1 // SPDX-License-Identifier: GPL-2.0-only
3 * Generic hugetlb support.
4 * (C) Nadia Yvette Chambers, April 2004
6 #include <linux/list.h>
7 #include <linux/init.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/sched/mm.h>
23 #include <linux/mmdebug.h>
24 #include <linux/sched/signal.h>
25 #include <linux/rmap.h>
26 #include <linux/string_helpers.h>
27 #include <linux/swap.h>
28 #include <linux/swapops.h>
29 #include <linux/jhash.h>
30 #include <linux/numa.h>
31 #include <linux/llist.h>
32 #include <linux/cma.h>
35 #include <asm/pgalloc.h>
39 #include <linux/hugetlb.h>
40 #include <linux/hugetlb_cgroup.h>
41 #include <linux/node.h>
42 #include <linux/userfaultfd_k.h>
43 #include <linux/page_owner.h>
46 int hugetlb_max_hstate __read_mostly;
47 unsigned int default_hstate_idx;
48 struct hstate hstates[HUGE_MAX_HSTATE];
51 static struct cma *hugetlb_cma[MAX_NUMNODES];
53 static unsigned long hugetlb_cma_size __initdata;
56 * Minimum page order among possible hugepage sizes, set to a proper value
59 static unsigned int minimum_order __read_mostly = UINT_MAX;
61 __initdata LIST_HEAD(huge_boot_pages);
63 /* for command line parsing */
64 static struct hstate * __initdata parsed_hstate;
65 static unsigned long __initdata default_hstate_max_huge_pages;
66 static bool __initdata parsed_valid_hugepagesz = true;
67 static bool __initdata parsed_default_hugepagesz;
70 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
71 * free_huge_pages, and surplus_huge_pages.
73 DEFINE_SPINLOCK(hugetlb_lock);
76 * Serializes faults on the same logical page. This is used to
77 * prevent spurious OOMs when the hugepage pool is fully utilized.
79 static int num_fault_mutexes;
80 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
82 static inline bool PageHugeFreed(struct page *head)
84 return page_private(head + 4) == -1UL;
87 static inline void SetPageHugeFreed(struct page *head)
89 set_page_private(head + 4, -1UL);
92 static inline void ClearPageHugeFreed(struct page *head)
94 set_page_private(head + 4, 0);
97 /* Forward declaration */
98 static int hugetlb_acct_memory(struct hstate *h, long delta);
100 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
102 bool free = (spool->count == 0) && (spool->used_hpages == 0);
104 spin_unlock(&spool->lock);
106 /* If no pages are used, and no other handles to the subpool
107 * remain, give up any reservations based on minimum size and
108 * free the subpool */
110 if (spool->min_hpages != -1)
111 hugetlb_acct_memory(spool->hstate,
117 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
120 struct hugepage_subpool *spool;
122 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
126 spin_lock_init(&spool->lock);
128 spool->max_hpages = max_hpages;
130 spool->min_hpages = min_hpages;
132 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
136 spool->rsv_hpages = min_hpages;
141 void hugepage_put_subpool(struct hugepage_subpool *spool)
143 spin_lock(&spool->lock);
144 BUG_ON(!spool->count);
146 unlock_or_release_subpool(spool);
150 * Subpool accounting for allocating and reserving pages.
151 * Return -ENOMEM if there are not enough resources to satisfy the
152 * request. Otherwise, return the number of pages by which the
153 * global pools must be adjusted (upward). The returned value may
154 * only be different than the passed value (delta) in the case where
155 * a subpool minimum size must be maintained.
157 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
165 spin_lock(&spool->lock);
167 if (spool->max_hpages != -1) { /* maximum size accounting */
168 if ((spool->used_hpages + delta) <= spool->max_hpages)
169 spool->used_hpages += delta;
176 /* minimum size accounting */
177 if (spool->min_hpages != -1 && spool->rsv_hpages) {
178 if (delta > spool->rsv_hpages) {
180 * Asking for more reserves than those already taken on
181 * behalf of subpool. Return difference.
183 ret = delta - spool->rsv_hpages;
184 spool->rsv_hpages = 0;
186 ret = 0; /* reserves already accounted for */
187 spool->rsv_hpages -= delta;
192 spin_unlock(&spool->lock);
197 * Subpool accounting for freeing and unreserving pages.
198 * Return the number of global page reservations that must be dropped.
199 * The return value may only be different than the passed value (delta)
200 * in the case where a subpool minimum size must be maintained.
202 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
210 spin_lock(&spool->lock);
212 if (spool->max_hpages != -1) /* maximum size accounting */
213 spool->used_hpages -= delta;
215 /* minimum size accounting */
216 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
217 if (spool->rsv_hpages + delta <= spool->min_hpages)
220 ret = spool->rsv_hpages + delta - spool->min_hpages;
222 spool->rsv_hpages += delta;
223 if (spool->rsv_hpages > spool->min_hpages)
224 spool->rsv_hpages = spool->min_hpages;
228 * If hugetlbfs_put_super couldn't free spool due to an outstanding
229 * quota reference, free it now.
231 unlock_or_release_subpool(spool);
236 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
238 return HUGETLBFS_SB(inode->i_sb)->spool;
241 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
243 return subpool_inode(file_inode(vma->vm_file));
246 /* Helper that removes a struct file_region from the resv_map cache and returns
249 static struct file_region *
250 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
252 struct file_region *nrg = NULL;
254 VM_BUG_ON(resv->region_cache_count <= 0);
256 resv->region_cache_count--;
257 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
258 list_del(&nrg->link);
266 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
267 struct file_region *rg)
269 #ifdef CONFIG_CGROUP_HUGETLB
270 nrg->reservation_counter = rg->reservation_counter;
277 /* Helper that records hugetlb_cgroup uncharge info. */
278 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
280 struct resv_map *resv,
281 struct file_region *nrg)
283 #ifdef CONFIG_CGROUP_HUGETLB
285 nrg->reservation_counter =
286 &h_cg->rsvd_hugepage[hstate_index(h)];
287 nrg->css = &h_cg->css;
289 * The caller will hold exactly one h_cg->css reference for the
290 * whole contiguous reservation region. But this area might be
291 * scattered when there are already some file_regions reside in
292 * it. As a result, many file_regions may share only one css
293 * reference. In order to ensure that one file_region must hold
294 * exactly one h_cg->css reference, we should do css_get for
295 * each file_region and leave the reference held by caller
299 if (!resv->pages_per_hpage)
300 resv->pages_per_hpage = pages_per_huge_page(h);
301 /* pages_per_hpage should be the same for all entries in
304 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
306 nrg->reservation_counter = NULL;
312 static void put_uncharge_info(struct file_region *rg)
314 #ifdef CONFIG_CGROUP_HUGETLB
320 static bool has_same_uncharge_info(struct file_region *rg,
321 struct file_region *org)
323 #ifdef CONFIG_CGROUP_HUGETLB
325 rg->reservation_counter == org->reservation_counter &&
333 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
335 struct file_region *nrg = NULL, *prg = NULL;
337 prg = list_prev_entry(rg, link);
338 if (&prg->link != &resv->regions && prg->to == rg->from &&
339 has_same_uncharge_info(prg, rg)) {
343 put_uncharge_info(rg);
349 nrg = list_next_entry(rg, link);
350 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
351 has_same_uncharge_info(nrg, rg)) {
352 nrg->from = rg->from;
355 put_uncharge_info(rg);
361 * Must be called with resv->lock held.
363 * Calling this with regions_needed != NULL will count the number of pages
364 * to be added but will not modify the linked list. And regions_needed will
365 * indicate the number of file_regions needed in the cache to carry out to add
366 * the regions for this range.
368 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
369 struct hugetlb_cgroup *h_cg,
370 struct hstate *h, long *regions_needed)
373 struct list_head *head = &resv->regions;
374 long last_accounted_offset = f;
375 struct file_region *rg = NULL, *trg = NULL, *nrg = NULL;
380 /* In this loop, we essentially handle an entry for the range
381 * [last_accounted_offset, rg->from), at every iteration, with some
384 list_for_each_entry_safe(rg, trg, head, link) {
385 /* Skip irrelevant regions that start before our range. */
387 /* If this region ends after the last accounted offset,
388 * then we need to update last_accounted_offset.
390 if (rg->to > last_accounted_offset)
391 last_accounted_offset = rg->to;
395 /* When we find a region that starts beyond our range, we've
401 /* Add an entry for last_accounted_offset -> rg->from, and
402 * update last_accounted_offset.
404 if (rg->from > last_accounted_offset) {
405 add += rg->from - last_accounted_offset;
406 if (!regions_needed) {
407 nrg = get_file_region_entry_from_cache(
408 resv, last_accounted_offset, rg->from);
409 record_hugetlb_cgroup_uncharge_info(h_cg, h,
411 list_add(&nrg->link, rg->link.prev);
412 coalesce_file_region(resv, nrg);
414 *regions_needed += 1;
417 last_accounted_offset = rg->to;
420 /* Handle the case where our range extends beyond
421 * last_accounted_offset.
423 if (last_accounted_offset < t) {
424 add += t - last_accounted_offset;
425 if (!regions_needed) {
426 nrg = get_file_region_entry_from_cache(
427 resv, last_accounted_offset, t);
428 record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg);
429 list_add(&nrg->link, rg->link.prev);
430 coalesce_file_region(resv, nrg);
432 *regions_needed += 1;
439 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
441 static int allocate_file_region_entries(struct resv_map *resv,
443 __must_hold(&resv->lock)
445 struct list_head allocated_regions;
446 int to_allocate = 0, i = 0;
447 struct file_region *trg = NULL, *rg = NULL;
449 VM_BUG_ON(regions_needed < 0);
451 INIT_LIST_HEAD(&allocated_regions);
454 * Check for sufficient descriptors in the cache to accommodate
455 * the number of in progress add operations plus regions_needed.
457 * This is a while loop because when we drop the lock, some other call
458 * to region_add or region_del may have consumed some region_entries,
459 * so we keep looping here until we finally have enough entries for
460 * (adds_in_progress + regions_needed).
462 while (resv->region_cache_count <
463 (resv->adds_in_progress + regions_needed)) {
464 to_allocate = resv->adds_in_progress + regions_needed -
465 resv->region_cache_count;
467 /* At this point, we should have enough entries in the cache
468 * for all the existings adds_in_progress. We should only be
469 * needing to allocate for regions_needed.
471 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
473 spin_unlock(&resv->lock);
474 for (i = 0; i < to_allocate; i++) {
475 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
478 list_add(&trg->link, &allocated_regions);
481 spin_lock(&resv->lock);
483 list_splice(&allocated_regions, &resv->region_cache);
484 resv->region_cache_count += to_allocate;
490 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
498 * Add the huge page range represented by [f, t) to the reserve
499 * map. Regions will be taken from the cache to fill in this range.
500 * Sufficient regions should exist in the cache due to the previous
501 * call to region_chg with the same range, but in some cases the cache will not
502 * have sufficient entries due to races with other code doing region_add or
503 * region_del. The extra needed entries will be allocated.
505 * regions_needed is the out value provided by a previous call to region_chg.
507 * Return the number of new huge pages added to the map. This number is greater
508 * than or equal to zero. If file_region entries needed to be allocated for
509 * this operation and we were not able to allocate, it returns -ENOMEM.
510 * region_add of regions of length 1 never allocate file_regions and cannot
511 * fail; region_chg will always allocate at least 1 entry and a region_add for
512 * 1 page will only require at most 1 entry.
514 static long region_add(struct resv_map *resv, long f, long t,
515 long in_regions_needed, struct hstate *h,
516 struct hugetlb_cgroup *h_cg)
518 long add = 0, actual_regions_needed = 0;
520 spin_lock(&resv->lock);
523 /* Count how many regions are actually needed to execute this add. */
524 add_reservation_in_range(resv, f, t, NULL, NULL,
525 &actual_regions_needed);
528 * Check for sufficient descriptors in the cache to accommodate
529 * this add operation. Note that actual_regions_needed may be greater
530 * than in_regions_needed, as the resv_map may have been modified since
531 * the region_chg call. In this case, we need to make sure that we
532 * allocate extra entries, such that we have enough for all the
533 * existing adds_in_progress, plus the excess needed for this
536 if (actual_regions_needed > in_regions_needed &&
537 resv->region_cache_count <
538 resv->adds_in_progress +
539 (actual_regions_needed - in_regions_needed)) {
540 /* region_add operation of range 1 should never need to
541 * allocate file_region entries.
543 VM_BUG_ON(t - f <= 1);
545 if (allocate_file_region_entries(
546 resv, actual_regions_needed - in_regions_needed)) {
553 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
555 resv->adds_in_progress -= in_regions_needed;
557 spin_unlock(&resv->lock);
563 * Examine the existing reserve map and determine how many
564 * huge pages in the specified range [f, t) are NOT currently
565 * represented. This routine is called before a subsequent
566 * call to region_add that will actually modify the reserve
567 * map to add the specified range [f, t). region_chg does
568 * not change the number of huge pages represented by the
569 * map. A number of new file_region structures is added to the cache as a
570 * placeholder, for the subsequent region_add call to use. At least 1
571 * file_region structure is added.
573 * out_regions_needed is the number of regions added to the
574 * resv->adds_in_progress. This value needs to be provided to a follow up call
575 * to region_add or region_abort for proper accounting.
577 * Returns the number of huge pages that need to be added to the existing
578 * reservation map for the range [f, t). This number is greater or equal to
579 * zero. -ENOMEM is returned if a new file_region structure or cache entry
580 * is needed and can not be allocated.
582 static long region_chg(struct resv_map *resv, long f, long t,
583 long *out_regions_needed)
587 spin_lock(&resv->lock);
589 /* Count how many hugepages in this range are NOT represented. */
590 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
593 if (*out_regions_needed == 0)
594 *out_regions_needed = 1;
596 if (allocate_file_region_entries(resv, *out_regions_needed))
599 resv->adds_in_progress += *out_regions_needed;
601 spin_unlock(&resv->lock);
606 * Abort the in progress add operation. The adds_in_progress field
607 * of the resv_map keeps track of the operations in progress between
608 * calls to region_chg and region_add. Operations are sometimes
609 * aborted after the call to region_chg. In such cases, region_abort
610 * is called to decrement the adds_in_progress counter. regions_needed
611 * is the value returned by the region_chg call, it is used to decrement
612 * the adds_in_progress counter.
614 * NOTE: The range arguments [f, t) are not needed or used in this
615 * routine. They are kept to make reading the calling code easier as
616 * arguments will match the associated region_chg call.
618 static void region_abort(struct resv_map *resv, long f, long t,
621 spin_lock(&resv->lock);
622 VM_BUG_ON(!resv->region_cache_count);
623 resv->adds_in_progress -= regions_needed;
624 spin_unlock(&resv->lock);
628 * Delete the specified range [f, t) from the reserve map. If the
629 * t parameter is LONG_MAX, this indicates that ALL regions after f
630 * should be deleted. Locate the regions which intersect [f, t)
631 * and either trim, delete or split the existing regions.
633 * Returns the number of huge pages deleted from the reserve map.
634 * In the normal case, the return value is zero or more. In the
635 * case where a region must be split, a new region descriptor must
636 * be allocated. If the allocation fails, -ENOMEM will be returned.
637 * NOTE: If the parameter t == LONG_MAX, then we will never split
638 * a region and possibly return -ENOMEM. Callers specifying
639 * t == LONG_MAX do not need to check for -ENOMEM error.
641 static long region_del(struct resv_map *resv, long f, long t)
643 struct list_head *head = &resv->regions;
644 struct file_region *rg, *trg;
645 struct file_region *nrg = NULL;
649 spin_lock(&resv->lock);
650 list_for_each_entry_safe(rg, trg, head, link) {
652 * Skip regions before the range to be deleted. file_region
653 * ranges are normally of the form [from, to). However, there
654 * may be a "placeholder" entry in the map which is of the form
655 * (from, to) with from == to. Check for placeholder entries
656 * at the beginning of the range to be deleted.
658 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
664 if (f > rg->from && t < rg->to) { /* Must split region */
666 * Check for an entry in the cache before dropping
667 * lock and attempting allocation.
670 resv->region_cache_count > resv->adds_in_progress) {
671 nrg = list_first_entry(&resv->region_cache,
674 list_del(&nrg->link);
675 resv->region_cache_count--;
679 spin_unlock(&resv->lock);
680 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
687 hugetlb_cgroup_uncharge_file_region(
688 resv, rg, t - f, false);
690 /* New entry for end of split region */
694 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
696 INIT_LIST_HEAD(&nrg->link);
698 /* Original entry is trimmed */
701 list_add(&nrg->link, &rg->link);
706 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
707 del += rg->to - rg->from;
708 hugetlb_cgroup_uncharge_file_region(resv, rg,
709 rg->to - rg->from, true);
715 if (f <= rg->from) { /* Trim beginning of region */
716 hugetlb_cgroup_uncharge_file_region(resv, rg,
717 t - rg->from, false);
721 } else { /* Trim end of region */
722 hugetlb_cgroup_uncharge_file_region(resv, rg,
730 spin_unlock(&resv->lock);
736 * A rare out of memory error was encountered which prevented removal of
737 * the reserve map region for a page. The huge page itself was free'ed
738 * and removed from the page cache. This routine will adjust the subpool
739 * usage count, and the global reserve count if needed. By incrementing
740 * these counts, the reserve map entry which could not be deleted will
741 * appear as a "reserved" entry instead of simply dangling with incorrect
744 void hugetlb_fix_reserve_counts(struct inode *inode)
746 struct hugepage_subpool *spool = subpool_inode(inode);
748 bool reserved = false;
750 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
751 if (rsv_adjust > 0) {
752 struct hstate *h = hstate_inode(inode);
754 if (!hugetlb_acct_memory(h, 1))
756 } else if (!rsv_adjust) {
761 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
765 * Count and return the number of huge pages in the reserve map
766 * that intersect with the range [f, t).
768 static long region_count(struct resv_map *resv, long f, long t)
770 struct list_head *head = &resv->regions;
771 struct file_region *rg;
774 spin_lock(&resv->lock);
775 /* Locate each segment we overlap with, and count that overlap. */
776 list_for_each_entry(rg, head, link) {
785 seg_from = max(rg->from, f);
786 seg_to = min(rg->to, t);
788 chg += seg_to - seg_from;
790 spin_unlock(&resv->lock);
796 * Convert the address within this vma to the page offset within
797 * the mapping, in pagecache page units; huge pages here.
799 static pgoff_t vma_hugecache_offset(struct hstate *h,
800 struct vm_area_struct *vma, unsigned long address)
802 return ((address - vma->vm_start) >> huge_page_shift(h)) +
803 (vma->vm_pgoff >> huge_page_order(h));
806 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
807 unsigned long address)
809 return vma_hugecache_offset(hstate_vma(vma), vma, address);
811 EXPORT_SYMBOL_GPL(linear_hugepage_index);
814 * Return the size of the pages allocated when backing a VMA. In the majority
815 * cases this will be same size as used by the page table entries.
817 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
819 if (vma->vm_ops && vma->vm_ops->pagesize)
820 return vma->vm_ops->pagesize(vma);
823 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
826 * Return the page size being used by the MMU to back a VMA. In the majority
827 * of cases, the page size used by the kernel matches the MMU size. On
828 * architectures where it differs, an architecture-specific 'strong'
829 * version of this symbol is required.
831 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
833 return vma_kernel_pagesize(vma);
837 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
838 * bits of the reservation map pointer, which are always clear due to
841 #define HPAGE_RESV_OWNER (1UL << 0)
842 #define HPAGE_RESV_UNMAPPED (1UL << 1)
843 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
846 * These helpers are used to track how many pages are reserved for
847 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
848 * is guaranteed to have their future faults succeed.
850 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
851 * the reserve counters are updated with the hugetlb_lock held. It is safe
852 * to reset the VMA at fork() time as it is not in use yet and there is no
853 * chance of the global counters getting corrupted as a result of the values.
855 * The private mapping reservation is represented in a subtly different
856 * manner to a shared mapping. A shared mapping has a region map associated
857 * with the underlying file, this region map represents the backing file
858 * pages which have ever had a reservation assigned which this persists even
859 * after the page is instantiated. A private mapping has a region map
860 * associated with the original mmap which is attached to all VMAs which
861 * reference it, this region map represents those offsets which have consumed
862 * reservation ie. where pages have been instantiated.
864 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
866 return (unsigned long)vma->vm_private_data;
869 static void set_vma_private_data(struct vm_area_struct *vma,
872 vma->vm_private_data = (void *)value;
876 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
877 struct hugetlb_cgroup *h_cg,
880 #ifdef CONFIG_CGROUP_HUGETLB
882 resv_map->reservation_counter = NULL;
883 resv_map->pages_per_hpage = 0;
884 resv_map->css = NULL;
886 resv_map->reservation_counter =
887 &h_cg->rsvd_hugepage[hstate_index(h)];
888 resv_map->pages_per_hpage = pages_per_huge_page(h);
889 resv_map->css = &h_cg->css;
894 struct resv_map *resv_map_alloc(void)
896 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
897 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
899 if (!resv_map || !rg) {
905 kref_init(&resv_map->refs);
906 spin_lock_init(&resv_map->lock);
907 INIT_LIST_HEAD(&resv_map->regions);
909 resv_map->adds_in_progress = 0;
911 * Initialize these to 0. On shared mappings, 0's here indicate these
912 * fields don't do cgroup accounting. On private mappings, these will be
913 * re-initialized to the proper values, to indicate that hugetlb cgroup
914 * reservations are to be un-charged from here.
916 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
918 INIT_LIST_HEAD(&resv_map->region_cache);
919 list_add(&rg->link, &resv_map->region_cache);
920 resv_map->region_cache_count = 1;
925 void resv_map_release(struct kref *ref)
927 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
928 struct list_head *head = &resv_map->region_cache;
929 struct file_region *rg, *trg;
931 /* Clear out any active regions before we release the map. */
932 region_del(resv_map, 0, LONG_MAX);
934 /* ... and any entries left in the cache */
935 list_for_each_entry_safe(rg, trg, head, link) {
940 VM_BUG_ON(resv_map->adds_in_progress);
945 static inline struct resv_map *inode_resv_map(struct inode *inode)
948 * At inode evict time, i_mapping may not point to the original
949 * address space within the inode. This original address space
950 * contains the pointer to the resv_map. So, always use the
951 * address space embedded within the inode.
952 * The VERY common case is inode->mapping == &inode->i_data but,
953 * this may not be true for device special inodes.
955 return (struct resv_map *)(&inode->i_data)->private_data;
958 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
960 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
961 if (vma->vm_flags & VM_MAYSHARE) {
962 struct address_space *mapping = vma->vm_file->f_mapping;
963 struct inode *inode = mapping->host;
965 return inode_resv_map(inode);
968 return (struct resv_map *)(get_vma_private_data(vma) &
973 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
975 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
976 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
978 set_vma_private_data(vma, (get_vma_private_data(vma) &
979 HPAGE_RESV_MASK) | (unsigned long)map);
982 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
984 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
985 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
987 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
990 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
992 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
994 return (get_vma_private_data(vma) & flag) != 0;
997 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
998 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
1000 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1001 if (!(vma->vm_flags & VM_MAYSHARE))
1002 vma->vm_private_data = (void *)0;
1005 /* Returns true if the VMA has associated reserve pages */
1006 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1008 if (vma->vm_flags & VM_NORESERVE) {
1010 * This address is already reserved by other process(chg == 0),
1011 * so, we should decrement reserved count. Without decrementing,
1012 * reserve count remains after releasing inode, because this
1013 * allocated page will go into page cache and is regarded as
1014 * coming from reserved pool in releasing step. Currently, we
1015 * don't have any other solution to deal with this situation
1016 * properly, so add work-around here.
1018 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1024 /* Shared mappings always use reserves */
1025 if (vma->vm_flags & VM_MAYSHARE) {
1027 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1028 * be a region map for all pages. The only situation where
1029 * there is no region map is if a hole was punched via
1030 * fallocate. In this case, there really are no reserves to
1031 * use. This situation is indicated if chg != 0.
1040 * Only the process that called mmap() has reserves for
1043 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1045 * Like the shared case above, a hole punch or truncate
1046 * could have been performed on the private mapping.
1047 * Examine the value of chg to determine if reserves
1048 * actually exist or were previously consumed.
1049 * Very Subtle - The value of chg comes from a previous
1050 * call to vma_needs_reserves(). The reserve map for
1051 * private mappings has different (opposite) semantics
1052 * than that of shared mappings. vma_needs_reserves()
1053 * has already taken this difference in semantics into
1054 * account. Therefore, the meaning of chg is the same
1055 * as in the shared case above. Code could easily be
1056 * combined, but keeping it separate draws attention to
1057 * subtle differences.
1068 static void enqueue_huge_page(struct hstate *h, struct page *page)
1070 int nid = page_to_nid(page);
1071 list_move(&page->lru, &h->hugepage_freelists[nid]);
1072 h->free_huge_pages++;
1073 h->free_huge_pages_node[nid]++;
1074 SetPageHugeFreed(page);
1077 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1080 bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA);
1082 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1083 if (nocma && is_migrate_cma_page(page))
1086 if (PageHWPoison(page))
1089 list_move(&page->lru, &h->hugepage_activelist);
1090 set_page_refcounted(page);
1091 ClearPageHugeFreed(page);
1092 h->free_huge_pages--;
1093 h->free_huge_pages_node[nid]--;
1100 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1103 unsigned int cpuset_mems_cookie;
1104 struct zonelist *zonelist;
1107 int node = NUMA_NO_NODE;
1109 zonelist = node_zonelist(nid, gfp_mask);
1112 cpuset_mems_cookie = read_mems_allowed_begin();
1113 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1116 if (!cpuset_zone_allowed(zone, gfp_mask))
1119 * no need to ask again on the same node. Pool is node rather than
1122 if (zone_to_nid(zone) == node)
1124 node = zone_to_nid(zone);
1126 page = dequeue_huge_page_node_exact(h, node);
1130 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1136 static struct page *dequeue_huge_page_vma(struct hstate *h,
1137 struct vm_area_struct *vma,
1138 unsigned long address, int avoid_reserve,
1142 struct mempolicy *mpol;
1144 nodemask_t *nodemask;
1148 * A child process with MAP_PRIVATE mappings created by their parent
1149 * have no page reserves. This check ensures that reservations are
1150 * not "stolen". The child may still get SIGKILLed
1152 if (!vma_has_reserves(vma, chg) &&
1153 h->free_huge_pages - h->resv_huge_pages == 0)
1156 /* If reserves cannot be used, ensure enough pages are in the pool */
1157 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1160 gfp_mask = htlb_alloc_mask(h);
1161 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1162 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1163 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1164 SetPagePrivate(page);
1165 h->resv_huge_pages--;
1168 mpol_cond_put(mpol);
1176 * common helper functions for hstate_next_node_to_{alloc|free}.
1177 * We may have allocated or freed a huge page based on a different
1178 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1179 * be outside of *nodes_allowed. Ensure that we use an allowed
1180 * node for alloc or free.
1182 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1184 nid = next_node_in(nid, *nodes_allowed);
1185 VM_BUG_ON(nid >= MAX_NUMNODES);
1190 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1192 if (!node_isset(nid, *nodes_allowed))
1193 nid = next_node_allowed(nid, nodes_allowed);
1198 * returns the previously saved node ["this node"] from which to
1199 * allocate a persistent huge page for the pool and advance the
1200 * next node from which to allocate, handling wrap at end of node
1203 static int hstate_next_node_to_alloc(struct hstate *h,
1204 nodemask_t *nodes_allowed)
1208 VM_BUG_ON(!nodes_allowed);
1210 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1211 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1217 * helper for free_pool_huge_page() - return the previously saved
1218 * node ["this node"] from which to free a huge page. Advance the
1219 * next node id whether or not we find a free huge page to free so
1220 * that the next attempt to free addresses the next node.
1222 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1226 VM_BUG_ON(!nodes_allowed);
1228 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1229 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1234 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1235 for (nr_nodes = nodes_weight(*mask); \
1237 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1240 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1241 for (nr_nodes = nodes_weight(*mask); \
1243 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1246 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1247 static void destroy_compound_gigantic_page(struct page *page,
1251 int nr_pages = 1 << order;
1252 struct page *p = page + 1;
1254 atomic_set(compound_mapcount_ptr(page), 0);
1255 atomic_set(compound_pincount_ptr(page), 0);
1257 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1258 clear_compound_head(p);
1259 set_page_refcounted(p);
1262 set_compound_order(page, 0);
1263 page[1].compound_nr = 0;
1264 __ClearPageHead(page);
1267 static void free_gigantic_page(struct page *page, unsigned int order)
1270 * If the page isn't allocated using the cma allocator,
1271 * cma_release() returns false.
1274 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1278 free_contig_range(page_to_pfn(page), 1 << order);
1281 #ifdef CONFIG_CONTIG_ALLOC
1282 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1283 int nid, nodemask_t *nodemask)
1285 unsigned long nr_pages = 1UL << huge_page_order(h);
1286 if (nid == NUMA_NO_NODE)
1287 nid = numa_mem_id();
1294 if (hugetlb_cma[nid]) {
1295 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1296 huge_page_order(h), true);
1301 if (!(gfp_mask & __GFP_THISNODE)) {
1302 for_each_node_mask(node, *nodemask) {
1303 if (node == nid || !hugetlb_cma[node])
1306 page = cma_alloc(hugetlb_cma[node], nr_pages,
1307 huge_page_order(h), true);
1315 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1318 #else /* !CONFIG_CONTIG_ALLOC */
1319 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1320 int nid, nodemask_t *nodemask)
1324 #endif /* CONFIG_CONTIG_ALLOC */
1326 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1327 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1328 int nid, nodemask_t *nodemask)
1332 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1333 static inline void destroy_compound_gigantic_page(struct page *page,
1334 unsigned int order) { }
1337 static void update_and_free_page(struct hstate *h, struct page *page)
1340 struct page *subpage = page;
1342 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1346 h->nr_huge_pages_node[page_to_nid(page)]--;
1347 for (i = 0; i < pages_per_huge_page(h);
1348 i++, subpage = mem_map_next(subpage, page, i)) {
1349 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1350 1 << PG_referenced | 1 << PG_dirty |
1351 1 << PG_active | 1 << PG_private |
1354 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1355 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1356 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1357 set_page_refcounted(page);
1358 if (hstate_is_gigantic(h)) {
1360 * Temporarily drop the hugetlb_lock, because
1361 * we might block in free_gigantic_page().
1363 spin_unlock(&hugetlb_lock);
1364 destroy_compound_gigantic_page(page, huge_page_order(h));
1365 free_gigantic_page(page, huge_page_order(h));
1366 spin_lock(&hugetlb_lock);
1368 __free_pages(page, huge_page_order(h));
1372 struct hstate *size_to_hstate(unsigned long size)
1376 for_each_hstate(h) {
1377 if (huge_page_size(h) == size)
1384 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1385 * to hstate->hugepage_activelist.)
1387 * This function can be called for tail pages, but never returns true for them.
1389 bool page_huge_active(struct page *page)
1391 return PageHeadHuge(page) && PagePrivate(&page[1]);
1394 /* never called for tail page */
1395 void set_page_huge_active(struct page *page)
1397 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1398 SetPagePrivate(&page[1]);
1401 static void clear_page_huge_active(struct page *page)
1403 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1404 ClearPagePrivate(&page[1]);
1408 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1411 static inline bool PageHugeTemporary(struct page *page)
1413 if (!PageHuge(page))
1416 return (unsigned long)page[2].mapping == -1U;
1419 static inline void SetPageHugeTemporary(struct page *page)
1421 page[2].mapping = (void *)-1U;
1424 static inline void ClearPageHugeTemporary(struct page *page)
1426 page[2].mapping = NULL;
1429 static void __free_huge_page(struct page *page)
1432 * Can't pass hstate in here because it is called from the
1433 * compound page destructor.
1435 struct hstate *h = page_hstate(page);
1436 int nid = page_to_nid(page);
1437 struct hugepage_subpool *spool =
1438 (struct hugepage_subpool *)page_private(page);
1439 bool restore_reserve;
1441 VM_BUG_ON_PAGE(page_count(page), page);
1442 VM_BUG_ON_PAGE(page_mapcount(page), page);
1444 set_page_private(page, 0);
1445 page->mapping = NULL;
1446 restore_reserve = PagePrivate(page);
1447 ClearPagePrivate(page);
1450 * If PagePrivate() was set on page, page allocation consumed a
1451 * reservation. If the page was associated with a subpool, there
1452 * would have been a page reserved in the subpool before allocation
1453 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1454 * reservtion, do not call hugepage_subpool_put_pages() as this will
1455 * remove the reserved page from the subpool.
1457 if (!restore_reserve) {
1459 * A return code of zero implies that the subpool will be
1460 * under its minimum size if the reservation is not restored
1461 * after page is free. Therefore, force restore_reserve
1464 if (hugepage_subpool_put_pages(spool, 1) == 0)
1465 restore_reserve = true;
1468 spin_lock(&hugetlb_lock);
1469 clear_page_huge_active(page);
1470 hugetlb_cgroup_uncharge_page(hstate_index(h),
1471 pages_per_huge_page(h), page);
1472 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1473 pages_per_huge_page(h), page);
1474 if (restore_reserve)
1475 h->resv_huge_pages++;
1477 if (PageHugeTemporary(page)) {
1478 list_del(&page->lru);
1479 ClearPageHugeTemporary(page);
1480 update_and_free_page(h, page);
1481 } else if (h->surplus_huge_pages_node[nid]) {
1482 /* remove the page from active list */
1483 list_del(&page->lru);
1484 update_and_free_page(h, page);
1485 h->surplus_huge_pages--;
1486 h->surplus_huge_pages_node[nid]--;
1488 arch_clear_hugepage_flags(page);
1489 enqueue_huge_page(h, page);
1491 spin_unlock(&hugetlb_lock);
1495 * As free_huge_page() can be called from a non-task context, we have
1496 * to defer the actual freeing in a workqueue to prevent potential
1497 * hugetlb_lock deadlock.
1499 * free_hpage_workfn() locklessly retrieves the linked list of pages to
1500 * be freed and frees them one-by-one. As the page->mapping pointer is
1501 * going to be cleared in __free_huge_page() anyway, it is reused as the
1502 * llist_node structure of a lockless linked list of huge pages to be freed.
1504 static LLIST_HEAD(hpage_freelist);
1506 static void free_hpage_workfn(struct work_struct *work)
1508 struct llist_node *node;
1511 node = llist_del_all(&hpage_freelist);
1514 page = container_of((struct address_space **)node,
1515 struct page, mapping);
1517 __free_huge_page(page);
1520 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1522 void free_huge_page(struct page *page)
1525 * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1529 * Only call schedule_work() if hpage_freelist is previously
1530 * empty. Otherwise, schedule_work() had been called but the
1531 * workfn hasn't retrieved the list yet.
1533 if (llist_add((struct llist_node *)&page->mapping,
1535 schedule_work(&free_hpage_work);
1539 __free_huge_page(page);
1542 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1544 INIT_LIST_HEAD(&page->lru);
1545 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1546 set_hugetlb_cgroup(page, NULL);
1547 set_hugetlb_cgroup_rsvd(page, NULL);
1548 spin_lock(&hugetlb_lock);
1550 h->nr_huge_pages_node[nid]++;
1551 ClearPageHugeFreed(page);
1552 spin_unlock(&hugetlb_lock);
1555 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1558 int nr_pages = 1 << order;
1559 struct page *p = page + 1;
1561 /* we rely on prep_new_huge_page to set the destructor */
1562 set_compound_order(page, order);
1563 __ClearPageReserved(page);
1564 __SetPageHead(page);
1565 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1567 * For gigantic hugepages allocated through bootmem at
1568 * boot, it's safer to be consistent with the not-gigantic
1569 * hugepages and clear the PG_reserved bit from all tail pages
1570 * too. Otherwise drivers using get_user_pages() to access tail
1571 * pages may get the reference counting wrong if they see
1572 * PG_reserved set on a tail page (despite the head page not
1573 * having PG_reserved set). Enforcing this consistency between
1574 * head and tail pages allows drivers to optimize away a check
1575 * on the head page when they need know if put_page() is needed
1576 * after get_user_pages().
1578 __ClearPageReserved(p);
1579 set_page_count(p, 0);
1580 set_compound_head(p, page);
1582 atomic_set(compound_mapcount_ptr(page), -1);
1583 atomic_set(compound_pincount_ptr(page), 0);
1587 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1588 * transparent huge pages. See the PageTransHuge() documentation for more
1591 int PageHuge(struct page *page)
1593 if (!PageCompound(page))
1596 page = compound_head(page);
1597 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1599 EXPORT_SYMBOL_GPL(PageHuge);
1602 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1603 * normal or transparent huge pages.
1605 int PageHeadHuge(struct page *page_head)
1607 if (!PageHead(page_head))
1610 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1614 * Find and lock address space (mapping) in write mode.
1616 * Upon entry, the page is locked which means that page_mapping() is
1617 * stable. Due to locking order, we can only trylock_write. If we can
1618 * not get the lock, simply return NULL to caller.
1620 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1622 struct address_space *mapping = page_mapping(hpage);
1627 if (i_mmap_trylock_write(mapping))
1633 pgoff_t hugetlb_basepage_index(struct page *page)
1635 struct page *page_head = compound_head(page);
1636 pgoff_t index = page_index(page_head);
1637 unsigned long compound_idx;
1639 if (compound_order(page_head) >= MAX_ORDER)
1640 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1642 compound_idx = page - page_head;
1644 return (index << compound_order(page_head)) + compound_idx;
1647 static struct page *alloc_buddy_huge_page(struct hstate *h,
1648 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1649 nodemask_t *node_alloc_noretry)
1651 int order = huge_page_order(h);
1653 bool alloc_try_hard = true;
1656 * By default we always try hard to allocate the page with
1657 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1658 * a loop (to adjust global huge page counts) and previous allocation
1659 * failed, do not continue to try hard on the same node. Use the
1660 * node_alloc_noretry bitmap to manage this state information.
1662 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1663 alloc_try_hard = false;
1664 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1666 gfp_mask |= __GFP_RETRY_MAYFAIL;
1667 if (nid == NUMA_NO_NODE)
1668 nid = numa_mem_id();
1669 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1671 __count_vm_event(HTLB_BUDDY_PGALLOC);
1673 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1676 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1677 * indicates an overall state change. Clear bit so that we resume
1678 * normal 'try hard' allocations.
1680 if (node_alloc_noretry && page && !alloc_try_hard)
1681 node_clear(nid, *node_alloc_noretry);
1684 * If we tried hard to get a page but failed, set bit so that
1685 * subsequent attempts will not try as hard until there is an
1686 * overall state change.
1688 if (node_alloc_noretry && !page && alloc_try_hard)
1689 node_set(nid, *node_alloc_noretry);
1695 * Common helper to allocate a fresh hugetlb page. All specific allocators
1696 * should use this function to get new hugetlb pages
1698 static struct page *alloc_fresh_huge_page(struct hstate *h,
1699 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1700 nodemask_t *node_alloc_noretry)
1704 if (hstate_is_gigantic(h))
1705 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1707 page = alloc_buddy_huge_page(h, gfp_mask,
1708 nid, nmask, node_alloc_noretry);
1712 if (hstate_is_gigantic(h))
1713 prep_compound_gigantic_page(page, huge_page_order(h));
1714 prep_new_huge_page(h, page, page_to_nid(page));
1720 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1723 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1724 nodemask_t *node_alloc_noretry)
1728 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1730 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1731 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1732 node_alloc_noretry);
1740 put_page(page); /* free it into the hugepage allocator */
1746 * Free huge page from pool from next node to free.
1747 * Attempt to keep persistent huge pages more or less
1748 * balanced over allowed nodes.
1749 * Called with hugetlb_lock locked.
1751 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1757 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1759 * If we're returning unused surplus pages, only examine
1760 * nodes with surplus pages.
1762 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1763 !list_empty(&h->hugepage_freelists[node])) {
1765 list_entry(h->hugepage_freelists[node].next,
1767 list_del(&page->lru);
1768 h->free_huge_pages--;
1769 h->free_huge_pages_node[node]--;
1771 h->surplus_huge_pages--;
1772 h->surplus_huge_pages_node[node]--;
1774 update_and_free_page(h, page);
1784 * Dissolve a given free hugepage into free buddy pages. This function does
1785 * nothing for in-use hugepages and non-hugepages.
1786 * This function returns values like below:
1788 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1789 * (allocated or reserved.)
1790 * 0: successfully dissolved free hugepages or the page is not a
1791 * hugepage (considered as already dissolved)
1793 int dissolve_free_huge_page(struct page *page)
1798 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1799 if (!PageHuge(page))
1802 spin_lock(&hugetlb_lock);
1803 if (!PageHuge(page)) {
1808 if (!page_count(page)) {
1809 struct page *head = compound_head(page);
1810 struct hstate *h = page_hstate(head);
1811 int nid = page_to_nid(head);
1812 if (h->free_huge_pages - h->resv_huge_pages == 0)
1816 * We should make sure that the page is already on the free list
1817 * when it is dissolved.
1819 if (unlikely(!PageHugeFreed(head))) {
1820 spin_unlock(&hugetlb_lock);
1824 * Theoretically, we should return -EBUSY when we
1825 * encounter this race. In fact, we have a chance
1826 * to successfully dissolve the page if we do a
1827 * retry. Because the race window is quite small.
1828 * If we seize this opportunity, it is an optimization
1829 * for increasing the success rate of dissolving page.
1835 * Move PageHWPoison flag from head page to the raw error page,
1836 * which makes any subpages rather than the error page reusable.
1838 if (PageHWPoison(head) && page != head) {
1839 SetPageHWPoison(page);
1840 ClearPageHWPoison(head);
1842 list_del(&head->lru);
1843 h->free_huge_pages--;
1844 h->free_huge_pages_node[nid]--;
1845 h->max_huge_pages--;
1846 update_and_free_page(h, head);
1850 spin_unlock(&hugetlb_lock);
1855 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1856 * make specified memory blocks removable from the system.
1857 * Note that this will dissolve a free gigantic hugepage completely, if any
1858 * part of it lies within the given range.
1859 * Also note that if dissolve_free_huge_page() returns with an error, all
1860 * free hugepages that were dissolved before that error are lost.
1862 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1868 if (!hugepages_supported())
1871 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1872 page = pfn_to_page(pfn);
1873 rc = dissolve_free_huge_page(page);
1882 * Allocates a fresh surplus page from the page allocator.
1884 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1885 int nid, nodemask_t *nmask)
1887 struct page *page = NULL;
1889 if (hstate_is_gigantic(h))
1892 spin_lock(&hugetlb_lock);
1893 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1895 spin_unlock(&hugetlb_lock);
1897 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1901 spin_lock(&hugetlb_lock);
1903 * We could have raced with the pool size change.
1904 * Double check that and simply deallocate the new page
1905 * if we would end up overcommiting the surpluses. Abuse
1906 * temporary page to workaround the nasty free_huge_page
1909 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1910 SetPageHugeTemporary(page);
1911 spin_unlock(&hugetlb_lock);
1915 h->surplus_huge_pages++;
1916 h->surplus_huge_pages_node[page_to_nid(page)]++;
1920 spin_unlock(&hugetlb_lock);
1925 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1926 int nid, nodemask_t *nmask)
1930 if (hstate_is_gigantic(h))
1933 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1938 * We do not account these pages as surplus because they are only
1939 * temporary and will be released properly on the last reference
1941 SetPageHugeTemporary(page);
1947 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1950 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1951 struct vm_area_struct *vma, unsigned long addr)
1954 struct mempolicy *mpol;
1955 gfp_t gfp_mask = htlb_alloc_mask(h);
1957 nodemask_t *nodemask;
1959 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1960 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1961 mpol_cond_put(mpol);
1966 /* page migration callback function */
1967 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1968 nodemask_t *nmask, gfp_t gfp_mask)
1970 spin_lock(&hugetlb_lock);
1971 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1974 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1976 spin_unlock(&hugetlb_lock);
1980 spin_unlock(&hugetlb_lock);
1982 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1985 /* mempolicy aware migration callback */
1986 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1987 unsigned long address)
1989 struct mempolicy *mpol;
1990 nodemask_t *nodemask;
1995 gfp_mask = htlb_alloc_mask(h);
1996 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1997 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
1998 mpol_cond_put(mpol);
2004 * Increase the hugetlb pool such that it can accommodate a reservation
2007 static int gather_surplus_pages(struct hstate *h, int delta)
2008 __must_hold(&hugetlb_lock)
2010 struct list_head surplus_list;
2011 struct page *page, *tmp;
2013 int needed, allocated;
2014 bool alloc_ok = true;
2016 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2018 h->resv_huge_pages += delta;
2023 INIT_LIST_HEAD(&surplus_list);
2027 spin_unlock(&hugetlb_lock);
2028 for (i = 0; i < needed; i++) {
2029 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2030 NUMA_NO_NODE, NULL);
2035 list_add(&page->lru, &surplus_list);
2041 * After retaking hugetlb_lock, we need to recalculate 'needed'
2042 * because either resv_huge_pages or free_huge_pages may have changed.
2044 spin_lock(&hugetlb_lock);
2045 needed = (h->resv_huge_pages + delta) -
2046 (h->free_huge_pages + allocated);
2051 * We were not able to allocate enough pages to
2052 * satisfy the entire reservation so we free what
2053 * we've allocated so far.
2058 * The surplus_list now contains _at_least_ the number of extra pages
2059 * needed to accommodate the reservation. Add the appropriate number
2060 * of pages to the hugetlb pool and free the extras back to the buddy
2061 * allocator. Commit the entire reservation here to prevent another
2062 * process from stealing the pages as they are added to the pool but
2063 * before they are reserved.
2065 needed += allocated;
2066 h->resv_huge_pages += delta;
2069 /* Free the needed pages to the hugetlb pool */
2070 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2074 * This page is now managed by the hugetlb allocator and has
2075 * no users -- drop the buddy allocator's reference.
2077 put_page_testzero(page);
2078 VM_BUG_ON_PAGE(page_count(page), page);
2079 enqueue_huge_page(h, page);
2082 spin_unlock(&hugetlb_lock);
2084 /* Free unnecessary surplus pages to the buddy allocator */
2085 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2087 spin_lock(&hugetlb_lock);
2093 * This routine has two main purposes:
2094 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2095 * in unused_resv_pages. This corresponds to the prior adjustments made
2096 * to the associated reservation map.
2097 * 2) Free any unused surplus pages that may have been allocated to satisfy
2098 * the reservation. As many as unused_resv_pages may be freed.
2100 * Called with hugetlb_lock held. However, the lock could be dropped (and
2101 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
2102 * we must make sure nobody else can claim pages we are in the process of
2103 * freeing. Do this by ensuring resv_huge_page always is greater than the
2104 * number of huge pages we plan to free when dropping the lock.
2106 static void return_unused_surplus_pages(struct hstate *h,
2107 unsigned long unused_resv_pages)
2109 unsigned long nr_pages;
2111 /* Cannot return gigantic pages currently */
2112 if (hstate_is_gigantic(h))
2116 * Part (or even all) of the reservation could have been backed
2117 * by pre-allocated pages. Only free surplus pages.
2119 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2122 * We want to release as many surplus pages as possible, spread
2123 * evenly across all nodes with memory. Iterate across these nodes
2124 * until we can no longer free unreserved surplus pages. This occurs
2125 * when the nodes with surplus pages have no free pages.
2126 * free_pool_huge_page() will balance the freed pages across the
2127 * on-line nodes with memory and will handle the hstate accounting.
2129 * Note that we decrement resv_huge_pages as we free the pages. If
2130 * we drop the lock, resv_huge_pages will still be sufficiently large
2131 * to cover subsequent pages we may free.
2133 while (nr_pages--) {
2134 h->resv_huge_pages--;
2135 unused_resv_pages--;
2136 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
2138 cond_resched_lock(&hugetlb_lock);
2142 /* Fully uncommit the reservation */
2143 h->resv_huge_pages -= unused_resv_pages;
2148 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2149 * are used by the huge page allocation routines to manage reservations.
2151 * vma_needs_reservation is called to determine if the huge page at addr
2152 * within the vma has an associated reservation. If a reservation is
2153 * needed, the value 1 is returned. The caller is then responsible for
2154 * managing the global reservation and subpool usage counts. After
2155 * the huge page has been allocated, vma_commit_reservation is called
2156 * to add the page to the reservation map. If the page allocation fails,
2157 * the reservation must be ended instead of committed. vma_end_reservation
2158 * is called in such cases.
2160 * In the normal case, vma_commit_reservation returns the same value
2161 * as the preceding vma_needs_reservation call. The only time this
2162 * is not the case is if a reserve map was changed between calls. It
2163 * is the responsibility of the caller to notice the difference and
2164 * take appropriate action.
2166 * vma_add_reservation is used in error paths where a reservation must
2167 * be restored when a newly allocated huge page must be freed. It is
2168 * to be called after calling vma_needs_reservation to determine if a
2169 * reservation exists.
2171 enum vma_resv_mode {
2177 static long __vma_reservation_common(struct hstate *h,
2178 struct vm_area_struct *vma, unsigned long addr,
2179 enum vma_resv_mode mode)
2181 struct resv_map *resv;
2184 long dummy_out_regions_needed;
2186 resv = vma_resv_map(vma);
2190 idx = vma_hugecache_offset(h, vma, addr);
2192 case VMA_NEEDS_RESV:
2193 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2194 /* We assume that vma_reservation_* routines always operate on
2195 * 1 page, and that adding to resv map a 1 page entry can only
2196 * ever require 1 region.
2198 VM_BUG_ON(dummy_out_regions_needed != 1);
2200 case VMA_COMMIT_RESV:
2201 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2202 /* region_add calls of range 1 should never fail. */
2206 region_abort(resv, idx, idx + 1, 1);
2210 if (vma->vm_flags & VM_MAYSHARE) {
2211 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2212 /* region_add calls of range 1 should never fail. */
2215 region_abort(resv, idx, idx + 1, 1);
2216 ret = region_del(resv, idx, idx + 1);
2223 if (vma->vm_flags & VM_MAYSHARE)
2225 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
2227 * In most cases, reserves always exist for private mappings.
2228 * However, a file associated with mapping could have been
2229 * hole punched or truncated after reserves were consumed.
2230 * As subsequent fault on such a range will not use reserves.
2231 * Subtle - The reserve map for private mappings has the
2232 * opposite meaning than that of shared mappings. If NO
2233 * entry is in the reserve map, it means a reservation exists.
2234 * If an entry exists in the reserve map, it means the
2235 * reservation has already been consumed. As a result, the
2236 * return value of this routine is the opposite of the
2237 * value returned from reserve map manipulation routines above.
2245 return ret < 0 ? ret : 0;
2248 static long vma_needs_reservation(struct hstate *h,
2249 struct vm_area_struct *vma, unsigned long addr)
2251 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2254 static long vma_commit_reservation(struct hstate *h,
2255 struct vm_area_struct *vma, unsigned long addr)
2257 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2260 static void vma_end_reservation(struct hstate *h,
2261 struct vm_area_struct *vma, unsigned long addr)
2263 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2266 static long vma_add_reservation(struct hstate *h,
2267 struct vm_area_struct *vma, unsigned long addr)
2269 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2273 * This routine is called to restore a reservation on error paths. In the
2274 * specific error paths, a huge page was allocated (via alloc_huge_page)
2275 * and is about to be freed. If a reservation for the page existed,
2276 * alloc_huge_page would have consumed the reservation and set PagePrivate
2277 * in the newly allocated page. When the page is freed via free_huge_page,
2278 * the global reservation count will be incremented if PagePrivate is set.
2279 * However, free_huge_page can not adjust the reserve map. Adjust the
2280 * reserve map here to be consistent with global reserve count adjustments
2281 * to be made by free_huge_page.
2283 static void restore_reserve_on_error(struct hstate *h,
2284 struct vm_area_struct *vma, unsigned long address,
2287 if (unlikely(PagePrivate(page))) {
2288 long rc = vma_needs_reservation(h, vma, address);
2290 if (unlikely(rc < 0)) {
2292 * Rare out of memory condition in reserve map
2293 * manipulation. Clear PagePrivate so that
2294 * global reserve count will not be incremented
2295 * by free_huge_page. This will make it appear
2296 * as though the reservation for this page was
2297 * consumed. This may prevent the task from
2298 * faulting in the page at a later time. This
2299 * is better than inconsistent global huge page
2300 * accounting of reserve counts.
2302 ClearPagePrivate(page);
2304 rc = vma_add_reservation(h, vma, address);
2305 if (unlikely(rc < 0))
2307 * See above comment about rare out of
2310 ClearPagePrivate(page);
2312 vma_end_reservation(h, vma, address);
2316 struct page *alloc_huge_page(struct vm_area_struct *vma,
2317 unsigned long addr, int avoid_reserve)
2319 struct hugepage_subpool *spool = subpool_vma(vma);
2320 struct hstate *h = hstate_vma(vma);
2322 long map_chg, map_commit;
2325 struct hugetlb_cgroup *h_cg;
2326 bool deferred_reserve;
2328 idx = hstate_index(h);
2330 * Examine the region/reserve map to determine if the process
2331 * has a reservation for the page to be allocated. A return
2332 * code of zero indicates a reservation exists (no change).
2334 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2336 return ERR_PTR(-ENOMEM);
2339 * Processes that did not create the mapping will have no
2340 * reserves as indicated by the region/reserve map. Check
2341 * that the allocation will not exceed the subpool limit.
2342 * Allocations for MAP_NORESERVE mappings also need to be
2343 * checked against any subpool limit.
2345 if (map_chg || avoid_reserve) {
2346 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2348 vma_end_reservation(h, vma, addr);
2349 return ERR_PTR(-ENOSPC);
2353 * Even though there was no reservation in the region/reserve
2354 * map, there could be reservations associated with the
2355 * subpool that can be used. This would be indicated if the
2356 * return value of hugepage_subpool_get_pages() is zero.
2357 * However, if avoid_reserve is specified we still avoid even
2358 * the subpool reservations.
2364 /* If this allocation is not consuming a reservation, charge it now.
2366 deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma);
2367 if (deferred_reserve) {
2368 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2369 idx, pages_per_huge_page(h), &h_cg);
2371 goto out_subpool_put;
2374 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2376 goto out_uncharge_cgroup_reservation;
2378 spin_lock(&hugetlb_lock);
2380 * glb_chg is passed to indicate whether or not a page must be taken
2381 * from the global free pool (global change). gbl_chg == 0 indicates
2382 * a reservation exists for the allocation.
2384 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2386 spin_unlock(&hugetlb_lock);
2387 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2389 goto out_uncharge_cgroup;
2390 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2391 SetPagePrivate(page);
2392 h->resv_huge_pages--;
2394 spin_lock(&hugetlb_lock);
2395 list_add(&page->lru, &h->hugepage_activelist);
2398 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2399 /* If allocation is not consuming a reservation, also store the
2400 * hugetlb_cgroup pointer on the page.
2402 if (deferred_reserve) {
2403 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2407 spin_unlock(&hugetlb_lock);
2409 set_page_private(page, (unsigned long)spool);
2411 map_commit = vma_commit_reservation(h, vma, addr);
2412 if (unlikely(map_chg > map_commit)) {
2414 * The page was added to the reservation map between
2415 * vma_needs_reservation and vma_commit_reservation.
2416 * This indicates a race with hugetlb_reserve_pages.
2417 * Adjust for the subpool count incremented above AND
2418 * in hugetlb_reserve_pages for the same page. Also,
2419 * the reservation count added in hugetlb_reserve_pages
2420 * no longer applies.
2424 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2425 hugetlb_acct_memory(h, -rsv_adjust);
2426 if (deferred_reserve)
2427 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2428 pages_per_huge_page(h), page);
2432 out_uncharge_cgroup:
2433 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2434 out_uncharge_cgroup_reservation:
2435 if (deferred_reserve)
2436 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2439 if (map_chg || avoid_reserve)
2440 hugepage_subpool_put_pages(spool, 1);
2441 vma_end_reservation(h, vma, addr);
2442 return ERR_PTR(-ENOSPC);
2445 int alloc_bootmem_huge_page(struct hstate *h)
2446 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2447 int __alloc_bootmem_huge_page(struct hstate *h)
2449 struct huge_bootmem_page *m;
2452 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2455 addr = memblock_alloc_try_nid_raw(
2456 huge_page_size(h), huge_page_size(h),
2457 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2460 * Use the beginning of the huge page to store the
2461 * huge_bootmem_page struct (until gather_bootmem
2462 * puts them into the mem_map).
2471 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2472 /* Put them into a private list first because mem_map is not up yet */
2473 INIT_LIST_HEAD(&m->list);
2474 list_add(&m->list, &huge_boot_pages);
2480 * Put bootmem huge pages into the standard lists after mem_map is up.
2481 * Note: This only applies to gigantic (order > MAX_ORDER) pages.
2483 static void __init gather_bootmem_prealloc(void)
2485 struct huge_bootmem_page *m;
2487 list_for_each_entry(m, &huge_boot_pages, list) {
2488 struct page *page = virt_to_page(m);
2489 struct hstate *h = m->hstate;
2491 VM_BUG_ON(!hstate_is_gigantic(h));
2492 WARN_ON(page_count(page) != 1);
2493 prep_compound_gigantic_page(page, huge_page_order(h));
2494 WARN_ON(PageReserved(page));
2495 prep_new_huge_page(h, page, page_to_nid(page));
2496 put_page(page); /* free it into the hugepage allocator */
2499 * We need to restore the 'stolen' pages to totalram_pages
2500 * in order to fix confusing memory reports from free(1) and
2501 * other side-effects, like CommitLimit going negative.
2503 adjust_managed_page_count(page, pages_per_huge_page(h));
2508 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2511 nodemask_t *node_alloc_noretry;
2513 if (!hstate_is_gigantic(h)) {
2515 * Bit mask controlling how hard we retry per-node allocations.
2516 * Ignore errors as lower level routines can deal with
2517 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2518 * time, we are likely in bigger trouble.
2520 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2523 /* allocations done at boot time */
2524 node_alloc_noretry = NULL;
2527 /* bit mask controlling how hard we retry per-node allocations */
2528 if (node_alloc_noretry)
2529 nodes_clear(*node_alloc_noretry);
2531 for (i = 0; i < h->max_huge_pages; ++i) {
2532 if (hstate_is_gigantic(h)) {
2533 if (hugetlb_cma_size) {
2534 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2537 if (!alloc_bootmem_huge_page(h))
2539 } else if (!alloc_pool_huge_page(h,
2540 &node_states[N_MEMORY],
2541 node_alloc_noretry))
2545 if (i < h->max_huge_pages) {
2548 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2549 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2550 h->max_huge_pages, buf, i);
2551 h->max_huge_pages = i;
2554 kfree(node_alloc_noretry);
2557 static void __init hugetlb_init_hstates(void)
2561 for_each_hstate(h) {
2562 if (minimum_order > huge_page_order(h))
2563 minimum_order = huge_page_order(h);
2565 /* oversize hugepages were init'ed in early boot */
2566 if (!hstate_is_gigantic(h))
2567 hugetlb_hstate_alloc_pages(h);
2569 VM_BUG_ON(minimum_order == UINT_MAX);
2572 static void __init report_hugepages(void)
2576 for_each_hstate(h) {
2579 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2580 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2581 buf, h->free_huge_pages);
2585 #ifdef CONFIG_HIGHMEM
2586 static void try_to_free_low(struct hstate *h, unsigned long count,
2587 nodemask_t *nodes_allowed)
2591 if (hstate_is_gigantic(h))
2594 for_each_node_mask(i, *nodes_allowed) {
2595 struct page *page, *next;
2596 struct list_head *freel = &h->hugepage_freelists[i];
2597 list_for_each_entry_safe(page, next, freel, lru) {
2598 if (count >= h->nr_huge_pages)
2600 if (PageHighMem(page))
2602 list_del(&page->lru);
2603 update_and_free_page(h, page);
2604 h->free_huge_pages--;
2605 h->free_huge_pages_node[page_to_nid(page)]--;
2610 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2611 nodemask_t *nodes_allowed)
2617 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2618 * balanced by operating on them in a round-robin fashion.
2619 * Returns 1 if an adjustment was made.
2621 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2626 VM_BUG_ON(delta != -1 && delta != 1);
2629 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2630 if (h->surplus_huge_pages_node[node])
2634 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2635 if (h->surplus_huge_pages_node[node] <
2636 h->nr_huge_pages_node[node])
2643 h->surplus_huge_pages += delta;
2644 h->surplus_huge_pages_node[node] += delta;
2648 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2649 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2650 nodemask_t *nodes_allowed)
2652 unsigned long min_count, ret;
2653 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2656 * Bit mask controlling how hard we retry per-node allocations.
2657 * If we can not allocate the bit mask, do not attempt to allocate
2658 * the requested huge pages.
2660 if (node_alloc_noretry)
2661 nodes_clear(*node_alloc_noretry);
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 NODEMASK_FREE(node_alloc_noretry);
2700 /* Fall through to decrease pool */
2704 * Increase the pool size
2705 * First take pages out of surplus state. Then make up the
2706 * remaining difference by allocating fresh huge pages.
2708 * We might race with alloc_surplus_huge_page() here and be unable
2709 * to convert a surplus huge page to a normal huge page. That is
2710 * not critical, though, it just means the overall size of the
2711 * pool might be one hugepage larger than it needs to be, but
2712 * within all the constraints specified by the sysctls.
2714 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2715 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2719 while (count > persistent_huge_pages(h)) {
2721 * If this allocation races such that we no longer need the
2722 * page, free_huge_page will handle it by freeing the page
2723 * and reducing the surplus.
2725 spin_unlock(&hugetlb_lock);
2727 /* yield cpu to avoid soft lockup */
2730 ret = alloc_pool_huge_page(h, nodes_allowed,
2731 node_alloc_noretry);
2732 spin_lock(&hugetlb_lock);
2736 /* Bail for signals. Probably ctrl-c from user */
2737 if (signal_pending(current))
2742 * Decrease the pool size
2743 * First return free pages to the buddy allocator (being careful
2744 * to keep enough around to satisfy reservations). Then place
2745 * pages into surplus state as needed so the pool will shrink
2746 * to the desired size as pages become free.
2748 * By placing pages into the surplus state independent of the
2749 * overcommit value, we are allowing the surplus pool size to
2750 * exceed overcommit. There are few sane options here. Since
2751 * alloc_surplus_huge_page() is checking the global counter,
2752 * though, we'll note that we're not allowed to exceed surplus
2753 * and won't grow the pool anywhere else. Not until one of the
2754 * sysctls are changed, or the surplus pages go out of use.
2756 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2757 min_count = max(count, min_count);
2758 try_to_free_low(h, min_count, nodes_allowed);
2759 while (min_count < persistent_huge_pages(h)) {
2760 if (!free_pool_huge_page(h, nodes_allowed, 0))
2762 cond_resched_lock(&hugetlb_lock);
2764 while (count < persistent_huge_pages(h)) {
2765 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2769 h->max_huge_pages = persistent_huge_pages(h);
2770 spin_unlock(&hugetlb_lock);
2772 NODEMASK_FREE(node_alloc_noretry);
2777 #define HSTATE_ATTR_RO(_name) \
2778 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2780 #define HSTATE_ATTR(_name) \
2781 static struct kobj_attribute _name##_attr = \
2782 __ATTR(_name, 0644, _name##_show, _name##_store)
2784 static struct kobject *hugepages_kobj;
2785 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2787 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2789 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2793 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2794 if (hstate_kobjs[i] == kobj) {
2796 *nidp = NUMA_NO_NODE;
2800 return kobj_to_node_hstate(kobj, nidp);
2803 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2804 struct kobj_attribute *attr, char *buf)
2807 unsigned long nr_huge_pages;
2810 h = kobj_to_hstate(kobj, &nid);
2811 if (nid == NUMA_NO_NODE)
2812 nr_huge_pages = h->nr_huge_pages;
2814 nr_huge_pages = h->nr_huge_pages_node[nid];
2816 return sprintf(buf, "%lu\n", nr_huge_pages);
2819 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2820 struct hstate *h, int nid,
2821 unsigned long count, size_t len)
2824 nodemask_t nodes_allowed, *n_mask;
2826 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2829 if (nid == NUMA_NO_NODE) {
2831 * global hstate attribute
2833 if (!(obey_mempolicy &&
2834 init_nodemask_of_mempolicy(&nodes_allowed)))
2835 n_mask = &node_states[N_MEMORY];
2837 n_mask = &nodes_allowed;
2840 * Node specific request. count adjustment happens in
2841 * set_max_huge_pages() after acquiring hugetlb_lock.
2843 init_nodemask_of_node(&nodes_allowed, nid);
2844 n_mask = &nodes_allowed;
2847 err = set_max_huge_pages(h, count, nid, n_mask);
2849 return err ? err : len;
2852 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2853 struct kobject *kobj, const char *buf,
2857 unsigned long count;
2861 err = kstrtoul(buf, 10, &count);
2865 h = kobj_to_hstate(kobj, &nid);
2866 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2869 static ssize_t nr_hugepages_show(struct kobject *kobj,
2870 struct kobj_attribute *attr, char *buf)
2872 return nr_hugepages_show_common(kobj, attr, buf);
2875 static ssize_t nr_hugepages_store(struct kobject *kobj,
2876 struct kobj_attribute *attr, const char *buf, size_t len)
2878 return nr_hugepages_store_common(false, kobj, buf, len);
2880 HSTATE_ATTR(nr_hugepages);
2885 * hstate attribute for optionally mempolicy-based constraint on persistent
2886 * huge page alloc/free.
2888 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2889 struct kobj_attribute *attr, char *buf)
2891 return nr_hugepages_show_common(kobj, attr, buf);
2894 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2895 struct kobj_attribute *attr, const char *buf, size_t len)
2897 return nr_hugepages_store_common(true, kobj, buf, len);
2899 HSTATE_ATTR(nr_hugepages_mempolicy);
2903 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2904 struct kobj_attribute *attr, char *buf)
2906 struct hstate *h = kobj_to_hstate(kobj, NULL);
2907 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2910 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2911 struct kobj_attribute *attr, const char *buf, size_t count)
2914 unsigned long input;
2915 struct hstate *h = kobj_to_hstate(kobj, NULL);
2917 if (hstate_is_gigantic(h))
2920 err = kstrtoul(buf, 10, &input);
2924 spin_lock(&hugetlb_lock);
2925 h->nr_overcommit_huge_pages = input;
2926 spin_unlock(&hugetlb_lock);
2930 HSTATE_ATTR(nr_overcommit_hugepages);
2932 static ssize_t free_hugepages_show(struct kobject *kobj,
2933 struct kobj_attribute *attr, char *buf)
2936 unsigned long free_huge_pages;
2939 h = kobj_to_hstate(kobj, &nid);
2940 if (nid == NUMA_NO_NODE)
2941 free_huge_pages = h->free_huge_pages;
2943 free_huge_pages = h->free_huge_pages_node[nid];
2945 return sprintf(buf, "%lu\n", free_huge_pages);
2947 HSTATE_ATTR_RO(free_hugepages);
2949 static ssize_t resv_hugepages_show(struct kobject *kobj,
2950 struct kobj_attribute *attr, char *buf)
2952 struct hstate *h = kobj_to_hstate(kobj, NULL);
2953 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2955 HSTATE_ATTR_RO(resv_hugepages);
2957 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2958 struct kobj_attribute *attr, char *buf)
2961 unsigned long surplus_huge_pages;
2964 h = kobj_to_hstate(kobj, &nid);
2965 if (nid == NUMA_NO_NODE)
2966 surplus_huge_pages = h->surplus_huge_pages;
2968 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2970 return sprintf(buf, "%lu\n", surplus_huge_pages);
2972 HSTATE_ATTR_RO(surplus_hugepages);
2974 static struct attribute *hstate_attrs[] = {
2975 &nr_hugepages_attr.attr,
2976 &nr_overcommit_hugepages_attr.attr,
2977 &free_hugepages_attr.attr,
2978 &resv_hugepages_attr.attr,
2979 &surplus_hugepages_attr.attr,
2981 &nr_hugepages_mempolicy_attr.attr,
2986 static const struct attribute_group hstate_attr_group = {
2987 .attrs = hstate_attrs,
2990 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2991 struct kobject **hstate_kobjs,
2992 const struct attribute_group *hstate_attr_group)
2995 int hi = hstate_index(h);
2997 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2998 if (!hstate_kobjs[hi])
3001 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3003 kobject_put(hstate_kobjs[hi]);
3004 hstate_kobjs[hi] = NULL;
3010 static void __init hugetlb_sysfs_init(void)
3015 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3016 if (!hugepages_kobj)
3019 for_each_hstate(h) {
3020 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3021 hstate_kobjs, &hstate_attr_group);
3023 pr_err("HugeTLB: Unable to add hstate %s", h->name);
3030 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3031 * with node devices in node_devices[] using a parallel array. The array
3032 * index of a node device or _hstate == node id.
3033 * This is here to avoid any static dependency of the node device driver, in
3034 * the base kernel, on the hugetlb module.
3036 struct node_hstate {
3037 struct kobject *hugepages_kobj;
3038 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3040 static struct node_hstate node_hstates[MAX_NUMNODES];
3043 * A subset of global hstate attributes for node devices
3045 static struct attribute *per_node_hstate_attrs[] = {
3046 &nr_hugepages_attr.attr,
3047 &free_hugepages_attr.attr,
3048 &surplus_hugepages_attr.attr,
3052 static const struct attribute_group per_node_hstate_attr_group = {
3053 .attrs = per_node_hstate_attrs,
3057 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3058 * Returns node id via non-NULL nidp.
3060 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3064 for (nid = 0; nid < nr_node_ids; nid++) {
3065 struct node_hstate *nhs = &node_hstates[nid];
3067 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3068 if (nhs->hstate_kobjs[i] == kobj) {
3080 * Unregister hstate attributes from a single node device.
3081 * No-op if no hstate attributes attached.
3083 static void hugetlb_unregister_node(struct node *node)
3086 struct node_hstate *nhs = &node_hstates[node->dev.id];
3088 if (!nhs->hugepages_kobj)
3089 return; /* no hstate attributes */
3091 for_each_hstate(h) {
3092 int idx = hstate_index(h);
3093 if (nhs->hstate_kobjs[idx]) {
3094 kobject_put(nhs->hstate_kobjs[idx]);
3095 nhs->hstate_kobjs[idx] = NULL;
3099 kobject_put(nhs->hugepages_kobj);
3100 nhs->hugepages_kobj = NULL;
3105 * Register hstate attributes for a single node device.
3106 * No-op if attributes already registered.
3108 static void hugetlb_register_node(struct node *node)
3111 struct node_hstate *nhs = &node_hstates[node->dev.id];
3114 if (nhs->hugepages_kobj)
3115 return; /* already allocated */
3117 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3119 if (!nhs->hugepages_kobj)
3122 for_each_hstate(h) {
3123 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3125 &per_node_hstate_attr_group);
3127 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3128 h->name, node->dev.id);
3129 hugetlb_unregister_node(node);
3136 * hugetlb init time: register hstate attributes for all registered node
3137 * devices of nodes that have memory. All on-line nodes should have
3138 * registered their associated device by this time.
3140 static void __init hugetlb_register_all_nodes(void)
3144 for_each_node_state(nid, N_MEMORY) {
3145 struct node *node = node_devices[nid];
3146 if (node->dev.id == nid)
3147 hugetlb_register_node(node);
3151 * Let the node device driver know we're here so it can
3152 * [un]register hstate attributes on node hotplug.
3154 register_hugetlbfs_with_node(hugetlb_register_node,
3155 hugetlb_unregister_node);
3157 #else /* !CONFIG_NUMA */
3159 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3167 static void hugetlb_register_all_nodes(void) { }
3171 static int __init hugetlb_init(void)
3175 if (!hugepages_supported()) {
3176 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3177 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3182 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3183 * architectures depend on setup being done here.
3185 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3186 if (!parsed_default_hugepagesz) {
3188 * If we did not parse a default huge page size, set
3189 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3190 * number of huge pages for this default size was implicitly
3191 * specified, set that here as well.
3192 * Note that the implicit setting will overwrite an explicit
3193 * setting. A warning will be printed in this case.
3195 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3196 if (default_hstate_max_huge_pages) {
3197 if (default_hstate.max_huge_pages) {
3200 string_get_size(huge_page_size(&default_hstate),
3201 1, STRING_UNITS_2, buf, 32);
3202 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3203 default_hstate.max_huge_pages, buf);
3204 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3205 default_hstate_max_huge_pages);
3207 default_hstate.max_huge_pages =
3208 default_hstate_max_huge_pages;
3212 hugetlb_cma_check();
3213 hugetlb_init_hstates();
3214 gather_bootmem_prealloc();
3217 hugetlb_sysfs_init();
3218 hugetlb_register_all_nodes();
3219 hugetlb_cgroup_file_init();
3222 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3224 num_fault_mutexes = 1;
3226 hugetlb_fault_mutex_table =
3227 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3229 BUG_ON(!hugetlb_fault_mutex_table);
3231 for (i = 0; i < num_fault_mutexes; i++)
3232 mutex_init(&hugetlb_fault_mutex_table[i]);
3235 subsys_initcall(hugetlb_init);
3237 /* Overwritten by architectures with more huge page sizes */
3238 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3240 return size == HPAGE_SIZE;
3243 void __init hugetlb_add_hstate(unsigned int order)
3248 if (size_to_hstate(PAGE_SIZE << order)) {
3251 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3253 h = &hstates[hugetlb_max_hstate++];
3255 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
3256 h->nr_huge_pages = 0;
3257 h->free_huge_pages = 0;
3258 for (i = 0; i < MAX_NUMNODES; ++i)
3259 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3260 INIT_LIST_HEAD(&h->hugepage_activelist);
3261 h->next_nid_to_alloc = first_memory_node;
3262 h->next_nid_to_free = first_memory_node;
3263 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3264 huge_page_size(h)/1024);
3270 * hugepages command line processing
3271 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3272 * specification. If not, ignore the hugepages value. hugepages can also
3273 * be the first huge page command line option in which case it implicitly
3274 * specifies the number of huge pages for the default size.
3276 static int __init hugepages_setup(char *s)
3279 static unsigned long *last_mhp;
3281 if (!parsed_valid_hugepagesz) {
3282 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3283 parsed_valid_hugepagesz = true;
3288 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3289 * yet, so this hugepages= parameter goes to the "default hstate".
3290 * Otherwise, it goes with the previously parsed hugepagesz or
3291 * default_hugepagesz.
3293 else if (!hugetlb_max_hstate)
3294 mhp = &default_hstate_max_huge_pages;
3296 mhp = &parsed_hstate->max_huge_pages;
3298 if (mhp == last_mhp) {
3299 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3303 if (sscanf(s, "%lu", mhp) <= 0)
3307 * Global state is always initialized later in hugetlb_init.
3308 * But we need to allocate >= MAX_ORDER hstates here early to still
3309 * use the bootmem allocator.
3311 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
3312 hugetlb_hstate_alloc_pages(parsed_hstate);
3318 __setup("hugepages=", hugepages_setup);
3321 * hugepagesz command line processing
3322 * A specific huge page size can only be specified once with hugepagesz.
3323 * hugepagesz is followed by hugepages on the command line. The global
3324 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3325 * hugepagesz argument was valid.
3327 static int __init hugepagesz_setup(char *s)
3332 parsed_valid_hugepagesz = false;
3333 size = (unsigned long)memparse(s, NULL);
3335 if (!arch_hugetlb_valid_size(size)) {
3336 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3340 h = size_to_hstate(size);
3343 * hstate for this size already exists. This is normally
3344 * an error, but is allowed if the existing hstate is the
3345 * default hstate. More specifically, it is only allowed if
3346 * the number of huge pages for the default hstate was not
3347 * previously specified.
3349 if (!parsed_default_hugepagesz || h != &default_hstate ||
3350 default_hstate.max_huge_pages) {
3351 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3356 * No need to call hugetlb_add_hstate() as hstate already
3357 * exists. But, do set parsed_hstate so that a following
3358 * hugepages= parameter will be applied to this hstate.
3361 parsed_valid_hugepagesz = true;
3365 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3366 parsed_valid_hugepagesz = true;
3369 __setup("hugepagesz=", hugepagesz_setup);
3372 * default_hugepagesz command line input
3373 * Only one instance of default_hugepagesz allowed on command line.
3375 static int __init default_hugepagesz_setup(char *s)
3379 parsed_valid_hugepagesz = false;
3380 if (parsed_default_hugepagesz) {
3381 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3385 size = (unsigned long)memparse(s, NULL);
3387 if (!arch_hugetlb_valid_size(size)) {
3388 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3392 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3393 parsed_valid_hugepagesz = true;
3394 parsed_default_hugepagesz = true;
3395 default_hstate_idx = hstate_index(size_to_hstate(size));
3398 * The number of default huge pages (for this size) could have been
3399 * specified as the first hugetlb parameter: hugepages=X. If so,
3400 * then default_hstate_max_huge_pages is set. If the default huge
3401 * page size is gigantic (>= MAX_ORDER), then the pages must be
3402 * allocated here from bootmem allocator.
3404 if (default_hstate_max_huge_pages) {
3405 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3406 if (hstate_is_gigantic(&default_hstate))
3407 hugetlb_hstate_alloc_pages(&default_hstate);
3408 default_hstate_max_huge_pages = 0;
3413 __setup("default_hugepagesz=", default_hugepagesz_setup);
3415 static unsigned int allowed_mems_nr(struct hstate *h)
3418 unsigned int nr = 0;
3419 nodemask_t *mpol_allowed;
3420 unsigned int *array = h->free_huge_pages_node;
3421 gfp_t gfp_mask = htlb_alloc_mask(h);
3423 mpol_allowed = policy_nodemask_current(gfp_mask);
3425 for_each_node_mask(node, cpuset_current_mems_allowed) {
3426 if (!mpol_allowed ||
3427 (mpol_allowed && node_isset(node, *mpol_allowed)))
3434 #ifdef CONFIG_SYSCTL
3435 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3436 void *buffer, size_t *length,
3437 loff_t *ppos, unsigned long *out)
3439 struct ctl_table dup_table;
3442 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3443 * can duplicate the @table and alter the duplicate of it.
3446 dup_table.data = out;
3448 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3451 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3452 struct ctl_table *table, int write,
3453 void *buffer, size_t *length, loff_t *ppos)
3455 struct hstate *h = &default_hstate;
3456 unsigned long tmp = h->max_huge_pages;
3459 if (!hugepages_supported())
3462 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3468 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3469 NUMA_NO_NODE, tmp, *length);
3474 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3475 void *buffer, size_t *length, loff_t *ppos)
3478 return hugetlb_sysctl_handler_common(false, table, write,
3479 buffer, length, ppos);
3483 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3484 void *buffer, size_t *length, loff_t *ppos)
3486 return hugetlb_sysctl_handler_common(true, table, write,
3487 buffer, length, ppos);
3489 #endif /* CONFIG_NUMA */
3491 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3492 void *buffer, size_t *length, loff_t *ppos)
3494 struct hstate *h = &default_hstate;
3498 if (!hugepages_supported())
3501 tmp = h->nr_overcommit_huge_pages;
3503 if (write && hstate_is_gigantic(h))
3506 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3512 spin_lock(&hugetlb_lock);
3513 h->nr_overcommit_huge_pages = tmp;
3514 spin_unlock(&hugetlb_lock);
3520 #endif /* CONFIG_SYSCTL */
3522 void hugetlb_report_meminfo(struct seq_file *m)
3525 unsigned long total = 0;
3527 if (!hugepages_supported())
3530 for_each_hstate(h) {
3531 unsigned long count = h->nr_huge_pages;
3533 total += (PAGE_SIZE << huge_page_order(h)) * count;
3535 if (h == &default_hstate)
3537 "HugePages_Total: %5lu\n"
3538 "HugePages_Free: %5lu\n"
3539 "HugePages_Rsvd: %5lu\n"
3540 "HugePages_Surp: %5lu\n"
3541 "Hugepagesize: %8lu kB\n",
3545 h->surplus_huge_pages,
3546 (PAGE_SIZE << huge_page_order(h)) / 1024);
3549 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3552 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
3554 struct hstate *h = &default_hstate;
3556 if (!hugepages_supported())
3559 return sysfs_emit_at(buf, len,
3560 "Node %d HugePages_Total: %5u\n"
3561 "Node %d HugePages_Free: %5u\n"
3562 "Node %d HugePages_Surp: %5u\n",
3563 nid, h->nr_huge_pages_node[nid],
3564 nid, h->free_huge_pages_node[nid],
3565 nid, h->surplus_huge_pages_node[nid]);
3568 void hugetlb_show_meminfo(void)
3573 if (!hugepages_supported())
3576 for_each_node_state(nid, N_MEMORY)
3578 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3580 h->nr_huge_pages_node[nid],
3581 h->free_huge_pages_node[nid],
3582 h->surplus_huge_pages_node[nid],
3583 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3586 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3588 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3589 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3592 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3593 unsigned long hugetlb_total_pages(void)
3596 unsigned long nr_total_pages = 0;
3599 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3600 return nr_total_pages;
3603 static int hugetlb_acct_memory(struct hstate *h, long delta)
3607 spin_lock(&hugetlb_lock);
3609 * When cpuset is configured, it breaks the strict hugetlb page
3610 * reservation as the accounting is done on a global variable. Such
3611 * reservation is completely rubbish in the presence of cpuset because
3612 * the reservation is not checked against page availability for the
3613 * current cpuset. Application can still potentially OOM'ed by kernel
3614 * with lack of free htlb page in cpuset that the task is in.
3615 * Attempt to enforce strict accounting with cpuset is almost
3616 * impossible (or too ugly) because cpuset is too fluid that
3617 * task or memory node can be dynamically moved between cpusets.
3619 * The change of semantics for shared hugetlb mapping with cpuset is
3620 * undesirable. However, in order to preserve some of the semantics,
3621 * we fall back to check against current free page availability as
3622 * a best attempt and hopefully to minimize the impact of changing
3623 * semantics that cpuset has.
3625 * Apart from cpuset, we also have memory policy mechanism that
3626 * also determines from which node the kernel will allocate memory
3627 * in a NUMA system. So similar to cpuset, we also should consider
3628 * the memory policy of the current task. Similar to the description
3632 if (gather_surplus_pages(h, delta) < 0)
3635 if (delta > allowed_mems_nr(h)) {
3636 return_unused_surplus_pages(h, delta);
3643 return_unused_surplus_pages(h, (unsigned long) -delta);
3646 spin_unlock(&hugetlb_lock);
3650 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3652 struct resv_map *resv = vma_resv_map(vma);
3655 * This new VMA should share its siblings reservation map if present.
3656 * The VMA will only ever have a valid reservation map pointer where
3657 * it is being copied for another still existing VMA. As that VMA
3658 * has a reference to the reservation map it cannot disappear until
3659 * after this open call completes. It is therefore safe to take a
3660 * new reference here without additional locking.
3662 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
3663 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
3664 kref_get(&resv->refs);
3668 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3670 struct hstate *h = hstate_vma(vma);
3671 struct resv_map *resv = vma_resv_map(vma);
3672 struct hugepage_subpool *spool = subpool_vma(vma);
3673 unsigned long reserve, start, end;
3676 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3679 start = vma_hugecache_offset(h, vma, vma->vm_start);
3680 end = vma_hugecache_offset(h, vma, vma->vm_end);
3682 reserve = (end - start) - region_count(resv, start, end);
3683 hugetlb_cgroup_uncharge_counter(resv, start, end);
3686 * Decrement reserve counts. The global reserve count may be
3687 * adjusted if the subpool has a minimum size.
3689 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3690 hugetlb_acct_memory(h, -gbl_reserve);
3693 kref_put(&resv->refs, resv_map_release);
3696 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3698 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3703 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3705 struct hstate *hstate = hstate_vma(vma);
3707 return 1UL << huge_page_shift(hstate);
3711 * We cannot handle pagefaults against hugetlb pages at all. They cause
3712 * handle_mm_fault() to try to instantiate regular-sized pages in the
3713 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3716 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3723 * When a new function is introduced to vm_operations_struct and added
3724 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3725 * This is because under System V memory model, mappings created via
3726 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3727 * their original vm_ops are overwritten with shm_vm_ops.
3729 const struct vm_operations_struct hugetlb_vm_ops = {
3730 .fault = hugetlb_vm_op_fault,
3731 .open = hugetlb_vm_op_open,
3732 .close = hugetlb_vm_op_close,
3733 .split = hugetlb_vm_op_split,
3734 .pagesize = hugetlb_vm_op_pagesize,
3737 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3743 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3744 vma->vm_page_prot)));
3746 entry = huge_pte_wrprotect(mk_huge_pte(page,
3747 vma->vm_page_prot));
3749 entry = pte_mkyoung(entry);
3750 entry = pte_mkhuge(entry);
3751 entry = arch_make_huge_pte(entry, vma, page, writable);
3756 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3757 unsigned long address, pte_t *ptep)
3761 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3762 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3763 update_mmu_cache(vma, address, ptep);
3766 bool is_hugetlb_entry_migration(pte_t pte)
3770 if (huge_pte_none(pte) || pte_present(pte))
3772 swp = pte_to_swp_entry(pte);
3773 if (is_migration_entry(swp))
3779 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
3783 if (huge_pte_none(pte) || pte_present(pte))
3785 swp = pte_to_swp_entry(pte);
3786 if (is_hwpoison_entry(swp))
3792 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3793 struct vm_area_struct *vma)
3795 pte_t *src_pte, *dst_pte, entry, dst_entry;
3796 struct page *ptepage;
3799 struct hstate *h = hstate_vma(vma);
3800 unsigned long sz = huge_page_size(h);
3801 struct address_space *mapping = vma->vm_file->f_mapping;
3802 struct mmu_notifier_range range;
3805 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3808 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3811 mmu_notifier_invalidate_range_start(&range);
3814 * For shared mappings i_mmap_rwsem must be held to call
3815 * huge_pte_alloc, otherwise the returned ptep could go
3816 * away if part of a shared pmd and another thread calls
3819 i_mmap_lock_read(mapping);
3822 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3823 spinlock_t *src_ptl, *dst_ptl;
3824 src_pte = huge_pte_offset(src, addr, sz);
3827 dst_pte = huge_pte_alloc(dst, addr, sz);
3834 * If the pagetables are shared don't copy or take references.
3835 * dst_pte == src_pte is the common case of src/dest sharing.
3837 * However, src could have 'unshared' and dst shares with
3838 * another vma. If dst_pte !none, this implies sharing.
3839 * Check here before taking page table lock, and once again
3840 * after taking the lock below.
3842 dst_entry = huge_ptep_get(dst_pte);
3843 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3846 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3847 src_ptl = huge_pte_lockptr(h, src, src_pte);
3848 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3849 entry = huge_ptep_get(src_pte);
3850 dst_entry = huge_ptep_get(dst_pte);
3851 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3853 * Skip if src entry none. Also, skip in the
3854 * unlikely case dst entry !none as this implies
3855 * sharing with another vma.
3858 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3859 is_hugetlb_entry_hwpoisoned(entry))) {
3860 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3862 if (is_write_migration_entry(swp_entry) && cow) {
3864 * COW mappings require pages in both
3865 * parent and child to be set to read.
3867 make_migration_entry_read(&swp_entry);
3868 entry = swp_entry_to_pte(swp_entry);
3869 set_huge_swap_pte_at(src, addr, src_pte,
3872 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3876 * No need to notify as we are downgrading page
3877 * table protection not changing it to point
3880 * See Documentation/vm/mmu_notifier.rst
3882 huge_ptep_set_wrprotect(src, addr, src_pte);
3884 entry = huge_ptep_get(src_pte);
3885 ptepage = pte_page(entry);
3887 page_dup_rmap(ptepage, true);
3888 set_huge_pte_at(dst, addr, dst_pte, entry);
3889 hugetlb_count_add(pages_per_huge_page(h), dst);
3891 spin_unlock(src_ptl);
3892 spin_unlock(dst_ptl);
3896 mmu_notifier_invalidate_range_end(&range);
3898 i_mmap_unlock_read(mapping);
3903 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3904 unsigned long start, unsigned long end,
3905 struct page *ref_page)
3907 struct mm_struct *mm = vma->vm_mm;
3908 unsigned long address;
3913 struct hstate *h = hstate_vma(vma);
3914 unsigned long sz = huge_page_size(h);
3915 struct mmu_notifier_range range;
3916 bool force_flush = false;
3918 WARN_ON(!is_vm_hugetlb_page(vma));
3919 BUG_ON(start & ~huge_page_mask(h));
3920 BUG_ON(end & ~huge_page_mask(h));
3923 * This is a hugetlb vma, all the pte entries should point
3926 tlb_change_page_size(tlb, sz);
3927 tlb_start_vma(tlb, vma);
3930 * If sharing possible, alert mmu notifiers of worst case.
3932 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3934 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3935 mmu_notifier_invalidate_range_start(&range);
3937 for (; address < end; address += sz) {
3938 ptep = huge_pte_offset(mm, address, sz);
3942 ptl = huge_pte_lock(h, mm, ptep);
3943 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
3945 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
3950 pte = huge_ptep_get(ptep);
3951 if (huge_pte_none(pte)) {
3957 * Migrating hugepage or HWPoisoned hugepage is already
3958 * unmapped and its refcount is dropped, so just clear pte here.
3960 if (unlikely(!pte_present(pte))) {
3961 huge_pte_clear(mm, address, ptep, sz);
3966 page = pte_page(pte);
3968 * If a reference page is supplied, it is because a specific
3969 * page is being unmapped, not a range. Ensure the page we
3970 * are about to unmap is the actual page of interest.
3973 if (page != ref_page) {
3978 * Mark the VMA as having unmapped its page so that
3979 * future faults in this VMA will fail rather than
3980 * looking like data was lost
3982 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3985 pte = huge_ptep_get_and_clear(mm, address, ptep);
3986 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3987 if (huge_pte_dirty(pte))
3988 set_page_dirty(page);
3990 hugetlb_count_sub(pages_per_huge_page(h), mm);
3991 page_remove_rmap(page, true);
3994 tlb_remove_page_size(tlb, page, huge_page_size(h));
3996 * Bail out after unmapping reference page if supplied
4001 mmu_notifier_invalidate_range_end(&range);
4002 tlb_end_vma(tlb, vma);
4005 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
4006 * could defer the flush until now, since by holding i_mmap_rwsem we
4007 * guaranteed that the last refernece would not be dropped. But we must
4008 * do the flushing before we return, as otherwise i_mmap_rwsem will be
4009 * dropped and the last reference to the shared PMDs page might be
4012 * In theory we could defer the freeing of the PMD pages as well, but
4013 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
4014 * detect sharing, so we cannot defer the release of the page either.
4015 * Instead, do flush now.
4018 tlb_flush_mmu_tlbonly(tlb);
4021 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4022 struct vm_area_struct *vma, unsigned long start,
4023 unsigned long end, struct page *ref_page)
4025 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4028 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4029 * test will fail on a vma being torn down, and not grab a page table
4030 * on its way out. We're lucky that the flag has such an appropriate
4031 * name, and can in fact be safely cleared here. We could clear it
4032 * before the __unmap_hugepage_range above, but all that's necessary
4033 * is to clear it before releasing the i_mmap_rwsem. This works
4034 * because in the context this is called, the VMA is about to be
4035 * destroyed and the i_mmap_rwsem is held.
4037 vma->vm_flags &= ~VM_MAYSHARE;
4040 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
4041 unsigned long end, struct page *ref_page)
4043 struct mm_struct *mm;
4044 struct mmu_gather tlb;
4045 unsigned long tlb_start = start;
4046 unsigned long tlb_end = end;
4049 * If shared PMDs were possibly used within this vma range, adjust
4050 * start/end for worst case tlb flushing.
4051 * Note that we can not be sure if PMDs are shared until we try to
4052 * unmap pages. However, we want to make sure TLB flushing covers
4053 * the largest possible range.
4055 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
4059 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
4060 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
4061 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
4065 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4066 * mappping it owns the reserve page for. The intention is to unmap the page
4067 * from other VMAs and let the children be SIGKILLed if they are faulting the
4070 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4071 struct page *page, unsigned long address)
4073 struct hstate *h = hstate_vma(vma);
4074 struct vm_area_struct *iter_vma;
4075 struct address_space *mapping;
4079 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4080 * from page cache lookup which is in HPAGE_SIZE units.
4082 address = address & huge_page_mask(h);
4083 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4085 mapping = vma->vm_file->f_mapping;
4088 * Take the mapping lock for the duration of the table walk. As
4089 * this mapping should be shared between all the VMAs,
4090 * __unmap_hugepage_range() is called as the lock is already held
4092 i_mmap_lock_write(mapping);
4093 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
4094 /* Do not unmap the current VMA */
4095 if (iter_vma == vma)
4099 * Shared VMAs have their own reserves and do not affect
4100 * MAP_PRIVATE accounting but it is possible that a shared
4101 * VMA is using the same page so check and skip such VMAs.
4103 if (iter_vma->vm_flags & VM_MAYSHARE)
4107 * Unmap the page from other VMAs without their own reserves.
4108 * They get marked to be SIGKILLed if they fault in these
4109 * areas. This is because a future no-page fault on this VMA
4110 * could insert a zeroed page instead of the data existing
4111 * from the time of fork. This would look like data corruption
4113 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
4114 unmap_hugepage_range(iter_vma, address,
4115 address + huge_page_size(h), page);
4117 i_mmap_unlock_write(mapping);
4121 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4122 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4123 * cannot race with other handlers or page migration.
4124 * Keep the pte_same checks anyway to make transition from the mutex easier.
4126 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
4127 unsigned long address, pte_t *ptep,
4128 struct page *pagecache_page, spinlock_t *ptl)
4131 struct hstate *h = hstate_vma(vma);
4132 struct page *old_page, *new_page;
4133 int outside_reserve = 0;
4135 unsigned long haddr = address & huge_page_mask(h);
4136 struct mmu_notifier_range range;
4138 pte = huge_ptep_get(ptep);
4139 old_page = pte_page(pte);
4142 /* If no-one else is actually using this page, avoid the copy
4143 * and just make the page writable */
4144 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
4145 page_move_anon_rmap(old_page, vma);
4146 set_huge_ptep_writable(vma, haddr, ptep);
4151 * If the process that created a MAP_PRIVATE mapping is about to
4152 * perform a COW due to a shared page count, attempt to satisfy
4153 * the allocation without using the existing reserves. The pagecache
4154 * page is used to determine if the reserve at this address was
4155 * consumed or not. If reserves were used, a partial faulted mapping
4156 * at the time of fork() could consume its reserves on COW instead
4157 * of the full address range.
4159 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4160 old_page != pagecache_page)
4161 outside_reserve = 1;
4166 * Drop page table lock as buddy allocator may be called. It will
4167 * be acquired again before returning to the caller, as expected.
4170 new_page = alloc_huge_page(vma, haddr, outside_reserve);
4172 if (IS_ERR(new_page)) {
4174 * If a process owning a MAP_PRIVATE mapping fails to COW,
4175 * it is due to references held by a child and an insufficient
4176 * huge page pool. To guarantee the original mappers
4177 * reliability, unmap the page from child processes. The child
4178 * may get SIGKILLed if it later faults.
4180 if (outside_reserve) {
4181 struct address_space *mapping = vma->vm_file->f_mapping;
4186 BUG_ON(huge_pte_none(pte));
4188 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4189 * unmapping. unmapping needs to hold i_mmap_rwsem
4190 * in write mode. Dropping i_mmap_rwsem in read mode
4191 * here is OK as COW mappings do not interact with
4194 * Reacquire both after unmap operation.
4196 idx = vma_hugecache_offset(h, vma, haddr);
4197 hash = hugetlb_fault_mutex_hash(mapping, idx);
4198 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4199 i_mmap_unlock_read(mapping);
4201 unmap_ref_private(mm, vma, old_page, haddr);
4203 i_mmap_lock_read(mapping);
4204 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4206 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4208 pte_same(huge_ptep_get(ptep), pte)))
4209 goto retry_avoidcopy;
4211 * race occurs while re-acquiring page table
4212 * lock, and our job is done.
4217 ret = vmf_error(PTR_ERR(new_page));
4218 goto out_release_old;
4222 * When the original hugepage is shared one, it does not have
4223 * anon_vma prepared.
4225 if (unlikely(anon_vma_prepare(vma))) {
4227 goto out_release_all;
4230 copy_user_huge_page(new_page, old_page, address, vma,
4231 pages_per_huge_page(h));
4232 __SetPageUptodate(new_page);
4234 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4235 haddr + huge_page_size(h));
4236 mmu_notifier_invalidate_range_start(&range);
4239 * Retake the page table lock to check for racing updates
4240 * before the page tables are altered
4243 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4244 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4245 ClearPagePrivate(new_page);
4248 huge_ptep_clear_flush(vma, haddr, ptep);
4249 mmu_notifier_invalidate_range(mm, range.start, range.end);
4250 set_huge_pte_at(mm, haddr, ptep,
4251 make_huge_pte(vma, new_page, 1));
4252 page_remove_rmap(old_page, true);
4253 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4254 set_page_huge_active(new_page);
4255 /* Make the old page be freed below */
4256 new_page = old_page;
4259 mmu_notifier_invalidate_range_end(&range);
4261 restore_reserve_on_error(h, vma, haddr, new_page);
4266 spin_lock(ptl); /* Caller expects lock to be held */
4270 /* Return the pagecache page at a given address within a VMA */
4271 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4272 struct vm_area_struct *vma, unsigned long address)
4274 struct address_space *mapping;
4277 mapping = vma->vm_file->f_mapping;
4278 idx = vma_hugecache_offset(h, vma, address);
4280 return find_lock_page(mapping, idx);
4284 * Return whether there is a pagecache page to back given address within VMA.
4285 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4287 static bool hugetlbfs_pagecache_present(struct hstate *h,
4288 struct vm_area_struct *vma, unsigned long address)
4290 struct address_space *mapping;
4294 mapping = vma->vm_file->f_mapping;
4295 idx = vma_hugecache_offset(h, vma, address);
4297 page = find_get_page(mapping, idx);
4300 return page != NULL;
4303 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4306 struct inode *inode = mapping->host;
4307 struct hstate *h = hstate_inode(inode);
4308 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4312 ClearPagePrivate(page);
4315 * set page dirty so that it will not be removed from cache/file
4316 * by non-hugetlbfs specific code paths.
4318 set_page_dirty(page);
4320 spin_lock(&inode->i_lock);
4321 inode->i_blocks += blocks_per_huge_page(h);
4322 spin_unlock(&inode->i_lock);
4326 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4327 struct vm_area_struct *vma,
4328 struct address_space *mapping, pgoff_t idx,
4329 unsigned long address, pte_t *ptep, unsigned int flags)
4331 struct hstate *h = hstate_vma(vma);
4332 vm_fault_t ret = VM_FAULT_SIGBUS;
4338 unsigned long haddr = address & huge_page_mask(h);
4339 bool new_page = false;
4342 * Currently, we are forced to kill the process in the event the
4343 * original mapper has unmapped pages from the child due to a failed
4344 * COW. Warn that such a situation has occurred as it may not be obvious
4346 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4347 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4353 * We can not race with truncation due to holding i_mmap_rwsem.
4354 * i_size is modified when holding i_mmap_rwsem, so check here
4355 * once for faults beyond end of file.
4357 size = i_size_read(mapping->host) >> huge_page_shift(h);
4362 page = find_lock_page(mapping, idx);
4365 * Check for page in userfault range
4367 if (userfaultfd_missing(vma)) {
4369 struct vm_fault vmf = {
4374 * Hard to debug if it ends up being
4375 * used by a callee that assumes
4376 * something about the other
4377 * uninitialized fields... same as in
4383 * hugetlb_fault_mutex and i_mmap_rwsem must be
4384 * dropped before handling userfault. Reacquire
4385 * after handling fault to make calling code simpler.
4387 hash = hugetlb_fault_mutex_hash(mapping, idx);
4388 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4389 i_mmap_unlock_read(mapping);
4390 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
4391 i_mmap_lock_read(mapping);
4392 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4396 page = alloc_huge_page(vma, haddr, 0);
4399 * Returning error will result in faulting task being
4400 * sent SIGBUS. The hugetlb fault mutex prevents two
4401 * tasks from racing to fault in the same page which
4402 * could result in false unable to allocate errors.
4403 * Page migration does not take the fault mutex, but
4404 * does a clear then write of pte's under page table
4405 * lock. Page fault code could race with migration,
4406 * notice the clear pte and try to allocate a page
4407 * here. Before returning error, get ptl and make
4408 * sure there really is no pte entry.
4410 ptl = huge_pte_lock(h, mm, ptep);
4411 if (!huge_pte_none(huge_ptep_get(ptep))) {
4417 ret = vmf_error(PTR_ERR(page));
4420 clear_huge_page(page, address, pages_per_huge_page(h));
4421 __SetPageUptodate(page);
4424 if (vma->vm_flags & VM_MAYSHARE) {
4425 int err = huge_add_to_page_cache(page, mapping, idx);
4434 if (unlikely(anon_vma_prepare(vma))) {
4436 goto backout_unlocked;
4442 * If memory error occurs between mmap() and fault, some process
4443 * don't have hwpoisoned swap entry for errored virtual address.
4444 * So we need to block hugepage fault by PG_hwpoison bit check.
4446 if (unlikely(PageHWPoison(page))) {
4447 ret = VM_FAULT_HWPOISON_LARGE |
4448 VM_FAULT_SET_HINDEX(hstate_index(h));
4449 goto backout_unlocked;
4454 * If we are going to COW a private mapping later, we examine the
4455 * pending reservations for this page now. This will ensure that
4456 * any allocations necessary to record that reservation occur outside
4459 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4460 if (vma_needs_reservation(h, vma, haddr) < 0) {
4462 goto backout_unlocked;
4464 /* Just decrements count, does not deallocate */
4465 vma_end_reservation(h, vma, haddr);
4468 ptl = huge_pte_lock(h, mm, ptep);
4470 if (!huge_pte_none(huge_ptep_get(ptep)))
4474 ClearPagePrivate(page);
4475 hugepage_add_new_anon_rmap(page, vma, haddr);
4477 page_dup_rmap(page, true);
4478 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4479 && (vma->vm_flags & VM_SHARED)));
4480 set_huge_pte_at(mm, haddr, ptep, new_pte);
4482 hugetlb_count_add(pages_per_huge_page(h), mm);
4483 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4484 /* Optimization, do the COW without a second fault */
4485 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4491 * Only make newly allocated pages active. Existing pages found
4492 * in the pagecache could be !page_huge_active() if they have been
4493 * isolated for migration.
4496 set_page_huge_active(page);
4506 restore_reserve_on_error(h, vma, haddr, page);
4512 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4514 unsigned long key[2];
4517 key[0] = (unsigned long) mapping;
4520 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4522 return hash & (num_fault_mutexes - 1);
4526 * For uniprocesor systems we always use a single mutex, so just
4527 * return 0 and avoid the hashing overhead.
4529 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4535 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4536 unsigned long address, unsigned int flags)
4543 struct page *page = NULL;
4544 struct page *pagecache_page = NULL;
4545 struct hstate *h = hstate_vma(vma);
4546 struct address_space *mapping;
4547 int need_wait_lock = 0;
4548 unsigned long haddr = address & huge_page_mask(h);
4550 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4553 * Since we hold no locks, ptep could be stale. That is
4554 * OK as we are only making decisions based on content and
4555 * not actually modifying content here.
4557 entry = huge_ptep_get(ptep);
4558 if (unlikely(is_hugetlb_entry_migration(entry))) {
4559 migration_entry_wait_huge(vma, mm, ptep);
4561 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4562 return VM_FAULT_HWPOISON_LARGE |
4563 VM_FAULT_SET_HINDEX(hstate_index(h));
4567 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4568 * until finished with ptep. This serves two purposes:
4569 * 1) It prevents huge_pmd_unshare from being called elsewhere
4570 * and making the ptep no longer valid.
4571 * 2) It synchronizes us with i_size modifications during truncation.
4573 * ptep could have already be assigned via huge_pte_offset. That
4574 * is OK, as huge_pte_alloc will return the same value unless
4575 * something has changed.
4577 mapping = vma->vm_file->f_mapping;
4578 i_mmap_lock_read(mapping);
4579 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4581 i_mmap_unlock_read(mapping);
4582 return VM_FAULT_OOM;
4586 * Serialize hugepage allocation and instantiation, so that we don't
4587 * get spurious allocation failures if two CPUs race to instantiate
4588 * the same page in the page cache.
4590 idx = vma_hugecache_offset(h, vma, haddr);
4591 hash = hugetlb_fault_mutex_hash(mapping, idx);
4592 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4594 entry = huge_ptep_get(ptep);
4595 if (huge_pte_none(entry)) {
4596 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4603 * entry could be a migration/hwpoison entry at this point, so this
4604 * check prevents the kernel from going below assuming that we have
4605 * an active hugepage in pagecache. This goto expects the 2nd page
4606 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4607 * properly handle it.
4609 if (!pte_present(entry))
4613 * If we are going to COW the mapping later, we examine the pending
4614 * reservations for this page now. This will ensure that any
4615 * allocations necessary to record that reservation occur outside the
4616 * spinlock. For private mappings, we also lookup the pagecache
4617 * page now as it is used to determine if a reservation has been
4620 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4621 if (vma_needs_reservation(h, vma, haddr) < 0) {
4625 /* Just decrements count, does not deallocate */
4626 vma_end_reservation(h, vma, haddr);
4628 if (!(vma->vm_flags & VM_MAYSHARE))
4629 pagecache_page = hugetlbfs_pagecache_page(h,
4633 ptl = huge_pte_lock(h, mm, ptep);
4635 /* Check for a racing update before calling hugetlb_cow */
4636 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4640 * hugetlb_cow() requires page locks of pte_page(entry) and
4641 * pagecache_page, so here we need take the former one
4642 * when page != pagecache_page or !pagecache_page.
4644 page = pte_page(entry);
4645 if (page != pagecache_page)
4646 if (!trylock_page(page)) {
4653 if (flags & FAULT_FLAG_WRITE) {
4654 if (!huge_pte_write(entry)) {
4655 ret = hugetlb_cow(mm, vma, address, ptep,
4656 pagecache_page, ptl);
4659 entry = huge_pte_mkdirty(entry);
4661 entry = pte_mkyoung(entry);
4662 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4663 flags & FAULT_FLAG_WRITE))
4664 update_mmu_cache(vma, haddr, ptep);
4666 if (page != pagecache_page)
4672 if (pagecache_page) {
4673 unlock_page(pagecache_page);
4674 put_page(pagecache_page);
4677 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4678 i_mmap_unlock_read(mapping);
4680 * Generally it's safe to hold refcount during waiting page lock. But
4681 * here we just wait to defer the next page fault to avoid busy loop and
4682 * the page is not used after unlocked before returning from the current
4683 * page fault. So we are safe from accessing freed page, even if we wait
4684 * here without taking refcount.
4687 wait_on_page_locked(page);
4692 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4693 * modifications for huge pages.
4695 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4697 struct vm_area_struct *dst_vma,
4698 unsigned long dst_addr,
4699 unsigned long src_addr,
4700 struct page **pagep)
4702 struct address_space *mapping;
4705 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4706 struct hstate *h = hstate_vma(dst_vma);
4713 /* If a page already exists, then it's UFFDIO_COPY for
4714 * a non-missing case. Return -EEXIST.
4717 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
4722 page = alloc_huge_page(dst_vma, dst_addr, 0);
4728 ret = copy_huge_page_from_user(page,
4729 (const void __user *) src_addr,
4730 pages_per_huge_page(h), false);
4732 /* fallback to copy_from_user outside mmap_lock */
4733 if (unlikely(ret)) {
4736 /* don't free the page */
4745 * The memory barrier inside __SetPageUptodate makes sure that
4746 * preceding stores to the page contents become visible before
4747 * the set_pte_at() write.
4749 __SetPageUptodate(page);
4751 mapping = dst_vma->vm_file->f_mapping;
4752 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4755 * If shared, add to page cache
4758 size = i_size_read(mapping->host) >> huge_page_shift(h);
4761 goto out_release_nounlock;
4764 * Serialization between remove_inode_hugepages() and
4765 * huge_add_to_page_cache() below happens through the
4766 * hugetlb_fault_mutex_table that here must be hold by
4769 ret = huge_add_to_page_cache(page, mapping, idx);
4771 goto out_release_nounlock;
4774 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4778 * Recheck the i_size after holding PT lock to make sure not
4779 * to leave any page mapped (as page_mapped()) beyond the end
4780 * of the i_size (remove_inode_hugepages() is strict about
4781 * enforcing that). If we bail out here, we'll also leave a
4782 * page in the radix tree in the vm_shared case beyond the end
4783 * of the i_size, but remove_inode_hugepages() will take care
4784 * of it as soon as we drop the hugetlb_fault_mutex_table.
4786 size = i_size_read(mapping->host) >> huge_page_shift(h);
4789 goto out_release_unlock;
4792 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4793 goto out_release_unlock;
4796 page_dup_rmap(page, true);
4798 ClearPagePrivate(page);
4799 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4802 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4803 if (dst_vma->vm_flags & VM_WRITE)
4804 _dst_pte = huge_pte_mkdirty(_dst_pte);
4805 _dst_pte = pte_mkyoung(_dst_pte);
4807 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4809 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4810 dst_vma->vm_flags & VM_WRITE);
4811 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4813 /* No need to invalidate - it was non-present before */
4814 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4817 set_page_huge_active(page);
4827 out_release_nounlock:
4832 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4833 struct page **pages, struct vm_area_struct **vmas,
4834 unsigned long *position, unsigned long *nr_pages,
4835 long i, unsigned int flags, int *locked)
4837 unsigned long pfn_offset;
4838 unsigned long vaddr = *position;
4839 unsigned long remainder = *nr_pages;
4840 struct hstate *h = hstate_vma(vma);
4843 while (vaddr < vma->vm_end && remainder) {
4845 spinlock_t *ptl = NULL;
4850 * If we have a pending SIGKILL, don't keep faulting pages and
4851 * potentially allocating memory.
4853 if (fatal_signal_pending(current)) {
4859 * Some archs (sparc64, sh*) have multiple pte_ts to
4860 * each hugepage. We have to make sure we get the
4861 * first, for the page indexing below to work.
4863 * Note that page table lock is not held when pte is null.
4865 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4868 ptl = huge_pte_lock(h, mm, pte);
4869 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4872 * When coredumping, it suits get_dump_page if we just return
4873 * an error where there's an empty slot with no huge pagecache
4874 * to back it. This way, we avoid allocating a hugepage, and
4875 * the sparse dumpfile avoids allocating disk blocks, but its
4876 * huge holes still show up with zeroes where they need to be.
4878 if (absent && (flags & FOLL_DUMP) &&
4879 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4887 * We need call hugetlb_fault for both hugepages under migration
4888 * (in which case hugetlb_fault waits for the migration,) and
4889 * hwpoisoned hugepages (in which case we need to prevent the
4890 * caller from accessing to them.) In order to do this, we use
4891 * here is_swap_pte instead of is_hugetlb_entry_migration and
4892 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4893 * both cases, and because we can't follow correct pages
4894 * directly from any kind of swap entries.
4896 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4897 ((flags & FOLL_WRITE) &&
4898 !huge_pte_write(huge_ptep_get(pte)))) {
4900 unsigned int fault_flags = 0;
4904 if (flags & FOLL_WRITE)
4905 fault_flags |= FAULT_FLAG_WRITE;
4907 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4908 FAULT_FLAG_KILLABLE;
4909 if (flags & FOLL_NOWAIT)
4910 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4911 FAULT_FLAG_RETRY_NOWAIT;
4912 if (flags & FOLL_TRIED) {
4914 * Note: FAULT_FLAG_ALLOW_RETRY and
4915 * FAULT_FLAG_TRIED can co-exist
4917 fault_flags |= FAULT_FLAG_TRIED;
4919 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4920 if (ret & VM_FAULT_ERROR) {
4921 err = vm_fault_to_errno(ret, flags);
4925 if (ret & VM_FAULT_RETRY) {
4927 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4931 * VM_FAULT_RETRY must not return an
4932 * error, it will return zero
4935 * No need to update "position" as the
4936 * caller will not check it after
4937 * *nr_pages is set to 0.
4944 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4945 page = pte_page(huge_ptep_get(pte));
4948 * If subpage information not requested, update counters
4949 * and skip the same_page loop below.
4951 if (!pages && !vmas && !pfn_offset &&
4952 (vaddr + huge_page_size(h) < vma->vm_end) &&
4953 (remainder >= pages_per_huge_page(h))) {
4954 vaddr += huge_page_size(h);
4955 remainder -= pages_per_huge_page(h);
4956 i += pages_per_huge_page(h);
4963 pages[i] = mem_map_offset(page, pfn_offset);
4965 * try_grab_page() should always succeed here, because:
4966 * a) we hold the ptl lock, and b) we've just checked
4967 * that the huge page is present in the page tables. If
4968 * the huge page is present, then the tail pages must
4969 * also be present. The ptl prevents the head page and
4970 * tail pages from being rearranged in any way. So this
4971 * page must be available at this point, unless the page
4972 * refcount overflowed:
4974 if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) {
4989 if (vaddr < vma->vm_end && remainder &&
4990 pfn_offset < pages_per_huge_page(h)) {
4992 * We use pfn_offset to avoid touching the pageframes
4993 * of this compound page.
4999 *nr_pages = remainder;
5001 * setting position is actually required only if remainder is
5002 * not zero but it's faster not to add a "if (remainder)"
5010 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
5012 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
5015 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
5018 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
5019 unsigned long address, unsigned long end, pgprot_t newprot)
5021 struct mm_struct *mm = vma->vm_mm;
5022 unsigned long start = address;
5025 struct hstate *h = hstate_vma(vma);
5026 unsigned long pages = 0;
5027 bool shared_pmd = false;
5028 struct mmu_notifier_range range;
5031 * In the case of shared PMDs, the area to flush could be beyond
5032 * start/end. Set range.start/range.end to cover the maximum possible
5033 * range if PMD sharing is possible.
5035 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5036 0, vma, mm, start, end);
5037 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5039 BUG_ON(address >= end);
5040 flush_cache_range(vma, range.start, range.end);
5042 mmu_notifier_invalidate_range_start(&range);
5043 i_mmap_lock_write(vma->vm_file->f_mapping);
5044 for (; address < end; address += huge_page_size(h)) {
5046 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5049 ptl = huge_pte_lock(h, mm, ptep);
5050 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
5056 pte = huge_ptep_get(ptep);
5057 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5061 if (unlikely(is_hugetlb_entry_migration(pte))) {
5062 swp_entry_t entry = pte_to_swp_entry(pte);
5064 if (is_write_migration_entry(entry)) {
5067 make_migration_entry_read(&entry);
5068 newpte = swp_entry_to_pte(entry);
5069 set_huge_swap_pte_at(mm, address, ptep,
5070 newpte, huge_page_size(h));
5076 if (!huge_pte_none(pte)) {
5079 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5080 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
5081 pte = arch_make_huge_pte(pte, vma, NULL, 0);
5082 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
5088 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5089 * may have cleared our pud entry and done put_page on the page table:
5090 * once we release i_mmap_rwsem, another task can do the final put_page
5091 * and that page table be reused and filled with junk. If we actually
5092 * did unshare a page of pmds, flush the range corresponding to the pud.
5095 flush_hugetlb_tlb_range(vma, range.start, range.end);
5097 flush_hugetlb_tlb_range(vma, start, end);
5099 * No need to call mmu_notifier_invalidate_range() we are downgrading
5100 * page table protection not changing it to point to a new page.
5102 * See Documentation/vm/mmu_notifier.rst
5104 i_mmap_unlock_write(vma->vm_file->f_mapping);
5105 mmu_notifier_invalidate_range_end(&range);
5107 return pages << h->order;
5110 int hugetlb_reserve_pages(struct inode *inode,
5112 struct vm_area_struct *vma,
5113 vm_flags_t vm_flags)
5115 long ret, chg, add = -1;
5116 struct hstate *h = hstate_inode(inode);
5117 struct hugepage_subpool *spool = subpool_inode(inode);
5118 struct resv_map *resv_map;
5119 struct hugetlb_cgroup *h_cg = NULL;
5120 long gbl_reserve, regions_needed = 0;
5122 /* This should never happen */
5124 VM_WARN(1, "%s called with a negative range\n", __func__);
5129 * Only apply hugepage reservation if asked. At fault time, an
5130 * attempt will be made for VM_NORESERVE to allocate a page
5131 * without using reserves
5133 if (vm_flags & VM_NORESERVE)
5137 * Shared mappings base their reservation on the number of pages that
5138 * are already allocated on behalf of the file. Private mappings need
5139 * to reserve the full area even if read-only as mprotect() may be
5140 * called to make the mapping read-write. Assume !vma is a shm mapping
5142 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5144 * resv_map can not be NULL as hugetlb_reserve_pages is only
5145 * called for inodes for which resv_maps were created (see
5146 * hugetlbfs_get_inode).
5148 resv_map = inode_resv_map(inode);
5150 chg = region_chg(resv_map, from, to, ®ions_needed);
5153 /* Private mapping. */
5154 resv_map = resv_map_alloc();
5160 set_vma_resv_map(vma, resv_map);
5161 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5169 ret = hugetlb_cgroup_charge_cgroup_rsvd(
5170 hstate_index(h), chg * pages_per_huge_page(h), &h_cg);
5177 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5178 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5181 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5185 * There must be enough pages in the subpool for the mapping. If
5186 * the subpool has a minimum size, there may be some global
5187 * reservations already in place (gbl_reserve).
5189 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5190 if (gbl_reserve < 0) {
5192 goto out_uncharge_cgroup;
5196 * Check enough hugepages are available for the reservation.
5197 * Hand the pages back to the subpool if there are not
5199 ret = hugetlb_acct_memory(h, gbl_reserve);
5205 * Account for the reservations made. Shared mappings record regions
5206 * that have reservations as they are shared by multiple VMAs.
5207 * When the last VMA disappears, the region map says how much
5208 * the reservation was and the page cache tells how much of
5209 * the reservation was consumed. Private mappings are per-VMA and
5210 * only the consumed reservations are tracked. When the VMA
5211 * disappears, the original reservation is the VMA size and the
5212 * consumed reservations are stored in the map. Hence, nothing
5213 * else has to be done for private mappings here
5215 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5216 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5218 if (unlikely(add < 0)) {
5219 hugetlb_acct_memory(h, -gbl_reserve);
5222 } else if (unlikely(chg > add)) {
5224 * pages in this range were added to the reserve
5225 * map between region_chg and region_add. This
5226 * indicates a race with alloc_huge_page. Adjust
5227 * the subpool and reserve counts modified above
5228 * based on the difference.
5233 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
5234 * reference to h_cg->css. See comment below for detail.
5236 hugetlb_cgroup_uncharge_cgroup_rsvd(
5238 (chg - add) * pages_per_huge_page(h), h_cg);
5240 rsv_adjust = hugepage_subpool_put_pages(spool,
5242 hugetlb_acct_memory(h, -rsv_adjust);
5245 * The file_regions will hold their own reference to
5246 * h_cg->css. So we should release the reference held
5247 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
5250 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
5255 /* put back original number of pages, chg */
5256 (void)hugepage_subpool_put_pages(spool, chg);
5257 out_uncharge_cgroup:
5258 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5259 chg * pages_per_huge_page(h), h_cg);
5261 if (!vma || vma->vm_flags & VM_MAYSHARE)
5262 /* Only call region_abort if the region_chg succeeded but the
5263 * region_add failed or didn't run.
5265 if (chg >= 0 && add < 0)
5266 region_abort(resv_map, from, to, regions_needed);
5267 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5268 kref_put(&resv_map->refs, resv_map_release);
5272 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5275 struct hstate *h = hstate_inode(inode);
5276 struct resv_map *resv_map = inode_resv_map(inode);
5278 struct hugepage_subpool *spool = subpool_inode(inode);
5282 * Since this routine can be called in the evict inode path for all
5283 * hugetlbfs inodes, resv_map could be NULL.
5286 chg = region_del(resv_map, start, end);
5288 * region_del() can fail in the rare case where a region
5289 * must be split and another region descriptor can not be
5290 * allocated. If end == LONG_MAX, it will not fail.
5296 spin_lock(&inode->i_lock);
5297 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5298 spin_unlock(&inode->i_lock);
5301 * If the subpool has a minimum size, the number of global
5302 * reservations to be released may be adjusted.
5304 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5305 hugetlb_acct_memory(h, -gbl_reserve);
5310 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5311 static unsigned long page_table_shareable(struct vm_area_struct *svma,
5312 struct vm_area_struct *vma,
5313 unsigned long addr, pgoff_t idx)
5315 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5317 unsigned long sbase = saddr & PUD_MASK;
5318 unsigned long s_end = sbase + PUD_SIZE;
5320 /* Allow segments to share if only one is marked locked */
5321 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5322 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5325 * match the virtual addresses, permission and the alignment of the
5328 if (pmd_index(addr) != pmd_index(saddr) ||
5329 vm_flags != svm_flags ||
5330 sbase < svma->vm_start || svma->vm_end < s_end)
5336 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5338 unsigned long base = addr & PUD_MASK;
5339 unsigned long end = base + PUD_SIZE;
5342 * check on proper vm_flags and page table alignment
5344 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5350 * Determine if start,end range within vma could be mapped by shared pmd.
5351 * If yes, adjust start and end to cover range associated with possible
5352 * shared pmd mappings.
5354 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5355 unsigned long *start, unsigned long *end)
5357 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
5358 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
5361 * vma need span at least one aligned PUD size and the start,end range
5362 * must at least partialy within it.
5364 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
5365 (*end <= v_start) || (*start >= v_end))
5368 /* Extend the range to be PUD aligned for a worst case scenario */
5369 if (*start > v_start)
5370 *start = ALIGN_DOWN(*start, PUD_SIZE);
5373 *end = ALIGN(*end, PUD_SIZE);
5377 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5378 * and returns the corresponding pte. While this is not necessary for the
5379 * !shared pmd case because we can allocate the pmd later as well, it makes the
5380 * code much cleaner.
5382 * This routine must be called with i_mmap_rwsem held in at least read mode if
5383 * sharing is possible. For hugetlbfs, this prevents removal of any page
5384 * table entries associated with the address space. This is important as we
5385 * are setting up sharing based on existing page table entries (mappings).
5387 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5388 * huge_pte_alloc know that sharing is not possible and do not take
5389 * i_mmap_rwsem as a performance optimization. This is handled by the
5390 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5391 * only required for subsequent processing.
5393 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5395 struct vm_area_struct *vma = find_vma(mm, addr);
5396 struct address_space *mapping = vma->vm_file->f_mapping;
5397 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5399 struct vm_area_struct *svma;
5400 unsigned long saddr;
5405 if (!vma_shareable(vma, addr))
5406 return (pte_t *)pmd_alloc(mm, pud, addr);
5408 i_mmap_assert_locked(mapping);
5409 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5413 saddr = page_table_shareable(svma, vma, addr, idx);
5415 spte = huge_pte_offset(svma->vm_mm, saddr,
5416 vma_mmu_pagesize(svma));
5418 get_page(virt_to_page(spte));
5427 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
5428 if (pud_none(*pud)) {
5429 pud_populate(mm, pud,
5430 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5433 put_page(virt_to_page(spte));
5437 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5442 * unmap huge page backed by shared pte.
5444 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5445 * indicated by page_count > 1, unmap is achieved by clearing pud and
5446 * decrementing the ref count. If count == 1, the pte page is not shared.
5448 * Called with page table lock held and i_mmap_rwsem held in write mode.
5450 * returns: 1 successfully unmapped a shared pte page
5451 * 0 the underlying pte page is not shared, or it is the last user
5453 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5454 unsigned long *addr, pte_t *ptep)
5456 pgd_t *pgd = pgd_offset(mm, *addr);
5457 p4d_t *p4d = p4d_offset(pgd, *addr);
5458 pud_t *pud = pud_offset(p4d, *addr);
5460 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
5461 BUG_ON(page_count(virt_to_page(ptep)) == 0);
5462 if (page_count(virt_to_page(ptep)) == 1)
5466 put_page(virt_to_page(ptep));
5468 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5471 #define want_pmd_share() (1)
5472 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5473 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5478 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5479 unsigned long *addr, pte_t *ptep)
5484 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5485 unsigned long *start, unsigned long *end)
5488 #define want_pmd_share() (0)
5489 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5491 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5492 pte_t *huge_pte_alloc(struct mm_struct *mm,
5493 unsigned long addr, unsigned long sz)
5500 pgd = pgd_offset(mm, addr);
5501 p4d = p4d_alloc(mm, pgd, addr);
5504 pud = pud_alloc(mm, p4d, addr);
5506 if (sz == PUD_SIZE) {
5509 BUG_ON(sz != PMD_SIZE);
5510 if (want_pmd_share() && pud_none(*pud))
5511 pte = huge_pmd_share(mm, addr, pud);
5513 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5516 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
5522 * huge_pte_offset() - Walk the page table to resolve the hugepage
5523 * entry at address @addr
5525 * Return: Pointer to page table entry (PUD or PMD) for
5526 * address @addr, or NULL if a !p*d_present() entry is encountered and the
5527 * size @sz doesn't match the hugepage size at this level of the page
5530 pte_t *huge_pte_offset(struct mm_struct *mm,
5531 unsigned long addr, unsigned long sz)
5538 pgd = pgd_offset(mm, addr);
5539 if (!pgd_present(*pgd))
5541 p4d = p4d_offset(pgd, addr);
5542 if (!p4d_present(*p4d))
5545 pud = pud_offset(p4d, addr);
5547 /* must be pud huge, non-present or none */
5548 return (pte_t *)pud;
5549 if (!pud_present(*pud))
5551 /* must have a valid entry and size to go further */
5553 pmd = pmd_offset(pud, addr);
5554 /* must be pmd huge, non-present or none */
5555 return (pte_t *)pmd;
5558 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5561 * These functions are overwritable if your architecture needs its own
5564 struct page * __weak
5565 follow_huge_addr(struct mm_struct *mm, unsigned long address,
5568 return ERR_PTR(-EINVAL);
5571 struct page * __weak
5572 follow_huge_pd(struct vm_area_struct *vma,
5573 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5575 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5579 struct page * __weak
5580 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
5581 pmd_t *pmd, int flags)
5583 struct page *page = NULL;
5587 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5588 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
5589 (FOLL_PIN | FOLL_GET)))
5593 ptl = pmd_lockptr(mm, pmd);
5596 * make sure that the address range covered by this pmd is not
5597 * unmapped from other threads.
5599 if (!pmd_huge(*pmd))
5601 pte = huge_ptep_get((pte_t *)pmd);
5602 if (pte_present(pte)) {
5603 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
5605 * try_grab_page() should always succeed here, because: a) we
5606 * hold the pmd (ptl) lock, and b) we've just checked that the
5607 * huge pmd (head) page is present in the page tables. The ptl
5608 * prevents the head page and tail pages from being rearranged
5609 * in any way. So this page must be available at this point,
5610 * unless the page refcount overflowed:
5612 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5617 if (is_hugetlb_entry_migration(pte)) {
5619 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5623 * hwpoisoned entry is treated as no_page_table in
5624 * follow_page_mask().
5632 struct page * __weak
5633 follow_huge_pud(struct mm_struct *mm, unsigned long address,
5634 pud_t *pud, int flags)
5636 if (flags & (FOLL_GET | FOLL_PIN))
5639 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5642 struct page * __weak
5643 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5645 if (flags & (FOLL_GET | FOLL_PIN))
5648 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5651 bool isolate_huge_page(struct page *page, struct list_head *list)
5655 spin_lock(&hugetlb_lock);
5656 if (!PageHeadHuge(page) || !page_huge_active(page) ||
5657 !get_page_unless_zero(page)) {
5661 clear_page_huge_active(page);
5662 list_move_tail(&page->lru, list);
5664 spin_unlock(&hugetlb_lock);
5668 void putback_active_hugepage(struct page *page)
5670 VM_BUG_ON_PAGE(!PageHead(page), page);
5671 spin_lock(&hugetlb_lock);
5672 set_page_huge_active(page);
5673 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5674 spin_unlock(&hugetlb_lock);
5678 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5680 struct hstate *h = page_hstate(oldpage);
5682 hugetlb_cgroup_migrate(oldpage, newpage);
5683 set_page_owner_migrate_reason(newpage, reason);
5686 * transfer temporary state of the new huge page. This is
5687 * reverse to other transitions because the newpage is going to
5688 * be final while the old one will be freed so it takes over
5689 * the temporary status.
5691 * Also note that we have to transfer the per-node surplus state
5692 * here as well otherwise the global surplus count will not match
5695 if (PageHugeTemporary(newpage)) {
5696 int old_nid = page_to_nid(oldpage);
5697 int new_nid = page_to_nid(newpage);
5699 SetPageHugeTemporary(oldpage);
5700 ClearPageHugeTemporary(newpage);
5702 spin_lock(&hugetlb_lock);
5703 if (h->surplus_huge_pages_node[old_nid]) {
5704 h->surplus_huge_pages_node[old_nid]--;
5705 h->surplus_huge_pages_node[new_nid]++;
5707 spin_unlock(&hugetlb_lock);
5712 static bool cma_reserve_called __initdata;
5714 static int __init cmdline_parse_hugetlb_cma(char *p)
5716 hugetlb_cma_size = memparse(p, &p);
5720 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
5722 void __init hugetlb_cma_reserve(int order)
5724 unsigned long size, reserved, per_node;
5727 cma_reserve_called = true;
5729 if (!hugetlb_cma_size)
5732 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
5733 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
5734 (PAGE_SIZE << order) / SZ_1M);
5739 * If 3 GB area is requested on a machine with 4 numa nodes,
5740 * let's allocate 1 GB on first three nodes and ignore the last one.
5742 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
5743 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
5744 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
5747 for_each_node_state(nid, N_ONLINE) {
5749 char name[CMA_MAX_NAME];
5751 size = min(per_node, hugetlb_cma_size - reserved);
5752 size = round_up(size, PAGE_SIZE << order);
5754 snprintf(name, sizeof(name), "hugetlb%d", nid);
5755 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
5757 &hugetlb_cma[nid], nid);
5759 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
5765 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
5768 if (reserved >= hugetlb_cma_size)
5773 void __init hugetlb_cma_check(void)
5775 if (!hugetlb_cma_size || cma_reserve_called)
5778 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
5781 #endif /* CONFIG_CMA */