false, kobj, buf, len);
2867 HSTATE_ATTR(nr_hugepages);
2872 * hstate attribute for optionally mempolicy-based constraint on persistent
2873 * huge page alloc/free.
2875 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2876 struct kobj_attribute *attr, char *buf)
2878 return nr_hugepages_show_common(kobj, attr, buf);
2881 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2882 struct kobj_attribute *attr, const char *buf, size_t len)
2884 return nr_hugepages_store_common(true, kobj, buf, len);
2886 HSTATE_ATTR(nr_hugepages_mempolicy);
2890 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2891 struct kobj_attribute *attr, char *buf)
2893 struct hstate *h = kobj_to_hstate(kobj, NULL);
2894 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2897 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2898 struct kobj_attribute *attr, const char *buf, size_t count)
2901 unsigned long input;
2902 struct hstate *h = kobj_to_hstate(kobj, NULL);
2904 if (hstate_is_gigantic(h))
2907 err = kstrtoul(buf, 10, &input);
2911 spin_lock(&hugetlb_lock);
2912 h->nr_overcommit_huge_pages = input;
2913 spin_unlock(&hugetlb_lock);
2917 HSTATE_ATTR(nr_overcommit_hugepages);
2919 static ssize_t free_hugepages_show(struct kobject *kobj,
2920 struct kobj_attribute *attr, char *buf)
2923 unsigned long free_huge_pages;
2926 h = kobj_to_hstate(kobj, &nid);
2927 if (nid == NUMA_NO_NODE)
2928 free_huge_pages = h->free_huge_pages;
2930 free_huge_pages = h->free_huge_pages_node[nid];
2932 return sprintf(buf, "%lu\n", free_huge_pages);
2934 HSTATE_ATTR_RO(free_hugepages);
2936 static ssize_t resv_hugepages_show(struct kobject *kobj,
2937 struct kobj_attribute *attr, char *buf)
2939 struct hstate *h = kobj_to_hstate(kobj, NULL);
2940 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2942 HSTATE_ATTR_RO(resv_hugepages);
2944 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2945 struct kobj_attribute *attr, char *buf)
2948 unsigned long surplus_huge_pages;
2951 h = kobj_to_hstate(kobj, &nid);
2952 if (nid == NUMA_NO_NODE)
2953 surplus_huge_pages = h->surplus_huge_pages;
2955 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2957 return sprintf(buf, "%lu\n", surplus_huge_pages);
2959 HSTATE_ATTR_RO(surplus_hugepages);
2961 static struct attribute *hstate_attrs[] = {
2962 &nr_hugepages_attr.attr,
2963 &nr_overcommit_hugepages_attr.attr,
2964 &free_hugepages_attr.attr,
2965 &resv_hugepages_attr.attr,
2966 &surplus_hugepages_attr.attr,
2968 &nr_hugepages_mempolicy_attr.attr,
2973 static const struct attribute_group hstate_attr_group = {
2974 .attrs = hstate_attrs,
2977 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2978 struct kobject **hstate_kobjs,
2979 const struct attribute_group *hstate_attr_group)
2982 int hi = hstate_index(h);
2984 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2985 if (!hstate_kobjs[hi])
2988 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2990 kobject_put(hstate_kobjs[hi]);
2991 hstate_kobjs[hi] = NULL;
2997 static void __init hugetlb_sysfs_init(void)
3002 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3003 if (!hugepages_kobj)
3006 for_each_hstate(h) {
3007 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3008 hstate_kobjs, &hstate_attr_group);
3010 pr_err("HugeTLB: Unable to add hstate %s", h->name);
3017 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3018 * with node devices in node_devices[] using a parallel array. The array
3019 * index of a node device or _hstate == node id.
3020 * This is here to avoid any static dependency of the node device driver, in
3021 * the base kernel, on the hugetlb module.
3023 struct node_hstate {
3024 struct kobject *hugepages_kobj;
3025 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3027 static struct node_hstate node_hstates[MAX_NUMNODES];
3030 * A subset of global hstate attributes for node devices
3032 static struct attribute *per_node_hstate_attrs[] = {
3033 &nr_hugepages_attr.attr,
3034 &free_hugepages_attr.attr,
3035 &surplus_hugepages_attr.attr,
3039 static const struct attribute_group per_node_hstate_attr_group = {
3040 .attrs = per_node_hstate_attrs,
3044 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3045 * Returns node id via non-NULL nidp.
3047 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3051 for (nid = 0; nid < nr_node_ids; nid++) {
3052 struct node_hstate *nhs = &node_hstates[nid];
3054 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3055 if (nhs->hstate_kobjs[i] == kobj) {
3067 * Unregister hstate attributes from a single node device.
3068 * No-op if no hstate attributes attached.
3070 static void hugetlb_unregister_node(struct node *node)
3073 struct node_hstate *nhs = &node_hstates[node->dev.id];
3075 if (!nhs->hugepages_kobj)
3076 return; /* no hstate attributes */
3078 for_each_hstate(h) {
3079 int idx = hstate_index(h);
3080 if (nhs->hstate_kobjs[idx]) {
3081 kobject_put(nhs->hstate_kobjs[idx]);
3082 nhs->hstate_kobjs[idx] = NULL;
3086 kobject_put(nhs->hugepages_kobj);
3087 nhs->hugepages_kobj = NULL;
3092 * Register hstate attributes for a single node device.
3093 * No-op if attributes already registered.
3095 static void hugetlb_register_node(struct node *node)
3098 struct node_hstate *nhs = &node_hstates[node->dev.id];
3101 if (nhs->hugepages_kobj)
3102 return; /* already allocated */
3104 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3106 if (!nhs->hugepages_kobj)
3109 for_each_hstate(h) {
3110 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3112 &per_node_hstate_attr_group);
3114 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3115 h->name, node->dev.id);
3116 hugetlb_unregister_node(node);
3123 * hugetlb init time: register hstate attributes for all registered node
3124 * devices of nodes that have memory. All on-line nodes should have
3125 * registered their associated device by this time.
3127 static void __init hugetlb_register_all_nodes(void)
3131 for_each_node_state(nid, N_MEMORY) {
3132 struct node *node = node_devices[nid];
3133 if (node->dev.id == nid)
3134 hugetlb_register_node(node);
3138 * Let the node device driver know we're here so it can
3139 * [un]register hstate attributes on node hotplug.
3141 register_hugetlbfs_with_node(hugetlb_register_node,
3142 hugetlb_unregister_node);
3144 #else /* !CONFIG_NUMA */
3146 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3154 static void hugetlb_register_all_nodes(void) { }
3158 static int __init hugetlb_init(void)
3162 if (!hugepages_supported()) {
3163 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3164 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3169 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3170 * architectures depend on setup being done here.
3172 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3173 if (!parsed_default_hugepagesz) {
3175 * If we did not parse a default huge page size, set
3176 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3177 * number of huge pages for this default size was implicitly
3178 * specified, set that here as well.
3179 * Note that the implicit setting will overwrite an explicit
3180 * setting. A warning will be printed in this case.
3182 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3183 if (default_hstate_max_huge_pages) {
3184 if (default_hstate.max_huge_pages) {
3187 string_get_size(huge_page_size(&default_hstate),
3188 1, STRING_UNITS_2, buf, 32);
3189 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3190 default_hstate.max_huge_pages, buf);
3191 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3192 default_hstate_max_huge_pages);
3194 default_hstate.max_huge_pages =
3195 default_hstate_max_huge_pages;
3199 hugetlb_cma_check();
3200 hugetlb_init_hstates();
3201 gather_bootmem_prealloc();
3204 hugetlb_sysfs_init();
3205 hugetlb_register_all_nodes();
3206 hugetlb_cgroup_file_init();
3209 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3211 num_fault_mutexes = 1;
3213 hugetlb_fault_mutex_table =
3214 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3216 BUG_ON(!hugetlb_fault_mutex_table);
3218 for (i = 0; i < num_fault_mutexes; i++)
3219 mutex_init(&hugetlb_fault_mutex_table[i]);
3222 subsys_initcall(hugetlb_init);
3224 /* Overwritten by architectures with more huge page sizes */
3225 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3227 return size == HPAGE_SIZE;
3230 void __init hugetlb_add_hstate(unsigned int order)
3235 if (size_to_hstate(PAGE_SIZE << order)) {
3238 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3240 h = &hstates[hugetlb_max_hstate++];
3242 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
3243 h->nr_huge_pages = 0;
3244 h->free_huge_pages = 0;
3245 for (i = 0; i < MAX_NUMNODES; ++i)
3246 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3247 INIT_LIST_HEAD(&h->hugepage_activelist);
3248 h->next_nid_to_alloc = first_memory_node;
3249 h->next_nid_to_free = first_memory_node;
3250 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3251 huge_page_size(h)/1024);
3257 * hugepages command line processing
3258 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3259 * specification. If not, ignore the hugepages value. hugepages can also
3260 * be the first huge page command line option in which case it implicitly
3261 * specifies the number of huge pages for the default size.
3263 static int __init hugepages_setup(char *s)
3266 static unsigned long *last_mhp;
3268 if (!parsed_valid_hugepagesz) {
3269 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3270 parsed_valid_hugepagesz = true;
3275 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3276 * yet, so this hugepages= parameter goes to the "default hstate".
3277 * Otherwise, it goes with the previously parsed hugepagesz or
3278 * default_hugepagesz.
3280 else if (!hugetlb_max_hstate)
3281 mhp = &default_hstate_max_huge_pages;
3283 mhp = &parsed_hstate->max_huge_pages;
3285 if (mhp == last_mhp) {
3286 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3290 if (sscanf(s, "%lu", mhp) <= 0)
3294 * Global state is always initialized later in hugetlb_init.
3295 * But we need to allocate >= MAX_ORDER hstates here early to still
3296 * use the bootmem allocator.
3298 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
3299 hugetlb_hstate_alloc_pages(parsed_hstate);
3305 __setup("hugepages=", hugepages_setup);
3308 * hugepagesz command line processing
3309 * A specific huge page size can only be specified once with hugepagesz.
3310 * hugepagesz is followed by hugepages on the command line. The global
3311 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3312 * hugepagesz argument was valid.
3314 static int __init hugepagesz_setup(char *s)
3319 parsed_valid_hugepagesz = false;
3320 size = (unsigned long)memparse(s, NULL);
3322 if (!arch_hugetlb_valid_size(size)) {
3323 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3327 h = size_to_hstate(size);
3330 * hstate for this size already exists. This is normally
3331 * an error, but is allowed if the existing hstate is the
3332 * default hstate. More specifically, it is only allowed if
3333 * the number of huge pages for the default hstate was not
3334 * previously specified.
3336 if (!parsed_default_hugepagesz || h != &default_hstate ||
3337 default_hstate.max_huge_pages) {
3338 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3343 * No need to call hugetlb_add_hstate() as hstate already
3344 * exists. But, do set parsed_hstate so that a following
3345 * hugepages= parameter will be applied to this hstate.
3348 parsed_valid_hugepagesz = true;
3352 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3353 parsed_valid_hugepagesz = true;
3356 __setup("hugepagesz=", hugepagesz_setup);
3359 * default_hugepagesz command line input
3360 * Only one instance of default_hugepagesz allowed on command line.
3362 static int __init default_hugepagesz_setup(char *s)
3366 parsed_valid_hugepagesz = false;
3367 if (parsed_default_hugepagesz) {
3368 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3372 size = (unsigned long)memparse(s, NULL);
3374 if (!arch_hugetlb_valid_size(size)) {
3375 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3379 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3380 parsed_valid_hugepagesz = true;
3381 parsed_default_hugepagesz = true;
3382 default_hstate_idx = hstate_index(size_to_hstate(size));
3385 * The number of default huge pages (for this size) could have been
3386 * specified as the first hugetlb parameter: hugepages=X. If so,
3387 * then default_hstate_max_huge_pages is set. If the default huge
3388 * page size is gigantic (>= MAX_ORDER), then the pages must be
3389 * allocated here from bootmem allocator.
3391 if (default_hstate_max_huge_pages) {
3392 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3393 if (hstate_is_gigantic(&default_hstate))
3394 hugetlb_hstate_alloc_pages(&default_hstate);
3395 default_hstate_max_huge_pages = 0;
3400 __setup("default_hugepagesz=", default_hugepagesz_setup);
3402 static unsigned int allowed_mems_nr(struct hstate *h)
3405 unsigned int nr = 0;
3406 nodemask_t *mpol_allowed;
3407 unsigned int *array = h->free_huge_pages_node;
3408 gfp_t gfp_mask = htlb_alloc_mask(h);
3410 mpol_allowed = policy_nodemask_current(gfp_mask);
3412 for_each_node_mask(node, cpuset_current_mems_allowed) {
3413 if (!mpol_allowed ||
3414 (mpol_allowed && node_isset(node, *mpol_allowed)))
3421 #ifdef CONFIG_SYSCTL
3422 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3423 void *buffer, size_t *length,
3424 loff_t *ppos, unsigned long *out)
3426 struct ctl_table dup_table;
3429 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3430 * can duplicate the @table and alter the duplicate of it.
3433 dup_table.data = out;
3435 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3438 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3439 struct ctl_table *table, int write,
3440 void *buffer, size_t *length, loff_t *ppos)
3442 struct hstate *h = &default_hstate;
3443 unsigned long tmp = h->max_huge_pages;
3446 if (!hugepages_supported())
3449 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3455 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3456 NUMA_NO_NODE, tmp, *length);
3461 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3462 void *buffer, size_t *length, loff_t *ppos)
3465 return hugetlb_sysctl_handler_common(false, table, write,
3466 buffer, length, ppos);
3470 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3471 void *buffer, size_t *length, loff_t *ppos)
3473 return hugetlb_sysctl_handler_common(true, table, write,
3474 buffer, length, ppos);
3476 #endif /* CONFIG_NUMA */
3478 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3479 void *buffer, size_t *length, loff_t *ppos)
3481 struct hstate *h = &default_hstate;
3485 if (!hugepages_supported())
3488 tmp = h->nr_overcommit_huge_pages;
3490 if (write && hstate_is_gigantic(h))
3493 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3499 spin_lock(&hugetlb_lock);
3500 h->nr_overcommit_huge_pages = tmp;
3501 spin_unlock(&hugetlb_lock);
3507 #endif /* CONFIG_SYSCTL */
3509 void hugetlb_report_meminfo(struct seq_file *m)
3512 unsigned long total = 0;
3514 if (!hugepages_supported())
3517 for_each_hstate(h) {
3518 unsigned long count = h->nr_huge_pages;
3520 total += (PAGE_SIZE << huge_page_order(h)) * count;
3522 if (h == &default_hstate)
3524 "HugePages_Total: %5lu\n"
3525 "HugePages_Free: %5lu\n"
3526 "HugePages_Rsvd: %5lu\n"
3527 "HugePages_Surp: %5lu\n"
3528 "Hugepagesize: %8lu kB\n",
3532 h->surplus_huge_pages,
3533 (PAGE_SIZE << huge_page_order(h)) / 1024);
3536 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3539 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
3541 struct hstate *h = &default_hstate;
3543 if (!hugepages_supported())
3546 return sysfs_emit_at(buf, len,
3547 "Node %d HugePages_Total: %5u\n"
3548 "Node %d HugePages_Free: %5u\n"
3549 "Node %d HugePages_Surp: %5u\n",
3550 nid, h->nr_huge_pages_node[nid],
3551 nid, h->free_huge_pages_node[nid],
3552 nid, h->surplus_huge_pages_node[nid]);
3555 void hugetlb_show_meminfo(void)
3560 if (!hugepages_supported())
3563 for_each_node_state(nid, N_MEMORY)
3565 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3567 h->nr_huge_pages_node[nid],
3568 h->free_huge_pages_node[nid],
3569 h->surplus_huge_pages_node[nid],
3570 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3573 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3575 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3576 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3579 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3580 unsigned long hugetlb_total_pages(void)
3583 unsigned long nr_total_pages = 0;
3586 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3587 return nr_total_pages;
3590 static int hugetlb_acct_memory(struct hstate *h, long delta)
3594 spin_lock(&hugetlb_lock);
3596 * When cpuset is configured, it breaks the strict hugetlb page
3597 * reservation as the accounting is done on a global variable. Such
3598 * reservation is completely rubbish in the presence of cpuset because
3599 * the reservation is not checked against page availability for the
3600 * current cpuset. Application can still potentially OOM'ed by kernel
3601 * with lack of free htlb page in cpuset that the task is in.
3602 * Attempt to enforce strict accounting with cpuset is almost
3603 * impossible (or too ugly) because cpuset is too fluid that
3604 * task or memory node can be dynamically moved between cpusets.
3606 * The change of semantics for shared hugetlb mapping with cpuset is
3607 * undesirable. However, in order to preserve some of the semantics,
3608 * we fall back to check against current free page availability as
3609 * a best attempt and hopefully to minimize the impact of changing
3610 * semantics that cpuset has.
3612 * Apart from cpuset, we also have memory policy mechanism that
3613 * also determines from which node the kernel will allocate memory
3614 * in a NUMA system. So similar to cpuset, we also should consider
3615 * the memory policy of the current task. Similar to the description
3619 if (gather_surplus_pages(h, delta) < 0)
3622 if (delta > allowed_mems_nr(h)) {
3623 return_unused_surplus_pages(h, delta);
3630 return_unused_surplus_pages(h, (unsigned long) -delta);
3633 spin_unlock(&hugetlb_lock);
3637 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3639 struct resv_map *resv = vma_resv_map(vma);
3642 * This new VMA should share its siblings reservation map if present.
3643 * The VMA will only ever have a valid reservation map pointer where
3644 * it is being copied for another still existing VMA. As that VMA
3645 * has a reference to the reservation map it cannot disappear until
3646 * after this open call completes. It is therefore safe to take a
3647 * new reference here without additional locking.
3649 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3650 kref_get(&resv->refs);
3653 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3655 struct hstate *h = hstate_vma(vma);
3656 struct resv_map *resv = vma_resv_map(vma);
3657 struct hugepage_subpool *spool = subpool_vma(vma);
3658 unsigned long reserve, start, end;
3661 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3664 start = vma_hugecache_offset(h, vma, vma->vm_start);
3665 end = vma_hugecache_offset(h, vma, vma->vm_end);
3667 reserve = (end - start) - region_count(resv, start, end);
3668 hugetlb_cgroup_uncharge_counter(resv, start, end);
3671 * Decrement reserve counts. The global reserve count may be
3672 * adjusted if the subpool has a minimum size.
3674 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3675 hugetlb_acct_memory(h, -gbl_reserve);
3678 kref_put(&resv->refs, resv_map_release);
3681 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3683 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3688 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3690 struct hstate *hstate = hstate_vma(vma);
3692 return 1UL << huge_page_shift(hstate);
3696 * We cannot handle pagefaults against hugetlb pages at all. They cause
3697 * handle_mm_fault() to try to instantiate regular-sized pages in the
3698 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3701 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3708 * When a new function is introduced to vm_operations_struct and added
3709 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3710 * This is because under System V memory model, mappings created via
3711 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3712 * their original vm_ops are overwritten with shm_vm_ops.
3714 const struct vm_operations_struct hugetlb_vm_ops = {
3715 .fault = hugetlb_vm_op_fault,
3716 .open = hugetlb_vm_op_open,
3717 .close = hugetlb_vm_op_close,
3718 .split = hugetlb_vm_op_split,
3719 .pagesize = hugetlb_vm_op_pagesize,
3722 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3728 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3729 vma->vm_page_prot)));
3731 entry = huge_pte_wrprotect(mk_huge_pte(page,
3732 vma->vm_page_prot));
3734 entry = pte_mkyoung(entry);
3735 entry = pte_mkhuge(entry);
3736 entry = arch_make_huge_pte(entry, vma, page, writable);
3741 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3742 unsigned long address, pte_t *ptep)
3746 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3747 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3748 update_mmu_cache(vma, address, ptep);
3751 bool is_hugetlb_entry_migration(pte_t pte)
3755 if (huge_pte_none(pte) || pte_present(pte))
3757 swp = pte_to_swp_entry(pte);
3758 if (is_migration_entry(swp))
3764 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
3768 if (huge_pte_none(pte) || pte_present(pte))
3770 swp = pte_to_swp_entry(pte);
3771 if (is_hwpoison_entry(swp))
3777 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3778 struct vm_area_struct *vma)
3780 pte_t *src_pte, *dst_pte, entry, dst_entry;
3781 struct page *ptepage;
3784 struct hstate *h = hstate_vma(vma);
3785 unsigned long sz = huge_page_size(h);
3786 struct address_space *mapping = vma->vm_file->f_mapping;
3787 struct mmu_notifier_range range;
3790 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3793 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3796 mmu_notifier_invalidate_range_start(&range);
3799 * For shared mappings i_mmap_rwsem must be held to call
3800 * huge_pte_alloc, otherwise the returned ptep could go
3801 * away if part of a shared pmd and another thread calls
3804 i_mmap_lock_read(mapping);
3807 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3808 spinlock_t *src_ptl, *dst_ptl;
3809 src_pte = huge_pte_offset(src, addr, sz);
3812 dst_pte = huge_pte_alloc(dst, addr, sz);
3819 * If the pagetables are shared don't copy or take references.
3820 * dst_pte == src_pte is the common case of src/dest sharing.
3822 * However, src could have 'unshared' and dst shares with
3823 * another vma. If dst_pte !none, this implies sharing.
3824 * Check here before taking page table lock, and once again
3825 * after taking the lock below.
3827 dst_entry = huge_ptep_get(dst_pte);
3828 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3831 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3832 src_ptl = huge_pte_lockptr(h, src, src_pte);
3833 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3834 entry = huge_ptep_get(src_pte);
3835 dst_entry = huge_ptep_get(dst_pte);
3836 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3838 * Skip if src entry none. Also, skip in the
3839 * unlikely case dst entry !none as this implies
3840 * sharing with another vma.
3843 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3844 is_hugetlb_entry_hwpoisoned(entry))) {
3845 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3847 if (is_write_migration_entry(swp_entry) && cow) {
3849 * COW mappings require pages in both
3850 * parent and child to be set to read.
3852 make_migration_entry_read(&swp_entry);
3853 entry = swp_entry_to_pte(swp_entry);
3854 set_huge_swap_pte_at(src, addr, src_pte,
3857 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3861 * No need to notify as we are downgrading page
3862 * table protection not changing it to point
3865 * See Documentation/vm/mmu_notifier.rst
3867 huge_ptep_set_wrprotect(src, addr, src_pte);
3869 entry = huge_ptep_get(src_pte);
3870 ptepage = pte_page(entry);
3872 page_dup_rmap(ptepage, true);
3873 set_huge_pte_at(dst, addr, dst_pte, entry);
3874 hugetlb_count_add(pages_per_huge_page(h), dst);
3876 spin_unlock(src_ptl);
3877 spin_unlock(dst_ptl);
3881 mmu_notifier_invalidate_range_end(&range);
3883 i_mmap_unlock_read(mapping);
3888 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3889 unsigned long start, unsigned long end,
3890 struct page *ref_page)
3892 struct mm_struct *mm = vma->vm_mm;
3893 unsigned long address;
3898 struct hstate *h = hstate_vma(vma);
3899 unsigned long sz = huge_page_size(h);
3900 struct mmu_notifier_range range;
3902 WARN_ON(!is_vm_hugetlb_page(vma));
3903 BUG_ON(start & ~huge_page_mask(h));
3904 BUG_ON(end & ~huge_page_mask(h));
3907 * This is a hugetlb vma, all the pte entries should point
3910 tlb_change_page_size(tlb, sz);
3911 tlb_start_vma(tlb, vma);
3914 * If sharing possible, alert mmu notifiers of worst case.
3916 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3918 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3919 mmu_notifier_invalidate_range_start(&range);
3921 for (; address < end; address += sz) {
3922 ptep = huge_pte_offset(mm, address, sz);
3926 ptl = huge_pte_lock(h, mm, ptep);
3927 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
3930 * We just unmapped a page of PMDs by clearing a PUD.
3931 * The caller's TLB flush range should cover this area.
3936 pte = huge_ptep_get(ptep);
3937 if (huge_pte_none(pte)) {
3943 * Migrating hugepage or HWPoisoned hugepage is already
3944 * unmapped and its refcount is dropped, so just clear pte here.
3946 if (unlikely(!pte_present(pte))) {
3947 huge_pte_clear(mm, address, ptep, sz);
3952 page = pte_page(pte);
3954 * If a reference page is supplied, it is because a specific
3955 * page is being unmapped, not a range. Ensure the page we
3956 * are about to unmap is the actual page of interest.
3959 if (page != ref_page) {
3964 * Mark the VMA as having unmapped its page so that
3965 * future faults in this VMA will fail rather than
3966 * looking like data was lost
3968 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3971 pte = huge_ptep_get_and_clear(mm, address, ptep);
3972 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3973 if (huge_pte_dirty(pte))
3974 set_page_dirty(page);
3976 hugetlb_count_sub(pages_per_huge_page(h), mm);
3977 page_remove_rmap(page, true);
3980 tlb_remove_page_size(tlb, page, huge_page_size(h));
3982 * Bail out after unmapping reference page if supplied
3987 mmu_notifier_invalidate_range_end(&range);
3988 tlb_end_vma(tlb, vma);
3991 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3992 struct vm_area_struct *vma, unsigned long start,
3993 unsigned long end, struct page *ref_page)
3995 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3998 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3999 * test will fail on a vma being torn down, and not grab a page table
4000 * on its way out. We're lucky that the flag has such an appropriate
4001 * name, and can in fact be safely cleared here. We could clear it
4002 * before the __unmap_hugepage_range above, but all that's necessary
4003 * is to clear it before releasing the i_mmap_rwsem. This works
4004 * because in the context this is called, the VMA is about to be
4005 * destroyed and the i_mmap_rwsem is held.
4007 vma->vm_flags &= ~VM_MAYSHARE;
4010 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
4011 unsigned long end, struct page *ref_page)
4013 struct mm_struct *mm;
4014 struct mmu_gather tlb;
4015 unsigned long tlb_start = start;
4016 unsigned long tlb_end = end;
4019 * If shared PMDs were possibly used within this vma range, adjust
4020 * start/end for worst case tlb flushing.
4021 * Note that we can not be sure if PMDs are shared until we try to
4022 * unmap pages. However, we want to make sure TLB flushing covers
4023 * the largest possible range.
4025 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
4029 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
4030 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
4031 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
4035 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4036 * mappping it owns the reserve page for. The intention is to unmap the page
4037 * from other VMAs and let the children be SIGKILLed if they are faulting the
4040 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4041 struct page *page, unsigned long address)
4043 struct hstate *h = hstate_vma(vma);
4044 struct vm_area_struct *iter_vma;
4045 struct address_space *mapping;
4049 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4050 * from page cache lookup which is in HPAGE_SIZE units.
4052 address = address & huge_page_mask(h);
4053 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4055 mapping = vma->vm_file->f_mapping;
4058 * Take the mapping lock for the duration of the table walk. As
4059 * this mapping should be shared between all the VMAs,
4060 * __unmap_hugepage_range() is called as the lock is already held
4062 i_mmap_lock_write(mapping);
4063 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
4064 /* Do not unmap the current VMA */
4065 if (iter_vma == vma)
4069 * Shared VMAs have their own reserves and do not affect
4070 * MAP_PRIVATE accounting but it is possible that a shared
4071 * VMA is using the same page so check and skip such VMAs.
4073 if (iter_vma->vm_flags & VM_MAYSHARE)
4077 * Unmap the page from other VMAs without their own reserves.
4078 * They get marked to be SIGKILLed if they fault in these
4079 * areas. This is because a future no-page fault on this VMA
4080 * could insert a zeroed page instead of the data existing
4081 * from the time of fork. This would look like data corruption
4083 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
4084 unmap_hugepage_range(iter_vma, address,
4085 address + huge_page_size(h), page);
4087 i_mmap_unlock_write(mapping);
4091 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4092 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4093 * cannot race with other handlers or page migration.
4094 * Keep the pte_same checks anyway to make transition from the mutex easier.
4096 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
4097 unsigned long address, pte_t *ptep,
4098 struct page *pagecache_page, spinlock_t *ptl)
4101 struct hstate *h = hstate_vma(vma);
4102 struct page *old_page, *new_page;
4103 int outside_reserve = 0;
4105 unsigned long haddr = address & huge_page_mask(h);
4106 struct mmu_notifier_range range;
4108 pte = huge_ptep_get(ptep);
4109 old_page = pte_page(pte);
4112 /* If no-one else is actually using this page, avoid the copy
4113 * and just make the page writable */
4114 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
4115 page_move_anon_rmap(old_page, vma);
4116 set_huge_ptep_writable(vma, haddr, ptep);
4121 * If the process that created a MAP_PRIVATE mapping is about to
4122 * perform a COW due to a shared page count, attempt to satisfy
4123 * the allocation without using the existing reserves. The pagecache
4124 * page is used to determine if the reserve at this address was
4125 * consumed or not. If reserves were used, a partial faulted mapping
4126 * at the time of fork() could consume its reserves on COW instead
4127 * of the full address range.
4129 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4130 old_page != pagecache_page)
4131 outside_reserve = 1;
4136 * Drop page table lock as buddy allocator may be called. It will
4137 * be acquired again before returning to the caller, as expected.
4140 new_page = alloc_huge_page(vma, haddr, outside_reserve);
4142 if (IS_ERR(new_page)) {
4144 * If a process owning a MAP_PRIVATE mapping fails to COW,
4145 * it is due to references held by a child and an insufficient
4146 * huge page pool. To guarantee the original mappers
4147 * reliability, unmap the page from child processes. The child
4148 * may get SIGKILLed if it later faults.
4150 if (outside_reserve) {
4151 struct address_space *mapping = vma->vm_file->f_mapping;
4156 BUG_ON(huge_pte_none(pte));
4158 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4159 * unmapping. unmapping needs to hold i_mmap_rwsem
4160 * in write mode. Dropping i_mmap_rwsem in read mode
4161 * here is OK as COW mappings do not interact with
4164 * Reacquire both after unmap operation.
4166 idx = vma_hugecache_offset(h, vma, haddr);
4167 hash = hugetlb_fault_mutex_hash(mapping, idx);
4168 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4169 i_mmap_unlock_read(mapping);
4171 unmap_ref_private(mm, vma, old_page, haddr);
4173 i_mmap_lock_read(mapping);
4174 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4176 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4178 pte_same(huge_ptep_get(ptep), pte)))
4179 goto retry_avoidcopy;
4181 * race occurs while re-acquiring page table
4182 * lock, and our job is done.
4187 ret = vmf_error(PTR_ERR(new_page));
4188 goto out_release_old;
4192 * When the original hugepage is shared one, it does not have
4193 * anon_vma prepared.
4195 if (unlikely(anon_vma_prepare(vma))) {
4197 goto out_release_all;
4200 copy_user_huge_page(new_page, old_page, address, vma,
4201 pages_per_huge_page(h));
4202 __SetPageUptodate(new_page);
4204 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4205 haddr + huge_page_size(h));
4206 mmu_notifier_invalidate_range_start(&range);
4209 * Retake the page table lock to check for racing updates
4210 * before the page tables are altered
4213 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4214 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4215 ClearPagePrivate(new_page);
4218 huge_ptep_clear_flush(vma, haddr, ptep);
4219 mmu_notifier_invalidate_range(mm, range.start, range.end);
4220 set_huge_pte_at(mm, haddr, ptep,
4221 make_huge_pte(vma, new_page, 1));
4222 page_remove_rmap(old_page, true);
4223 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4224 set_page_huge_active(new_page);
4225 /* Make the old page be freed below */
4226 new_page = old_page;
4229 mmu_notifier_invalidate_range_end(&range);
4231 restore_reserve_on_error(h, vma, haddr, new_page);
4236 spin_lock(ptl); /* Caller expects lock to be held */
4240 /* Return the pagecache page at a given address within a VMA */
4241 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4242 struct vm_area_struct *vma, unsigned long address)
4244 struct address_space *mapping;
4247 mapping = vma->vm_file->f_mapping;
4248 idx = vma_hugecache_offset(h, vma, address);
4250 return find_lock_page(mapping, idx);
4254 * Return whether there is a pagecache page to back given address within VMA.
4255 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4257 static bool hugetlbfs_pagecache_present(struct hstate *h,
4258 struct vm_area_struct *vma, unsigned long address)
4260 struct address_space *mapping;
4264 mapping = vma->vm_file->f_mapping;
4265 idx = vma_hugecache_offset(h, vma, address);
4267 page = find_get_page(mapping, idx);
4270 return page != NULL;
4273 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4276 struct inode *inode = mapping->host;
4277 struct hstate *h = hstate_inode(inode);
4278 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4282 ClearPagePrivate(page);
4285 * set page dirty so that it will not be removed from cache/file
4286 * by non-hugetlbfs specific code paths.
4288 set_page_dirty(page);
4290 spin_lock(&inode->i_lock);
4291 inode->i_blocks += blocks_per_huge_page(h);
4292 spin_unlock(&inode->i_lock);
4296 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4297 struct vm_area_struct *vma,
4298 struct address_space *mapping, pgoff_t idx,
4299 unsigned long address, pte_t *ptep, unsigned int flags)
4301 struct hstate *h = hstate_vma(vma);
4302 vm_fault_t ret = VM_FAULT_SIGBUS;
4308 unsigned long haddr = address & huge_page_mask(h);
4309 bool new_page = false;
4312 * Currently, we are forced to kill the process in the event the
4313 * original mapper has unmapped pages from the child due to a failed
4314 * COW. Warn that such a situation has occurred as it may not be obvious
4316 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4317 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4323 * We can not race with truncation due to holding i_mmap_rwsem.
4324 * i_size is modified when holding i_mmap_rwsem, so check here
4325 * once for faults beyond end of file.
4327 size = i_size_read(mapping->host) >> huge_page_shift(h);
4332 page = find_lock_page(mapping, idx);
4335 * Check for page in userfault range
4337 if (userfaultfd_missing(vma)) {
4339 struct vm_fault vmf = {
4344 * Hard to debug if it ends up being
4345 * used by a callee that assumes
4346 * something about the other
4347 * uninitialized fields... same as in
4353 * hugetlb_fault_mutex and i_mmap_rwsem must be
4354 * dropped before handling userfault. Reacquire
4355 * after handling fault to make calling code simpler.
4357 hash = hugetlb_fault_mutex_hash(mapping, idx);
4358 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4359 i_mmap_unlock_read(mapping);
4360 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
4361 i_mmap_lock_read(mapping);
4362 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4366 page = alloc_huge_page(vma, haddr, 0);
4369 * Returning error will result in faulting task being
4370 * sent SIGBUS. The hugetlb fault mutex prevents two
4371 * tasks from racing to fault in the same page which
4372 * could result in false unable to allocate errors.
4373 * Page migration does not take the fault mutex, but
4374 * does a clear then write of pte's under page table
4375 * lock. Page fault code could race with migration,
4376 * notice the clear pte and try to allocate a page
4377 * here. Before returning error, get ptl and make
4378 * sure there really is no pte entry.
4380 ptl = huge_pte_lock(h, mm, ptep);
4381 if (!huge_pte_none(huge_ptep_get(ptep))) {
4387 ret = vmf_error(PTR_ERR(page));
4390 clear_huge_page(page, address, pages_per_huge_page(h));
4391 __SetPageUptodate(page);
4394 if (vma->vm_flags & VM_MAYSHARE) {
4395 int err = huge_add_to_page_cache(page, mapping, idx);
4404 if (unlikely(anon_vma_prepare(vma))) {
4406 goto backout_unlocked;
4412 * If memory error occurs between mmap() and fault, some process
4413 * don't have hwpoisoned swap entry for errored virtual address.
4414 * So we need to block hugepage fault by PG_hwpoison bit check.
4416 if (unlikely(PageHWPoison(page))) {
4417 ret = VM_FAULT_HWPOISON_LARGE |
4418 VM_FAULT_SET_HINDEX(hstate_index(h));
4419 goto backout_unlocked;
4424 * If we are going to COW a private mapping later, we examine the
4425 * pending reservations for this page now. This will ensure that
4426 * any allocations necessary to record that reservation occur outside
4429 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4430 if (vma_needs_reservation(h, vma, haddr) < 0) {
4432 goto backout_unlocked;
4434 /* Just decrements count, does not deallocate */
4435 vma_end_reservation(h, vma, haddr);
4438 ptl = huge_pte_lock(h, mm, ptep);
4440 if (!huge_pte_none(huge_ptep_get(ptep)))
4444 ClearPagePrivate(page);
4445 hugepage_add_new_anon_rmap(page, vma, haddr);
4447 page_dup_rmap(page, true);
4448 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4449 && (vma->vm_flags & VM_SHARED)));
4450 set_huge_pte_at(mm, haddr, ptep, new_pte);
4452 hugetlb_count_add(pages_per_huge_page(h), mm);
4453 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4454 /* Optimization, do the COW without a second fault */
4455 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4461 * Only make newly allocated pages active. Existing pages found
4462 * in the pagecache could be !page_huge_active() if they have been
4463 * isolated for migration.
4466 set_page_huge_active(page);
4476 restore_reserve_on_error(h, vma, haddr, page);
4482 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4484 unsigned long key[2];
4487 key[0] = (unsigned long) mapping;
4490 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4492 return hash & (num_fault_mutexes - 1);
4496 * For uniprocesor systems we always use a single mutex, so just
4497 * return 0 and avoid the hashing overhead.
4499 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4505 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4506 unsigned long address, unsigned int flags)
4513 struct page *page = NULL;
4514 struct page *pagecache_page = NULL;
4515 struct hstate *h = hstate_vma(vma);
4516 struct address_space *mapping;
4517 int need_wait_lock = 0;
4518 unsigned long haddr = address & huge_page_mask(h);
4520 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4523 * Since we hold no locks, ptep could be stale. That is
4524 * OK as we are only making decisions based on content and
4525 * not actually modifying content here.
4527 entry = huge_ptep_get(ptep);
4528 if (unlikely(is_hugetlb_entry_migration(entry))) {
4529 migration_entry_wait_huge(vma, mm, ptep);
4531 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4532 return VM_FAULT_HWPOISON_LARGE |
4533 VM_FAULT_SET_HINDEX(hstate_index(h));
4537 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4538 * until finished with ptep. This serves two purposes:
4539 * 1) It prevents huge_pmd_unshare from being called elsewhere
4540 * and making the ptep no longer valid.
4541 * 2) It synchronizes us with i_size modifications during truncation.
4543 * ptep could have already be assigned via huge_pte_offset. That
4544 * is OK, as huge_pte_alloc will return the same value unless
4545 * something has changed.
4547 mapping = vma->vm_file->f_mapping;
4548 i_mmap_lock_read(mapping);
4549 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4551 i_mmap_unlock_read(mapping);
4552 return VM_FAULT_OOM;
4556 * Serialize hugepage allocation and instantiation, so that we don't
4557 * get spurious allocation failures if two CPUs race to instantiate
4558 * the same page in the page cache.
4560 idx = vma_hugecache_offset(h, vma, haddr);
4561 hash = hugetlb_fault_mutex_hash(mapping, idx);
4562 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4564 entry = huge_ptep_get(ptep);
4565 if (huge_pte_none(entry)) {
4566 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4573 * entry could be a migration/hwpoison entry at this point, so this
4574 * check prevents the kernel from going below assuming that we have
4575 * an active hugepage in pagecache. This goto expects the 2nd page
4576 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4577 * properly handle it.
4579 if (!pte_present(entry))
4583 * If we are going to COW the mapping later, we examine the pending
4584 * reservations for this page now. This will ensure that any
4585 * allocations necessary to record that reservation occur outside the
4586 * spinlock. For private mappings, we also lookup the pagecache
4587 * page now as it is used to determine if a reservation has been
4590 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4591 if (vma_needs_reservation(h, vma, haddr) < 0) {
4595 /* Just decrements count, does not deallocate */
4596 vma_end_reservation(h, vma, haddr);
4598 if (!(vma->vm_flags & VM_MAYSHARE))
4599 pagecache_page = hugetlbfs_pagecache_page(h,
4603 ptl = huge_pte_lock(h, mm, ptep);
4605 /* Check for a racing update before calling hugetlb_cow */
4606 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4610 * hugetlb_cow() requires page locks of pte_page(entry) and
4611 * pagecache_page, so here we need take the former one
4612 * when page != pagecache_page or !pagecache_page.
4614 page = pte_page(entry);
4615 if (page != pagecache_page)
4616 if (!trylock_page(page)) {
4623 if (flags & FAULT_FLAG_WRITE) {
4624 if (!huge_pte_write(entry)) {
4625 ret = hugetlb_cow(mm, vma, address, ptep,
4626 pagecache_page, ptl);
4629 entry = huge_pte_mkdirty(entry);
4631 entry = pte_mkyoung(entry);
4632 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4633 flags & FAULT_FLAG_WRITE))
4634 update_mmu_cache(vma, haddr, ptep);
4636 if (page != pagecache_page)
4642 if (pagecache_page) {
4643 unlock_page(pagecache_page);
4644 put_page(pagecache_page);
4647 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4648 i_mmap_unlock_read(mapping);
4650 * Generally it's safe to hold refcount during waiting page lock. But
4651 * here we just wait to defer the next page fault to avoid busy loop and
4652 * the page is not used after unlocked before returning from the current
4653 * page fault. So we are safe from accessing freed page, even if we wait
4654 * here without taking refcount.
4657 wait_on_page_locked(page);
4662 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4663 * modifications for huge pages.
4665 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4667 struct vm_area_struct *dst_vma,
4668 unsigned long dst_addr,
4669 unsigned long src_addr,
4670 struct page **pagep)
4672 struct address_space *mapping;
4675 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4676 struct hstate *h = hstate_vma(dst_vma);
4684 page = alloc_huge_page(dst_vma, dst_addr, 0);
4688 ret = copy_huge_page_from_user(page,
4689 (const void __user *) src_addr,
4690 pages_per_huge_page(h), false);
4692 /* fallback to copy_from_user outside mmap_lock */
4693 if (unlikely(ret)) {
4696 /* don't free the page */
4705 * The memory barrier inside __SetPageUptodate makes sure that
4706 * preceding stores to the page contents become visible before
4707 * the set_pte_at() write.
4709 __SetPageUptodate(page);
4711 mapping = dst_vma->vm_file->f_mapping;
4712 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4715 * If shared, add to page cache
4718 size = i_size_read(mapping->host) >> huge_page_shift(h);
4721 goto out_release_nounlock;
4724 * Serialization between remove_inode_hugepages() and
4725 * huge_add_to_page_cache() below happens through the
4726 * hugetlb_fault_mutex_table that here must be hold by
4729 ret = huge_add_to_page_cache(page, mapping, idx);
4731 goto out_release_nounlock;
4734 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4738 * Recheck the i_size after holding PT lock to make sure not
4739 * to leave any page mapped (as page_mapped()) beyond the end
4740 * of the i_size (remove_inode_hugepages() is strict about
4741 * enforcing that). If we bail out here, we'll also leave a
4742 * page in the radix tree in the vm_shared case beyond the end
4743 * of the i_size, but remove_inode_hugepages() will take care
4744 * of it as soon as we drop the hugetlb_fault_mutex_table.
4746 size = i_size_read(mapping->host) >> huge_page_shift(h);
4749 goto out_release_unlock;
4752 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4753 goto out_release_unlock;
4756 page_dup_rmap(page, true);
4758 ClearPagePrivate(page);
4759 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4762 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4763 if (dst_vma->vm_flags & VM_WRITE)
4764 _dst_pte = huge_pte_mkdirty(_dst_pte);
4765 _dst_pte = pte_mkyoung(_dst_pte);
4767 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4769 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4770 dst_vma->vm_flags & VM_WRITE);
4771 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4773 /* No need to invalidate - it was non-present before */
4774 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4777 set_page_huge_active(page);
4787 out_release_nounlock:
4792 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4793 struct page **pages, struct vm_area_struct **vmas,
4794 unsigned long *position, unsigned long *nr_pages,
4795 long i, unsigned int flags, int *locked)
4797 unsigned long pfn_offset;
4798 unsigned long vaddr = *position;
4799 unsigned long remainder = *nr_pages;
4800 struct hstate *h = hstate_vma(vma);
4803 while (vaddr < vma->vm_end && remainder) {
4805 spinlock_t *ptl = NULL;
4810 * If we have a pending SIGKILL, don't keep faulting pages and
4811 * potentially allocating memory.
4813 if (fatal_signal_pending(current)) {
4819 * Some archs (sparc64, sh*) have multiple pte_ts to
4820 * each hugepage. We have to make sure we get the
4821 * first, for the page indexing below to work.
4823 * Note that page table lock is not held when pte is null.
4825 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4828 ptl = huge_pte_lock(h, mm, pte);
4829 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4832 * When coredumping, it suits get_dump_page if we just return
4833 * an error where there's an empty slot with no huge pagecache
4834 * to back it. This way, we avoid allocating a hugepage, and
4835 * the sparse dumpfile avoids allocating disk blocks, but its
4836 * huge holes still show up with zeroes where they need to be.
4838 if (absent && (flags & FOLL_DUMP) &&
4839 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4847 * We need call hugetlb_fault for both hugepages under migration
4848 * (in which case hugetlb_fault waits for the migration,) and
4849 * hwpoisoned hugepages (in which case we need to prevent the
4850 * caller from accessing to them.) In order to do this, we use
4851 * here is_swap_pte instead of is_hugetlb_entry_migration and
4852 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4853 * both cases, and because we can't follow correct pages
4854 * directly from any kind of swap entries.
4856 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4857 ((flags & FOLL_WRITE) &&
4858 !huge_pte_write(huge_ptep_get(pte)))) {
4860 unsigned int fault_flags = 0;
4864 if (flags & FOLL_WRITE)
4865 fault_flags |= FAULT_FLAG_WRITE;
4867 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4868 FAULT_FLAG_KILLABLE;
4869 if (flags & FOLL_NOWAIT)
4870 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4871 FAULT_FLAG_RETRY_NOWAIT;
4872 if (flags & FOLL_TRIED) {
4874 * Note: FAULT_FLAG_ALLOW_RETRY and
4875 * FAULT_FLAG_TRIED can co-exist
4877 fault_flags |= FAULT_FLAG_TRIED;
4879 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4880 if (ret & VM_FAULT_ERROR) {
4881 err = vm_fault_to_errno(ret, flags);
4885 if (ret & VM_FAULT_RETRY) {
4887 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4891 * VM_FAULT_RETRY must not return an
4892 * error, it will return zero
4895 * No need to update "position" as the
4896 * caller will not check it after
4897 * *nr_pages is set to 0.
4904 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4905 page = pte_page(huge_ptep_get(pte));
4908 * If subpage information not requested, update counters
4909 * and skip the same_page loop below.
4911 if (!pages && !vmas && !pfn_offset &&
4912 (vaddr + huge_page_size(h) < vma->vm_end) &&
4913 (remainder >= pages_per_huge_page(h))) {
4914 vaddr += huge_page_size(h);
4915 remainder -= pages_per_huge_page(h);
4916 i += pages_per_huge_page(h);
4923 pages[i] = mem_map_offset(page, pfn_offset);
4925 * try_grab_page() should always succeed here, because:
4926 * a) we hold the ptl lock, and b) we've just checked
4927 * that the huge page is present in the page tables. If
4928 * the huge page is present, then the tail pages must
4929 * also be present. The ptl prevents the head page and
4930 * tail pages from being rearranged in any way. So this
4931 * page must be available at this point, unless the page
4932 * refcount overflowed:
4934 if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) {
4949 if (vaddr < vma->vm_end && remainder &&
4950 pfn_offset < pages_per_huge_page(h)) {
4952 * We use pfn_offset to avoid touching the pageframes
4953 * of this compound page.
4959 *nr_pages = remainder;
4961 * setting position is actually required only if remainder is
4962 * not zero but it's faster not to add a "if (remainder)"
4970 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4972 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4975 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4978 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4979 unsigned long address, unsigned long end, pgprot_t newprot)
4981 struct mm_struct *mm = vma->vm_mm;
4982 unsigned long start = address;
4985 struct hstate *h = hstate_vma(vma);
4986 unsigned long pages = 0;
4987 bool shared_pmd = false;
4988 struct mmu_notifier_range range;
4991 * In the case of shared PMDs, the area to flush could be beyond
4992 * start/end. Set range.start/range.end to cover the maximum possible
4993 * range if PMD sharing is possible.
4995 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
4996 0, vma, mm, start, end);
4997 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4999 BUG_ON(address >= end);
5000 flush_cache_range(vma, range.start, range.end);
5002 mmu_notifier_invalidate_range_start(&range);
5003 i_mmap_lock_write(vma->vm_file->f_mapping);
5004 for (; address < end; address += huge_page_size(h)) {
5006 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5009 ptl = huge_pte_lock(h, mm, ptep);
5010 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
5016 pte = huge_ptep_get(ptep);
5017 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5021 if (unlikely(is_hugetlb_entry_migration(pte))) {
5022 swp_entry_t entry = pte_to_swp_entry(pte);
5024 if (is_write_migration_entry(entry)) {
5027 make_migration_entry_read(&entry);
5028 newpte = swp_entry_to_pte(entry);
5029 set_huge_swap_pte_at(mm, address, ptep,
5030 newpte, huge_page_size(h));
5036 if (!huge_pte_none(pte)) {
5039 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5040 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
5041 pte = arch_make_huge_pte(pte, vma, NULL, 0);
5042 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
5048 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5049 * may have cleared our pud entry and done put_page on the page table:
5050 * once we release i_mmap_rwsem, another task can do the final put_page
5051 * and that page table be reused and filled with junk. If we actually
5052 * did unshare a page of pmds, flush the range corresponding to the pud.
5055 flush_hugetlb_tlb_range(vma, range.start, range.end);
5057 flush_hugetlb_tlb_range(vma, start, end);
5059 * No need to call mmu_notifier_invalidate_range() we are downgrading
5060 * page table protection not changing it to point to a new page.
5062 * See Documentation/vm/mmu_notifier.rst
5064 i_mmap_unlock_write(vma->vm_file->f_mapping);
5065 mmu_notifier_invalidate_range_end(&range);
5067 return pages << h->order;
5070 int hugetlb_reserve_pages(struct inode *inode,
5072 struct vm_area_struct *vma,
5073 vm_flags_t vm_flags)
5075 long ret, chg, add = -1;
5076 struct hstate *h = hstate_inode(inode);
5077 struct hugepage_subpool *spool = subpool_inode(inode);
5078 struct resv_map *resv_map;
5079 struct hugetlb_cgroup *h_cg = NULL;
5080 long gbl_reserve, regions_needed = 0;
5082 /* This should never happen */
5084 VM_WARN(1, "%s called with a negative range\n", __func__);
5089 * Only apply hugepage reservation if asked. At fault time, an
5090 * attempt will be made for VM_NORESERVE to allocate a page
5091 * without using reserves
5093 if (vm_flags & VM_NORESERVE)
5097 * Shared mappings base their reservation on the number of pages that
5098 * are already allocated on behalf of the file. Private mappings need
5099 * to reserve the full area even if read-only as mprotect() may be
5100 * called to make the mapping read-write. Assume !vma is a shm mapping
5102 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5104 * resv_map can not be NULL as hugetlb_reserve_pages is only
5105 * called for inodes for which resv_maps were created (see
5106 * hugetlbfs_get_inode).
5108 resv_map = inode_resv_map(inode);
5110 chg = region_chg(resv_map, from, to, ®ions_needed);
5113 /* Private mapping. */
5114 resv_map = resv_map_alloc();
5120 set_vma_resv_map(vma, resv_map);
5121 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5129 ret = hugetlb_cgroup_charge_cgroup_rsvd(
5130 hstate_index(h), chg * pages_per_huge_page(h), &h_cg);
5137 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5138 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5141 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5145 * There must be enough pages in the subpool for the mapping. If
5146 * the subpool has a minimum size, there may be some global
5147 * reservations already in place (gbl_reserve).
5149 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5150 if (gbl_reserve < 0) {
5152 goto out_uncharge_cgroup;
5156 * Check enough hugepages are available for the reservation.
5157 * Hand the pages back to the subpool if there are not
5159 ret = hugetlb_acct_memory(h, gbl_reserve);
5165 * Account for the reservations made. Shared mappings record regions
5166 * that have reservations as they are shared by multiple VMAs.
5167 * When the last VMA disappears, the region map says how much
5168 * the reservation was and the page cache tells how much of
5169 * the reservation was consumed. Private mappings are per-VMA and
5170 * only the consumed reservations are tracked. When the VMA
5171 * disappears, the original reservation is the VMA size and the
5172 * consumed reservations are stored in the map. Hence, nothing
5173 * else has to be done for private mappings here
5175 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5176 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5178 if (unlikely(add < 0)) {
5179 hugetlb_acct_memory(h, -gbl_reserve);
5182 } else if (unlikely(chg > add)) {
5184 * pages in this range were added to the reserve
5185 * map between region_chg and region_add. This
5186 * indicates a race with alloc_huge_page. Adjust
5187 * the subpool and reserve counts modified above
5188 * based on the difference.
5192 hugetlb_cgroup_uncharge_cgroup_rsvd(
5194 (chg - add) * pages_per_huge_page(h), h_cg);
5196 rsv_adjust = hugepage_subpool_put_pages(spool,
5198 hugetlb_acct_memory(h, -rsv_adjust);
5203 /* put back original number of pages, chg */
5204 (void)hugepage_subpool_put_pages(spool, chg);
5205 out_uncharge_cgroup:
5206 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5207 chg * pages_per_huge_page(h), h_cg);
5209 if (!vma || vma->vm_flags & VM_MAYSHARE)
5210 /* Only call region_abort if the region_chg succeeded but the
5211 * region_add failed or didn't run.
5213 if (chg >= 0 && add < 0)
5214 region_abort(resv_map, from, to, regions_needed);
5215 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5216 kref_put(&resv_map->refs, resv_map_release);
5220 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5223 struct hstate *h = hstate_inode(inode);
5224 struct resv_map *resv_map = inode_resv_map(inode);
5226 struct hugepage_subpool *spool = subpool_inode(inode);
5230 * Since this routine can be called in the evict inode path for all
5231 * hugetlbfs inodes, resv_map could be NULL.
5234 chg = region_del(resv_map, start, end);
5236 * region_del() can fail in the rare case where a region
5237 * must be split and another region descriptor can not be
5238 * allocated. If end == LONG_MAX, it will not fail.
5244 spin_lock(&inode->i_lock);
5245 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5246 spin_unlock(&inode->i_lock);
5249 * If the subpool has a minimum size, the number of global
5250 * reservations to be released may be adjusted.
5252 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5253 hugetlb_acct_memory(h, -gbl_reserve);
5258 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5259 static unsigned long page_table_shareable(struct vm_area_struct *svma,
5260 struct vm_area_struct *vma,
5261 unsigned long addr, pgoff_t idx)
5263 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5265 unsigned long sbase = saddr & PUD_MASK;
5266 unsigned long s_end = sbase + PUD_SIZE;
5268 /* Allow segments to share if only one is marked locked */
5269 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5270 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5273 * match the virtual addresses, permission and the alignment of the
5276 if (pmd_index(addr) != pmd_index(saddr) ||
5277 vm_flags != svm_flags ||
5278 sbase < svma->vm_start || svma->vm_end < s_end)
5284 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5286 unsigned long base = addr & PUD_MASK;
5287 unsigned long end = base + PUD_SIZE;
5290 * check on proper vm_flags and page table alignment
5292 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5298 * Determine if start,end range within vma could be mapped by shared pmd.
5299 * If yes, adjust start and end to cover range associated with possible
5300 * shared pmd mappings.
5302 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5303 unsigned long *start, unsigned long *end)
5305 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
5306 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
5309 * vma need span at least one aligned PUD size and the start,end range
5310 * must at least partialy within it.
5312 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
5313 (*end <= v_start) || (*start >= v_end))
5316 /* Extend the range to be PUD aligned for a worst case scenario */
5317 if (*start > v_start)
5318 *start = ALIGN_DOWN(*start, PUD_SIZE);
5321 *end = ALIGN(*end, PUD_SIZE);
5325 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5326 * and returns the corresponding pte. While this is not necessary for the
5327 * !shared pmd case because we can allocate the pmd later as well, it makes the
5328 * code much cleaner.
5330 * This routine must be called with i_mmap_rwsem held in at least read mode if
5331 * sharing is possible. For hugetlbfs, this prevents removal of any page
5332 * table entries associated with the address space. This is important as we
5333 * are setting up sharing based on existing page table entries (mappings).
5335 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5336 * huge_pte_alloc know that sharing is not possible and do not take
5337 * i_mmap_rwsem as a performance optimization. This is handled by the
5338 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5339 * only required for subsequent processing.
5341 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5343 struct vm_area_struct *vma = find_vma(mm, addr);
5344 struct address_space *mapping = vma->vm_file->f_mapping;
5345 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5347 struct vm_area_struct *svma;
5348 unsigned long saddr;
5353 if (!vma_shareable(vma, addr))
5354 return (pte_t *)pmd_alloc(mm, pud, addr);
5356 i_mmap_assert_locked(mapping);
5357 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5361 saddr = page_table_shareable(svma, vma, addr, idx);
5363 spte = huge_pte_offset(svma->vm_mm, saddr,
5364 vma_mmu_pagesize(svma));
5366 get_page(virt_to_page(spte));
5375 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
5376 if (pud_none(*pud)) {
5377 pud_populate(mm, pud,
5378 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5381 put_page(virt_to_page(spte));
5385 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5390 * unmap huge page backed by shared pte.
5392 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5393 * indicated by page_count > 1, unmap is achieved by clearing pud and
5394 * decrementing the ref count. If count == 1, the pte page is not shared.
5396 * Called with page table lock held and i_mmap_rwsem held in write mode.
5398 * returns: 1 successfully unmapped a shared pte page
5399 * 0 the underlying pte page is not shared, or it is the last user
5401 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5402 unsigned long *addr, pte_t *ptep)
5404 pgd_t *pgd = pgd_offset(mm, *addr);
5405 p4d_t *p4d = p4d_offset(pgd, *addr);
5406 pud_t *pud = pud_offset(p4d, *addr);
5408 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
5409 BUG_ON(page_count(virt_to_page(ptep)) == 0);
5410 if (page_count(virt_to_page(ptep)) == 1)
5414 put_page(virt_to_page(ptep));
5416 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5419 #define want_pmd_share() (1)
5420 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5421 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5426 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5427 unsigned long *addr, pte_t *ptep)
5432 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5433 unsigned long *start, unsigned long *end)
5436 #define want_pmd_share() (0)
5437 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5439 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5440 pte_t *huge_pte_alloc(struct mm_struct *mm,
5441 unsigned long addr, unsigned long sz)
5448 pgd = pgd_offset(mm, addr);
5449 p4d = p4d_alloc(mm, pgd, addr);
5452 pud = pud_alloc(mm, p4d, addr);
5454 if (sz == PUD_SIZE) {
5457 BUG_ON(sz != PMD_SIZE);
5458 if (want_pmd_share() && pud_none(*pud))
5459 pte = huge_pmd_share(mm, addr, pud);
5461 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5464 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
5470 * huge_pte_offset() - Walk the page table to resolve the hugepage
5471 * entry at address @addr
5473 * Return: Pointer to page table entry (PUD or PMD) for
5474 * address @addr, or NULL if a !p*d_present() entry is encountered and the
5475 * size @sz doesn't match the hugepage size at this level of the page
5478 pte_t *huge_pte_offset(struct mm_struct *mm,
5479 unsigned long addr, unsigned long sz)
5486 pgd = pgd_offset(mm, addr);
5487 if (!pgd_present(*pgd))
5489 p4d = p4d_offset(pgd, addr);
5490 if (!p4d_present(*p4d))
5493 pud = pud_offset(p4d, addr);
5495 /* must be pud huge, non-present or none */
5496 return (pte_t *)pud;
5497 if (!pud_present(*pud))
5499 /* must have a valid entry and size to go further */
5501 pmd = pmd_offset(pud, addr);
5502 /* must be pmd huge, non-present or none */
5503 return (pte_t *)pmd;
5506 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5509 * These functions are overwritable if your architecture needs its own
5512 struct page * __weak
5513 follow_huge_addr(struct mm_struct *mm, unsigned long address,
5516 return ERR_PTR(-EINVAL);
5519 struct page * __weak
5520 follow_huge_pd(struct vm_area_struct *vma,
5521 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5523 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5527 struct page * __weak
5528 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
5529 pmd_t *pmd, int flags)
5531 struct page *page = NULL;
5535 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5536 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
5537 (FOLL_PIN | FOLL_GET)))
5541 ptl = pmd_lockptr(mm, pmd);
5544 * make sure that the address range covered by this pmd is not
5545 * unmapped from other threads.
5547 if (!pmd_huge(*pmd))
5549 pte = huge_ptep_get((pte_t *)pmd);
5550 if (pte_present(pte)) {
5551 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
5553 * try_grab_page() should always succeed here, because: a) we
5554 * hold the pmd (ptl) lock, and b) we've just checked that the
5555 * huge pmd (head) page is present in the page tables. The ptl
5556 * prevents the head page and tail pages from being rearranged
5557 * in any way. So this page must be available at this point,
5558 * unless the page refcount overflowed:
5560 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5565 if (is_hugetlb_entry_migration(pte)) {
5567 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5571 * hwpoisoned entry is treated as no_page_table in
5572 * follow_page_mask().
5580 struct page * __weak
5581 follow_huge_pud(struct mm_struct *mm, unsigned long address,
5582 pud_t *pud, int flags)
5584 if (flags & (FOLL_GET | FOLL_PIN))
5587 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5590 struct page * __weak
5591 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5593 if (flags & (FOLL_GET | FOLL_PIN))
5596 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5599 bool isolate_huge_page(struct page *page, struct list_head *list)
5603 spin_lock(&hugetlb_lock);
5604 if (!PageHeadHuge(page) || !page_huge_active(page) ||
5605 !get_page_unless_zero(page)) {
5609 clear_page_huge_active(page);
5610 list_move_tail(&page->lru, list);
5612 spin_unlock(&hugetlb_lock);
5616 void putback_active_hugepage(struct page *page)
5618 VM_BUG_ON_PAGE(!PageHead(page), page);
5619 spin_lock(&hugetlb_lock);
5620 set_page_huge_active(page);
5621 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5622 spin_unlock(&hugetlb_lock);
5626 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5628 struct hstate *h = page_hstate(oldpage);
5630 hugetlb_cgroup_migrate(oldpage, newpage);
5631 set_page_owner_migrate_reason(newpage, reason);
5634 * transfer temporary state of the new huge page. This is
5635 * reverse to other transitions because the newpage is going to
5636 * be final while the old one will be freed so it takes over
5637 * the temporary status.
5639 * Also note that we have to transfer the per-node surplus state
5640 * here as well otherwise the global surplus count will not match
5643 if (PageHugeTemporary(newpage)) {
5644 int old_nid = page_to_nid(oldpage);
5645 int new_nid = page_to_nid(newpage);
5647 SetPageHugeTemporary(oldpage);
5648 ClearPageHugeTemporary(newpage);
5650 spin_lock(&hugetlb_lock);
5651 if (h->surplus_huge_pages_node[old_nid]) {
5652 h->surplus_huge_pages_node[old_nid]--;
5653 h->surplus_huge_pages_node[new_nid]++;
5655 spin_unlock(&hugetlb_lock);
5660 static bool cma_reserve_called __initdata;
5662 static int __init cmdline_parse_hugetlb_cma(char *p)
5664 hugetlb_cma_size = memparse(p, &p);
5668 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
5670 void __init hugetlb_cma_reserve(int order)
5672 unsigned long size, reserved, per_node;
5675 cma_reserve_called = true;
5677 if (!hugetlb_cma_size)
5680 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
5681 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
5682 (PAGE_SIZE << order) / SZ_1M);
5687 * If 3 GB area is requested on a machine with 4 numa nodes,
5688 * let's allocate 1 GB on first three nodes and ignore the last one.
5690 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
5691 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
5692 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
5695 for_each_node_state(nid, N_ONLINE) {
5697 char name[CMA_MAX_NAME];
5699 size = min(per_node, hugetlb_cma_size - reserved);
5700 size = round_up(size, PAGE_SIZE << order);
5702 snprintf(name, sizeof(name), "hugetlb%d", nid);
5703 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
5705 &hugetlb_cma[nid], nid);
5707 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
5713 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
5716 if (reserved >= hugetlb_cma_size)
5721 void __init hugetlb_cma_check(void)
5723 if (!hugetlb_cma_size || cma_reserve_called)
5726 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
5729 #endif /* CONFIG_CMA */