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 /* Forward declaration */
83 static int hugetlb_acct_memory(struct hstate *h, long delta);
85 static inline bool subpool_is_free(struct hugepage_subpool *spool)
89 if (spool->max_hpages != -1)
90 return spool->used_hpages == 0;
91 if (spool->min_hpages != -1)
92 return spool->rsv_hpages == spool->min_hpages;
97 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
99 spin_unlock(&spool->lock);
101 /* If no pages are used, and no other handles to the subpool
102 * remain, give up any reservations based on minimum size and
103 * free the subpool */
104 if (subpool_is_free(spool)) {
105 if (spool->min_hpages != -1)
106 hugetlb_acct_memory(spool->hstate,
112 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
115 struct hugepage_subpool *spool;
117 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
121 spin_lock_init(&spool->lock);
123 spool->max_hpages = max_hpages;
125 spool->min_hpages = min_hpages;
127 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
131 spool->rsv_hpages = min_hpages;
136 void hugepage_put_subpool(struct hugepage_subpool *spool)
138 spin_lock(&spool->lock);
139 BUG_ON(!spool->count);
141 unlock_or_release_subpool(spool);
145 * Subpool accounting for allocating and reserving pages.
146 * Return -ENOMEM if there are not enough resources to satisfy the
147 * request. Otherwise, return the number of pages by which the
148 * global pools must be adjusted (upward). The returned value may
149 * only be different than the passed value (delta) in the case where
150 * a subpool minimum size must be maintained.
152 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
160 spin_lock(&spool->lock);
162 if (spool->max_hpages != -1) { /* maximum size accounting */
163 if ((spool->used_hpages + delta) <= spool->max_hpages)
164 spool->used_hpages += delta;
171 /* minimum size accounting */
172 if (spool->min_hpages != -1 && spool->rsv_hpages) {
173 if (delta > spool->rsv_hpages) {
175 * Asking for more reserves than those already taken on
176 * behalf of subpool. Return difference.
178 ret = delta - spool->rsv_hpages;
179 spool->rsv_hpages = 0;
181 ret = 0; /* reserves already accounted for */
182 spool->rsv_hpages -= delta;
187 spin_unlock(&spool->lock);
192 * Subpool accounting for freeing and unreserving pages.
193 * Return the number of global page reservations that must be dropped.
194 * The return value may only be different than the passed value (delta)
195 * in the case where a subpool minimum size must be maintained.
197 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
205 spin_lock(&spool->lock);
207 if (spool->max_hpages != -1) /* maximum size accounting */
208 spool->used_hpages -= delta;
210 /* minimum size accounting */
211 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
212 if (spool->rsv_hpages + delta <= spool->min_hpages)
215 ret = spool->rsv_hpages + delta - spool->min_hpages;
217 spool->rsv_hpages += delta;
218 if (spool->rsv_hpages > spool->min_hpages)
219 spool->rsv_hpages = spool->min_hpages;
223 * If hugetlbfs_put_super couldn't free spool due to an outstanding
224 * quota reference, free it now.
226 unlock_or_release_subpool(spool);
231 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
233 return HUGETLBFS_SB(inode->i_sb)->spool;
236 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
238 return subpool_inode(file_inode(vma->vm_file));
241 /* Helper that removes a struct file_region from the resv_map cache and returns
244 static struct file_region *
245 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
247 struct file_region *nrg = NULL;
249 VM_BUG_ON(resv->region_cache_count <= 0);
251 resv->region_cache_count--;
252 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
253 list_del(&nrg->link);
261 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
262 struct file_region *rg)
264 #ifdef CONFIG_CGROUP_HUGETLB
265 nrg->reservation_counter = rg->reservation_counter;
272 /* Helper that records hugetlb_cgroup uncharge info. */
273 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
275 struct resv_map *resv,
276 struct file_region *nrg)
278 #ifdef CONFIG_CGROUP_HUGETLB
280 nrg->reservation_counter =
281 &h_cg->rsvd_hugepage[hstate_index(h)];
282 nrg->css = &h_cg->css;
284 * The caller will hold exactly one h_cg->css reference for the
285 * whole contiguous reservation region. But this area might be
286 * scattered when there are already some file_regions reside in
287 * it. As a result, many file_regions may share only one css
288 * reference. In order to ensure that one file_region must hold
289 * exactly one h_cg->css reference, we should do css_get for
290 * each file_region and leave the reference held by caller
294 if (!resv->pages_per_hpage)
295 resv->pages_per_hpage = pages_per_huge_page(h);
296 /* pages_per_hpage should be the same for all entries in
299 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
301 nrg->reservation_counter = NULL;
307 static void put_uncharge_info(struct file_region *rg)
309 #ifdef CONFIG_CGROUP_HUGETLB
315 static bool has_same_uncharge_info(struct file_region *rg,
316 struct file_region *org)
318 #ifdef CONFIG_CGROUP_HUGETLB
320 rg->reservation_counter == org->reservation_counter &&
328 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
330 struct file_region *nrg = NULL, *prg = NULL;
332 prg = list_prev_entry(rg, link);
333 if (&prg->link != &resv->regions && prg->to == rg->from &&
334 has_same_uncharge_info(prg, rg)) {
338 put_uncharge_info(rg);
344 nrg = list_next_entry(rg, link);
345 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
346 has_same_uncharge_info(nrg, rg)) {
347 nrg->from = rg->from;
350 put_uncharge_info(rg);
356 hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from,
357 long to, struct hstate *h, struct hugetlb_cgroup *cg,
358 long *regions_needed)
360 struct file_region *nrg;
362 if (!regions_needed) {
363 nrg = get_file_region_entry_from_cache(map, from, to);
364 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
365 list_add(&nrg->link, rg->link.prev);
366 coalesce_file_region(map, nrg);
368 *regions_needed += 1;
374 * Must be called with resv->lock held.
376 * Calling this with regions_needed != NULL will count the number of pages
377 * to be added but will not modify the linked list. And regions_needed will
378 * indicate the number of file_regions needed in the cache to carry out to add
379 * the regions for this range.
381 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
382 struct hugetlb_cgroup *h_cg,
383 struct hstate *h, long *regions_needed)
386 struct list_head *head = &resv->regions;
387 long last_accounted_offset = f;
388 struct file_region *rg = NULL, *trg = NULL;
393 /* In this loop, we essentially handle an entry for the range
394 * [last_accounted_offset, rg->from), at every iteration, with some
397 list_for_each_entry_safe(rg, trg, head, link) {
398 /* Skip irrelevant regions that start before our range. */
400 /* If this region ends after the last accounted offset,
401 * then we need to update last_accounted_offset.
403 if (rg->to > last_accounted_offset)
404 last_accounted_offset = rg->to;
408 /* When we find a region that starts beyond our range, we've
414 /* Add an entry for last_accounted_offset -> rg->from, and
415 * update last_accounted_offset.
417 if (rg->from > last_accounted_offset)
418 add += hugetlb_resv_map_add(resv, rg,
419 last_accounted_offset,
423 last_accounted_offset = rg->to;
426 /* Handle the case where our range extends beyond
427 * last_accounted_offset.
429 if (last_accounted_offset < t)
430 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
431 t, h, h_cg, regions_needed);
437 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
439 static int allocate_file_region_entries(struct resv_map *resv,
441 __must_hold(&resv->lock)
443 struct list_head allocated_regions;
444 int to_allocate = 0, i = 0;
445 struct file_region *trg = NULL, *rg = NULL;
447 VM_BUG_ON(regions_needed < 0);
449 INIT_LIST_HEAD(&allocated_regions);
452 * Check for sufficient descriptors in the cache to accommodate
453 * the number of in progress add operations plus regions_needed.
455 * This is a while loop because when we drop the lock, some other call
456 * to region_add or region_del may have consumed some region_entries,
457 * so we keep looping here until we finally have enough entries for
458 * (adds_in_progress + regions_needed).
460 while (resv->region_cache_count <
461 (resv->adds_in_progress + regions_needed)) {
462 to_allocate = resv->adds_in_progress + regions_needed -
463 resv->region_cache_count;
465 /* At this point, we should have enough entries in the cache
466 * for all the existings adds_in_progress. We should only be
467 * needing to allocate for regions_needed.
469 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
471 spin_unlock(&resv->lock);
472 for (i = 0; i < to_allocate; i++) {
473 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
476 list_add(&trg->link, &allocated_regions);
479 spin_lock(&resv->lock);
481 list_splice(&allocated_regions, &resv->region_cache);
482 resv->region_cache_count += to_allocate;
488 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
496 * Add the huge page range represented by [f, t) to the reserve
497 * map. Regions will be taken from the cache to fill in this range.
498 * Sufficient regions should exist in the cache due to the previous
499 * call to region_chg with the same range, but in some cases the cache will not
500 * have sufficient entries due to races with other code doing region_add or
501 * region_del. The extra needed entries will be allocated.
503 * regions_needed is the out value provided by a previous call to region_chg.
505 * Return the number of new huge pages added to the map. This number is greater
506 * than or equal to zero. If file_region entries needed to be allocated for
507 * this operation and we were not able to allocate, it returns -ENOMEM.
508 * region_add of regions of length 1 never allocate file_regions and cannot
509 * fail; region_chg will always allocate at least 1 entry and a region_add for
510 * 1 page will only require at most 1 entry.
512 static long region_add(struct resv_map *resv, long f, long t,
513 long in_regions_needed, struct hstate *h,
514 struct hugetlb_cgroup *h_cg)
516 long add = 0, actual_regions_needed = 0;
518 spin_lock(&resv->lock);
521 /* Count how many regions are actually needed to execute this add. */
522 add_reservation_in_range(resv, f, t, NULL, NULL,
523 &actual_regions_needed);
526 * Check for sufficient descriptors in the cache to accommodate
527 * this add operation. Note that actual_regions_needed may be greater
528 * than in_regions_needed, as the resv_map may have been modified since
529 * the region_chg call. In this case, we need to make sure that we
530 * allocate extra entries, such that we have enough for all the
531 * existing adds_in_progress, plus the excess needed for this
534 if (actual_regions_needed > in_regions_needed &&
535 resv->region_cache_count <
536 resv->adds_in_progress +
537 (actual_regions_needed - in_regions_needed)) {
538 /* region_add operation of range 1 should never need to
539 * allocate file_region entries.
541 VM_BUG_ON(t - f <= 1);
543 if (allocate_file_region_entries(
544 resv, actual_regions_needed - in_regions_needed)) {
551 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
553 resv->adds_in_progress -= in_regions_needed;
555 spin_unlock(&resv->lock);
560 * Examine the existing reserve map and determine how many
561 * huge pages in the specified range [f, t) are NOT currently
562 * represented. This routine is called before a subsequent
563 * call to region_add that will actually modify the reserve
564 * map to add the specified range [f, t). region_chg does
565 * not change the number of huge pages represented by the
566 * map. A number of new file_region structures is added to the cache as a
567 * placeholder, for the subsequent region_add call to use. At least 1
568 * file_region structure is added.
570 * out_regions_needed is the number of regions added to the
571 * resv->adds_in_progress. This value needs to be provided to a follow up call
572 * to region_add or region_abort for proper accounting.
574 * Returns the number of huge pages that need to be added to the existing
575 * reservation map for the range [f, t). This number is greater or equal to
576 * zero. -ENOMEM is returned if a new file_region structure or cache entry
577 * is needed and can not be allocated.
579 static long region_chg(struct resv_map *resv, long f, long t,
580 long *out_regions_needed)
584 spin_lock(&resv->lock);
586 /* Count how many hugepages in this range are NOT represented. */
587 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
590 if (*out_regions_needed == 0)
591 *out_regions_needed = 1;
593 if (allocate_file_region_entries(resv, *out_regions_needed))
596 resv->adds_in_progress += *out_regions_needed;
598 spin_unlock(&resv->lock);
603 * Abort the in progress add operation. The adds_in_progress field
604 * of the resv_map keeps track of the operations in progress between
605 * calls to region_chg and region_add. Operations are sometimes
606 * aborted after the call to region_chg. In such cases, region_abort
607 * is called to decrement the adds_in_progress counter. regions_needed
608 * is the value returned by the region_chg call, it is used to decrement
609 * the adds_in_progress counter.
611 * NOTE: The range arguments [f, t) are not needed or used in this
612 * routine. They are kept to make reading the calling code easier as
613 * arguments will match the associated region_chg call.
615 static void region_abort(struct resv_map *resv, long f, long t,
618 spin_lock(&resv->lock);
619 VM_BUG_ON(!resv->region_cache_count);
620 resv->adds_in_progress -= regions_needed;
621 spin_unlock(&resv->lock);
625 * Delete the specified range [f, t) from the reserve map. If the
626 * t parameter is LONG_MAX, this indicates that ALL regions after f
627 * should be deleted. Locate the regions which intersect [f, t)
628 * and either trim, delete or split the existing regions.
630 * Returns the number of huge pages deleted from the reserve map.
631 * In the normal case, the return value is zero or more. In the
632 * case where a region must be split, a new region descriptor must
633 * be allocated. If the allocation fails, -ENOMEM will be returned.
634 * NOTE: If the parameter t == LONG_MAX, then we will never split
635 * a region and possibly return -ENOMEM. Callers specifying
636 * t == LONG_MAX do not need to check for -ENOMEM error.
638 static long region_del(struct resv_map *resv, long f, long t)
640 struct list_head *head = &resv->regions;
641 struct file_region *rg, *trg;
642 struct file_region *nrg = NULL;
646 spin_lock(&resv->lock);
647 list_for_each_entry_safe(rg, trg, head, link) {
649 * Skip regions before the range to be deleted. file_region
650 * ranges are normally of the form [from, to). However, there
651 * may be a "placeholder" entry in the map which is of the form
652 * (from, to) with from == to. Check for placeholder entries
653 * at the beginning of the range to be deleted.
655 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
661 if (f > rg->from && t < rg->to) { /* Must split region */
663 * Check for an entry in the cache before dropping
664 * lock and attempting allocation.
667 resv->region_cache_count > resv->adds_in_progress) {
668 nrg = list_first_entry(&resv->region_cache,
671 list_del(&nrg->link);
672 resv->region_cache_count--;
676 spin_unlock(&resv->lock);
677 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
684 hugetlb_cgroup_uncharge_file_region(
685 resv, rg, t - f, false);
687 /* New entry for end of split region */
691 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
693 INIT_LIST_HEAD(&nrg->link);
695 /* Original entry is trimmed */
698 list_add(&nrg->link, &rg->link);
703 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
704 del += rg->to - rg->from;
705 hugetlb_cgroup_uncharge_file_region(resv, rg,
706 rg->to - rg->from, true);
712 if (f <= rg->from) { /* Trim beginning of region */
713 hugetlb_cgroup_uncharge_file_region(resv, rg,
714 t - rg->from, false);
718 } else { /* Trim end of region */
719 hugetlb_cgroup_uncharge_file_region(resv, rg,
727 spin_unlock(&resv->lock);
733 * A rare out of memory error was encountered which prevented removal of
734 * the reserve map region for a page. The huge page itself was free'ed
735 * and removed from the page cache. This routine will adjust the subpool
736 * usage count, and the global reserve count if needed. By incrementing
737 * these counts, the reserve map entry which could not be deleted will
738 * appear as a "reserved" entry instead of simply dangling with incorrect
741 void hugetlb_fix_reserve_counts(struct inode *inode)
743 struct hugepage_subpool *spool = subpool_inode(inode);
745 bool reserved = false;
747 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
748 if (rsv_adjust > 0) {
749 struct hstate *h = hstate_inode(inode);
751 if (!hugetlb_acct_memory(h, 1))
753 } else if (!rsv_adjust) {
758 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
762 * Count and return the number of huge pages in the reserve map
763 * that intersect with the range [f, t).
765 static long region_count(struct resv_map *resv, long f, long t)
767 struct list_head *head = &resv->regions;
768 struct file_region *rg;
771 spin_lock(&resv->lock);
772 /* Locate each segment we overlap with, and count that overlap. */
773 list_for_each_entry(rg, head, link) {
782 seg_from = max(rg->from, f);
783 seg_to = min(rg->to, t);
785 chg += seg_to - seg_from;
787 spin_unlock(&resv->lock);
793 * Convert the address within this vma to the page offset within
794 * the mapping, in pagecache page units; huge pages here.
796 static pgoff_t vma_hugecache_offset(struct hstate *h,
797 struct vm_area_struct *vma, unsigned long address)
799 return ((address - vma->vm_start) >> huge_page_shift(h)) +
800 (vma->vm_pgoff >> huge_page_order(h));
803 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
804 unsigned long address)
806 return vma_hugecache_offset(hstate_vma(vma), vma, address);
808 EXPORT_SYMBOL_GPL(linear_hugepage_index);
811 * Return the size of the pages allocated when backing a VMA. In the majority
812 * cases this will be same size as used by the page table entries.
814 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
816 if (vma->vm_ops && vma->vm_ops->pagesize)
817 return vma->vm_ops->pagesize(vma);
820 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
823 * Return the page size being used by the MMU to back a VMA. In the majority
824 * of cases, the page size used by the kernel matches the MMU size. On
825 * architectures where it differs, an architecture-specific 'strong'
826 * version of this symbol is required.
828 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
830 return vma_kernel_pagesize(vma);
834 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
835 * bits of the reservation map pointer, which are always clear due to
838 #define HPAGE_RESV_OWNER (1UL << 0)
839 #define HPAGE_RESV_UNMAPPED (1UL << 1)
840 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
843 * These helpers are used to track how many pages are reserved for
844 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
845 * is guaranteed to have their future faults succeed.
847 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
848 * the reserve counters are updated with the hugetlb_lock held. It is safe
849 * to reset the VMA at fork() time as it is not in use yet and there is no
850 * chance of the global counters getting corrupted as a result of the values.
852 * The private mapping reservation is represented in a subtly different
853 * manner to a shared mapping. A shared mapping has a region map associated
854 * with the underlying file, this region map represents the backing file
855 * pages which have ever had a reservation assigned which this persists even
856 * after the page is instantiated. A private mapping has a region map
857 * associated with the original mmap which is attached to all VMAs which
858 * reference it, this region map represents those offsets which have consumed
859 * reservation ie. where pages have been instantiated.
861 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
863 return (unsigned long)vma->vm_private_data;
866 static void set_vma_private_data(struct vm_area_struct *vma,
869 vma->vm_private_data = (void *)value;
873 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
874 struct hugetlb_cgroup *h_cg,
877 #ifdef CONFIG_CGROUP_HUGETLB
879 resv_map->reservation_counter = NULL;
880 resv_map->pages_per_hpage = 0;
881 resv_map->css = NULL;
883 resv_map->reservation_counter =
884 &h_cg->rsvd_hugepage[hstate_index(h)];
885 resv_map->pages_per_hpage = pages_per_huge_page(h);
886 resv_map->css = &h_cg->css;
891 struct resv_map *resv_map_alloc(void)
893 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
894 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
896 if (!resv_map || !rg) {
902 kref_init(&resv_map->refs);
903 spin_lock_init(&resv_map->lock);
904 INIT_LIST_HEAD(&resv_map->regions);
906 resv_map->adds_in_progress = 0;
908 * Initialize these to 0. On shared mappings, 0's here indicate these
909 * fields don't do cgroup accounting. On private mappings, these will be
910 * re-initialized to the proper values, to indicate that hugetlb cgroup
911 * reservations are to be un-charged from here.
913 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
915 INIT_LIST_HEAD(&resv_map->region_cache);
916 list_add(&rg->link, &resv_map->region_cache);
917 resv_map->region_cache_count = 1;
922 void resv_map_release(struct kref *ref)
924 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
925 struct list_head *head = &resv_map->region_cache;
926 struct file_region *rg, *trg;
928 /* Clear out any active regions before we release the map. */
929 region_del(resv_map, 0, LONG_MAX);
931 /* ... and any entries left in the cache */
932 list_for_each_entry_safe(rg, trg, head, link) {
937 VM_BUG_ON(resv_map->adds_in_progress);
942 static inline struct resv_map *inode_resv_map(struct inode *inode)
945 * At inode evict time, i_mapping may not point to the original
946 * address space within the inode. This original address space
947 * contains the pointer to the resv_map. So, always use the
948 * address space embedded within the inode.
949 * The VERY common case is inode->mapping == &inode->i_data but,
950 * this may not be true for device special inodes.
952 return (struct resv_map *)(&inode->i_data)->private_data;
955 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
957 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
958 if (vma->vm_flags & VM_MAYSHARE) {
959 struct address_space *mapping = vma->vm_file->f_mapping;
960 struct inode *inode = mapping->host;
962 return inode_resv_map(inode);
965 return (struct resv_map *)(get_vma_private_data(vma) &
970 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
972 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
973 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
975 set_vma_private_data(vma, (get_vma_private_data(vma) &
976 HPAGE_RESV_MASK) | (unsigned long)map);
979 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
981 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
982 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
984 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
987 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
989 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
991 return (get_vma_private_data(vma) & flag) != 0;
994 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
995 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
997 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
998 if (!(vma->vm_flags & VM_MAYSHARE))
999 vma->vm_private_data = (void *)0;
1002 /* Returns true if the VMA has associated reserve pages */
1003 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1005 if (vma->vm_flags & VM_NORESERVE) {
1007 * This address is already reserved by other process(chg == 0),
1008 * so, we should decrement reserved count. Without decrementing,
1009 * reserve count remains after releasing inode, because this
1010 * allocated page will go into page cache and is regarded as
1011 * coming from reserved pool in releasing step. Currently, we
1012 * don't have any other solution to deal with this situation
1013 * properly, so add work-around here.
1015 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1021 /* Shared mappings always use reserves */
1022 if (vma->vm_flags & VM_MAYSHARE) {
1024 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1025 * be a region map for all pages. The only situation where
1026 * there is no region map is if a hole was punched via
1027 * fallocate. In this case, there really are no reserves to
1028 * use. This situation is indicated if chg != 0.
1037 * Only the process that called mmap() has reserves for
1040 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1042 * Like the shared case above, a hole punch or truncate
1043 * could have been performed on the private mapping.
1044 * Examine the value of chg to determine if reserves
1045 * actually exist or were previously consumed.
1046 * Very Subtle - The value of chg comes from a previous
1047 * call to vma_needs_reserves(). The reserve map for
1048 * private mappings has different (opposite) semantics
1049 * than that of shared mappings. vma_needs_reserves()
1050 * has already taken this difference in semantics into
1051 * account. Therefore, the meaning of chg is the same
1052 * as in the shared case above. Code could easily be
1053 * combined, but keeping it separate draws attention to
1054 * subtle differences.
1065 static void enqueue_huge_page(struct hstate *h, struct page *page)
1067 int nid = page_to_nid(page);
1068 list_move(&page->lru, &h->hugepage_freelists[nid]);
1069 h->free_huge_pages++;
1070 h->free_huge_pages_node[nid]++;
1071 SetHPageFreed(page);
1074 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1077 bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA);
1079 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1080 if (nocma && is_migrate_cma_page(page))
1083 if (PageHWPoison(page))
1086 list_move(&page->lru, &h->hugepage_activelist);
1087 set_page_refcounted(page);
1088 ClearHPageFreed(page);
1089 h->free_huge_pages--;
1090 h->free_huge_pages_node[nid]--;
1097 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1100 unsigned int cpuset_mems_cookie;
1101 struct zonelist *zonelist;
1104 int node = NUMA_NO_NODE;
1106 zonelist = node_zonelist(nid, gfp_mask);
1109 cpuset_mems_cookie = read_mems_allowed_begin();
1110 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1113 if (!cpuset_zone_allowed(zone, gfp_mask))
1116 * no need to ask again on the same node. Pool is node rather than
1119 if (zone_to_nid(zone) == node)
1121 node = zone_to_nid(zone);
1123 page = dequeue_huge_page_node_exact(h, node);
1127 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1133 static struct page *dequeue_huge_page_vma(struct hstate *h,
1134 struct vm_area_struct *vma,
1135 unsigned long address, int avoid_reserve,
1139 struct mempolicy *mpol;
1141 nodemask_t *nodemask;
1145 * A child process with MAP_PRIVATE mappings created by their parent
1146 * have no page reserves. This check ensures that reservations are
1147 * not "stolen". The child may still get SIGKILLed
1149 if (!vma_has_reserves(vma, chg) &&
1150 h->free_huge_pages - h->resv_huge_pages == 0)
1153 /* If reserves cannot be used, ensure enough pages are in the pool */
1154 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1157 gfp_mask = htlb_alloc_mask(h);
1158 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1159 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1160 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1161 SetHPageRestoreReserve(page);
1162 h->resv_huge_pages--;
1165 mpol_cond_put(mpol);
1173 * common helper functions for hstate_next_node_to_{alloc|free}.
1174 * We may have allocated or freed a huge page based on a different
1175 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1176 * be outside of *nodes_allowed. Ensure that we use an allowed
1177 * node for alloc or free.
1179 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1181 nid = next_node_in(nid, *nodes_allowed);
1182 VM_BUG_ON(nid >= MAX_NUMNODES);
1187 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1189 if (!node_isset(nid, *nodes_allowed))
1190 nid = next_node_allowed(nid, nodes_allowed);
1195 * returns the previously saved node ["this node"] from which to
1196 * allocate a persistent huge page for the pool and advance the
1197 * next node from which to allocate, handling wrap at end of node
1200 static int hstate_next_node_to_alloc(struct hstate *h,
1201 nodemask_t *nodes_allowed)
1205 VM_BUG_ON(!nodes_allowed);
1207 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1208 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1214 * helper for free_pool_huge_page() - return the previously saved
1215 * node ["this node"] from which to free a huge page. Advance the
1216 * next node id whether or not we find a free huge page to free so
1217 * that the next attempt to free addresses the next node.
1219 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1223 VM_BUG_ON(!nodes_allowed);
1225 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1226 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1231 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1232 for (nr_nodes = nodes_weight(*mask); \
1234 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1237 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1238 for (nr_nodes = nodes_weight(*mask); \
1240 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1243 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1244 static void destroy_compound_gigantic_page(struct page *page,
1248 int nr_pages = 1 << order;
1249 struct page *p = page + 1;
1251 atomic_set(compound_mapcount_ptr(page), 0);
1252 atomic_set(compound_pincount_ptr(page), 0);
1254 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1255 clear_compound_head(p);
1256 set_page_refcounted(p);
1259 set_compound_order(page, 0);
1260 page[1].compound_nr = 0;
1261 __ClearPageHead(page);
1264 static void free_gigantic_page(struct page *page, unsigned int order)
1267 * If the page isn't allocated using the cma allocator,
1268 * cma_release() returns false.
1271 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1275 free_contig_range(page_to_pfn(page), 1 << order);
1278 #ifdef CONFIG_CONTIG_ALLOC
1279 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1280 int nid, nodemask_t *nodemask)
1282 unsigned long nr_pages = pages_per_huge_page(h);
1283 if (nid == NUMA_NO_NODE)
1284 nid = numa_mem_id();
1291 if (hugetlb_cma[nid]) {
1292 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1293 huge_page_order(h), true);
1298 if (!(gfp_mask & __GFP_THISNODE)) {
1299 for_each_node_mask(node, *nodemask) {
1300 if (node == nid || !hugetlb_cma[node])
1303 page = cma_alloc(hugetlb_cma[node], nr_pages,
1304 huge_page_order(h), true);
1312 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1315 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1316 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1317 #else /* !CONFIG_CONTIG_ALLOC */
1318 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1319 int nid, nodemask_t *nodemask)
1323 #endif /* CONFIG_CONTIG_ALLOC */
1325 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1326 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1327 int nid, nodemask_t *nodemask)
1331 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1332 static inline void destroy_compound_gigantic_page(struct page *page,
1333 unsigned int order) { }
1337 * Remove hugetlb page from lists, and update dtor so that page appears
1338 * as just a compound page. A reference is held on the page.
1340 * Must be called with hugetlb lock held.
1342 static void remove_hugetlb_page(struct hstate *h, struct page *page,
1343 bool adjust_surplus)
1345 int nid = page_to_nid(page);
1347 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1348 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1350 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1353 list_del(&page->lru);
1355 if (HPageFreed(page)) {
1356 h->free_huge_pages--;
1357 h->free_huge_pages_node[nid]--;
1359 if (adjust_surplus) {
1360 h->surplus_huge_pages--;
1361 h->surplus_huge_pages_node[nid]--;
1364 set_page_refcounted(page);
1365 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1368 h->nr_huge_pages_node[nid]--;
1371 static void update_and_free_page(struct hstate *h, struct page *page)
1374 struct page *subpage = page;
1376 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1379 for (i = 0; i < pages_per_huge_page(h);
1380 i++, subpage = mem_map_next(subpage, page, i)) {
1381 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1382 1 << PG_referenced | 1 << PG_dirty |
1383 1 << PG_active | 1 << PG_private |
1386 if (hstate_is_gigantic(h)) {
1387 destroy_compound_gigantic_page(page, huge_page_order(h));
1388 free_gigantic_page(page, huge_page_order(h));
1390 __free_pages(page, huge_page_order(h));
1394 struct hstate *size_to_hstate(unsigned long size)
1398 for_each_hstate(h) {
1399 if (huge_page_size(h) == size)
1405 static void __free_huge_page(struct page *page)
1408 * Can't pass hstate in here because it is called from the
1409 * compound page destructor.
1411 struct hstate *h = page_hstate(page);
1412 int nid = page_to_nid(page);
1413 struct hugepage_subpool *spool = hugetlb_page_subpool(page);
1414 bool restore_reserve;
1416 VM_BUG_ON_PAGE(page_count(page), page);
1417 VM_BUG_ON_PAGE(page_mapcount(page), page);
1419 hugetlb_set_page_subpool(page, NULL);
1420 page->mapping = NULL;
1421 restore_reserve = HPageRestoreReserve(page);
1422 ClearHPageRestoreReserve(page);
1425 * If HPageRestoreReserve 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 * reservation, 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 ClearHPageMigratable(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 (HPageTemporary(page)) {
1453 remove_hugetlb_page(h, page, false);
1454 spin_unlock(&hugetlb_lock);
1455 update_and_free_page(h, page);
1456 } else if (h->surplus_huge_pages_node[nid]) {
1457 /* remove the page from active list */
1458 remove_hugetlb_page(h, page, true);
1459 spin_unlock(&hugetlb_lock);
1460 update_and_free_page(h, page);
1462 arch_clear_hugepage_flags(page);
1463 enqueue_huge_page(h, page);
1464 spin_unlock(&hugetlb_lock);
1469 * As free_huge_page() can be called from a non-task context, we have
1470 * to defer the actual freeing in a workqueue to prevent potential
1471 * hugetlb_lock deadlock.
1473 * free_hpage_workfn() locklessly retrieves the linked list of pages to
1474 * be freed and frees them one-by-one. As the page->mapping pointer is
1475 * going to be cleared in __free_huge_page() anyway, it is reused as the
1476 * llist_node structure of a lockless linked list of huge pages to be freed.
1478 static LLIST_HEAD(hpage_freelist);
1480 static void free_hpage_workfn(struct work_struct *work)
1482 struct llist_node *node;
1485 node = llist_del_all(&hpage_freelist);
1488 page = container_of((struct address_space **)node,
1489 struct page, mapping);
1491 __free_huge_page(page);
1494 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1496 void free_huge_page(struct page *page)
1499 * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1503 * Only call schedule_work() if hpage_freelist is previously
1504 * empty. Otherwise, schedule_work() had been called but the
1505 * workfn hasn't retrieved the list yet.
1507 if (llist_add((struct llist_node *)&page->mapping,
1509 schedule_work(&free_hpage_work);
1513 __free_huge_page(page);
1516 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1518 INIT_LIST_HEAD(&page->lru);
1519 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1520 hugetlb_set_page_subpool(page, NULL);
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 ClearHPageFreed(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);
1558 atomic_set(compound_pincount_ptr(page), 0);
1562 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1563 * transparent huge pages. See the PageTransHuge() documentation for more
1566 int PageHuge(struct page *page)
1568 if (!PageCompound(page))
1571 page = compound_head(page);
1572 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1574 EXPORT_SYMBOL_GPL(PageHuge);
1577 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1578 * normal or transparent huge pages.
1580 int PageHeadHuge(struct page *page_head)
1582 if (!PageHead(page_head))
1585 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1589 * Find and lock address space (mapping) in write mode.
1591 * Upon entry, the page is locked which means that page_mapping() is
1592 * stable. Due to locking order, we can only trylock_write. If we can
1593 * not get the lock, simply return NULL to caller.
1595 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1597 struct address_space *mapping = page_mapping(hpage);
1602 if (i_mmap_trylock_write(mapping))
1608 pgoff_t __basepage_index(struct page *page)
1610 struct page *page_head = compound_head(page);
1611 pgoff_t index = page_index(page_head);
1612 unsigned long compound_idx;
1614 if (!PageHuge(page_head))
1615 return page_index(page);
1617 if (compound_order(page_head) >= MAX_ORDER)
1618 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1620 compound_idx = page - page_head;
1622 return (index << compound_order(page_head)) + compound_idx;
1625 static struct page *alloc_buddy_huge_page(struct hstate *h,
1626 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1627 nodemask_t *node_alloc_noretry)
1629 int order = huge_page_order(h);
1631 bool alloc_try_hard = true;
1634 * By default we always try hard to allocate the page with
1635 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1636 * a loop (to adjust global huge page counts) and previous allocation
1637 * failed, do not continue to try hard on the same node. Use the
1638 * node_alloc_noretry bitmap to manage this state information.
1640 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1641 alloc_try_hard = false;
1642 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1644 gfp_mask |= __GFP_RETRY_MAYFAIL;
1645 if (nid == NUMA_NO_NODE)
1646 nid = numa_mem_id();
1647 page = __alloc_pages(gfp_mask, order, nid, nmask);
1649 __count_vm_event(HTLB_BUDDY_PGALLOC);
1651 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1654 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1655 * indicates an overall state change. Clear bit so that we resume
1656 * normal 'try hard' allocations.
1658 if (node_alloc_noretry && page && !alloc_try_hard)
1659 node_clear(nid, *node_alloc_noretry);
1662 * If we tried hard to get a page but failed, set bit so that
1663 * subsequent attempts will not try as hard until there is an
1664 * overall state change.
1666 if (node_alloc_noretry && !page && alloc_try_hard)
1667 node_set(nid, *node_alloc_noretry);
1673 * Common helper to allocate a fresh hugetlb page. All specific allocators
1674 * should use this function to get new hugetlb pages
1676 static struct page *alloc_fresh_huge_page(struct hstate *h,
1677 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1678 nodemask_t *node_alloc_noretry)
1682 if (hstate_is_gigantic(h))
1683 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1685 page = alloc_buddy_huge_page(h, gfp_mask,
1686 nid, nmask, node_alloc_noretry);
1690 if (hstate_is_gigantic(h))
1691 prep_compound_gigantic_page(page, huge_page_order(h));
1692 prep_new_huge_page(h, page, page_to_nid(page));
1698 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1701 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1702 nodemask_t *node_alloc_noretry)
1706 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1708 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1709 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1710 node_alloc_noretry);
1718 put_page(page); /* free it into the hugepage allocator */
1724 * Free huge page from pool from next node to free.
1725 * Attempt to keep persistent huge pages more or less
1726 * balanced over allowed nodes.
1727 * Called with hugetlb_lock locked.
1729 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1735 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1737 * If we're returning unused surplus pages, only examine
1738 * nodes with surplus pages.
1740 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1741 !list_empty(&h->hugepage_freelists[node])) {
1743 list_entry(h->hugepage_freelists[node].next,
1745 remove_hugetlb_page(h, page, acct_surplus);
1747 * unlock/lock around update_and_free_page is temporary
1748 * and will be removed with subsequent patch.
1750 spin_unlock(&hugetlb_lock);
1751 update_and_free_page(h, page);
1752 spin_lock(&hugetlb_lock);
1762 * Dissolve a given free hugepage into free buddy pages. This function does
1763 * nothing for in-use hugepages and non-hugepages.
1764 * This function returns values like below:
1766 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1767 * (allocated or reserved.)
1768 * 0: successfully dissolved free hugepages or the page is not a
1769 * hugepage (considered as already dissolved)
1771 int dissolve_free_huge_page(struct page *page)
1776 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1777 if (!PageHuge(page))
1780 spin_lock(&hugetlb_lock);
1781 if (!PageHuge(page)) {
1786 if (!page_count(page)) {
1787 struct page *head = compound_head(page);
1788 struct hstate *h = page_hstate(head);
1789 if (h->free_huge_pages - h->resv_huge_pages == 0)
1793 * We should make sure that the page is already on the free list
1794 * when it is dissolved.
1796 if (unlikely(!HPageFreed(head))) {
1797 spin_unlock(&hugetlb_lock);
1801 * Theoretically, we should return -EBUSY when we
1802 * encounter this race. In fact, we have a chance
1803 * to successfully dissolve the page if we do a
1804 * retry. Because the race window is quite small.
1805 * If we seize this opportunity, it is an optimization
1806 * for increasing the success rate of dissolving page.
1812 * Move PageHWPoison flag from head page to the raw error page,
1813 * which makes any subpages rather than the error page reusable.
1815 if (PageHWPoison(head) && page != head) {
1816 SetPageHWPoison(page);
1817 ClearPageHWPoison(head);
1819 remove_hugetlb_page(h, page, false);
1820 h->max_huge_pages--;
1821 spin_unlock(&hugetlb_lock);
1822 update_and_free_page(h, head);
1826 spin_unlock(&hugetlb_lock);
1831 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1832 * make specified memory blocks removable from the system.
1833 * Note that this will dissolve a free gigantic hugepage completely, if any
1834 * part of it lies within the given range.
1835 * Also note that if dissolve_free_huge_page() returns with an error, all
1836 * free hugepages that were dissolved before that error are lost.
1838 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1844 if (!hugepages_supported())
1847 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1848 page = pfn_to_page(pfn);
1849 rc = dissolve_free_huge_page(page);
1858 * Allocates a fresh surplus page from the page allocator.
1860 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1861 int nid, nodemask_t *nmask)
1863 struct page *page = NULL;
1865 if (hstate_is_gigantic(h))
1868 spin_lock(&hugetlb_lock);
1869 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1871 spin_unlock(&hugetlb_lock);
1873 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1877 spin_lock(&hugetlb_lock);
1879 * We could have raced with the pool size change.
1880 * Double check that and simply deallocate the new page
1881 * if we would end up overcommiting the surpluses. Abuse
1882 * temporary page to workaround the nasty free_huge_page
1885 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1886 SetHPageTemporary(page);
1887 spin_unlock(&hugetlb_lock);
1891 h->surplus_huge_pages++;
1892 h->surplus_huge_pages_node[page_to_nid(page)]++;
1896 spin_unlock(&hugetlb_lock);
1901 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1902 int nid, nodemask_t *nmask)
1906 if (hstate_is_gigantic(h))
1909 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1914 * We do not account these pages as surplus because they are only
1915 * temporary and will be released properly on the last reference
1917 SetHPageTemporary(page);
1923 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1926 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1927 struct vm_area_struct *vma, unsigned long addr)
1930 struct mempolicy *mpol;
1931 gfp_t gfp_mask = htlb_alloc_mask(h);
1933 nodemask_t *nodemask;
1935 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1936 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1937 mpol_cond_put(mpol);
1942 /* page migration callback function */
1943 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1944 nodemask_t *nmask, gfp_t gfp_mask)
1946 spin_lock(&hugetlb_lock);
1947 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1950 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1952 spin_unlock(&hugetlb_lock);
1956 spin_unlock(&hugetlb_lock);
1958 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1961 /* mempolicy aware migration callback */
1962 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1963 unsigned long address)
1965 struct mempolicy *mpol;
1966 nodemask_t *nodemask;
1971 gfp_mask = htlb_alloc_mask(h);
1972 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1973 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
1974 mpol_cond_put(mpol);
1980 * Increase the hugetlb pool such that it can accommodate a reservation
1983 static int gather_surplus_pages(struct hstate *h, long delta)
1984 __must_hold(&hugetlb_lock)
1986 struct list_head surplus_list;
1987 struct page *page, *tmp;
1990 long needed, allocated;
1991 bool alloc_ok = true;
1993 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1995 h->resv_huge_pages += delta;
2000 INIT_LIST_HEAD(&surplus_list);
2004 spin_unlock(&hugetlb_lock);
2005 for (i = 0; i < needed; i++) {
2006 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2007 NUMA_NO_NODE, NULL);
2012 list_add(&page->lru, &surplus_list);
2018 * After retaking hugetlb_lock, we need to recalculate 'needed'
2019 * because either resv_huge_pages or free_huge_pages may have changed.
2021 spin_lock(&hugetlb_lock);
2022 needed = (h->resv_huge_pages + delta) -
2023 (h->free_huge_pages + allocated);
2028 * We were not able to allocate enough pages to
2029 * satisfy the entire reservation so we free what
2030 * we've allocated so far.
2035 * The surplus_list now contains _at_least_ the number of extra pages
2036 * needed to accommodate the reservation. Add the appropriate number
2037 * of pages to the hugetlb pool and free the extras back to the buddy
2038 * allocator. Commit the entire reservation here to prevent another
2039 * process from stealing the pages as they are added to the pool but
2040 * before they are reserved.
2042 needed += allocated;
2043 h->resv_huge_pages += delta;
2046 /* Free the needed pages to the hugetlb pool */
2047 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2053 * This page is now managed by the hugetlb allocator and has
2054 * no users -- drop the buddy allocator's reference.
2056 zeroed = put_page_testzero(page);
2057 VM_BUG_ON_PAGE(!zeroed, page);
2058 enqueue_huge_page(h, page);
2061 spin_unlock(&hugetlb_lock);
2063 /* Free unnecessary surplus pages to the buddy allocator */
2064 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2066 spin_lock(&hugetlb_lock);
2072 * This routine has two main purposes:
2073 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2074 * in unused_resv_pages. This corresponds to the prior adjustments made
2075 * to the associated reservation map.
2076 * 2) Free any unused surplus pages that may have been allocated to satisfy
2077 * the reservation. As many as unused_resv_pages may be freed.
2079 * Called with hugetlb_lock held. However, the lock could be dropped (and
2080 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
2081 * we must make sure nobody else can claim pages we are in the process of
2082 * freeing. Do this by ensuring resv_huge_page always is greater than the
2083 * number of huge pages we plan to free when dropping the lock.
2085 static void return_unused_surplus_pages(struct hstate *h,
2086 unsigned long unused_resv_pages)
2088 unsigned long nr_pages;
2090 /* Cannot return gigantic pages currently */
2091 if (hstate_is_gigantic(h))
2095 * Part (or even all) of the reservation could have been backed
2096 * by pre-allocated pages. Only free surplus pages.
2098 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2101 * We want to release as many surplus pages as possible, spread
2102 * evenly across all nodes with memory. Iterate across these nodes
2103 * until we can no longer free unreserved surplus pages. This occurs
2104 * when the nodes with surplus pages have no free pages.
2105 * free_pool_huge_page() will balance the freed pages across the
2106 * on-line nodes with memory and will handle the hstate accounting.
2108 * Note that we decrement resv_huge_pages as we free the pages. If
2109 * we drop the lock, resv_huge_pages will still be sufficiently large
2110 * to cover subsequent pages we may free.
2112 while (nr_pages--) {
2113 h->resv_huge_pages--;
2114 unused_resv_pages--;
2115 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
2117 cond_resched_lock(&hugetlb_lock);
2121 /* Fully uncommit the reservation */
2122 h->resv_huge_pages -= unused_resv_pages;
2127 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2128 * are used by the huge page allocation routines to manage reservations.
2130 * vma_needs_reservation is called to determine if the huge page at addr
2131 * within the vma has an associated reservation. If a reservation is
2132 * needed, the value 1 is returned. The caller is then responsible for
2133 * managing the global reservation and subpool usage counts. After
2134 * the huge page has been allocated, vma_commit_reservation is called
2135 * to add the page to the reservation map. If the page allocation fails,
2136 * the reservation must be ended instead of committed. vma_end_reservation
2137 * is called in such cases.
2139 * In the normal case, vma_commit_reservation returns the same value
2140 * as the preceding vma_needs_reservation call. The only time this
2141 * is not the case is if a reserve map was changed between calls. It
2142 * is the responsibility of the caller to notice the difference and
2143 * take appropriate action.
2145 * vma_add_reservation is used in error paths where a reservation must
2146 * be restored when a newly allocated huge page must be freed. It is
2147 * to be called after calling vma_needs_reservation to determine if a
2148 * reservation exists.
2150 enum vma_resv_mode {
2156 static long __vma_reservation_common(struct hstate *h,
2157 struct vm_area_struct *vma, unsigned long addr,
2158 enum vma_resv_mode mode)
2160 struct resv_map *resv;
2163 long dummy_out_regions_needed;
2165 resv = vma_resv_map(vma);
2169 idx = vma_hugecache_offset(h, vma, addr);
2171 case VMA_NEEDS_RESV:
2172 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2173 /* We assume that vma_reservation_* routines always operate on
2174 * 1 page, and that adding to resv map a 1 page entry can only
2175 * ever require 1 region.
2177 VM_BUG_ON(dummy_out_regions_needed != 1);
2179 case VMA_COMMIT_RESV:
2180 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2181 /* region_add calls of range 1 should never fail. */
2185 region_abort(resv, idx, idx + 1, 1);
2189 if (vma->vm_flags & VM_MAYSHARE) {
2190 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2191 /* region_add calls of range 1 should never fail. */
2194 region_abort(resv, idx, idx + 1, 1);
2195 ret = region_del(resv, idx, idx + 1);
2202 if (vma->vm_flags & VM_MAYSHARE)
2205 * We know private mapping must have HPAGE_RESV_OWNER set.
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.
2226 static long vma_needs_reservation(struct hstate *h,
2227 struct vm_area_struct *vma, unsigned long addr)
2229 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2232 static long vma_commit_reservation(struct hstate *h,
2233 struct vm_area_struct *vma, unsigned long addr)
2235 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2238 static void vma_end_reservation(struct hstate *h,
2239 struct vm_area_struct *vma, unsigned long addr)
2241 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2244 static long vma_add_reservation(struct hstate *h,
2245 struct vm_area_struct *vma, unsigned long addr)
2247 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2251 * This routine is called to restore a reservation on error paths. In the
2252 * specific error paths, a huge page was allocated (via alloc_huge_page)
2253 * and is about to be freed. If a reservation for the page existed,
2254 * alloc_huge_page would have consumed the reservation and set
2255 * HPageRestoreReserve in the newly allocated page. When the page is freed
2256 * via free_huge_page, the global reservation count will be incremented if
2257 * HPageRestoreReserve is set. However, free_huge_page can not adjust the
2258 * reserve map. Adjust the reserve map here to be consistent with global
2259 * reserve count adjustments to be made by free_huge_page.
2261 static void restore_reserve_on_error(struct hstate *h,
2262 struct vm_area_struct *vma, unsigned long address,
2265 if (unlikely(HPageRestoreReserve(page))) {
2266 long rc = vma_needs_reservation(h, vma, address);
2268 if (unlikely(rc < 0)) {
2270 * Rare out of memory condition in reserve map
2271 * manipulation. Clear HPageRestoreReserve so that
2272 * global reserve count will not be incremented
2273 * by free_huge_page. This will make it appear
2274 * as though the reservation for this page was
2275 * consumed. This may prevent the task from
2276 * faulting in the page at a later time. This
2277 * is better than inconsistent global huge page
2278 * accounting of reserve counts.
2280 ClearHPageRestoreReserve(page);
2282 rc = vma_add_reservation(h, vma, address);
2283 if (unlikely(rc < 0))
2285 * See above comment about rare out of
2288 ClearHPageRestoreReserve(page);
2290 vma_end_reservation(h, vma, address);
2294 struct page *alloc_huge_page(struct vm_area_struct *vma,
2295 unsigned long addr, int avoid_reserve)
2297 struct hugepage_subpool *spool = subpool_vma(vma);
2298 struct hstate *h = hstate_vma(vma);
2300 long map_chg, map_commit;
2303 struct hugetlb_cgroup *h_cg;
2304 bool deferred_reserve;
2306 idx = hstate_index(h);
2308 * Examine the region/reserve map to determine if the process
2309 * has a reservation for the page to be allocated. A return
2310 * code of zero indicates a reservation exists (no change).
2312 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2314 return ERR_PTR(-ENOMEM);
2317 * Processes that did not create the mapping will have no
2318 * reserves as indicated by the region/reserve map. Check
2319 * that the allocation will not exceed the subpool limit.
2320 * Allocations for MAP_NORESERVE mappings also need to be
2321 * checked against any subpool limit.
2323 if (map_chg || avoid_reserve) {
2324 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2326 vma_end_reservation(h, vma, addr);
2327 return ERR_PTR(-ENOSPC);
2331 * Even though there was no reservation in the region/reserve
2332 * map, there could be reservations associated with the
2333 * subpool that can be used. This would be indicated if the
2334 * return value of hugepage_subpool_get_pages() is zero.
2335 * However, if avoid_reserve is specified we still avoid even
2336 * the subpool reservations.
2342 /* If this allocation is not consuming a reservation, charge it now.
2344 deferred_reserve = map_chg || avoid_reserve;
2345 if (deferred_reserve) {
2346 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2347 idx, pages_per_huge_page(h), &h_cg);
2349 goto out_subpool_put;
2352 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2354 goto out_uncharge_cgroup_reservation;
2356 spin_lock(&hugetlb_lock);
2358 * glb_chg is passed to indicate whether or not a page must be taken
2359 * from the global free pool (global change). gbl_chg == 0 indicates
2360 * a reservation exists for the allocation.
2362 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2364 spin_unlock(&hugetlb_lock);
2365 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2367 goto out_uncharge_cgroup;
2368 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2369 SetHPageRestoreReserve(page);
2370 h->resv_huge_pages--;
2372 spin_lock(&hugetlb_lock);
2373 list_add(&page->lru, &h->hugepage_activelist);
2376 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2377 /* If allocation is not consuming a reservation, also store the
2378 * hugetlb_cgroup pointer on the page.
2380 if (deferred_reserve) {
2381 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2385 spin_unlock(&hugetlb_lock);
2387 hugetlb_set_page_subpool(page, spool);
2389 map_commit = vma_commit_reservation(h, vma, addr);
2390 if (unlikely(map_chg > map_commit)) {
2392 * The page was added to the reservation map between
2393 * vma_needs_reservation and vma_commit_reservation.
2394 * This indicates a race with hugetlb_reserve_pages.
2395 * Adjust for the subpool count incremented above AND
2396 * in hugetlb_reserve_pages for the same page. Also,
2397 * the reservation count added in hugetlb_reserve_pages
2398 * no longer applies.
2402 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2403 hugetlb_acct_memory(h, -rsv_adjust);
2404 if (deferred_reserve)
2405 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2406 pages_per_huge_page(h), page);
2410 out_uncharge_cgroup:
2411 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2412 out_uncharge_cgroup_reservation:
2413 if (deferred_reserve)
2414 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2417 if (map_chg || avoid_reserve)
2418 hugepage_subpool_put_pages(spool, 1);
2419 vma_end_reservation(h, vma, addr);
2420 return ERR_PTR(-ENOSPC);
2423 int alloc_bootmem_huge_page(struct hstate *h)
2424 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2425 int __alloc_bootmem_huge_page(struct hstate *h)
2427 struct huge_bootmem_page *m;
2430 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2433 addr = memblock_alloc_try_nid_raw(
2434 huge_page_size(h), huge_page_size(h),
2435 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2438 * Use the beginning of the huge page to store the
2439 * huge_bootmem_page struct (until gather_bootmem
2440 * puts them into the mem_map).
2449 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2450 /* Put them into a private list first because mem_map is not up yet */
2451 INIT_LIST_HEAD(&m->list);
2452 list_add(&m->list, &huge_boot_pages);
2457 static void __init prep_compound_huge_page(struct page *page,
2460 if (unlikely(order > (MAX_ORDER - 1)))
2461 prep_compound_gigantic_page(page, order);
2463 prep_compound_page(page, order);
2466 /* Put bootmem huge pages into the standard lists after mem_map is up */
2467 static void __init gather_bootmem_prealloc(void)
2469 struct huge_bootmem_page *m;
2471 list_for_each_entry(m, &huge_boot_pages, list) {
2472 struct page *page = virt_to_page(m);
2473 struct hstate *h = m->hstate;
2475 WARN_ON(page_count(page) != 1);
2476 prep_compound_huge_page(page, huge_page_order(h));
2477 WARN_ON(PageReserved(page));
2478 prep_new_huge_page(h, page, page_to_nid(page));
2479 put_page(page); /* free it into the hugepage allocator */
2482 * If we had gigantic hugepages allocated at boot time, we need
2483 * to restore the 'stolen' pages to totalram_pages in order to
2484 * fix confusing memory reports from free(1) and another
2485 * side-effects, like CommitLimit going negative.
2487 if (hstate_is_gigantic(h))
2488 adjust_managed_page_count(page, pages_per_huge_page(h));
2493 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2496 nodemask_t *node_alloc_noretry;
2498 if (!hstate_is_gigantic(h)) {
2500 * Bit mask controlling how hard we retry per-node allocations.
2501 * Ignore errors as lower level routines can deal with
2502 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2503 * time, we are likely in bigger trouble.
2505 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2508 /* allocations done at boot time */
2509 node_alloc_noretry = NULL;
2512 /* bit mask controlling how hard we retry per-node allocations */
2513 if (node_alloc_noretry)
2514 nodes_clear(*node_alloc_noretry);
2516 for (i = 0; i < h->max_huge_pages; ++i) {
2517 if (hstate_is_gigantic(h)) {
2518 if (hugetlb_cma_size) {
2519 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2522 if (!alloc_bootmem_huge_page(h))
2524 } else if (!alloc_pool_huge_page(h,
2525 &node_states[N_MEMORY],
2526 node_alloc_noretry))
2530 if (i < h->max_huge_pages) {
2533 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2534 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2535 h->max_huge_pages, buf, i);
2536 h->max_huge_pages = i;
2539 kfree(node_alloc_noretry);
2542 static void __init hugetlb_init_hstates(void)
2546 for_each_hstate(h) {
2547 if (minimum_order > huge_page_order(h))
2548 minimum_order = huge_page_order(h);
2550 /* oversize hugepages were init'ed in early boot */
2551 if (!hstate_is_gigantic(h))
2552 hugetlb_hstate_alloc_pages(h);
2554 VM_BUG_ON(minimum_order == UINT_MAX);
2557 static void __init report_hugepages(void)
2561 for_each_hstate(h) {
2564 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2565 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2566 buf, h->free_huge_pages);
2570 #ifdef CONFIG_HIGHMEM
2571 static void try_to_free_low(struct hstate *h, unsigned long count,
2572 nodemask_t *nodes_allowed)
2575 struct page *page, *next;
2576 LIST_HEAD(page_list);
2578 if (hstate_is_gigantic(h))
2582 * Collect pages to be freed on a list, and free after dropping lock
2584 for_each_node_mask(i, *nodes_allowed) {
2585 struct list_head *freel = &h->hugepage_freelists[i];
2586 list_for_each_entry_safe(page, next, freel, lru) {
2587 if (count >= h->nr_huge_pages)
2589 if (PageHighMem(page))
2591 remove_hugetlb_page(h, page, false);
2592 list_add(&page->lru, &page_list);
2597 spin_unlock(&hugetlb_lock);
2598 list_for_each_entry_safe(page, next, &page_list, lru) {
2599 update_and_free_page(h, page);
2602 spin_lock(&hugetlb_lock);
2605 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2606 nodemask_t *nodes_allowed)
2612 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2613 * balanced by operating on them in a round-robin fashion.
2614 * Returns 1 if an adjustment was made.
2616 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2621 VM_BUG_ON(delta != -1 && delta != 1);
2624 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2625 if (h->surplus_huge_pages_node[node])
2629 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2630 if (h->surplus_huge_pages_node[node] <
2631 h->nr_huge_pages_node[node])
2638 h->surplus_huge_pages += delta;
2639 h->surplus_huge_pages_node[node] += delta;
2643 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2644 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2645 nodemask_t *nodes_allowed)
2647 unsigned long min_count, ret;
2648 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2651 * Bit mask controlling how hard we retry per-node allocations.
2652 * If we can not allocate the bit mask, do not attempt to allocate
2653 * the requested huge pages.
2655 if (node_alloc_noretry)
2656 nodes_clear(*node_alloc_noretry);
2661 * resize_lock mutex prevents concurrent adjustments to number of
2662 * pages in hstate via the proc/sysfs interfaces.
2664 mutex_lock(&h->resize_lock);
2665 spin_lock(&hugetlb_lock);
2668 * Check for a node specific request.
2669 * Changing node specific huge page count may require a corresponding
2670 * change to the global count. In any case, the passed node mask
2671 * (nodes_allowed) will restrict alloc/free to the specified node.
2673 if (nid != NUMA_NO_NODE) {
2674 unsigned long old_count = count;
2676 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2678 * User may have specified a large count value which caused the
2679 * above calculation to overflow. In this case, they wanted
2680 * to allocate as many huge pages as possible. Set count to
2681 * largest possible value to align with their intention.
2683 if (count < old_count)
2688 * Gigantic pages runtime allocation depend on the capability for large
2689 * page range allocation.
2690 * If the system does not provide this feature, return an error when
2691 * the user tries to allocate gigantic pages but let the user free the
2692 * boottime allocated gigantic pages.
2694 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2695 if (count > persistent_huge_pages(h)) {
2696 spin_unlock(&hugetlb_lock);
2697 mutex_unlock(&h->resize_lock);
2698 NODEMASK_FREE(node_alloc_noretry);
2701 /* Fall through to decrease pool */
2705 * Increase the pool size
2706 * First take pages out of surplus state. Then make up the
2707 * remaining difference by allocating fresh huge pages.
2709 * We might race with alloc_surplus_huge_page() here and be unable
2710 * to convert a surplus huge page to a normal huge page. That is
2711 * not critical, though, it just means the overall size of the
2712 * pool might be one hugepage larger than it needs to be, but
2713 * within all the constraints specified by the sysctls.
2715 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2716 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2720 while (count > persistent_huge_pages(h)) {
2722 * If this allocation races such that we no longer need the
2723 * page, free_huge_page will handle it by freeing the page
2724 * and reducing the surplus.
2726 spin_unlock(&hugetlb_lock);
2728 /* yield cpu to avoid soft lockup */
2731 ret = alloc_pool_huge_page(h, nodes_allowed,
2732 node_alloc_noretry);
2733 spin_lock(&hugetlb_lock);
2737 /* Bail for signals. Probably ctrl-c from user */
2738 if (signal_pending(current))
2743 * Decrease the pool size
2744 * First return free pages to the buddy allocator (being careful
2745 * to keep enough around to satisfy reservations). Then place
2746 * pages into surplus state as needed so the pool will shrink
2747 * to the desired size as pages become free.
2749 * By placing pages into the surplus state independent of the
2750 * overcommit value, we are allowing the surplus pool size to
2751 * exceed overcommit. There are few sane options here. Since
2752 * alloc_surplus_huge_page() is checking the global counter,
2753 * though, we'll note that we're not allowed to exceed surplus
2754 * and won't grow the pool anywhere else. Not until one of the
2755 * sysctls are changed, or the surplus pages go out of use.
2757 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2758 min_count = max(count, min_count);
2759 try_to_free_low(h, min_count, nodes_allowed);
2760 while (min_count < persistent_huge_pages(h)) {
2761 if (!free_pool_huge_page(h, nodes_allowed, 0))
2763 cond_resched_lock(&hugetlb_lock);
2765 while (count < persistent_huge_pages(h)) {
2766 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2770 h->max_huge_pages = persistent_huge_pages(h);
2771 spin_unlock(&hugetlb_lock);
2772 mutex_unlock(&h->resize_lock);
2774 NODEMASK_FREE(node_alloc_noretry);
2779 #define HSTATE_ATTR_RO(_name) \
2780 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2782 #define HSTATE_ATTR(_name) \
2783 static struct kobj_attribute _name##_attr = \
2784 __ATTR(_name, 0644, _name##_show, _name##_store)
2786 static struct kobject *hugepages_kobj;
2787 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2789 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2791 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2795 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2796 if (hstate_kobjs[i] == kobj) {
2798 *nidp = NUMA_NO_NODE;
2802 return kobj_to_node_hstate(kobj, nidp);
2805 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2806 struct kobj_attribute *attr, char *buf)
2809 unsigned long nr_huge_pages;
2812 h = kobj_to_hstate(kobj, &nid);
2813 if (nid == NUMA_NO_NODE)
2814 nr_huge_pages = h->nr_huge_pages;
2816 nr_huge_pages = h->nr_huge_pages_node[nid];
2818 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
2821 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2822 struct hstate *h, int nid,
2823 unsigned long count, size_t len)
2826 nodemask_t nodes_allowed, *n_mask;
2828 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2831 if (nid == NUMA_NO_NODE) {
2833 * global hstate attribute
2835 if (!(obey_mempolicy &&
2836 init_nodemask_of_mempolicy(&nodes_allowed)))
2837 n_mask = &node_states[N_MEMORY];
2839 n_mask = &nodes_allowed;
2842 * Node specific request. count adjustment happens in
2843 * set_max_huge_pages() after acquiring hugetlb_lock.
2845 init_nodemask_of_node(&nodes_allowed, nid);
2846 n_mask = &nodes_allowed;
2849 err = set_max_huge_pages(h, count, nid, n_mask);
2851 return err ? err : len;
2854 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2855 struct kobject *kobj, const char *buf,
2859 unsigned long count;
2863 err = kstrtoul(buf, 10, &count);
2867 h = kobj_to_hstate(kobj, &nid);
2868 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2871 static ssize_t nr_hugepages_show(struct kobject *kobj,
2872 struct kobj_attribute *attr, char *buf)
2874 return nr_hugepages_show_common(kobj, attr, buf);
2877 static ssize_t nr_hugepages_store(struct kobject *kobj,
2878 struct kobj_attribute *attr, const char *buf, size_t len)
2880 return nr_hugepages_store_common(false, kobj, buf, len);
2882 HSTATE_ATTR(nr_hugepages);
2887 * hstate attribute for optionally mempolicy-based constraint on persistent
2888 * huge page alloc/free.
2890 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2891 struct kobj_attribute *attr,
2894 return nr_hugepages_show_common(kobj, attr, buf);
2897 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2898 struct kobj_attribute *attr, const char *buf, size_t len)
2900 return nr_hugepages_store_common(true, kobj, buf, len);
2902 HSTATE_ATTR(nr_hugepages_mempolicy);
2906 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2907 struct kobj_attribute *attr, char *buf)
2909 struct hstate *h = kobj_to_hstate(kobj, NULL);
2910 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
2913 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2914 struct kobj_attribute *attr, const char *buf, size_t count)
2917 unsigned long input;
2918 struct hstate *h = kobj_to_hstate(kobj, NULL);
2920 if (hstate_is_gigantic(h))
2923 err = kstrtoul(buf, 10, &input);
2927 spin_lock(&hugetlb_lock);
2928 h->nr_overcommit_huge_pages = input;
2929 spin_unlock(&hugetlb_lock);
2933 HSTATE_ATTR(nr_overcommit_hugepages);
2935 static ssize_t free_hugepages_show(struct kobject *kobj,
2936 struct kobj_attribute *attr, char *buf)
2939 unsigned long free_huge_pages;
2942 h = kobj_to_hstate(kobj, &nid);
2943 if (nid == NUMA_NO_NODE)
2944 free_huge_pages = h->free_huge_pages;
2946 free_huge_pages = h->free_huge_pages_node[nid];
2948 return sysfs_emit(buf, "%lu\n", free_huge_pages);
2950 HSTATE_ATTR_RO(free_hugepages);
2952 static ssize_t resv_hugepages_show(struct kobject *kobj,
2953 struct kobj_attribute *attr, char *buf)
2955 struct hstate *h = kobj_to_hstate(kobj, NULL);
2956 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
2958 HSTATE_ATTR_RO(resv_hugepages);
2960 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2961 struct kobj_attribute *attr, char *buf)
2964 unsigned long surplus_huge_pages;
2967 h = kobj_to_hstate(kobj, &nid);
2968 if (nid == NUMA_NO_NODE)
2969 surplus_huge_pages = h->surplus_huge_pages;
2971 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2973 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
2975 HSTATE_ATTR_RO(surplus_hugepages);
2977 static struct attribute *hstate_attrs[] = {
2978 &nr_hugepages_attr.attr,
2979 &nr_overcommit_hugepages_attr.attr,
2980 &free_hugepages_attr.attr,
2981 &resv_hugepages_attr.attr,
2982 &surplus_hugepages_attr.attr,
2984 &nr_hugepages_mempolicy_attr.attr,
2989 static const struct attribute_group hstate_attr_group = {
2990 .attrs = hstate_attrs,
2993 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2994 struct kobject **hstate_kobjs,
2995 const struct attribute_group *hstate_attr_group)
2998 int hi = hstate_index(h);
3000 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3001 if (!hstate_kobjs[hi])
3004 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3006 kobject_put(hstate_kobjs[hi]);
3007 hstate_kobjs[hi] = NULL;
3013 static void __init hugetlb_sysfs_init(void)
3018 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3019 if (!hugepages_kobj)
3022 for_each_hstate(h) {
3023 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3024 hstate_kobjs, &hstate_attr_group);
3026 pr_err("HugeTLB: Unable to add hstate %s", h->name);
3033 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3034 * with node devices in node_devices[] using a parallel array. The array
3035 * index of a node device or _hstate == node id.
3036 * This is here to avoid any static dependency of the node device driver, in
3037 * the base kernel, on the hugetlb module.
3039 struct node_hstate {
3040 struct kobject *hugepages_kobj;
3041 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3043 static struct node_hstate node_hstates[MAX_NUMNODES];
3046 * A subset of global hstate attributes for node devices
3048 static struct attribute *per_node_hstate_attrs[] = {
3049 &nr_hugepages_attr.attr,
3050 &free_hugepages_attr.attr,
3051 &surplus_hugepages_attr.attr,
3055 static const struct attribute_group per_node_hstate_attr_group = {
3056 .attrs = per_node_hstate_attrs,
3060 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3061 * Returns node id via non-NULL nidp.
3063 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3067 for (nid = 0; nid < nr_node_ids; nid++) {
3068 struct node_hstate *nhs = &node_hstates[nid];
3070 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3071 if (nhs->hstate_kobjs[i] == kobj) {
3083 * Unregister hstate attributes from a single node device.
3084 * No-op if no hstate attributes attached.
3086 static void hugetlb_unregister_node(struct node *node)
3089 struct node_hstate *nhs = &node_hstates[node->dev.id];
3091 if (!nhs->hugepages_kobj)
3092 return; /* no hstate attributes */
3094 for_each_hstate(h) {
3095 int idx = hstate_index(h);
3096 if (nhs->hstate_kobjs[idx]) {
3097 kobject_put(nhs->hstate_kobjs[idx]);
3098 nhs->hstate_kobjs[idx] = NULL;
3102 kobject_put(nhs->hugepages_kobj);
3103 nhs->hugepages_kobj = NULL;
3108 * Register hstate attributes for a single node device.
3109 * No-op if attributes already registered.
3111 static void hugetlb_register_node(struct node *node)
3114 struct node_hstate *nhs = &node_hstates[node->dev.id];
3117 if (nhs->hugepages_kobj)
3118 return; /* already allocated */
3120 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3122 if (!nhs->hugepages_kobj)
3125 for_each_hstate(h) {
3126 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3128 &per_node_hstate_attr_group);
3130 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3131 h->name, node->dev.id);
3132 hugetlb_unregister_node(node);
3139 * hugetlb init time: register hstate attributes for all registered node
3140 * devices of nodes that have memory. All on-line nodes should have
3141 * registered their associated device by this time.
3143 static void __init hugetlb_register_all_nodes(void)
3147 for_each_node_state(nid, N_MEMORY) {
3148 struct node *node = node_devices[nid];
3149 if (node->dev.id == nid)
3150 hugetlb_register_node(node);
3154 * Let the node device driver know we're here so it can
3155 * [un]register hstate attributes on node hotplug.
3157 register_hugetlbfs_with_node(hugetlb_register_node,
3158 hugetlb_unregister_node);
3160 #else /* !CONFIG_NUMA */
3162 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3170 static void hugetlb_register_all_nodes(void) { }
3174 static int __init hugetlb_init(void)
3178 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
3181 if (!hugepages_supported()) {
3182 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3183 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3188 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3189 * architectures depend on setup being done here.
3191 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3192 if (!parsed_default_hugepagesz) {
3194 * If we did not parse a default huge page size, set
3195 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3196 * number of huge pages for this default size was implicitly
3197 * specified, set that here as well.
3198 * Note that the implicit setting will overwrite an explicit
3199 * setting. A warning will be printed in this case.
3201 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3202 if (default_hstate_max_huge_pages) {
3203 if (default_hstate.max_huge_pages) {
3206 string_get_size(huge_page_size(&default_hstate),
3207 1, STRING_UNITS_2, buf, 32);
3208 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3209 default_hstate.max_huge_pages, buf);
3210 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3211 default_hstate_max_huge_pages);
3213 default_hstate.max_huge_pages =
3214 default_hstate_max_huge_pages;
3218 hugetlb_cma_check();
3219 hugetlb_init_hstates();
3220 gather_bootmem_prealloc();
3223 hugetlb_sysfs_init();
3224 hugetlb_register_all_nodes();
3225 hugetlb_cgroup_file_init();
3228 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3230 num_fault_mutexes = 1;
3232 hugetlb_fault_mutex_table =
3233 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3235 BUG_ON(!hugetlb_fault_mutex_table);
3237 for (i = 0; i < num_fault_mutexes; i++)
3238 mutex_init(&hugetlb_fault_mutex_table[i]);
3241 subsys_initcall(hugetlb_init);
3243 /* Overwritten by architectures with more huge page sizes */
3244 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3246 return size == HPAGE_SIZE;
3249 void __init hugetlb_add_hstate(unsigned int order)
3254 if (size_to_hstate(PAGE_SIZE << order)) {
3257 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3259 h = &hstates[hugetlb_max_hstate++];
3260 mutex_init(&h->resize_lock);
3262 h->mask = ~(huge_page_size(h) - 1);
3263 for (i = 0; i < MAX_NUMNODES; ++i)
3264 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3265 INIT_LIST_HEAD(&h->hugepage_activelist);
3266 h->next_nid_to_alloc = first_memory_node;
3267 h->next_nid_to_free = first_memory_node;
3268 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3269 huge_page_size(h)/1024);
3275 * hugepages command line processing
3276 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3277 * specification. If not, ignore the hugepages value. hugepages can also
3278 * be the first huge page command line option in which case it implicitly
3279 * specifies the number of huge pages for the default size.
3281 static int __init hugepages_setup(char *s)
3284 static unsigned long *last_mhp;
3286 if (!parsed_valid_hugepagesz) {
3287 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3288 parsed_valid_hugepagesz = true;
3293 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3294 * yet, so this hugepages= parameter goes to the "default hstate".
3295 * Otherwise, it goes with the previously parsed hugepagesz or
3296 * default_hugepagesz.
3298 else if (!hugetlb_max_hstate)
3299 mhp = &default_hstate_max_huge_pages;
3301 mhp = &parsed_hstate->max_huge_pages;
3303 if (mhp == last_mhp) {
3304 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3308 if (sscanf(s, "%lu", mhp) <= 0)
3312 * Global state is always initialized later in hugetlb_init.
3313 * But we need to allocate gigantic hstates here early to still
3314 * use the bootmem allocator.
3316 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
3317 hugetlb_hstate_alloc_pages(parsed_hstate);
3323 __setup("hugepages=", hugepages_setup);
3326 * hugepagesz command line processing
3327 * A specific huge page size can only be specified once with hugepagesz.
3328 * hugepagesz is followed by hugepages on the command line. The global
3329 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3330 * hugepagesz argument was valid.
3332 static int __init hugepagesz_setup(char *s)
3337 parsed_valid_hugepagesz = false;
3338 size = (unsigned long)memparse(s, NULL);
3340 if (!arch_hugetlb_valid_size(size)) {
3341 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3345 h = size_to_hstate(size);
3348 * hstate for this size already exists. This is normally
3349 * an error, but is allowed if the existing hstate is the
3350 * default hstate. More specifically, it is only allowed if
3351 * the number of huge pages for the default hstate was not
3352 * previously specified.
3354 if (!parsed_default_hugepagesz || h != &default_hstate ||
3355 default_hstate.max_huge_pages) {
3356 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3361 * No need to call hugetlb_add_hstate() as hstate already
3362 * exists. But, do set parsed_hstate so that a following
3363 * hugepages= parameter will be applied to this hstate.
3366 parsed_valid_hugepagesz = true;
3370 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3371 parsed_valid_hugepagesz = true;
3374 __setup("hugepagesz=", hugepagesz_setup);
3377 * default_hugepagesz command line input
3378 * Only one instance of default_hugepagesz allowed on command line.
3380 static int __init default_hugepagesz_setup(char *s)
3384 parsed_valid_hugepagesz = false;
3385 if (parsed_default_hugepagesz) {
3386 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3390 size = (unsigned long)memparse(s, NULL);
3392 if (!arch_hugetlb_valid_size(size)) {
3393 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3397 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3398 parsed_valid_hugepagesz = true;
3399 parsed_default_hugepagesz = true;
3400 default_hstate_idx = hstate_index(size_to_hstate(size));
3403 * The number of default huge pages (for this size) could have been
3404 * specified as the first hugetlb parameter: hugepages=X. If so,
3405 * then default_hstate_max_huge_pages is set. If the default huge
3406 * page size is gigantic (>= MAX_ORDER), then the pages must be
3407 * allocated here from bootmem allocator.
3409 if (default_hstate_max_huge_pages) {
3410 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3411 if (hstate_is_gigantic(&default_hstate))
3412 hugetlb_hstate_alloc_pages(&default_hstate);
3413 default_hstate_max_huge_pages = 0;
3418 __setup("default_hugepagesz=", default_hugepagesz_setup);
3420 static unsigned int allowed_mems_nr(struct hstate *h)
3423 unsigned int nr = 0;
3424 nodemask_t *mpol_allowed;
3425 unsigned int *array = h->free_huge_pages_node;
3426 gfp_t gfp_mask = htlb_alloc_mask(h);
3428 mpol_allowed = policy_nodemask_current(gfp_mask);
3430 for_each_node_mask(node, cpuset_current_mems_allowed) {
3431 if (!mpol_allowed || node_isset(node, *mpol_allowed))
3438 #ifdef CONFIG_SYSCTL
3439 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3440 void *buffer, size_t *length,
3441 loff_t *ppos, unsigned long *out)
3443 struct ctl_table dup_table;
3446 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3447 * can duplicate the @table and alter the duplicate of it.
3450 dup_table.data = out;
3452 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3455 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3456 struct ctl_table *table, int write,
3457 void *buffer, size_t *length, loff_t *ppos)
3459 struct hstate *h = &default_hstate;
3460 unsigned long tmp = h->max_huge_pages;
3463 if (!hugepages_supported())
3466 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3472 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3473 NUMA_NO_NODE, tmp, *length);
3478 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3479 void *buffer, size_t *length, loff_t *ppos)
3482 return hugetlb_sysctl_handler_common(false, table, write,
3483 buffer, length, ppos);
3487 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3488 void *buffer, size_t *length, loff_t *ppos)
3490 return hugetlb_sysctl_handler_common(true, table, write,
3491 buffer, length, ppos);
3493 #endif /* CONFIG_NUMA */
3495 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3496 void *buffer, size_t *length, loff_t *ppos)
3498 struct hstate *h = &default_hstate;
3502 if (!hugepages_supported())
3505 tmp = h->nr_overcommit_huge_pages;
3507 if (write && hstate_is_gigantic(h))
3510 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3516 spin_lock(&hugetlb_lock);
3517 h->nr_overcommit_huge_pages = tmp;
3518 spin_unlock(&hugetlb_lock);
3524 #endif /* CONFIG_SYSCTL */
3526 void hugetlb_report_meminfo(struct seq_file *m)
3529 unsigned long total = 0;
3531 if (!hugepages_supported())
3534 for_each_hstate(h) {
3535 unsigned long count = h->nr_huge_pages;
3537 total += huge_page_size(h) * count;
3539 if (h == &default_hstate)
3541 "HugePages_Total: %5lu\n"
3542 "HugePages_Free: %5lu\n"
3543 "HugePages_Rsvd: %5lu\n"
3544 "HugePages_Surp: %5lu\n"
3545 "Hugepagesize: %8lu kB\n",
3549 h->surplus_huge_pages,
3550 huge_page_size(h) / SZ_1K);
3553 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
3556 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
3558 struct hstate *h = &default_hstate;
3560 if (!hugepages_supported())
3563 return sysfs_emit_at(buf, len,
3564 "Node %d HugePages_Total: %5u\n"
3565 "Node %d HugePages_Free: %5u\n"
3566 "Node %d HugePages_Surp: %5u\n",
3567 nid, h->nr_huge_pages_node[nid],
3568 nid, h->free_huge_pages_node[nid],
3569 nid, h->surplus_huge_pages_node[nid]);
3572 void hugetlb_show_meminfo(void)
3577 if (!hugepages_supported())
3580 for_each_node_state(nid, N_MEMORY)
3582 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3584 h->nr_huge_pages_node[nid],
3585 h->free_huge_pages_node[nid],
3586 h->surplus_huge_pages_node[nid],
3587 huge_page_size(h) / SZ_1K);
3590 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3592 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3593 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3596 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3597 unsigned long hugetlb_total_pages(void)
3600 unsigned long nr_total_pages = 0;
3603 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3604 return nr_total_pages;
3607 static int hugetlb_acct_memory(struct hstate *h, long delta)
3614 spin_lock(&hugetlb_lock);
3616 * When cpuset is configured, it breaks the strict hugetlb page
3617 * reservation as the accounting is done on a global variable. Such
3618 * reservation is completely rubbish in the presence of cpuset because
3619 * the reservation is not checked against page availability for the
3620 * current cpuset. Application can still potentially OOM'ed by kernel
3621 * with lack of free htlb page in cpuset that the task is in.
3622 * Attempt to enforce strict accounting with cpuset is almost
3623 * impossible (or too ugly) because cpuset is too fluid that
3624 * task or memory node can be dynamically moved between cpusets.
3626 * The change of semantics for shared hugetlb mapping with cpuset is
3627 * undesirable. However, in order to preserve some of the semantics,
3628 * we fall back to check against current free page availability as
3629 * a best attempt and hopefully to minimize the impact of changing
3630 * semantics that cpuset has.
3632 * Apart from cpuset, we also have memory policy mechanism that
3633 * also determines from which node the kernel will allocate memory
3634 * in a NUMA system. So similar to cpuset, we also should consider
3635 * the memory policy of the current task. Similar to the description
3639 if (gather_surplus_pages(h, delta) < 0)
3642 if (delta > allowed_mems_nr(h)) {
3643 return_unused_surplus_pages(h, delta);
3650 return_unused_surplus_pages(h, (unsigned long) -delta);
3653 spin_unlock(&hugetlb_lock);
3657 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3659 struct resv_map *resv = vma_resv_map(vma);
3662 * This new VMA should share its siblings reservation map if present.
3663 * The VMA will only ever have a valid reservation map pointer where
3664 * it is being copied for another still existing VMA. As that VMA
3665 * has a reference to the reservation map it cannot disappear until
3666 * after this open call completes. It is therefore safe to take a
3667 * new reference here without additional locking.
3669 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3670 kref_get(&resv->refs);
3673 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3675 struct hstate *h = hstate_vma(vma);
3676 struct resv_map *resv = vma_resv_map(vma);
3677 struct hugepage_subpool *spool = subpool_vma(vma);
3678 unsigned long reserve, start, end;
3681 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3684 start = vma_hugecache_offset(h, vma, vma->vm_start);
3685 end = vma_hugecache_offset(h, vma, vma->vm_end);
3687 reserve = (end - start) - region_count(resv, start, end);
3688 hugetlb_cgroup_uncharge_counter(resv, start, end);
3691 * Decrement reserve counts. The global reserve count may be
3692 * adjusted if the subpool has a minimum size.
3694 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3695 hugetlb_acct_memory(h, -gbl_reserve);
3698 kref_put(&resv->refs, resv_map_release);
3701 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3703 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3708 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3710 return huge_page_size(hstate_vma(vma));
3714 * We cannot handle pagefaults against hugetlb pages at all. They cause
3715 * handle_mm_fault() to try to instantiate regular-sized pages in the
3716 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3719 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3726 * When a new function is introduced to vm_operations_struct and added
3727 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3728 * This is because under System V memory model, mappings created via
3729 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3730 * their original vm_ops are overwritten with shm_vm_ops.
3732 const struct vm_operations_struct hugetlb_vm_ops = {
3733 .fault = hugetlb_vm_op_fault,
3734 .open = hugetlb_vm_op_open,
3735 .close = hugetlb_vm_op_close,
3736 .may_split = hugetlb_vm_op_split,
3737 .pagesize = hugetlb_vm_op_pagesize,
3740 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3746 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3747 vma->vm_page_prot)));
3749 entry = huge_pte_wrprotect(mk_huge_pte(page,
3750 vma->vm_page_prot));
3752 entry = pte_mkyoung(entry);
3753 entry = pte_mkhuge(entry);
3754 entry = arch_make_huge_pte(entry, vma, page, writable);
3759 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3760 unsigned long address, pte_t *ptep)
3764 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3765 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3766 update_mmu_cache(vma, address, ptep);
3769 bool is_hugetlb_entry_migration(pte_t pte)
3773 if (huge_pte_none(pte) || pte_present(pte))
3775 swp = pte_to_swp_entry(pte);
3776 if (is_migration_entry(swp))
3782 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
3786 if (huge_pte_none(pte) || pte_present(pte))
3788 swp = pte_to_swp_entry(pte);
3789 if (is_hwpoison_entry(swp))
3796 hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
3797 struct page *new_page)
3799 __SetPageUptodate(new_page);
3800 set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
3801 hugepage_add_new_anon_rmap(new_page, vma, addr);
3802 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
3803 ClearHPageRestoreReserve(new_page);
3804 SetHPageMigratable(new_page);
3807 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3808 struct vm_area_struct *vma)
3810 pte_t *src_pte, *dst_pte, entry, dst_entry;
3811 struct page *ptepage;
3813 bool cow = is_cow_mapping(vma->vm_flags);
3814 struct hstate *h = hstate_vma(vma);
3815 unsigned long sz = huge_page_size(h);
3816 unsigned long npages = pages_per_huge_page(h);
3817 struct address_space *mapping = vma->vm_file->f_mapping;
3818 struct mmu_notifier_range range;
3822 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3825 mmu_notifier_invalidate_range_start(&range);
3828 * For shared mappings i_mmap_rwsem must be held to call
3829 * huge_pte_alloc, otherwise the returned ptep could go
3830 * away if part of a shared pmd and another thread calls
3833 i_mmap_lock_read(mapping);
3836 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3837 spinlock_t *src_ptl, *dst_ptl;
3838 src_pte = huge_pte_offset(src, addr, sz);
3841 dst_pte = huge_pte_alloc(dst, vma, addr, sz);
3848 * If the pagetables are shared don't copy or take references.
3849 * dst_pte == src_pte is the common case of src/dest sharing.
3851 * However, src could have 'unshared' and dst shares with
3852 * another vma. If dst_pte !none, this implies sharing.
3853 * Check here before taking page table lock, and once again
3854 * after taking the lock below.
3856 dst_entry = huge_ptep_get(dst_pte);
3857 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3860 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3861 src_ptl = huge_pte_lockptr(h, src, src_pte);
3862 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3863 entry = huge_ptep_get(src_pte);
3864 dst_entry = huge_ptep_get(dst_pte);
3866 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3868 * Skip if src entry none. Also, skip in the
3869 * unlikely case dst entry !none as this implies
3870 * sharing with another vma.
3873 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3874 is_hugetlb_entry_hwpoisoned(entry))) {
3875 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3877 if (is_write_migration_entry(swp_entry) && cow) {
3879 * COW mappings require pages in both
3880 * parent and child to be set to read.
3882 make_migration_entry_read(&swp_entry);
3883 entry = swp_entry_to_pte(swp_entry);
3884 set_huge_swap_pte_at(src, addr, src_pte,
3887 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3889 entry = huge_ptep_get(src_pte);
3890 ptepage = pte_page(entry);
3894 * This is a rare case where we see pinned hugetlb
3895 * pages while they're prone to COW. We need to do the
3896 * COW earlier during fork.
3898 * When pre-allocating the page or copying data, we
3899 * need to be without the pgtable locks since we could
3900 * sleep during the process.
3902 if (unlikely(page_needs_cow_for_dma(vma, ptepage))) {
3903 pte_t src_pte_old = entry;
3906 spin_unlock(src_ptl);
3907 spin_unlock(dst_ptl);
3908 /* Do not use reserve as it's private owned */
3909 new = alloc_huge_page(vma, addr, 1);
3915 copy_user_huge_page(new, ptepage, addr, vma,
3919 /* Install the new huge page if src pte stable */
3920 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3921 src_ptl = huge_pte_lockptr(h, src, src_pte);
3922 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3923 entry = huge_ptep_get(src_pte);
3924 if (!pte_same(src_pte_old, entry)) {
3926 /* dst_entry won't change as in child */
3929 hugetlb_install_page(vma, dst_pte, addr, new);
3930 spin_unlock(src_ptl);
3931 spin_unlock(dst_ptl);
3937 * No need to notify as we are downgrading page
3938 * table protection not changing it to point
3941 * See Documentation/vm/mmu_notifier.rst
3943 huge_ptep_set_wrprotect(src, addr, src_pte);
3946 page_dup_rmap(ptepage, true);
3947 set_huge_pte_at(dst, addr, dst_pte, entry);
3948 hugetlb_count_add(npages, dst);
3950 spin_unlock(src_ptl);
3951 spin_unlock(dst_ptl);
3955 mmu_notifier_invalidate_range_end(&range);
3957 i_mmap_unlock_read(mapping);
3962 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3963 unsigned long start, unsigned long end,
3964 struct page *ref_page)
3966 struct mm_struct *mm = vma->vm_mm;
3967 unsigned long address;
3972 struct hstate *h = hstate_vma(vma);
3973 unsigned long sz = huge_page_size(h);
3974 struct mmu_notifier_range range;
3976 WARN_ON(!is_vm_hugetlb_page(vma));
3977 BUG_ON(start & ~huge_page_mask(h));
3978 BUG_ON(end & ~huge_page_mask(h));
3981 * This is a hugetlb vma, all the pte entries should point
3984 tlb_change_page_size(tlb, sz);
3985 tlb_start_vma(tlb, vma);
3988 * If sharing possible, alert mmu notifiers of worst case.
3990 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3992 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3993 mmu_notifier_invalidate_range_start(&range);
3995 for (; address < end; address += sz) {
3996 ptep = huge_pte_offset(mm, address, sz);
4000 ptl = huge_pte_lock(h, mm, ptep);
4001 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
4004 * We just unmapped a page of PMDs by clearing a PUD.
4005 * The caller's TLB flush range should cover this area.
4010 pte = huge_ptep_get(ptep);
4011 if (huge_pte_none(pte)) {
4017 * Migrating hugepage or HWPoisoned hugepage is already
4018 * unmapped and its refcount is dropped, so just clear pte here.
4020 if (unlikely(!pte_present(pte))) {
4021 huge_pte_clear(mm, address, ptep, sz);
4026 page = pte_page(pte);
4028 * If a reference page is supplied, it is because a specific
4029 * page is being unmapped, not a range. Ensure the page we
4030 * are about to unmap is the actual page of interest.
4033 if (page != ref_page) {
4038 * Mark the VMA as having unmapped its page so that
4039 * future faults in this VMA will fail rather than
4040 * looking like data was lost
4042 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
4045 pte = huge_ptep_get_and_clear(mm, address, ptep);
4046 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
4047 if (huge_pte_dirty(pte))
4048 set_page_dirty(page);
4050 hugetlb_count_sub(pages_per_huge_page(h), mm);
4051 page_remove_rmap(page, true);
4054 tlb_remove_page_size(tlb, page, huge_page_size(h));
4056 * Bail out after unmapping reference page if supplied
4061 mmu_notifier_invalidate_range_end(&range);
4062 tlb_end_vma(tlb, vma);
4065 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4066 struct vm_area_struct *vma, unsigned long start,
4067 unsigned long end, struct page *ref_page)
4069 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4072 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4073 * test will fail on a vma being torn down, and not grab a page table
4074 * on its way out. We're lucky that the flag has such an appropriate
4075 * name, and can in fact be safely cleared here. We could clear it
4076 * before the __unmap_hugepage_range above, but all that's necessary
4077 * is to clear it before releasing the i_mmap_rwsem. This works
4078 * because in the context this is called, the VMA is about to be
4079 * destroyed and the i_mmap_rwsem is held.
4081 vma->vm_flags &= ~VM_MAYSHARE;
4084 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
4085 unsigned long end, struct page *ref_page)
4087 struct mmu_gather tlb;
4089 tlb_gather_mmu(&tlb, vma->vm_mm);
4090 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
4091 tlb_finish_mmu(&tlb);
4095 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4096 * mapping it owns the reserve page for. The intention is to unmap the page
4097 * from other VMAs and let the children be SIGKILLed if they are faulting the
4100 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4101 struct page *page, unsigned long address)
4103 struct hstate *h = hstate_vma(vma);
4104 struct vm_area_struct *iter_vma;
4105 struct address_space *mapping;
4109 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4110 * from page cache lookup which is in HPAGE_SIZE units.
4112 address = address & huge_page_mask(h);
4113 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4115 mapping = vma->vm_file->f_mapping;
4118 * Take the mapping lock for the duration of the table walk. As
4119 * this mapping should be shared between all the VMAs,
4120 * __unmap_hugepage_range() is called as the lock is already held
4122 i_mmap_lock_write(mapping);
4123 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
4124 /* Do not unmap the current VMA */
4125 if (iter_vma == vma)
4129 * Shared VMAs have their own reserves and do not affect
4130 * MAP_PRIVATE accounting but it is possible that a shared
4131 * VMA is using the same page so check and skip such VMAs.
4133 if (iter_vma->vm_flags & VM_MAYSHARE)
4137 * Unmap the page from other VMAs without their own reserves.
4138 * They get marked to be SIGKILLed if they fault in these
4139 * areas. This is because a future no-page fault on this VMA
4140 * could insert a zeroed page instead of the data existing
4141 * from the time of fork. This would look like data corruption
4143 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
4144 unmap_hugepage_range(iter_vma, address,
4145 address + huge_page_size(h), page);
4147 i_mmap_unlock_write(mapping);
4151 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4152 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4153 * cannot race with other handlers or page migration.
4154 * Keep the pte_same checks anyway to make transition from the mutex easier.
4156 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
4157 unsigned long address, pte_t *ptep,
4158 struct page *pagecache_page, spinlock_t *ptl)
4161 struct hstate *h = hstate_vma(vma);
4162 struct page *old_page, *new_page;
4163 int outside_reserve = 0;
4165 unsigned long haddr = address & huge_page_mask(h);
4166 struct mmu_notifier_range range;
4168 pte = huge_ptep_get(ptep);
4169 old_page = pte_page(pte);
4172 /* If no-one else is actually using this page, avoid the copy
4173 * and just make the page writable */
4174 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
4175 page_move_anon_rmap(old_page, vma);
4176 set_huge_ptep_writable(vma, haddr, ptep);
4181 * If the process that created a MAP_PRIVATE mapping is about to
4182 * perform a COW due to a shared page count, attempt to satisfy
4183 * the allocation without using the existing reserves. The pagecache
4184 * page is used to determine if the reserve at this address was
4185 * consumed or not. If reserves were used, a partial faulted mapping
4186 * at the time of fork() could consume its reserves on COW instead
4187 * of the full address range.
4189 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4190 old_page != pagecache_page)
4191 outside_reserve = 1;
4196 * Drop page table lock as buddy allocator may be called. It will
4197 * be acquired again before returning to the caller, as expected.
4200 new_page = alloc_huge_page(vma, haddr, outside_reserve);
4202 if (IS_ERR(new_page)) {
4204 * If a process owning a MAP_PRIVATE mapping fails to COW,
4205 * it is due to references held by a child and an insufficient
4206 * huge page pool. To guarantee the original mappers
4207 * reliability, unmap the page from child processes. The child
4208 * may get SIGKILLed if it later faults.
4210 if (outside_reserve) {
4211 struct address_space *mapping = vma->vm_file->f_mapping;
4216 BUG_ON(huge_pte_none(pte));
4218 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4219 * unmapping. unmapping needs to hold i_mmap_rwsem
4220 * in write mode. Dropping i_mmap_rwsem in read mode
4221 * here is OK as COW mappings do not interact with
4224 * Reacquire both after unmap operation.
4226 idx = vma_hugecache_offset(h, vma, haddr);
4227 hash = hugetlb_fault_mutex_hash(mapping, idx);
4228 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4229 i_mmap_unlock_read(mapping);
4231 unmap_ref_private(mm, vma, old_page, haddr);
4233 i_mmap_lock_read(mapping);
4234 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4236 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4238 pte_same(huge_ptep_get(ptep), pte)))
4239 goto retry_avoidcopy;
4241 * race occurs while re-acquiring page table
4242 * lock, and our job is done.
4247 ret = vmf_error(PTR_ERR(new_page));
4248 goto out_release_old;
4252 * When the original hugepage is shared one, it does not have
4253 * anon_vma prepared.
4255 if (unlikely(anon_vma_prepare(vma))) {
4257 goto out_release_all;
4260 copy_user_huge_page(new_page, old_page, address, vma,
4261 pages_per_huge_page(h));
4262 __SetPageUptodate(new_page);
4264 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4265 haddr + huge_page_size(h));
4266 mmu_notifier_invalidate_range_start(&range);
4269 * Retake the page table lock to check for racing updates
4270 * before the page tables are altered
4273 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4274 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4275 ClearHPageRestoreReserve(new_page);
4278 huge_ptep_clear_flush(vma, haddr, ptep);
4279 mmu_notifier_invalidate_range(mm, range.start, range.end);
4280 set_huge_pte_at(mm, haddr, ptep,
4281 make_huge_pte(vma, new_page, 1));
4282 page_remove_rmap(old_page, true);
4283 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4284 SetHPageMigratable(new_page);
4285 /* Make the old page be freed below */
4286 new_page = old_page;
4289 mmu_notifier_invalidate_range_end(&range);
4291 restore_reserve_on_error(h, vma, haddr, new_page);
4296 spin_lock(ptl); /* Caller expects lock to be held */
4300 /* Return the pagecache page at a given address within a VMA */
4301 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4302 struct vm_area_struct *vma, unsigned long address)
4304 struct address_space *mapping;
4307 mapping = vma->vm_file->f_mapping;
4308 idx = vma_hugecache_offset(h, vma, address);
4310 return find_lock_page(mapping, idx);
4314 * Return whether there is a pagecache page to back given address within VMA.
4315 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4317 static bool hugetlbfs_pagecache_present(struct hstate *h,
4318 struct vm_area_struct *vma, unsigned long address)
4320 struct address_space *mapping;
4324 mapping = vma->vm_file->f_mapping;
4325 idx = vma_hugecache_offset(h, vma, address);
4327 page = find_get_page(mapping, idx);
4330 return page != NULL;
4333 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4336 struct inode *inode = mapping->host;
4337 struct hstate *h = hstate_inode(inode);
4338 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4342 ClearHPageRestoreReserve(page);
4345 * set page dirty so that it will not be removed from cache/file
4346 * by non-hugetlbfs specific code paths.
4348 set_page_dirty(page);
4350 spin_lock(&inode->i_lock);
4351 inode->i_blocks += blocks_per_huge_page(h);
4352 spin_unlock(&inode->i_lock);
4356 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4357 struct vm_area_struct *vma,
4358 struct address_space *mapping, pgoff_t idx,
4359 unsigned long address, pte_t *ptep, unsigned int flags)
4361 struct hstate *h = hstate_vma(vma);
4362 vm_fault_t ret = VM_FAULT_SIGBUS;
4368 unsigned long haddr = address & huge_page_mask(h);
4369 bool new_page = false;
4372 * Currently, we are forced to kill the process in the event the
4373 * original mapper has unmapped pages from the child due to a failed
4374 * COW. Warn that such a situation has occurred as it may not be obvious
4376 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4377 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4383 * We can not race with truncation due to holding i_mmap_rwsem.
4384 * i_size is modified when holding i_mmap_rwsem, so check here
4385 * once for faults beyond end of file.
4387 size = i_size_read(mapping->host) >> huge_page_shift(h);
4392 page = find_lock_page(mapping, idx);
4395 * Check for page in userfault range
4397 if (userfaultfd_missing(vma)) {
4399 struct vm_fault vmf = {
4404 * Hard to debug if it ends up being
4405 * used by a callee that assumes
4406 * something about the other
4407 * uninitialized fields... same as in
4413 * hugetlb_fault_mutex and i_mmap_rwsem must be
4414 * dropped before handling userfault. Reacquire
4415 * after handling fault to make calling code simpler.
4417 hash = hugetlb_fault_mutex_hash(mapping, idx);
4418 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4419 i_mmap_unlock_read(mapping);
4420 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
4421 i_mmap_lock_read(mapping);
4422 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4426 page = alloc_huge_page(vma, haddr, 0);
4429 * Returning error will result in faulting task being
4430 * sent SIGBUS. The hugetlb fault mutex prevents two
4431 * tasks from racing to fault in the same page which
4432 * could result in false unable to allocate errors.
4433 * Page migration does not take the fault mutex, but
4434 * does a clear then write of pte's under page table
4435 * lock. Page fault code could race with migration,
4436 * notice the clear pte and try to allocate a page
4437 * here. Before returning error, get ptl and make
4438 * sure there really is no pte entry.
4440 ptl = huge_pte_lock(h, mm, ptep);
4442 if (huge_pte_none(huge_ptep_get(ptep)))
4443 ret = vmf_error(PTR_ERR(page));
4447 clear_huge_page(page, address, pages_per_huge_page(h));
4448 __SetPageUptodate(page);
4451 if (vma->vm_flags & VM_MAYSHARE) {
4452 int err = huge_add_to_page_cache(page, mapping, idx);
4461 if (unlikely(anon_vma_prepare(vma))) {
4463 goto backout_unlocked;
4469 * If memory error occurs between mmap() and fault, some process
4470 * don't have hwpoisoned swap entry for errored virtual address.
4471 * So we need to block hugepage fault by PG_hwpoison bit check.
4473 if (unlikely(PageHWPoison(page))) {
4474 ret = VM_FAULT_HWPOISON_LARGE |
4475 VM_FAULT_SET_HINDEX(hstate_index(h));
4476 goto backout_unlocked;
4481 * If we are going to COW a private mapping later, we examine the
4482 * pending reservations for this page now. This will ensure that
4483 * any allocations necessary to record that reservation occur outside
4486 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4487 if (vma_needs_reservation(h, vma, haddr) < 0) {
4489 goto backout_unlocked;
4491 /* Just decrements count, does not deallocate */
4492 vma_end_reservation(h, vma, haddr);
4495 ptl = huge_pte_lock(h, mm, ptep);
4497 if (!huge_pte_none(huge_ptep_get(ptep)))
4501 ClearHPageRestoreReserve(page);
4502 hugepage_add_new_anon_rmap(page, vma, haddr);
4504 page_dup_rmap(page, true);
4505 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4506 && (vma->vm_flags & VM_SHARED)));
4507 set_huge_pte_at(mm, haddr, ptep, new_pte);
4509 hugetlb_count_add(pages_per_huge_page(h), mm);
4510 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4511 /* Optimization, do the COW without a second fault */
4512 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4518 * Only set HPageMigratable in newly allocated pages. Existing pages
4519 * found in the pagecache may not have HPageMigratableset if they have
4520 * been isolated for migration.
4523 SetHPageMigratable(page);
4533 restore_reserve_on_error(h, vma, haddr, page);
4539 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4541 unsigned long key[2];
4544 key[0] = (unsigned long) mapping;
4547 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4549 return hash & (num_fault_mutexes - 1);
4553 * For uniprocessor systems we always use a single mutex, so just
4554 * return 0 and avoid the hashing overhead.
4556 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4562 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4563 unsigned long address, unsigned int flags)
4570 struct page *page = NULL;
4571 struct page *pagecache_page = NULL;
4572 struct hstate *h = hstate_vma(vma);
4573 struct address_space *mapping;
4574 int need_wait_lock = 0;
4575 unsigned long haddr = address & huge_page_mask(h);
4577 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4580 * Since we hold no locks, ptep could be stale. That is
4581 * OK as we are only making decisions based on content and
4582 * not actually modifying content here.
4584 entry = huge_ptep_get(ptep);
4585 if (unlikely(is_hugetlb_entry_migration(entry))) {
4586 migration_entry_wait_huge(vma, mm, ptep);
4588 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4589 return VM_FAULT_HWPOISON_LARGE |
4590 VM_FAULT_SET_HINDEX(hstate_index(h));
4594 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4595 * until finished with ptep. This serves two purposes:
4596 * 1) It prevents huge_pmd_unshare from being called elsewhere
4597 * and making the ptep no longer valid.
4598 * 2) It synchronizes us with i_size modifications during truncation.
4600 * ptep could have already be assigned via huge_pte_offset. That
4601 * is OK, as huge_pte_alloc will return the same value unless
4602 * something has changed.
4604 mapping = vma->vm_file->f_mapping;
4605 i_mmap_lock_read(mapping);
4606 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
4608 i_mmap_unlock_read(mapping);
4609 return VM_FAULT_OOM;
4613 * Serialize hugepage allocation and instantiation, so that we don't
4614 * get spurious allocation failures if two CPUs race to instantiate
4615 * the same page in the page cache.
4617 idx = vma_hugecache_offset(h, vma, haddr);
4618 hash = hugetlb_fault_mutex_hash(mapping, idx);
4619 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4621 entry = huge_ptep_get(ptep);
4622 if (huge_pte_none(entry)) {
4623 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4630 * entry could be a migration/hwpoison entry at this point, so this
4631 * check prevents the kernel from going below assuming that we have
4632 * an active hugepage in pagecache. This goto expects the 2nd page
4633 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4634 * properly handle it.
4636 if (!pte_present(entry))
4640 * If we are going to COW the mapping later, we examine the pending
4641 * reservations for this page now. This will ensure that any
4642 * allocations necessary to record that reservation occur outside the
4643 * spinlock. For private mappings, we also lookup the pagecache
4644 * page now as it is used to determine if a reservation has been
4647 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4648 if (vma_needs_reservation(h, vma, haddr) < 0) {
4652 /* Just decrements count, does not deallocate */
4653 vma_end_reservation(h, vma, haddr);
4655 if (!(vma->vm_flags & VM_MAYSHARE))
4656 pagecache_page = hugetlbfs_pagecache_page(h,
4660 ptl = huge_pte_lock(h, mm, ptep);
4662 /* Check for a racing update before calling hugetlb_cow */
4663 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4667 * hugetlb_cow() requires page locks of pte_page(entry) and
4668 * pagecache_page, so here we need take the former one
4669 * when page != pagecache_page or !pagecache_page.
4671 page = pte_page(entry);
4672 if (page != pagecache_page)
4673 if (!trylock_page(page)) {
4680 if (flags & FAULT_FLAG_WRITE) {
4681 if (!huge_pte_write(entry)) {
4682 ret = hugetlb_cow(mm, vma, address, ptep,
4683 pagecache_page, ptl);
4686 entry = huge_pte_mkdirty(entry);
4688 entry = pte_mkyoung(entry);
4689 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4690 flags & FAULT_FLAG_WRITE))
4691 update_mmu_cache(vma, haddr, ptep);
4693 if (page != pagecache_page)
4699 if (pagecache_page) {
4700 unlock_page(pagecache_page);
4701 put_page(pagecache_page);
4704 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4705 i_mmap_unlock_read(mapping);
4707 * Generally it's safe to hold refcount during waiting page lock. But
4708 * here we just wait to defer the next page fault to avoid busy loop and
4709 * the page is not used after unlocked before returning from the current
4710 * page fault. So we are safe from accessing freed page, even if we wait
4711 * here without taking refcount.
4714 wait_on_page_locked(page);
4719 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4720 * modifications for huge pages.
4722 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4724 struct vm_area_struct *dst_vma,
4725 unsigned long dst_addr,
4726 unsigned long src_addr,
4727 struct page **pagep)
4729 struct address_space *mapping;
4732 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4733 struct hstate *h = hstate_vma(dst_vma);
4741 page = alloc_huge_page(dst_vma, dst_addr, 0);
4745 ret = copy_huge_page_from_user(page,
4746 (const void __user *) src_addr,
4747 pages_per_huge_page(h), false);
4749 /* fallback to copy_from_user outside mmap_lock */
4750 if (unlikely(ret)) {
4753 /* don't free the page */
4762 * The memory barrier inside __SetPageUptodate makes sure that
4763 * preceding stores to the page contents become visible before
4764 * the set_pte_at() write.
4766 __SetPageUptodate(page);
4768 mapping = dst_vma->vm_file->f_mapping;
4769 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4772 * If shared, add to page cache
4775 size = i_size_read(mapping->host) >> huge_page_shift(h);
4778 goto out_release_nounlock;
4781 * Serialization between remove_inode_hugepages() and
4782 * huge_add_to_page_cache() below happens through the
4783 * hugetlb_fault_mutex_table that here must be hold by
4786 ret = huge_add_to_page_cache(page, mapping, idx);
4788 goto out_release_nounlock;
4791 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4795 * Recheck the i_size after holding PT lock to make sure not
4796 * to leave any page mapped (as page_mapped()) beyond the end
4797 * of the i_size (remove_inode_hugepages() is strict about
4798 * enforcing that). If we bail out here, we'll also leave a
4799 * page in the radix tree in the vm_shared case beyond the end
4800 * of the i_size, but remove_inode_hugepages() will take care
4801 * of it as soon as we drop the hugetlb_fault_mutex_table.
4803 size = i_size_read(mapping->host) >> huge_page_shift(h);
4806 goto out_release_unlock;
4809 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4810 goto out_release_unlock;
4813 page_dup_rmap(page, true);
4815 ClearHPageRestoreReserve(page);
4816 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4819 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4820 if (dst_vma->vm_flags & VM_WRITE)
4821 _dst_pte = huge_pte_mkdirty(_dst_pte);
4822 _dst_pte = pte_mkyoung(_dst_pte);
4824 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4826 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4827 dst_vma->vm_flags & VM_WRITE);
4828 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4830 /* No need to invalidate - it was non-present before */
4831 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4834 SetHPageMigratable(page);
4844 out_release_nounlock:
4849 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
4850 int refs, struct page **pages,
4851 struct vm_area_struct **vmas)
4855 for (nr = 0; nr < refs; nr++) {
4857 pages[nr] = mem_map_offset(page, nr);
4863 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4864 struct page **pages, struct vm_area_struct **vmas,
4865 unsigned long *position, unsigned long *nr_pages,
4866 long i, unsigned int flags, int *locked)
4868 unsigned long pfn_offset;
4869 unsigned long vaddr = *position;
4870 unsigned long remainder = *nr_pages;
4871 struct hstate *h = hstate_vma(vma);
4872 int err = -EFAULT, refs;
4874 while (vaddr < vma->vm_end && remainder) {
4876 spinlock_t *ptl = NULL;
4881 * If we have a pending SIGKILL, don't keep faulting pages and
4882 * potentially allocating memory.
4884 if (fatal_signal_pending(current)) {
4890 * Some archs (sparc64, sh*) have multiple pte_ts to
4891 * each hugepage. We have to make sure we get the
4892 * first, for the page indexing below to work.
4894 * Note that page table lock is not held when pte is null.
4896 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4899 ptl = huge_pte_lock(h, mm, pte);
4900 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4903 * When coredumping, it suits get_dump_page if we just return
4904 * an error where there's an empty slot with no huge pagecache
4905 * to back it. This way, we avoid allocating a hugepage, and
4906 * the sparse dumpfile avoids allocating disk blocks, but its
4907 * huge holes still show up with zeroes where they need to be.
4909 if (absent && (flags & FOLL_DUMP) &&
4910 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4918 * We need call hugetlb_fault for both hugepages under migration
4919 * (in which case hugetlb_fault waits for the migration,) and
4920 * hwpoisoned hugepages (in which case we need to prevent the
4921 * caller from accessing to them.) In order to do this, we use
4922 * here is_swap_pte instead of is_hugetlb_entry_migration and
4923 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4924 * both cases, and because we can't follow correct pages
4925 * directly from any kind of swap entries.
4927 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4928 ((flags & FOLL_WRITE) &&
4929 !huge_pte_write(huge_ptep_get(pte)))) {
4931 unsigned int fault_flags = 0;
4935 if (flags & FOLL_WRITE)
4936 fault_flags |= FAULT_FLAG_WRITE;
4938 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4939 FAULT_FLAG_KILLABLE;
4940 if (flags & FOLL_NOWAIT)
4941 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4942 FAULT_FLAG_RETRY_NOWAIT;
4943 if (flags & FOLL_TRIED) {
4945 * Note: FAULT_FLAG_ALLOW_RETRY and
4946 * FAULT_FLAG_TRIED can co-exist
4948 fault_flags |= FAULT_FLAG_TRIED;
4950 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4951 if (ret & VM_FAULT_ERROR) {
4952 err = vm_fault_to_errno(ret, flags);
4956 if (ret & VM_FAULT_RETRY) {
4958 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4962 * VM_FAULT_RETRY must not return an
4963 * error, it will return zero
4966 * No need to update "position" as the
4967 * caller will not check it after
4968 * *nr_pages is set to 0.
4975 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4976 page = pte_page(huge_ptep_get(pte));
4979 * If subpage information not requested, update counters
4980 * and skip the same_page loop below.
4982 if (!pages && !vmas && !pfn_offset &&
4983 (vaddr + huge_page_size(h) < vma->vm_end) &&
4984 (remainder >= pages_per_huge_page(h))) {
4985 vaddr += huge_page_size(h);
4986 remainder -= pages_per_huge_page(h);
4987 i += pages_per_huge_page(h);
4992 refs = min3(pages_per_huge_page(h) - pfn_offset,
4993 (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder);
4996 record_subpages_vmas(mem_map_offset(page, pfn_offset),
4998 likely(pages) ? pages + i : NULL,
4999 vmas ? vmas + i : NULL);
5003 * try_grab_compound_head() should always succeed here,
5004 * because: a) we hold the ptl lock, and b) we've just
5005 * checked that the huge page is present in the page
5006 * tables. If the huge page is present, then the tail
5007 * pages must also be present. The ptl prevents the
5008 * head page and tail pages from being rearranged in
5009 * any way. So this page must be available at this
5010 * point, unless the page refcount overflowed:
5012 if (WARN_ON_ONCE(!try_grab_compound_head(pages[i],
5022 vaddr += (refs << PAGE_SHIFT);
5028 *nr_pages = remainder;
5030 * setting position is actually required only if remainder is
5031 * not zero but it's faster not to add a "if (remainder)"
5039 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
5040 unsigned long address, unsigned long end, pgprot_t newprot)
5042 struct mm_struct *mm = vma->vm_mm;
5043 unsigned long start = address;
5046 struct hstate *h = hstate_vma(vma);
5047 unsigned long pages = 0;
5048 bool shared_pmd = false;
5049 struct mmu_notifier_range range;
5052 * In the case of shared PMDs, the area to flush could be beyond
5053 * start/end. Set range.start/range.end to cover the maximum possible
5054 * range if PMD sharing is possible.
5056 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5057 0, vma, mm, start, end);
5058 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5060 BUG_ON(address >= end);
5061 flush_cache_range(vma, range.start, range.end);
5063 mmu_notifier_invalidate_range_start(&range);
5064 i_mmap_lock_write(vma->vm_file->f_mapping);
5065 for (; address < end; address += huge_page_size(h)) {
5067 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5070 ptl = huge_pte_lock(h, mm, ptep);
5071 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
5077 pte = huge_ptep_get(ptep);
5078 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5082 if (unlikely(is_hugetlb_entry_migration(pte))) {
5083 swp_entry_t entry = pte_to_swp_entry(pte);
5085 if (is_write_migration_entry(entry)) {
5088 make_migration_entry_read(&entry);
5089 newpte = swp_entry_to_pte(entry);
5090 set_huge_swap_pte_at(mm, address, ptep,
5091 newpte, huge_page_size(h));
5097 if (!huge_pte_none(pte)) {
5100 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5101 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
5102 pte = arch_make_huge_pte(pte, vma, NULL, 0);
5103 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
5109 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5110 * may have cleared our pud entry and done put_page on the page table:
5111 * once we release i_mmap_rwsem, another task can do the final put_page
5112 * and that page table be reused and filled with junk. If we actually
5113 * did unshare a page of pmds, flush the range corresponding to the pud.
5116 flush_hugetlb_tlb_range(vma, range.start, range.end);
5118 flush_hugetlb_tlb_range(vma, start, end);
5120 * No need to call mmu_notifier_invalidate_range() we are downgrading
5121 * page table protection not changing it to point to a new page.
5123 * See Documentation/vm/mmu_notifier.rst
5125 i_mmap_unlock_write(vma->vm_file->f_mapping);
5126 mmu_notifier_invalidate_range_end(&range);
5128 return pages << h->order;
5131 /* Return true if reservation was successful, false otherwise. */
5132 bool hugetlb_reserve_pages(struct inode *inode,
5134 struct vm_area_struct *vma,
5135 vm_flags_t vm_flags)
5138 struct hstate *h = hstate_inode(inode);
5139 struct hugepage_subpool *spool = subpool_inode(inode);
5140 struct resv_map *resv_map;
5141 struct hugetlb_cgroup *h_cg = NULL;
5142 long gbl_reserve, regions_needed = 0;
5144 /* This should never happen */
5146 VM_WARN(1, "%s called with a negative range\n", __func__);
5151 * Only apply hugepage reservation if asked. At fault time, an
5152 * attempt will be made for VM_NORESERVE to allocate a page
5153 * without using reserves
5155 if (vm_flags & VM_NORESERVE)
5159 * Shared mappings base their reservation on the number of pages that
5160 * are already allocated on behalf of the file. Private mappings need
5161 * to reserve the full area even if read-only as mprotect() may be
5162 * called to make the mapping read-write. Assume !vma is a shm mapping
5164 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5166 * resv_map can not be NULL as hugetlb_reserve_pages is only
5167 * called for inodes for which resv_maps were created (see
5168 * hugetlbfs_get_inode).
5170 resv_map = inode_resv_map(inode);
5172 chg = region_chg(resv_map, from, to, ®ions_needed);
5175 /* Private mapping. */
5176 resv_map = resv_map_alloc();
5182 set_vma_resv_map(vma, resv_map);
5183 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5189 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
5190 chg * pages_per_huge_page(h), &h_cg) < 0)
5193 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5194 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5197 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5201 * There must be enough pages in the subpool for the mapping. If
5202 * the subpool has a minimum size, there may be some global
5203 * reservations already in place (gbl_reserve).
5205 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5206 if (gbl_reserve < 0)
5207 goto out_uncharge_cgroup;
5210 * Check enough hugepages are available for the reservation.
5211 * Hand the pages back to the subpool if there are not
5213 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
5217 * Account for the reservations made. Shared mappings record regions
5218 * that have reservations as they are shared by multiple VMAs.
5219 * When the last VMA disappears, the region map says how much
5220 * the reservation was and the page cache tells how much of
5221 * the reservation was consumed. Private mappings are per-VMA and
5222 * only the consumed reservations are tracked. When the VMA
5223 * disappears, the original reservation is the VMA size and the
5224 * consumed reservations are stored in the map. Hence, nothing
5225 * else has to be done for private mappings here
5227 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5228 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5230 if (unlikely(add < 0)) {
5231 hugetlb_acct_memory(h, -gbl_reserve);
5233 } else if (unlikely(chg > add)) {
5235 * pages in this range were added to the reserve
5236 * map between region_chg and region_add. This
5237 * indicates a race with alloc_huge_page. Adjust
5238 * the subpool and reserve counts modified above
5239 * based on the difference.
5244 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
5245 * reference to h_cg->css. See comment below for detail.
5247 hugetlb_cgroup_uncharge_cgroup_rsvd(
5249 (chg - add) * pages_per_huge_page(h), h_cg);
5251 rsv_adjust = hugepage_subpool_put_pages(spool,
5253 hugetlb_acct_memory(h, -rsv_adjust);
5256 * The file_regions will hold their own reference to
5257 * h_cg->css. So we should release the reference held
5258 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
5261 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
5267 /* put back original number of pages, chg */
5268 (void)hugepage_subpool_put_pages(spool, chg);
5269 out_uncharge_cgroup:
5270 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5271 chg * pages_per_huge_page(h), h_cg);
5273 if (!vma || vma->vm_flags & VM_MAYSHARE)
5274 /* Only call region_abort if the region_chg succeeded but the
5275 * region_add failed or didn't run.
5277 if (chg >= 0 && add < 0)
5278 region_abort(resv_map, from, to, regions_needed);
5279 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5280 kref_put(&resv_map->refs, resv_map_release);
5284 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5287 struct hstate *h = hstate_inode(inode);
5288 struct resv_map *resv_map = inode_resv_map(inode);
5290 struct hugepage_subpool *spool = subpool_inode(inode);
5294 * Since this routine can be called in the evict inode path for all
5295 * hugetlbfs inodes, resv_map could be NULL.
5298 chg = region_del(resv_map, start, end);
5300 * region_del() can fail in the rare case where a region
5301 * must be split and another region descriptor can not be
5302 * allocated. If end == LONG_MAX, it will not fail.
5308 spin_lock(&inode->i_lock);
5309 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5310 spin_unlock(&inode->i_lock);
5313 * If the subpool has a minimum size, the number of global
5314 * reservations to be released may be adjusted.
5316 * Note that !resv_map implies freed == 0. So (chg - freed)
5317 * won't go negative.
5319 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5320 hugetlb_acct_memory(h, -gbl_reserve);
5325 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5326 static unsigned long page_table_shareable(struct vm_area_struct *svma,
5327 struct vm_area_struct *vma,
5328 unsigned long addr, pgoff_t idx)
5330 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5332 unsigned long sbase = saddr & PUD_MASK;
5333 unsigned long s_end = sbase + PUD_SIZE;
5335 /* Allow segments to share if only one is marked locked */
5336 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5337 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5340 * match the virtual addresses, permission and the alignment of the
5343 if (pmd_index(addr) != pmd_index(saddr) ||
5344 vm_flags != svm_flags ||
5345 !range_in_vma(svma, sbase, s_end))
5351 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5353 unsigned long base = addr & PUD_MASK;
5354 unsigned long end = base + PUD_SIZE;
5357 * check on proper vm_flags and page table alignment
5359 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5364 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5366 #ifdef CONFIG_USERFAULTFD
5367 if (uffd_disable_huge_pmd_share(vma))
5370 return vma_shareable(vma, addr);
5374 * Determine if start,end range within vma could be mapped by shared pmd.
5375 * If yes, adjust start and end to cover range associated with possible
5376 * shared pmd mappings.
5378 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5379 unsigned long *start, unsigned long *end)
5381 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
5382 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
5385 * vma need span at least one aligned PUD size and the start,end range
5386 * must at least partialy within it.
5388 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
5389 (*end <= v_start) || (*start >= v_end))
5392 /* Extend the range to be PUD aligned for a worst case scenario */
5393 if (*start > v_start)
5394 *start = ALIGN_DOWN(*start, PUD_SIZE);
5397 *end = ALIGN(*end, PUD_SIZE);
5401 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5402 * and returns the corresponding pte. While this is not necessary for the
5403 * !shared pmd case because we can allocate the pmd later as well, it makes the
5404 * code much cleaner.
5406 * This routine must be called with i_mmap_rwsem held in at least read mode if
5407 * sharing is possible. For hugetlbfs, this prevents removal of any page
5408 * table entries associated with the address space. This is important as we
5409 * are setting up sharing based on existing page table entries (mappings).
5411 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5412 * huge_pte_alloc know that sharing is not possible and do not take
5413 * i_mmap_rwsem as a performance optimization. This is handled by the
5414 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5415 * only required for subsequent processing.
5417 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5418 unsigned long addr, pud_t *pud)
5420 struct address_space *mapping = vma->vm_file->f_mapping;
5421 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5423 struct vm_area_struct *svma;
5424 unsigned long saddr;
5429 i_mmap_assert_locked(mapping);
5430 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5434 saddr = page_table_shareable(svma, vma, addr, idx);
5436 spte = huge_pte_offset(svma->vm_mm, saddr,
5437 vma_mmu_pagesize(svma));
5439 get_page(virt_to_page(spte));
5448 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
5449 if (pud_none(*pud)) {
5450 pud_populate(mm, pud,
5451 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5454 put_page(virt_to_page(spte));
5458 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5463 * unmap huge page backed by shared pte.
5465 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5466 * indicated by page_count > 1, unmap is achieved by clearing pud and
5467 * decrementing the ref count. If count == 1, the pte page is not shared.
5469 * Called with page table lock held and i_mmap_rwsem held in write mode.
5471 * returns: 1 successfully unmapped a shared pte page
5472 * 0 the underlying pte page is not shared, or it is the last user
5474 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5475 unsigned long *addr, pte_t *ptep)
5477 pgd_t *pgd = pgd_offset(mm, *addr);
5478 p4d_t *p4d = p4d_offset(pgd, *addr);
5479 pud_t *pud = pud_offset(p4d, *addr);
5481 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
5482 BUG_ON(page_count(virt_to_page(ptep)) == 0);
5483 if (page_count(virt_to_page(ptep)) == 1)
5487 put_page(virt_to_page(ptep));
5489 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5493 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5494 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5495 unsigned long addr, pud_t *pud)
5500 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5501 unsigned long *addr, pte_t *ptep)
5506 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5507 unsigned long *start, unsigned long *end)
5511 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5515 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5517 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5518 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
5519 unsigned long addr, unsigned long sz)
5526 pgd = pgd_offset(mm, addr);
5527 p4d = p4d_alloc(mm, pgd, addr);
5530 pud = pud_alloc(mm, p4d, addr);
5532 if (sz == PUD_SIZE) {
5535 BUG_ON(sz != PMD_SIZE);
5536 if (want_pmd_share(vma, addr) && pud_none(*pud))
5537 pte = huge_pmd_share(mm, vma, addr, pud);
5539 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5542 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
5548 * huge_pte_offset() - Walk the page table to resolve the hugepage
5549 * entry at address @addr
5551 * Return: Pointer to page table entry (PUD or PMD) for
5552 * address @addr, or NULL if a !p*d_present() entry is encountered and the
5553 * size @sz doesn't match the hugepage size at this level of the page
5556 pte_t *huge_pte_offset(struct mm_struct *mm,
5557 unsigned long addr, unsigned long sz)
5564 pgd = pgd_offset(mm, addr);
5565 if (!pgd_present(*pgd))
5567 p4d = p4d_offset(pgd, addr);
5568 if (!p4d_present(*p4d))
5571 pud = pud_offset(p4d, addr);
5573 /* must be pud huge, non-present or none */
5574 return (pte_t *)pud;
5575 if (!pud_present(*pud))
5577 /* must have a valid entry and size to go further */
5579 pmd = pmd_offset(pud, addr);
5580 /* must be pmd huge, non-present or none */
5581 return (pte_t *)pmd;
5584 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5587 * These functions are overwritable if your architecture needs its own
5590 struct page * __weak
5591 follow_huge_addr(struct mm_struct *mm, unsigned long address,
5594 return ERR_PTR(-EINVAL);
5597 struct page * __weak
5598 follow_huge_pd(struct vm_area_struct *vma,
5599 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5601 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5605 struct page * __weak
5606 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
5607 pmd_t *pmd, int flags)
5609 struct page *page = NULL;
5613 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5614 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
5615 (FOLL_PIN | FOLL_GET)))
5619 ptl = pmd_lockptr(mm, pmd);
5622 * make sure that the address range covered by this pmd is not
5623 * unmapped from other threads.
5625 if (!pmd_huge(*pmd))
5627 pte = huge_ptep_get((pte_t *)pmd);
5628 if (pte_present(pte)) {
5629 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
5631 * try_grab_page() should always succeed here, because: a) we
5632 * hold the pmd (ptl) lock, and b) we've just checked that the
5633 * huge pmd (head) page is present in the page tables. The ptl
5634 * prevents the head page and tail pages from being rearranged
5635 * in any way. So this page must be available at this point,
5636 * unless the page refcount overflowed:
5638 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5643 if (is_hugetlb_entry_migration(pte)) {
5645 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5649 * hwpoisoned entry is treated as no_page_table in
5650 * follow_page_mask().
5658 struct page * __weak
5659 follow_huge_pud(struct mm_struct *mm, unsigned long address,
5660 pud_t *pud, int flags)
5662 if (flags & (FOLL_GET | FOLL_PIN))
5665 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5668 struct page * __weak
5669 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5671 if (flags & (FOLL_GET | FOLL_PIN))
5674 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5677 bool isolate_huge_page(struct page *page, struct list_head *list)
5681 spin_lock(&hugetlb_lock);
5682 if (!PageHeadHuge(page) ||
5683 !HPageMigratable(page) ||
5684 !get_page_unless_zero(page)) {
5688 ClearHPageMigratable(page);
5689 list_move_tail(&page->lru, list);
5691 spin_unlock(&hugetlb_lock);
5695 void putback_active_hugepage(struct page *page)
5697 spin_lock(&hugetlb_lock);
5698 SetHPageMigratable(page);
5699 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5700 spin_unlock(&hugetlb_lock);
5704 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5706 struct hstate *h = page_hstate(oldpage);
5708 hugetlb_cgroup_migrate(oldpage, newpage);
5709 set_page_owner_migrate_reason(newpage, reason);
5712 * transfer temporary state of the new huge page. This is
5713 * reverse to other transitions because the newpage is going to
5714 * be final while the old one will be freed so it takes over
5715 * the temporary status.
5717 * Also note that we have to transfer the per-node surplus state
5718 * here as well otherwise the global surplus count will not match
5721 if (HPageTemporary(newpage)) {
5722 int old_nid = page_to_nid(oldpage);
5723 int new_nid = page_to_nid(newpage);
5725 SetHPageTemporary(oldpage);
5726 ClearHPageTemporary(newpage);
5729 * There is no need to transfer the per-node surplus state
5730 * when we do not cross the node.
5732 if (new_nid == old_nid)
5734 spin_lock(&hugetlb_lock);
5735 if (h->surplus_huge_pages_node[old_nid]) {
5736 h->surplus_huge_pages_node[old_nid]--;
5737 h->surplus_huge_pages_node[new_nid]++;
5739 spin_unlock(&hugetlb_lock);
5744 * This function will unconditionally remove all the shared pmd pgtable entries
5745 * within the specific vma for a hugetlbfs memory range.
5747 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
5749 struct hstate *h = hstate_vma(vma);
5750 unsigned long sz = huge_page_size(h);
5751 struct mm_struct *mm = vma->vm_mm;
5752 struct mmu_notifier_range range;
5753 unsigned long address, start, end;
5757 if (!(vma->vm_flags & VM_MAYSHARE))
5760 start = ALIGN(vma->vm_start, PUD_SIZE);
5761 end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
5767 * No need to call adjust_range_if_pmd_sharing_possible(), because
5768 * we have already done the PUD_SIZE alignment.
5770 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
5772 mmu_notifier_invalidate_range_start(&range);
5773 i_mmap_lock_write(vma->vm_file->f_mapping);
5774 for (address = start; address < end; address += PUD_SIZE) {
5775 unsigned long tmp = address;
5777 ptep = huge_pte_offset(mm, address, sz);
5780 ptl = huge_pte_lock(h, mm, ptep);
5781 /* We don't want 'address' to be changed */
5782 huge_pmd_unshare(mm, vma, &tmp, ptep);
5785 flush_hugetlb_tlb_range(vma, start, end);
5786 i_mmap_unlock_write(vma->vm_file->f_mapping);
5788 * No need to call mmu_notifier_invalidate_range(), see
5789 * Documentation/vm/mmu_notifier.rst.
5791 mmu_notifier_invalidate_range_end(&range);
5795 static bool cma_reserve_called __initdata;
5797 static int __init cmdline_parse_hugetlb_cma(char *p)
5799 hugetlb_cma_size = memparse(p, &p);
5803 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
5805 void __init hugetlb_cma_reserve(int order)
5807 unsigned long size, reserved, per_node;
5810 cma_reserve_called = true;
5812 if (!hugetlb_cma_size)
5815 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
5816 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
5817 (PAGE_SIZE << order) / SZ_1M);
5822 * If 3 GB area is requested on a machine with 4 numa nodes,
5823 * let's allocate 1 GB on first three nodes and ignore the last one.
5825 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
5826 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
5827 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
5830 for_each_node_state(nid, N_ONLINE) {
5832 char name[CMA_MAX_NAME];
5834 size = min(per_node, hugetlb_cma_size - reserved);
5835 size = round_up(size, PAGE_SIZE << order);
5837 snprintf(name, sizeof(name), "hugetlb%d", nid);
5838 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
5840 &hugetlb_cma[nid], nid);
5842 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
5848 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
5851 if (reserved >= hugetlb_cma_size)
5856 void __init hugetlb_cma_check(void)
5858 if (!hugetlb_cma_size || cma_reserve_called)
5861 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
5864 #endif /* CONFIG_CMA */