1 // SPDX-License-Identifier: GPL-2.0-only
3 * Generic hugetlb support.
4 * (C) Nadia Yvette Chambers, April 2004
6 #include <linux/list.h>
7 #include <linux/init.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/sched/mm.h>
23 #include <linux/mmdebug.h>
24 #include <linux/sched/signal.h>
25 #include <linux/rmap.h>
26 #include <linux/string_helpers.h>
27 #include <linux/swap.h>
28 #include <linux/swapops.h>
29 #include <linux/jhash.h>
30 #include <linux/numa.h>
31 #include <linux/llist.h>
32 #include <linux/cma.h>
35 #include <asm/pgalloc.h>
39 #include <linux/hugetlb.h>
40 #include <linux/hugetlb_cgroup.h>
41 #include <linux/node.h>
42 #include <linux/userfaultfd_k.h>
43 #include <linux/page_owner.h>
46 int hugetlb_max_hstate __read_mostly;
47 unsigned int default_hstate_idx;
48 struct hstate hstates[HUGE_MAX_HSTATE];
51 static struct cma *hugetlb_cma[MAX_NUMNODES];
53 static unsigned long hugetlb_cma_size __initdata;
56 * Minimum page order among possible hugepage sizes, set to a proper value
59 static unsigned int minimum_order __read_mostly = UINT_MAX;
61 __initdata LIST_HEAD(huge_boot_pages);
63 /* for command line parsing */
64 static struct hstate * __initdata parsed_hstate;
65 static unsigned long __initdata default_hstate_max_huge_pages;
66 static bool __initdata parsed_valid_hugepagesz = true;
67 static bool __initdata parsed_default_hugepagesz;
70 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
71 * free_huge_pages, and surplus_huge_pages.
73 DEFINE_SPINLOCK(hugetlb_lock);
76 * Serializes faults on the same logical page. This is used to
77 * prevent spurious OOMs when the hugepage pool is fully utilized.
79 static int num_fault_mutexes;
80 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
82 /* Forward declaration */
83 static int hugetlb_acct_memory(struct hstate *h, long delta);
85 static inline bool subpool_is_free(struct hugepage_subpool *spool)
89 if (spool->max_hpages != -1)
90 return spool->used_hpages == 0;
91 if (spool->min_hpages != -1)
92 return spool->rsv_hpages == spool->min_hpages;
97 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
98 unsigned long irq_flags)
100 spin_unlock_irqrestore(&spool->lock, irq_flags);
102 /* If no pages are used, and no other handles to the subpool
103 * remain, give up any reservations based on minimum size and
104 * free the subpool */
105 if (subpool_is_free(spool)) {
106 if (spool->min_hpages != -1)
107 hugetlb_acct_memory(spool->hstate,
113 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
116 struct hugepage_subpool *spool;
118 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
122 spin_lock_init(&spool->lock);
124 spool->max_hpages = max_hpages;
126 spool->min_hpages = min_hpages;
128 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
132 spool->rsv_hpages = min_hpages;
137 void hugepage_put_subpool(struct hugepage_subpool *spool)
141 spin_lock_irqsave(&spool->lock, flags);
142 BUG_ON(!spool->count);
144 unlock_or_release_subpool(spool, flags);
148 * Subpool accounting for allocating and reserving pages.
149 * Return -ENOMEM if there are not enough resources to satisfy the
150 * request. Otherwise, return the number of pages by which the
151 * global pools must be adjusted (upward). The returned value may
152 * only be different than the passed value (delta) in the case where
153 * a subpool minimum size must be maintained.
155 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
163 spin_lock_irq(&spool->lock);
165 if (spool->max_hpages != -1) { /* maximum size accounting */
166 if ((spool->used_hpages + delta) <= spool->max_hpages)
167 spool->used_hpages += delta;
174 /* minimum size accounting */
175 if (spool->min_hpages != -1 && spool->rsv_hpages) {
176 if (delta > spool->rsv_hpages) {
178 * Asking for more reserves than those already taken on
179 * behalf of subpool. Return difference.
181 ret = delta - spool->rsv_hpages;
182 spool->rsv_hpages = 0;
184 ret = 0; /* reserves already accounted for */
185 spool->rsv_hpages -= delta;
190 spin_unlock_irq(&spool->lock);
195 * Subpool accounting for freeing and unreserving pages.
196 * Return the number of global page reservations that must be dropped.
197 * The return value may only be different than the passed value (delta)
198 * in the case where a subpool minimum size must be maintained.
200 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
209 spin_lock_irqsave(&spool->lock, flags);
211 if (spool->max_hpages != -1) /* maximum size accounting */
212 spool->used_hpages -= delta;
214 /* minimum size accounting */
215 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
216 if (spool->rsv_hpages + delta <= spool->min_hpages)
219 ret = spool->rsv_hpages + delta - spool->min_hpages;
221 spool->rsv_hpages += delta;
222 if (spool->rsv_hpages > spool->min_hpages)
223 spool->rsv_hpages = spool->min_hpages;
227 * If hugetlbfs_put_super couldn't free spool due to an outstanding
228 * quota reference, free it now.
230 unlock_or_release_subpool(spool, flags);
235 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
237 return HUGETLBFS_SB(inode->i_sb)->spool;
240 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
242 return subpool_inode(file_inode(vma->vm_file));
245 /* Helper that removes a struct file_region from the resv_map cache and returns
248 static struct file_region *
249 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
251 struct file_region *nrg = NULL;
253 VM_BUG_ON(resv->region_cache_count <= 0);
255 resv->region_cache_count--;
256 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
257 list_del(&nrg->link);
265 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
266 struct file_region *rg)
268 #ifdef CONFIG_CGROUP_HUGETLB
269 nrg->reservation_counter = rg->reservation_counter;
276 /* Helper that records hugetlb_cgroup uncharge info. */
277 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
279 struct resv_map *resv,
280 struct file_region *nrg)
282 #ifdef CONFIG_CGROUP_HUGETLB
284 nrg->reservation_counter =
285 &h_cg->rsvd_hugepage[hstate_index(h)];
286 nrg->css = &h_cg->css;
288 * The caller will hold exactly one h_cg->css reference for the
289 * whole contiguous reservation region. But this area might be
290 * scattered when there are already some file_regions reside in
291 * it. As a result, many file_regions may share only one css
292 * reference. In order to ensure that one file_region must hold
293 * exactly one h_cg->css reference, we should do css_get for
294 * each file_region and leave the reference held by caller
298 if (!resv->pages_per_hpage)
299 resv->pages_per_hpage = pages_per_huge_page(h);
300 /* pages_per_hpage should be the same for all entries in
303 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
305 nrg->reservation_counter = NULL;
311 static void put_uncharge_info(struct file_region *rg)
313 #ifdef CONFIG_CGROUP_HUGETLB
319 static bool has_same_uncharge_info(struct file_region *rg,
320 struct file_region *org)
322 #ifdef CONFIG_CGROUP_HUGETLB
324 rg->reservation_counter == org->reservation_counter &&
332 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
334 struct file_region *nrg = NULL, *prg = NULL;
336 prg = list_prev_entry(rg, link);
337 if (&prg->link != &resv->regions && prg->to == rg->from &&
338 has_same_uncharge_info(prg, rg)) {
342 put_uncharge_info(rg);
348 nrg = list_next_entry(rg, link);
349 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
350 has_same_uncharge_info(nrg, rg)) {
351 nrg->from = rg->from;
354 put_uncharge_info(rg);
360 hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from,
361 long to, struct hstate *h, struct hugetlb_cgroup *cg,
362 long *regions_needed)
364 struct file_region *nrg;
366 if (!regions_needed) {
367 nrg = get_file_region_entry_from_cache(map, from, to);
368 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
369 list_add(&nrg->link, rg->link.prev);
370 coalesce_file_region(map, nrg);
372 *regions_needed += 1;
378 * Must be called with resv->lock held.
380 * Calling this with regions_needed != NULL will count the number of pages
381 * to be added but will not modify the linked list. And regions_needed will
382 * indicate the number of file_regions needed in the cache to carry out to add
383 * the regions for this range.
385 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
386 struct hugetlb_cgroup *h_cg,
387 struct hstate *h, long *regions_needed)
390 struct list_head *head = &resv->regions;
391 long last_accounted_offset = f;
392 struct file_region *rg = NULL, *trg = NULL;
397 /* In this loop, we essentially handle an entry for the range
398 * [last_accounted_offset, rg->from), at every iteration, with some
401 list_for_each_entry_safe(rg, trg, head, link) {
402 /* Skip irrelevant regions that start before our range. */
404 /* If this region ends after the last accounted offset,
405 * then we need to update last_accounted_offset.
407 if (rg->to > last_accounted_offset)
408 last_accounted_offset = rg->to;
412 /* When we find a region that starts beyond our range, we've
418 /* Add an entry for last_accounted_offset -> rg->from, and
419 * update last_accounted_offset.
421 if (rg->from > last_accounted_offset)
422 add += hugetlb_resv_map_add(resv, rg,
423 last_accounted_offset,
427 last_accounted_offset = rg->to;
430 /* Handle the case where our range extends beyond
431 * last_accounted_offset.
433 if (last_accounted_offset < t)
434 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
435 t, h, h_cg, regions_needed);
441 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
443 static int allocate_file_region_entries(struct resv_map *resv,
445 __must_hold(&resv->lock)
447 struct list_head allocated_regions;
448 int to_allocate = 0, i = 0;
449 struct file_region *trg = NULL, *rg = NULL;
451 VM_BUG_ON(regions_needed < 0);
453 INIT_LIST_HEAD(&allocated_regions);
456 * Check for sufficient descriptors in the cache to accommodate
457 * the number of in progress add operations plus regions_needed.
459 * This is a while loop because when we drop the lock, some other call
460 * to region_add or region_del may have consumed some region_entries,
461 * so we keep looping here until we finally have enough entries for
462 * (adds_in_progress + regions_needed).
464 while (resv->region_cache_count <
465 (resv->adds_in_progress + regions_needed)) {
466 to_allocate = resv->adds_in_progress + regions_needed -
467 resv->region_cache_count;
469 /* At this point, we should have enough entries in the cache
470 * for all the existings adds_in_progress. We should only be
471 * needing to allocate for regions_needed.
473 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
475 spin_unlock(&resv->lock);
476 for (i = 0; i < to_allocate; i++) {
477 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
480 list_add(&trg->link, &allocated_regions);
483 spin_lock(&resv->lock);
485 list_splice(&allocated_regions, &resv->region_cache);
486 resv->region_cache_count += to_allocate;
492 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
500 * Add the huge page range represented by [f, t) to the reserve
501 * map. Regions will be taken from the cache to fill in this range.
502 * Sufficient regions should exist in the cache due to the previous
503 * call to region_chg with the same range, but in some cases the cache will not
504 * have sufficient entries due to races with other code doing region_add or
505 * region_del. The extra needed entries will be allocated.
507 * regions_needed is the out value provided by a previous call to region_chg.
509 * Return the number of new huge pages added to the map. This number is greater
510 * than or equal to zero. If file_region entries needed to be allocated for
511 * this operation and we were not able to allocate, it returns -ENOMEM.
512 * region_add of regions of length 1 never allocate file_regions and cannot
513 * fail; region_chg will always allocate at least 1 entry and a region_add for
514 * 1 page will only require at most 1 entry.
516 static long region_add(struct resv_map *resv, long f, long t,
517 long in_regions_needed, struct hstate *h,
518 struct hugetlb_cgroup *h_cg)
520 long add = 0, actual_regions_needed = 0;
522 spin_lock(&resv->lock);
525 /* Count how many regions are actually needed to execute this add. */
526 add_reservation_in_range(resv, f, t, NULL, NULL,
527 &actual_regions_needed);
530 * Check for sufficient descriptors in the cache to accommodate
531 * this add operation. Note that actual_regions_needed may be greater
532 * than in_regions_needed, as the resv_map may have been modified since
533 * the region_chg call. In this case, we need to make sure that we
534 * allocate extra entries, such that we have enough for all the
535 * existing adds_in_progress, plus the excess needed for this
538 if (actual_regions_needed > in_regions_needed &&
539 resv->region_cache_count <
540 resv->adds_in_progress +
541 (actual_regions_needed - in_regions_needed)) {
542 /* region_add operation of range 1 should never need to
543 * allocate file_region entries.
545 VM_BUG_ON(t - f <= 1);
547 if (allocate_file_region_entries(
548 resv, actual_regions_needed - in_regions_needed)) {
555 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
557 resv->adds_in_progress -= in_regions_needed;
559 spin_unlock(&resv->lock);
564 * Examine the existing reserve map and determine how many
565 * huge pages in the specified range [f, t) are NOT currently
566 * represented. This routine is called before a subsequent
567 * call to region_add that will actually modify the reserve
568 * map to add the specified range [f, t). region_chg does
569 * not change the number of huge pages represented by the
570 * map. A number of new file_region structures is added to the cache as a
571 * placeholder, for the subsequent region_add call to use. At least 1
572 * file_region structure is added.
574 * out_regions_needed is the number of regions added to the
575 * resv->adds_in_progress. This value needs to be provided to a follow up call
576 * to region_add or region_abort for proper accounting.
578 * Returns the number of huge pages that need to be added to the existing
579 * reservation map for the range [f, t). This number is greater or equal to
580 * zero. -ENOMEM is returned if a new file_region structure or cache entry
581 * is needed and can not be allocated.
583 static long region_chg(struct resv_map *resv, long f, long t,
584 long *out_regions_needed)
588 spin_lock(&resv->lock);
590 /* Count how many hugepages in this range are NOT represented. */
591 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
594 if (*out_regions_needed == 0)
595 *out_regions_needed = 1;
597 if (allocate_file_region_entries(resv, *out_regions_needed))
600 resv->adds_in_progress += *out_regions_needed;
602 spin_unlock(&resv->lock);
607 * Abort the in progress add operation. The adds_in_progress field
608 * of the resv_map keeps track of the operations in progress between
609 * calls to region_chg and region_add. Operations are sometimes
610 * aborted after the call to region_chg. In such cases, region_abort
611 * is called to decrement the adds_in_progress counter. regions_needed
612 * is the value returned by the region_chg call, it is used to decrement
613 * the adds_in_progress counter.
615 * NOTE: The range arguments [f, t) are not needed or used in this
616 * routine. They are kept to make reading the calling code easier as
617 * arguments will match the associated region_chg call.
619 static void region_abort(struct resv_map *resv, long f, long t,
622 spin_lock(&resv->lock);
623 VM_BUG_ON(!resv->region_cache_count);
624 resv->adds_in_progress -= regions_needed;
625 spin_unlock(&resv->lock);
629 * Delete the specified range [f, t) from the reserve map. If the
630 * t parameter is LONG_MAX, this indicates that ALL regions after f
631 * should be deleted. Locate the regions which intersect [f, t)
632 * and either trim, delete or split the existing regions.
634 * Returns the number of huge pages deleted from the reserve map.
635 * In the normal case, the return value is zero or more. In the
636 * case where a region must be split, a new region descriptor must
637 * be allocated. If the allocation fails, -ENOMEM will be returned.
638 * NOTE: If the parameter t == LONG_MAX, then we will never split
639 * a region and possibly return -ENOMEM. Callers specifying
640 * t == LONG_MAX do not need to check for -ENOMEM error.
642 static long region_del(struct resv_map *resv, long f, long t)
644 struct list_head *head = &resv->regions;
645 struct file_region *rg, *trg;
646 struct file_region *nrg = NULL;
650 spin_lock(&resv->lock);
651 list_for_each_entry_safe(rg, trg, head, link) {
653 * Skip regions before the range to be deleted. file_region
654 * ranges are normally of the form [from, to). However, there
655 * may be a "placeholder" entry in the map which is of the form
656 * (from, to) with from == to. Check for placeholder entries
657 * at the beginning of the range to be deleted.
659 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
665 if (f > rg->from && t < rg->to) { /* Must split region */
667 * Check for an entry in the cache before dropping
668 * lock and attempting allocation.
671 resv->region_cache_count > resv->adds_in_progress) {
672 nrg = list_first_entry(&resv->region_cache,
675 list_del(&nrg->link);
676 resv->region_cache_count--;
680 spin_unlock(&resv->lock);
681 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
688 hugetlb_cgroup_uncharge_file_region(
689 resv, rg, t - f, false);
691 /* New entry for end of split region */
695 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
697 INIT_LIST_HEAD(&nrg->link);
699 /* Original entry is trimmed */
702 list_add(&nrg->link, &rg->link);
707 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
708 del += rg->to - rg->from;
709 hugetlb_cgroup_uncharge_file_region(resv, rg,
710 rg->to - rg->from, true);
716 if (f <= rg->from) { /* Trim beginning of region */
717 hugetlb_cgroup_uncharge_file_region(resv, rg,
718 t - rg->from, false);
722 } else { /* Trim end of region */
723 hugetlb_cgroup_uncharge_file_region(resv, rg,
731 spin_unlock(&resv->lock);
737 * A rare out of memory error was encountered which prevented removal of
738 * the reserve map region for a page. The huge page itself was free'ed
739 * and removed from the page cache. This routine will adjust the subpool
740 * usage count, and the global reserve count if needed. By incrementing
741 * these counts, the reserve map entry which could not be deleted will
742 * appear as a "reserved" entry instead of simply dangling with incorrect
745 void hugetlb_fix_reserve_counts(struct inode *inode)
747 struct hugepage_subpool *spool = subpool_inode(inode);
749 bool reserved = false;
751 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
752 if (rsv_adjust > 0) {
753 struct hstate *h = hstate_inode(inode);
755 if (!hugetlb_acct_memory(h, 1))
757 } else if (!rsv_adjust) {
762 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
766 * Count and return the number of huge pages in the reserve map
767 * that intersect with the range [f, t).
769 static long region_count(struct resv_map *resv, long f, long t)
771 struct list_head *head = &resv->regions;
772 struct file_region *rg;
775 spin_lock(&resv->lock);
776 /* Locate each segment we overlap with, and count that overlap. */
777 list_for_each_entry(rg, head, link) {
786 seg_from = max(rg->from, f);
787 seg_to = min(rg->to, t);
789 chg += seg_to - seg_from;
791 spin_unlock(&resv->lock);
797 * Convert the address within this vma to the page offset within
798 * the mapping, in pagecache page units; huge pages here.
800 static pgoff_t vma_hugecache_offset(struct hstate *h,
801 struct vm_area_struct *vma, unsigned long address)
803 return ((address - vma->vm_start) >> huge_page_shift(h)) +
804 (vma->vm_pgoff >> huge_page_order(h));
807 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
808 unsigned long address)
810 return vma_hugecache_offset(hstate_vma(vma), vma, address);
812 EXPORT_SYMBOL_GPL(linear_hugepage_index);
815 * Return the size of the pages allocated when backing a VMA. In the majority
816 * cases this will be same size as used by the page table entries.
818 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
820 if (vma->vm_ops && vma->vm_ops->pagesize)
821 return vma->vm_ops->pagesize(vma);
824 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
827 * Return the page size being used by the MMU to back a VMA. In the majority
828 * of cases, the page size used by the kernel matches the MMU size. On
829 * architectures where it differs, an architecture-specific 'strong'
830 * version of this symbol is required.
832 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
834 return vma_kernel_pagesize(vma);
838 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
839 * bits of the reservation map pointer, which are always clear due to
842 #define HPAGE_RESV_OWNER (1UL << 0)
843 #define HPAGE_RESV_UNMAPPED (1UL << 1)
844 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
847 * These helpers are used to track how many pages are reserved for
848 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
849 * is guaranteed to have their future faults succeed.
851 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
852 * the reserve counters are updated with the hugetlb_lock held. It is safe
853 * to reset the VMA at fork() time as it is not in use yet and there is no
854 * chance of the global counters getting corrupted as a result of the values.
856 * The private mapping reservation is represented in a subtly different
857 * manner to a shared mapping. A shared mapping has a region map associated
858 * with the underlying file, this region map represents the backing file
859 * pages which have ever had a reservation assigned which this persists even
860 * after the page is instantiated. A private mapping has a region map
861 * associated with the original mmap which is attached to all VMAs which
862 * reference it, this region map represents those offsets which have consumed
863 * reservation ie. where pages have been instantiated.
865 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
867 return (unsigned long)vma->vm_private_data;
870 static void set_vma_private_data(struct vm_area_struct *vma,
873 vma->vm_private_data = (void *)value;
877 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
878 struct hugetlb_cgroup *h_cg,
881 #ifdef CONFIG_CGROUP_HUGETLB
883 resv_map->reservation_counter = NULL;
884 resv_map->pages_per_hpage = 0;
885 resv_map->css = NULL;
887 resv_map->reservation_counter =
888 &h_cg->rsvd_hugepage[hstate_index(h)];
889 resv_map->pages_per_hpage = pages_per_huge_page(h);
890 resv_map->css = &h_cg->css;
895 struct resv_map *resv_map_alloc(void)
897 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
898 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
900 if (!resv_map || !rg) {
906 kref_init(&resv_map->refs);
907 spin_lock_init(&resv_map->lock);
908 INIT_LIST_HEAD(&resv_map->regions);
910 resv_map->adds_in_progress = 0;
912 * Initialize these to 0. On shared mappings, 0's here indicate these
913 * fields don't do cgroup accounting. On private mappings, these will be
914 * re-initialized to the proper values, to indicate that hugetlb cgroup
915 * reservations are to be un-charged from here.
917 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
919 INIT_LIST_HEAD(&resv_map->region_cache);
920 list_add(&rg->link, &resv_map->region_cache);
921 resv_map->region_cache_count = 1;
926 void resv_map_release(struct kref *ref)
928 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
929 struct list_head *head = &resv_map->region_cache;
930 struct file_region *rg, *trg;
932 /* Clear out any active regions before we release the map. */
933 region_del(resv_map, 0, LONG_MAX);
935 /* ... and any entries left in the cache */
936 list_for_each_entry_safe(rg, trg, head, link) {
941 VM_BUG_ON(resv_map->adds_in_progress);
946 static inline struct resv_map *inode_resv_map(struct inode *inode)
949 * At inode evict time, i_mapping may not point to the original
950 * address space within the inode. This original address space
951 * contains the pointer to the resv_map. So, always use the
952 * address space embedded within the inode.
953 * The VERY common case is inode->mapping == &inode->i_data but,
954 * this may not be true for device special inodes.
956 return (struct resv_map *)(&inode->i_data)->private_data;
959 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
961 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
962 if (vma->vm_flags & VM_MAYSHARE) {
963 struct address_space *mapping = vma->vm_file->f_mapping;
964 struct inode *inode = mapping->host;
966 return inode_resv_map(inode);
969 return (struct resv_map *)(get_vma_private_data(vma) &
974 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
976 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
977 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
979 set_vma_private_data(vma, (get_vma_private_data(vma) &
980 HPAGE_RESV_MASK) | (unsigned long)map);
983 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
985 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
986 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
988 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
991 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
993 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
995 return (get_vma_private_data(vma) & flag) != 0;
998 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
999 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
1001 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1002 if (!(vma->vm_flags & VM_MAYSHARE))
1003 vma->vm_private_data = (void *)0;
1006 /* Returns true if the VMA has associated reserve pages */
1007 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1009 if (vma->vm_flags & VM_NORESERVE) {
1011 * This address is already reserved by other process(chg == 0),
1012 * so, we should decrement reserved count. Without decrementing,
1013 * reserve count remains after releasing inode, because this
1014 * allocated page will go into page cache and is regarded as
1015 * coming from reserved pool in releasing step. Currently, we
1016 * don't have any other solution to deal with this situation
1017 * properly, so add work-around here.
1019 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1025 /* Shared mappings always use reserves */
1026 if (vma->vm_flags & VM_MAYSHARE) {
1028 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1029 * be a region map for all pages. The only situation where
1030 * there is no region map is if a hole was punched via
1031 * fallocate. In this case, there really are no reserves to
1032 * use. This situation is indicated if chg != 0.
1041 * Only the process that called mmap() has reserves for
1044 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1046 * Like the shared case above, a hole punch or truncate
1047 * could have been performed on the private mapping.
1048 * Examine the value of chg to determine if reserves
1049 * actually exist or were previously consumed.
1050 * Very Subtle - The value of chg comes from a previous
1051 * call to vma_needs_reserves(). The reserve map for
1052 * private mappings has different (opposite) semantics
1053 * than that of shared mappings. vma_needs_reserves()
1054 * has already taken this difference in semantics into
1055 * account. Therefore, the meaning of chg is the same
1056 * as in the shared case above. Code could easily be
1057 * combined, but keeping it separate draws attention to
1058 * subtle differences.
1069 static void enqueue_huge_page(struct hstate *h, struct page *page)
1071 int nid = page_to_nid(page);
1073 lockdep_assert_held(&hugetlb_lock);
1074 list_move(&page->lru, &h->hugepage_freelists[nid]);
1075 h->free_huge_pages++;
1076 h->free_huge_pages_node[nid]++;
1077 SetHPageFreed(page);
1080 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1083 bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA);
1085 lockdep_assert_held(&hugetlb_lock);
1086 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1087 if (nocma && is_migrate_cma_page(page))
1090 if (PageHWPoison(page))
1093 list_move(&page->lru, &h->hugepage_activelist);
1094 set_page_refcounted(page);
1095 ClearHPageFreed(page);
1096 h->free_huge_pages--;
1097 h->free_huge_pages_node[nid]--;
1104 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1107 unsigned int cpuset_mems_cookie;
1108 struct zonelist *zonelist;
1111 int node = NUMA_NO_NODE;
1113 zonelist = node_zonelist(nid, gfp_mask);
1116 cpuset_mems_cookie = read_mems_allowed_begin();
1117 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1120 if (!cpuset_zone_allowed(zone, gfp_mask))
1123 * no need to ask again on the same node. Pool is node rather than
1126 if (zone_to_nid(zone) == node)
1128 node = zone_to_nid(zone);
1130 page = dequeue_huge_page_node_exact(h, node);
1134 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1140 static struct page *dequeue_huge_page_vma(struct hstate *h,
1141 struct vm_area_struct *vma,
1142 unsigned long address, int avoid_reserve,
1146 struct mempolicy *mpol;
1148 nodemask_t *nodemask;
1152 * A child process with MAP_PRIVATE mappings created by their parent
1153 * have no page reserves. This check ensures that reservations are
1154 * not "stolen". The child may still get SIGKILLed
1156 if (!vma_has_reserves(vma, chg) &&
1157 h->free_huge_pages - h->resv_huge_pages == 0)
1160 /* If reserves cannot be used, ensure enough pages are in the pool */
1161 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1164 gfp_mask = htlb_alloc_mask(h);
1165 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1166 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1167 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1168 SetHPageRestoreReserve(page);
1169 h->resv_huge_pages--;
1172 mpol_cond_put(mpol);
1180 * common helper functions for hstate_next_node_to_{alloc|free}.
1181 * We may have allocated or freed a huge page based on a different
1182 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1183 * be outside of *nodes_allowed. Ensure that we use an allowed
1184 * node for alloc or free.
1186 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1188 nid = next_node_in(nid, *nodes_allowed);
1189 VM_BUG_ON(nid >= MAX_NUMNODES);
1194 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1196 if (!node_isset(nid, *nodes_allowed))
1197 nid = next_node_allowed(nid, nodes_allowed);
1202 * returns the previously saved node ["this node"] from which to
1203 * allocate a persistent huge page for the pool and advance the
1204 * next node from which to allocate, handling wrap at end of node
1207 static int hstate_next_node_to_alloc(struct hstate *h,
1208 nodemask_t *nodes_allowed)
1212 VM_BUG_ON(!nodes_allowed);
1214 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1215 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1221 * helper for remove_pool_huge_page() - return the previously saved
1222 * node ["this node"] from which to free a huge page. Advance the
1223 * next node id whether or not we find a free huge page to free so
1224 * that the next attempt to free addresses the next node.
1226 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1230 VM_BUG_ON(!nodes_allowed);
1232 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1233 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1238 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1239 for (nr_nodes = nodes_weight(*mask); \
1241 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1244 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1245 for (nr_nodes = nodes_weight(*mask); \
1247 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1250 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1251 static void destroy_compound_gigantic_page(struct page *page,
1255 int nr_pages = 1 << order;
1256 struct page *p = page + 1;
1258 atomic_set(compound_mapcount_ptr(page), 0);
1259 atomic_set(compound_pincount_ptr(page), 0);
1261 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1262 clear_compound_head(p);
1263 set_page_refcounted(p);
1266 set_compound_order(page, 0);
1267 page[1].compound_nr = 0;
1268 __ClearPageHead(page);
1271 static void free_gigantic_page(struct page *page, unsigned int order)
1274 * If the page isn't allocated using the cma allocator,
1275 * cma_release() returns false.
1278 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1282 free_contig_range(page_to_pfn(page), 1 << order);
1285 #ifdef CONFIG_CONTIG_ALLOC
1286 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1287 int nid, nodemask_t *nodemask)
1289 unsigned long nr_pages = pages_per_huge_page(h);
1290 if (nid == NUMA_NO_NODE)
1291 nid = numa_mem_id();
1298 if (hugetlb_cma[nid]) {
1299 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1300 huge_page_order(h), true);
1305 if (!(gfp_mask & __GFP_THISNODE)) {
1306 for_each_node_mask(node, *nodemask) {
1307 if (node == nid || !hugetlb_cma[node])
1310 page = cma_alloc(hugetlb_cma[node], nr_pages,
1311 huge_page_order(h), true);
1319 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1322 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1323 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1324 #else /* !CONFIG_CONTIG_ALLOC */
1325 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1326 int nid, nodemask_t *nodemask)
1330 #endif /* CONFIG_CONTIG_ALLOC */
1332 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1333 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1334 int nid, nodemask_t *nodemask)
1338 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1339 static inline void destroy_compound_gigantic_page(struct page *page,
1340 unsigned int order) { }
1344 * Remove hugetlb page from lists, and update dtor so that page appears
1345 * as just a compound page. A reference is held on the page.
1347 * Must be called with hugetlb lock held.
1349 static void remove_hugetlb_page(struct hstate *h, struct page *page,
1350 bool adjust_surplus)
1352 int nid = page_to_nid(page);
1354 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1355 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1357 lockdep_assert_held(&hugetlb_lock);
1358 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1361 list_del(&page->lru);
1363 if (HPageFreed(page)) {
1364 h->free_huge_pages--;
1365 h->free_huge_pages_node[nid]--;
1367 if (adjust_surplus) {
1368 h->surplus_huge_pages--;
1369 h->surplus_huge_pages_node[nid]--;
1372 set_page_refcounted(page);
1373 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1376 h->nr_huge_pages_node[nid]--;
1379 static void update_and_free_page(struct hstate *h, struct page *page)
1382 struct page *subpage = page;
1384 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1387 for (i = 0; i < pages_per_huge_page(h);
1388 i++, subpage = mem_map_next(subpage, page, i)) {
1389 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1390 1 << PG_referenced | 1 << PG_dirty |
1391 1 << PG_active | 1 << PG_private |
1394 if (hstate_is_gigantic(h)) {
1395 destroy_compound_gigantic_page(page, huge_page_order(h));
1396 free_gigantic_page(page, huge_page_order(h));
1398 __free_pages(page, huge_page_order(h));
1402 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1404 struct page *page, *t_page;
1406 list_for_each_entry_safe(page, t_page, list, lru) {
1407 update_and_free_page(h, page);
1412 struct hstate *size_to_hstate(unsigned long size)
1416 for_each_hstate(h) {
1417 if (huge_page_size(h) == size)
1423 void free_huge_page(struct page *page)
1426 * Can't pass hstate in here because it is called from the
1427 * compound page destructor.
1429 struct hstate *h = page_hstate(page);
1430 int nid = page_to_nid(page);
1431 struct hugepage_subpool *spool = hugetlb_page_subpool(page);
1432 bool restore_reserve;
1433 unsigned long flags;
1435 VM_BUG_ON_PAGE(page_count(page), page);
1436 VM_BUG_ON_PAGE(page_mapcount(page), page);
1438 hugetlb_set_page_subpool(page, NULL);
1439 page->mapping = NULL;
1440 restore_reserve = HPageRestoreReserve(page);
1441 ClearHPageRestoreReserve(page);
1444 * If HPageRestoreReserve was set on page, page allocation consumed a
1445 * reservation. If the page was associated with a subpool, there
1446 * would have been a page reserved in the subpool before allocation
1447 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1448 * reservation, do not call hugepage_subpool_put_pages() as this will
1449 * remove the reserved page from the subpool.
1451 if (!restore_reserve) {
1453 * A return code of zero implies that the subpool will be
1454 * under its minimum size if the reservation is not restored
1455 * after page is free. Therefore, force restore_reserve
1458 if (hugepage_subpool_put_pages(spool, 1) == 0)
1459 restore_reserve = true;
1462 spin_lock_irqsave(&hugetlb_lock, flags);
1463 ClearHPageMigratable(page);
1464 hugetlb_cgroup_uncharge_page(hstate_index(h),
1465 pages_per_huge_page(h), page);
1466 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1467 pages_per_huge_page(h), page);
1468 if (restore_reserve)
1469 h->resv_huge_pages++;
1471 if (HPageTemporary(page)) {
1472 remove_hugetlb_page(h, page, false);
1473 spin_unlock_irqrestore(&hugetlb_lock, flags);
1474 update_and_free_page(h, page);
1475 } else if (h->surplus_huge_pages_node[nid]) {
1476 /* remove the page from active list */
1477 remove_hugetlb_page(h, page, true);
1478 spin_unlock_irqrestore(&hugetlb_lock, flags);
1479 update_and_free_page(h, page);
1481 arch_clear_hugepage_flags(page);
1482 enqueue_huge_page(h, page);
1483 spin_unlock_irqrestore(&hugetlb_lock, flags);
1487 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1489 INIT_LIST_HEAD(&page->lru);
1490 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1491 hugetlb_set_page_subpool(page, NULL);
1492 set_hugetlb_cgroup(page, NULL);
1493 set_hugetlb_cgroup_rsvd(page, NULL);
1494 spin_lock_irq(&hugetlb_lock);
1496 h->nr_huge_pages_node[nid]++;
1497 ClearHPageFreed(page);
1498 spin_unlock_irq(&hugetlb_lock);
1501 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1504 int nr_pages = 1 << order;
1505 struct page *p = page + 1;
1507 /* we rely on prep_new_huge_page to set the destructor */
1508 set_compound_order(page, order);
1509 __ClearPageReserved(page);
1510 __SetPageHead(page);
1511 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1513 * For gigantic hugepages allocated through bootmem at
1514 * boot, it's safer to be consistent with the not-gigantic
1515 * hugepages and clear the PG_reserved bit from all tail pages
1516 * too. Otherwise drivers using get_user_pages() to access tail
1517 * pages may get the reference counting wrong if they see
1518 * PG_reserved set on a tail page (despite the head page not
1519 * having PG_reserved set). Enforcing this consistency between
1520 * head and tail pages allows drivers to optimize away a check
1521 * on the head page when they need know if put_page() is needed
1522 * after get_user_pages().
1524 __ClearPageReserved(p);
1525 set_page_count(p, 0);
1526 set_compound_head(p, page);
1528 atomic_set(compound_mapcount_ptr(page), -1);
1529 atomic_set(compound_pincount_ptr(page), 0);
1533 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1534 * transparent huge pages. See the PageTransHuge() documentation for more
1537 int PageHuge(struct page *page)
1539 if (!PageCompound(page))
1542 page = compound_head(page);
1543 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1545 EXPORT_SYMBOL_GPL(PageHuge);
1548 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1549 * normal or transparent huge pages.
1551 int PageHeadHuge(struct page *page_head)
1553 if (!PageHead(page_head))
1556 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1560 * Find and lock address space (mapping) in write mode.
1562 * Upon entry, the page is locked which means that page_mapping() is
1563 * stable. Due to locking order, we can only trylock_write. If we can
1564 * not get the lock, simply return NULL to caller.
1566 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1568 struct address_space *mapping = page_mapping(hpage);
1573 if (i_mmap_trylock_write(mapping))
1579 pgoff_t __basepage_index(struct page *page)
1581 struct page *page_head = compound_head(page);
1582 pgoff_t index = page_index(page_head);
1583 unsigned long compound_idx;
1585 if (!PageHuge(page_head))
1586 return page_index(page);
1588 if (compound_order(page_head) >= MAX_ORDER)
1589 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1591 compound_idx = page - page_head;
1593 return (index << compound_order(page_head)) + compound_idx;
1596 static struct page *alloc_buddy_huge_page(struct hstate *h,
1597 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1598 nodemask_t *node_alloc_noretry)
1600 int order = huge_page_order(h);
1602 bool alloc_try_hard = true;
1605 * By default we always try hard to allocate the page with
1606 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1607 * a loop (to adjust global huge page counts) and previous allocation
1608 * failed, do not continue to try hard on the same node. Use the
1609 * node_alloc_noretry bitmap to manage this state information.
1611 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1612 alloc_try_hard = false;
1613 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1615 gfp_mask |= __GFP_RETRY_MAYFAIL;
1616 if (nid == NUMA_NO_NODE)
1617 nid = numa_mem_id();
1618 page = __alloc_pages(gfp_mask, order, nid, nmask);
1620 __count_vm_event(HTLB_BUDDY_PGALLOC);
1622 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1625 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1626 * indicates an overall state change. Clear bit so that we resume
1627 * normal 'try hard' allocations.
1629 if (node_alloc_noretry && page && !alloc_try_hard)
1630 node_clear(nid, *node_alloc_noretry);
1633 * If we tried hard to get a page but failed, set bit so that
1634 * subsequent attempts will not try as hard until there is an
1635 * overall state change.
1637 if (node_alloc_noretry && !page && alloc_try_hard)
1638 node_set(nid, *node_alloc_noretry);
1644 * Common helper to allocate a fresh hugetlb page. All specific allocators
1645 * should use this function to get new hugetlb pages
1647 static struct page *alloc_fresh_huge_page(struct hstate *h,
1648 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1649 nodemask_t *node_alloc_noretry)
1653 if (hstate_is_gigantic(h))
1654 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1656 page = alloc_buddy_huge_page(h, gfp_mask,
1657 nid, nmask, node_alloc_noretry);
1661 if (hstate_is_gigantic(h))
1662 prep_compound_gigantic_page(page, huge_page_order(h));
1663 prep_new_huge_page(h, page, page_to_nid(page));
1669 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1672 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1673 nodemask_t *node_alloc_noretry)
1677 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1679 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1680 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1681 node_alloc_noretry);
1689 put_page(page); /* free it into the hugepage allocator */
1695 * Remove huge page from pool from next node to free. Attempt to keep
1696 * persistent huge pages more or less balanced over allowed nodes.
1697 * This routine only 'removes' the hugetlb page. The caller must make
1698 * an additional call to free the page to low level allocators.
1699 * Called with hugetlb_lock locked.
1701 static struct page *remove_pool_huge_page(struct hstate *h,
1702 nodemask_t *nodes_allowed,
1706 struct page *page = NULL;
1708 lockdep_assert_held(&hugetlb_lock);
1709 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1711 * If we're returning unused surplus pages, only examine
1712 * nodes with surplus pages.
1714 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1715 !list_empty(&h->hugepage_freelists[node])) {
1716 page = list_entry(h->hugepage_freelists[node].next,
1718 remove_hugetlb_page(h, page, acct_surplus);
1727 * Dissolve a given free hugepage into free buddy pages. This function does
1728 * nothing for in-use hugepages and non-hugepages.
1729 * This function returns values like below:
1731 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1732 * (allocated or reserved.)
1733 * 0: successfully dissolved free hugepages or the page is not a
1734 * hugepage (considered as already dissolved)
1736 int dissolve_free_huge_page(struct page *page)
1741 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1742 if (!PageHuge(page))
1745 spin_lock_irq(&hugetlb_lock);
1746 if (!PageHuge(page)) {
1751 if (!page_count(page)) {
1752 struct page *head = compound_head(page);
1753 struct hstate *h = page_hstate(head);
1754 if (h->free_huge_pages - h->resv_huge_pages == 0)
1758 * We should make sure that the page is already on the free list
1759 * when it is dissolved.
1761 if (unlikely(!HPageFreed(head))) {
1762 spin_unlock_irq(&hugetlb_lock);
1766 * Theoretically, we should return -EBUSY when we
1767 * encounter this race. In fact, we have a chance
1768 * to successfully dissolve the page if we do a
1769 * retry. Because the race window is quite small.
1770 * If we seize this opportunity, it is an optimization
1771 * for increasing the success rate of dissolving page.
1777 * Move PageHWPoison flag from head page to the raw error page,
1778 * which makes any subpages rather than the error page reusable.
1780 if (PageHWPoison(head) && page != head) {
1781 SetPageHWPoison(page);
1782 ClearPageHWPoison(head);
1784 remove_hugetlb_page(h, page, false);
1785 h->max_huge_pages--;
1786 spin_unlock_irq(&hugetlb_lock);
1787 update_and_free_page(h, head);
1791 spin_unlock_irq(&hugetlb_lock);
1796 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1797 * make specified memory blocks removable from the system.
1798 * Note that this will dissolve a free gigantic hugepage completely, if any
1799 * part of it lies within the given range.
1800 * Also note that if dissolve_free_huge_page() returns with an error, all
1801 * free hugepages that were dissolved before that error are lost.
1803 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1809 if (!hugepages_supported())
1812 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1813 page = pfn_to_page(pfn);
1814 rc = dissolve_free_huge_page(page);
1823 * Allocates a fresh surplus page from the page allocator.
1825 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1826 int nid, nodemask_t *nmask)
1828 struct page *page = NULL;
1830 if (hstate_is_gigantic(h))
1833 spin_lock_irq(&hugetlb_lock);
1834 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1836 spin_unlock_irq(&hugetlb_lock);
1838 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1842 spin_lock_irq(&hugetlb_lock);
1844 * We could have raced with the pool size change.
1845 * Double check that and simply deallocate the new page
1846 * if we would end up overcommiting the surpluses. Abuse
1847 * temporary page to workaround the nasty free_huge_page
1850 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1851 SetHPageTemporary(page);
1852 spin_unlock_irq(&hugetlb_lock);
1856 h->surplus_huge_pages++;
1857 h->surplus_huge_pages_node[page_to_nid(page)]++;
1861 spin_unlock_irq(&hugetlb_lock);
1866 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1867 int nid, nodemask_t *nmask)
1871 if (hstate_is_gigantic(h))
1874 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1879 * We do not account these pages as surplus because they are only
1880 * temporary and will be released properly on the last reference
1882 SetHPageTemporary(page);
1888 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1891 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1892 struct vm_area_struct *vma, unsigned long addr)
1895 struct mempolicy *mpol;
1896 gfp_t gfp_mask = htlb_alloc_mask(h);
1898 nodemask_t *nodemask;
1900 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1901 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1902 mpol_cond_put(mpol);
1907 /* page migration callback function */
1908 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1909 nodemask_t *nmask, gfp_t gfp_mask)
1911 spin_lock_irq(&hugetlb_lock);
1912 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1915 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1917 spin_unlock_irq(&hugetlb_lock);
1921 spin_unlock_irq(&hugetlb_lock);
1923 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1926 /* mempolicy aware migration callback */
1927 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1928 unsigned long address)
1930 struct mempolicy *mpol;
1931 nodemask_t *nodemask;
1936 gfp_mask = htlb_alloc_mask(h);
1937 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1938 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
1939 mpol_cond_put(mpol);
1945 * Increase the hugetlb pool such that it can accommodate a reservation
1948 static int gather_surplus_pages(struct hstate *h, long delta)
1949 __must_hold(&hugetlb_lock)
1951 struct list_head surplus_list;
1952 struct page *page, *tmp;
1955 long needed, allocated;
1956 bool alloc_ok = true;
1958 lockdep_assert_held(&hugetlb_lock);
1959 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1961 h->resv_huge_pages += delta;
1966 INIT_LIST_HEAD(&surplus_list);
1970 spin_unlock_irq(&hugetlb_lock);
1971 for (i = 0; i < needed; i++) {
1972 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1973 NUMA_NO_NODE, NULL);
1978 list_add(&page->lru, &surplus_list);
1984 * After retaking hugetlb_lock, we need to recalculate 'needed'
1985 * because either resv_huge_pages or free_huge_pages may have changed.
1987 spin_lock_irq(&hugetlb_lock);
1988 needed = (h->resv_huge_pages + delta) -
1989 (h->free_huge_pages + allocated);
1994 * We were not able to allocate enough pages to
1995 * satisfy the entire reservation so we free what
1996 * we've allocated so far.
2001 * The surplus_list now contains _at_least_ the number of extra pages
2002 * needed to accommodate the reservation. Add the appropriate number
2003 * of pages to the hugetlb pool and free the extras back to the buddy
2004 * allocator. Commit the entire reservation here to prevent another
2005 * process from stealing the pages as they are added to the pool but
2006 * before they are reserved.
2008 needed += allocated;
2009 h->resv_huge_pages += delta;
2012 /* Free the needed pages to the hugetlb pool */
2013 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2019 * This page is now managed by the hugetlb allocator and has
2020 * no users -- drop the buddy allocator's reference.
2022 zeroed = put_page_testzero(page);
2023 VM_BUG_ON_PAGE(!zeroed, page);
2024 enqueue_huge_page(h, page);
2027 spin_unlock_irq(&hugetlb_lock);
2029 /* Free unnecessary surplus pages to the buddy allocator */
2030 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2032 spin_lock_irq(&hugetlb_lock);
2038 * This routine has two main purposes:
2039 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2040 * in unused_resv_pages. This corresponds to the prior adjustments made
2041 * to the associated reservation map.
2042 * 2) Free any unused surplus pages that may have been allocated to satisfy
2043 * the reservation. As many as unused_resv_pages may be freed.
2045 static void return_unused_surplus_pages(struct hstate *h,
2046 unsigned long unused_resv_pages)
2048 unsigned long nr_pages;
2050 LIST_HEAD(page_list);
2052 lockdep_assert_held(&hugetlb_lock);
2053 /* Uncommit the reservation */
2054 h->resv_huge_pages -= unused_resv_pages;
2056 /* Cannot return gigantic pages currently */
2057 if (hstate_is_gigantic(h))
2061 * Part (or even all) of the reservation could have been backed
2062 * by pre-allocated pages. Only free surplus pages.
2064 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2067 * We want to release as many surplus pages as possible, spread
2068 * evenly across all nodes with memory. Iterate across these nodes
2069 * until we can no longer free unreserved surplus pages. This occurs
2070 * when the nodes with surplus pages have no free pages.
2071 * remove_pool_huge_page() will balance the freed pages across the
2072 * on-line nodes with memory and will handle the hstate accounting.
2074 while (nr_pages--) {
2075 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2079 list_add(&page->lru, &page_list);
2083 spin_unlock_irq(&hugetlb_lock);
2084 update_and_free_pages_bulk(h, &page_list);
2085 spin_lock_irq(&hugetlb_lock);
2090 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2091 * are used by the huge page allocation routines to manage reservations.
2093 * vma_needs_reservation is called to determine if the huge page at addr
2094 * within the vma has an associated reservation. If a reservation is
2095 * needed, the value 1 is returned. The caller is then responsible for
2096 * managing the global reservation and subpool usage counts. After
2097 * the huge page has been allocated, vma_commit_reservation is called
2098 * to add the page to the reservation map. If the page allocation fails,
2099 * the reservation must be ended instead of committed. vma_end_reservation
2100 * is called in such cases.
2102 * In the normal case, vma_commit_reservation returns the same value
2103 * as the preceding vma_needs_reservation call. The only time this
2104 * is not the case is if a reserve map was changed between calls. It
2105 * is the responsibility of the caller to notice the difference and
2106 * take appropriate action.
2108 * vma_add_reservation is used in error paths where a reservation must
2109 * be restored when a newly allocated huge page must be freed. It is
2110 * to be called after calling vma_needs_reservation to determine if a
2111 * reservation exists.
2113 enum vma_resv_mode {
2119 static long __vma_reservation_common(struct hstate *h,
2120 struct vm_area_struct *vma, unsigned long addr,
2121 enum vma_resv_mode mode)
2123 struct resv_map *resv;
2126 long dummy_out_regions_needed;
2128 resv = vma_resv_map(vma);
2132 idx = vma_hugecache_offset(h, vma, addr);
2134 case VMA_NEEDS_RESV:
2135 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2136 /* We assume that vma_reservation_* routines always operate on
2137 * 1 page, and that adding to resv map a 1 page entry can only
2138 * ever require 1 region.
2140 VM_BUG_ON(dummy_out_regions_needed != 1);
2142 case VMA_COMMIT_RESV:
2143 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2144 /* region_add calls of range 1 should never fail. */
2148 region_abort(resv, idx, idx + 1, 1);
2152 if (vma->vm_flags & VM_MAYSHARE) {
2153 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2154 /* region_add calls of range 1 should never fail. */
2157 region_abort(resv, idx, idx + 1, 1);
2158 ret = region_del(resv, idx, idx + 1);
2165 if (vma->vm_flags & VM_MAYSHARE)
2168 * We know private mapping must have HPAGE_RESV_OWNER set.
2170 * In most cases, reserves always exist for private mappings.
2171 * However, a file associated with mapping could have been
2172 * hole punched or truncated after reserves were consumed.
2173 * As subsequent fault on such a range will not use reserves.
2174 * Subtle - The reserve map for private mappings has the
2175 * opposite meaning than that of shared mappings. If NO
2176 * entry is in the reserve map, it means a reservation exists.
2177 * If an entry exists in the reserve map, it means the
2178 * reservation has already been consumed. As a result, the
2179 * return value of this routine is the opposite of the
2180 * value returned from reserve map manipulation routines above.
2189 static long vma_needs_reservation(struct hstate *h,
2190 struct vm_area_struct *vma, unsigned long addr)
2192 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2195 static long vma_commit_reservation(struct hstate *h,
2196 struct vm_area_struct *vma, unsigned long addr)
2198 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2201 static void vma_end_reservation(struct hstate *h,
2202 struct vm_area_struct *vma, unsigned long addr)
2204 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2207 static long vma_add_reservation(struct hstate *h,
2208 struct vm_area_struct *vma, unsigned long addr)
2210 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2214 * This routine is called to restore a reservation on error paths. In the
2215 * specific error paths, a huge page was allocated (via alloc_huge_page)
2216 * and is about to be freed. If a reservation for the page existed,
2217 * alloc_huge_page would have consumed the reservation and set
2218 * HPageRestoreReserve in the newly allocated page. When the page is freed
2219 * via free_huge_page, the global reservation count will be incremented if
2220 * HPageRestoreReserve is set. However, free_huge_page can not adjust the
2221 * reserve map. Adjust the reserve map here to be consistent with global
2222 * reserve count adjustments to be made by free_huge_page.
2224 static void restore_reserve_on_error(struct hstate *h,
2225 struct vm_area_struct *vma, unsigned long address,
2228 if (unlikely(HPageRestoreReserve(page))) {
2229 long rc = vma_needs_reservation(h, vma, address);
2231 if (unlikely(rc < 0)) {
2233 * Rare out of memory condition in reserve map
2234 * manipulation. Clear HPageRestoreReserve so that
2235 * global reserve count will not be incremented
2236 * by free_huge_page. This will make it appear
2237 * as though the reservation for this page was
2238 * consumed. This may prevent the task from
2239 * faulting in the page at a later time. This
2240 * is better than inconsistent global huge page
2241 * accounting of reserve counts.
2243 ClearHPageRestoreReserve(page);
2245 rc = vma_add_reservation(h, vma, address);
2246 if (unlikely(rc < 0))
2248 * See above comment about rare out of
2251 ClearHPageRestoreReserve(page);
2253 vma_end_reservation(h, vma, address);
2257 struct page *alloc_huge_page(struct vm_area_struct *vma,
2258 unsigned long addr, int avoid_reserve)
2260 struct hugepage_subpool *spool = subpool_vma(vma);
2261 struct hstate *h = hstate_vma(vma);
2263 long map_chg, map_commit;
2266 struct hugetlb_cgroup *h_cg;
2267 bool deferred_reserve;
2269 idx = hstate_index(h);
2271 * Examine the region/reserve map to determine if the process
2272 * has a reservation for the page to be allocated. A return
2273 * code of zero indicates a reservation exists (no change).
2275 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2277 return ERR_PTR(-ENOMEM);
2280 * Processes that did not create the mapping will have no
2281 * reserves as indicated by the region/reserve map. Check
2282 * that the allocation will not exceed the subpool limit.
2283 * Allocations for MAP_NORESERVE mappings also need to be
2284 * checked against any subpool limit.
2286 if (map_chg || avoid_reserve) {
2287 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2289 vma_end_reservation(h, vma, addr);
2290 return ERR_PTR(-ENOSPC);
2294 * Even though there was no reservation in the region/reserve
2295 * map, there could be reservations associated with the
2296 * subpool that can be used. This would be indicated if the
2297 * return value of hugepage_subpool_get_pages() is zero.
2298 * However, if avoid_reserve is specified we still avoid even
2299 * the subpool reservations.
2305 /* If this allocation is not consuming a reservation, charge it now.
2307 deferred_reserve = map_chg || avoid_reserve;
2308 if (deferred_reserve) {
2309 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2310 idx, pages_per_huge_page(h), &h_cg);
2312 goto out_subpool_put;
2315 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2317 goto out_uncharge_cgroup_reservation;
2319 spin_lock_irq(&hugetlb_lock);
2321 * glb_chg is passed to indicate whether or not a page must be taken
2322 * from the global free pool (global change). gbl_chg == 0 indicates
2323 * a reservation exists for the allocation.
2325 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2327 spin_unlock_irq(&hugetlb_lock);
2328 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2330 goto out_uncharge_cgroup;
2331 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2332 SetHPageRestoreReserve(page);
2333 h->resv_huge_pages--;
2335 spin_lock_irq(&hugetlb_lock);
2336 list_add(&page->lru, &h->hugepage_activelist);
2339 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2340 /* If allocation is not consuming a reservation, also store the
2341 * hugetlb_cgroup pointer on the page.
2343 if (deferred_reserve) {
2344 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2348 spin_unlock_irq(&hugetlb_lock);
2350 hugetlb_set_page_subpool(page, spool);
2352 map_commit = vma_commit_reservation(h, vma, addr);
2353 if (unlikely(map_chg > map_commit)) {
2355 * The page was added to the reservation map between
2356 * vma_needs_reservation and vma_commit_reservation.
2357 * This indicates a race with hugetlb_reserve_pages.
2358 * Adjust for the subpool count incremented above AND
2359 * in hugetlb_reserve_pages for the same page. Also,
2360 * the reservation count added in hugetlb_reserve_pages
2361 * no longer applies.
2365 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2366 hugetlb_acct_memory(h, -rsv_adjust);
2367 if (deferred_reserve)
2368 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2369 pages_per_huge_page(h), page);
2373 out_uncharge_cgroup:
2374 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2375 out_uncharge_cgroup_reservation:
2376 if (deferred_reserve)
2377 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2380 if (map_chg || avoid_reserve)
2381 hugepage_subpool_put_pages(spool, 1);
2382 vma_end_reservation(h, vma, addr);
2383 return ERR_PTR(-ENOSPC);
2386 int alloc_bootmem_huge_page(struct hstate *h)
2387 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2388 int __alloc_bootmem_huge_page(struct hstate *h)
2390 struct huge_bootmem_page *m;
2393 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2396 addr = memblock_alloc_try_nid_raw(
2397 huge_page_size(h), huge_page_size(h),
2398 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2401 * Use the beginning of the huge page to store the
2402 * huge_bootmem_page struct (until gather_bootmem
2403 * puts them into the mem_map).
2412 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2413 /* Put them into a private list first because mem_map is not up yet */
2414 INIT_LIST_HEAD(&m->list);
2415 list_add(&m->list, &huge_boot_pages);
2420 static void __init prep_compound_huge_page(struct page *page,
2423 if (unlikely(order > (MAX_ORDER - 1)))
2424 prep_compound_gigantic_page(page, order);
2426 prep_compound_page(page, order);
2429 /* Put bootmem huge pages into the standard lists after mem_map is up */
2430 static void __init gather_bootmem_prealloc(void)
2432 struct huge_bootmem_page *m;
2434 list_for_each_entry(m, &huge_boot_pages, list) {
2435 struct page *page = virt_to_page(m);
2436 struct hstate *h = m->hstate;
2438 WARN_ON(page_count(page) != 1);
2439 prep_compound_huge_page(page, huge_page_order(h));
2440 WARN_ON(PageReserved(page));
2441 prep_new_huge_page(h, page, page_to_nid(page));
2442 put_page(page); /* free it into the hugepage allocator */
2445 * If we had gigantic hugepages allocated at boot time, we need
2446 * to restore the 'stolen' pages to totalram_pages in order to
2447 * fix confusing memory reports from free(1) and another
2448 * side-effects, like CommitLimit going negative.
2450 if (hstate_is_gigantic(h))
2451 adjust_managed_page_count(page, pages_per_huge_page(h));
2456 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2459 nodemask_t *node_alloc_noretry;
2461 if (!hstate_is_gigantic(h)) {
2463 * Bit mask controlling how hard we retry per-node allocations.
2464 * Ignore errors as lower level routines can deal with
2465 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2466 * time, we are likely in bigger trouble.
2468 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2471 /* allocations done at boot time */
2472 node_alloc_noretry = NULL;
2475 /* bit mask controlling how hard we retry per-node allocations */
2476 if (node_alloc_noretry)
2477 nodes_clear(*node_alloc_noretry);
2479 for (i = 0; i < h->max_huge_pages; ++i) {
2480 if (hstate_is_gigantic(h)) {
2481 if (hugetlb_cma_size) {
2482 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2485 if (!alloc_bootmem_huge_page(h))
2487 } else if (!alloc_pool_huge_page(h,
2488 &node_states[N_MEMORY],
2489 node_alloc_noretry))
2493 if (i < h->max_huge_pages) {
2496 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2497 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2498 h->max_huge_pages, buf, i);
2499 h->max_huge_pages = i;
2502 kfree(node_alloc_noretry);
2505 static void __init hugetlb_init_hstates(void)
2509 for_each_hstate(h) {
2510 if (minimum_order > huge_page_order(h))
2511 minimum_order = huge_page_order(h);
2513 /* oversize hugepages were init'ed in early boot */
2514 if (!hstate_is_gigantic(h))
2515 hugetlb_hstate_alloc_pages(h);
2517 VM_BUG_ON(minimum_order == UINT_MAX);
2520 static void __init report_hugepages(void)
2524 for_each_hstate(h) {
2527 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2528 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2529 buf, h->free_huge_pages);
2533 #ifdef CONFIG_HIGHMEM
2534 static void try_to_free_low(struct hstate *h, unsigned long count,
2535 nodemask_t *nodes_allowed)
2538 LIST_HEAD(page_list);
2540 lockdep_assert_held(&hugetlb_lock);
2541 if (hstate_is_gigantic(h))
2545 * Collect pages to be freed on a list, and free after dropping lock
2547 for_each_node_mask(i, *nodes_allowed) {
2548 struct page *page, *next;
2549 struct list_head *freel = &h->hugepage_freelists[i];
2550 list_for_each_entry_safe(page, next, freel, lru) {
2551 if (count >= h->nr_huge_pages)
2553 if (PageHighMem(page))
2555 remove_hugetlb_page(h, page, false);
2556 list_add(&page->lru, &page_list);
2561 spin_unlock_irq(&hugetlb_lock);
2562 update_and_free_pages_bulk(h, &page_list);
2563 spin_lock_irq(&hugetlb_lock);
2566 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2567 nodemask_t *nodes_allowed)
2573 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2574 * balanced by operating on them in a round-robin fashion.
2575 * Returns 1 if an adjustment was made.
2577 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2582 lockdep_assert_held(&hugetlb_lock);
2583 VM_BUG_ON(delta != -1 && delta != 1);
2586 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2587 if (h->surplus_huge_pages_node[node])
2591 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2592 if (h->surplus_huge_pages_node[node] <
2593 h->nr_huge_pages_node[node])
2600 h->surplus_huge_pages += delta;
2601 h->surplus_huge_pages_node[node] += delta;
2605 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2606 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2607 nodemask_t *nodes_allowed)
2609 unsigned long min_count, ret;
2611 LIST_HEAD(page_list);
2612 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2615 * Bit mask controlling how hard we retry per-node allocations.
2616 * If we can not allocate the bit mask, do not attempt to allocate
2617 * the requested huge pages.
2619 if (node_alloc_noretry)
2620 nodes_clear(*node_alloc_noretry);
2625 * resize_lock mutex prevents concurrent adjustments to number of
2626 * pages in hstate via the proc/sysfs interfaces.
2628 mutex_lock(&h->resize_lock);
2629 spin_lock_irq(&hugetlb_lock);
2632 * Check for a node specific request.
2633 * Changing node specific huge page count may require a corresponding
2634 * change to the global count. In any case, the passed node mask
2635 * (nodes_allowed) will restrict alloc/free to the specified node.
2637 if (nid != NUMA_NO_NODE) {
2638 unsigned long old_count = count;
2640 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2642 * User may have specified a large count value which caused the
2643 * above calculation to overflow. In this case, they wanted
2644 * to allocate as many huge pages as possible. Set count to
2645 * largest possible value to align with their intention.
2647 if (count < old_count)
2652 * Gigantic pages runtime allocation depend on the capability for large
2653 * page range allocation.
2654 * If the system does not provide this feature, return an error when
2655 * the user tries to allocate gigantic pages but let the user free the
2656 * boottime allocated gigantic pages.
2658 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2659 if (count > persistent_huge_pages(h)) {
2660 spin_unlock_irq(&hugetlb_lock);
2661 mutex_unlock(&h->resize_lock);
2662 NODEMASK_FREE(node_alloc_noretry);
2665 /* Fall through to decrease pool */
2669 * Increase the pool size
2670 * First take pages out of surplus state. Then make up the
2671 * remaining difference by allocating fresh huge pages.
2673 * We might race with alloc_surplus_huge_page() here and be unable
2674 * to convert a surplus huge page to a normal huge page. That is
2675 * not critical, though, it just means the overall size of the
2676 * pool might be one hugepage larger than it needs to be, but
2677 * within all the constraints specified by the sysctls.
2679 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2680 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2684 while (count > persistent_huge_pages(h)) {
2686 * If this allocation races such that we no longer need the
2687 * page, free_huge_page will handle it by freeing the page
2688 * and reducing the surplus.
2690 spin_unlock_irq(&hugetlb_lock);
2692 /* yield cpu to avoid soft lockup */
2695 ret = alloc_pool_huge_page(h, nodes_allowed,
2696 node_alloc_noretry);
2697 spin_lock_irq(&hugetlb_lock);
2701 /* Bail for signals. Probably ctrl-c from user */
2702 if (signal_pending(current))
2707 * Decrease the pool size
2708 * First return free pages to the buddy allocator (being careful
2709 * to keep enough around to satisfy reservations). Then place
2710 * pages into surplus state as needed so the pool will shrink
2711 * to the desired size as pages become free.
2713 * By placing pages into the surplus state independent of the
2714 * overcommit value, we are allowing the surplus pool size to
2715 * exceed overcommit. There are few sane options here. Since
2716 * alloc_surplus_huge_page() is checking the global counter,
2717 * though, we'll note that we're not allowed to exceed surplus
2718 * and won't grow the pool anywhere else. Not until one of the
2719 * sysctls are changed, or the surplus pages go out of use.
2721 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2722 min_count = max(count, min_count);
2723 try_to_free_low(h, min_count, nodes_allowed);
2726 * Collect pages to be removed on list without dropping lock
2728 while (min_count < persistent_huge_pages(h)) {
2729 page = remove_pool_huge_page(h, nodes_allowed, 0);
2733 list_add(&page->lru, &page_list);
2735 /* free the pages after dropping lock */
2736 spin_unlock_irq(&hugetlb_lock);
2737 update_and_free_pages_bulk(h, &page_list);
2738 spin_lock_irq(&hugetlb_lock);
2740 while (count < persistent_huge_pages(h)) {
2741 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2745 h->max_huge_pages = persistent_huge_pages(h);
2746 spin_unlock_irq(&hugetlb_lock);
2747 mutex_unlock(&h->resize_lock);
2749 NODEMASK_FREE(node_alloc_noretry);
2754 #define HSTATE_ATTR_RO(_name) \
2755 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2757 #define HSTATE_ATTR(_name) \
2758 static struct kobj_attribute _name##_attr = \
2759 __ATTR(_name, 0644, _name##_show, _name##_store)
2761 static struct kobject *hugepages_kobj;
2762 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2764 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2766 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2770 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2771 if (hstate_kobjs[i] == kobj) {
2773 *nidp = NUMA_NO_NODE;
2777 return kobj_to_node_hstate(kobj, nidp);
2780 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2781 struct kobj_attribute *attr, char *buf)
2784 unsigned long nr_huge_pages;
2787 h = kobj_to_hstate(kobj, &nid);
2788 if (nid == NUMA_NO_NODE)
2789 nr_huge_pages = h->nr_huge_pages;
2791 nr_huge_pages = h->nr_huge_pages_node[nid];
2793 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
2796 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2797 struct hstate *h, int nid,
2798 unsigned long count, size_t len)
2801 nodemask_t nodes_allowed, *n_mask;
2803 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2806 if (nid == NUMA_NO_NODE) {
2808 * global hstate attribute
2810 if (!(obey_mempolicy &&
2811 init_nodemask_of_mempolicy(&nodes_allowed)))
2812 n_mask = &node_states[N_MEMORY];
2814 n_mask = &nodes_allowed;
2817 * Node specific request. count adjustment happens in
2818 * set_max_huge_pages() after acquiring hugetlb_lock.
2820 init_nodemask_of_node(&nodes_allowed, nid);
2821 n_mask = &nodes_allowed;
2824 err = set_max_huge_pages(h, count, nid, n_mask);
2826 return err ? err : len;
2829 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2830 struct kobject *kobj, const char *buf,
2834 unsigned long count;
2838 err = kstrtoul(buf, 10, &count);
2842 h = kobj_to_hstate(kobj, &nid);
2843 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2846 static ssize_t nr_hugepages_show(struct kobject *kobj,
2847 struct kobj_attribute *attr, char *buf)
2849 return nr_hugepages_show_common(kobj, attr, buf);
2852 static ssize_t nr_hugepages_store(struct kobject *kobj,
2853 struct kobj_attribute *attr, const char *buf, size_t len)
2855 return nr_hugepages_store_common(false, kobj, buf, len);
2857 HSTATE_ATTR(nr_hugepages);
2862 * hstate attribute for optionally mempolicy-based constraint on persistent
2863 * huge page alloc/free.
2865 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2866 struct kobj_attribute *attr,
2869 return nr_hugepages_show_common(kobj, attr, buf);
2872 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2873 struct kobj_attribute *attr, const char *buf, size_t len)
2875 return nr_hugepages_store_common(true, kobj, buf, len);
2877 HSTATE_ATTR(nr_hugepages_mempolicy);
2881 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2882 struct kobj_attribute *attr, char *buf)
2884 struct hstate *h = kobj_to_hstate(kobj, NULL);
2885 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
2888 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2889 struct kobj_attribute *attr, const char *buf, size_t count)
2892 unsigned long input;
2893 struct hstate *h = kobj_to_hstate(kobj, NULL);
2895 if (hstate_is_gigantic(h))
2898 err = kstrtoul(buf, 10, &input);
2902 spin_lock_irq(&hugetlb_lock);
2903 h->nr_overcommit_huge_pages = input;
2904 spin_unlock_irq(&hugetlb_lock);
2908 HSTATE_ATTR(nr_overcommit_hugepages);
2910 static ssize_t free_hugepages_show(struct kobject *kobj,
2911 struct kobj_attribute *attr, char *buf)
2914 unsigned long free_huge_pages;
2917 h = kobj_to_hstate(kobj, &nid);
2918 if (nid == NUMA_NO_NODE)
2919 free_huge_pages = h->free_huge_pages;
2921 free_huge_pages = h->free_huge_pages_node[nid];
2923 return sysfs_emit(buf, "%lu\n", free_huge_pages);
2925 HSTATE_ATTR_RO(free_hugepages);
2927 static ssize_t resv_hugepages_show(struct kobject *kobj,
2928 struct kobj_attribute *attr, char *buf)
2930 struct hstate *h = kobj_to_hstate(kobj, NULL);
2931 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
2933 HSTATE_ATTR_RO(resv_hugepages);
2935 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2936 struct kobj_attribute *attr, char *buf)
2939 unsigned long surplus_huge_pages;
2942 h = kobj_to_hstate(kobj, &nid);
2943 if (nid == NUMA_NO_NODE)
2944 surplus_huge_pages = h->surplus_huge_pages;
2946 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2948 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
2950 HSTATE_ATTR_RO(surplus_hugepages);
2952 static struct attribute *hstate_attrs[] = {
2953 &nr_hugepages_attr.attr,
2954 &nr_overcommit_hugepages_attr.attr,
2955 &free_hugepages_attr.attr,
2956 &resv_hugepages_attr.attr,
2957 &surplus_hugepages_attr.attr,
2959 &nr_hugepages_mempolicy_attr.attr,
2964 static const struct attribute_group hstate_attr_group = {
2965 .attrs = hstate_attrs,
2968 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2969 struct kobject **hstate_kobjs,
2970 const struct attribute_group *hstate_attr_group)
2973 int hi = hstate_index(h);
2975 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2976 if (!hstate_kobjs[hi])
2979 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2981 kobject_put(hstate_kobjs[hi]);
2982 hstate_kobjs[hi] = NULL;
2988 static void __init hugetlb_sysfs_init(void)
2993 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2994 if (!hugepages_kobj)
2997 for_each_hstate(h) {
2998 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2999 hstate_kobjs, &hstate_attr_group);
3001 pr_err("HugeTLB: Unable to add hstate %s", h->name);
3008 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3009 * with node devices in node_devices[] using a parallel array. The array
3010 * index of a node device or _hstate == node id.
3011 * This is here to avoid any static dependency of the node device driver, in
3012 * the base kernel, on the hugetlb module.
3014 struct node_hstate {
3015 struct kobject *hugepages_kobj;
3016 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3018 static struct node_hstate node_hstates[MAX_NUMNODES];
3021 * A subset of global hstate attributes for node devices
3023 static struct attribute *per_node_hstate_attrs[] = {
3024 &nr_hugepages_attr.attr,
3025 &free_hugepages_attr.attr,
3026 &surplus_hugepages_attr.attr,
3030 static const struct attribute_group per_node_hstate_attr_group = {
3031 .attrs = per_node_hstate_attrs,
3035 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3036 * Returns node id via non-NULL nidp.
3038 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3042 for (nid = 0; nid < nr_node_ids; nid++) {
3043 struct node_hstate *nhs = &node_hstates[nid];
3045 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3046 if (nhs->hstate_kobjs[i] == kobj) {
3058 * Unregister hstate attributes from a single node device.
3059 * No-op if no hstate attributes attached.
3061 static void hugetlb_unregister_node(struct node *node)
3064 struct node_hstate *nhs = &node_hstates[node->dev.id];
3066 if (!nhs->hugepages_kobj)
3067 return; /* no hstate attributes */
3069 for_each_hstate(h) {
3070 int idx = hstate_index(h);
3071 if (nhs->hstate_kobjs[idx]) {
3072 kobject_put(nhs->hstate_kobjs[idx]);
3073 nhs->hstate_kobjs[idx] = NULL;
3077 kobject_put(nhs->hugepages_kobj);
3078 nhs->hugepages_kobj = NULL;
3083 * Register hstate attributes for a single node device.
3084 * No-op if attributes already registered.
3086 static void hugetlb_register_node(struct node *node)
3089 struct node_hstate *nhs = &node_hstates[node->dev.id];
3092 if (nhs->hugepages_kobj)
3093 return; /* already allocated */
3095 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3097 if (!nhs->hugepages_kobj)
3100 for_each_hstate(h) {
3101 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3103 &per_node_hstate_attr_group);
3105 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3106 h->name, node->dev.id);
3107 hugetlb_unregister_node(node);
3114 * hugetlb init time: register hstate attributes for all registered node
3115 * devices of nodes that have memory. All on-line nodes should have
3116 * registered their associated device by this time.
3118 static void __init hugetlb_register_all_nodes(void)
3122 for_each_node_state(nid, N_MEMORY) {
3123 struct node *node = node_devices[nid];
3124 if (node->dev.id == nid)
3125 hugetlb_register_node(node);
3129 * Let the node device driver know we're here so it can
3130 * [un]register hstate attributes on node hotplug.
3132 register_hugetlbfs_with_node(hugetlb_register_node,
3133 hugetlb_unregister_node);
3135 #else /* !CONFIG_NUMA */
3137 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3145 static void hugetlb_register_all_nodes(void) { }
3149 static int __init hugetlb_init(void)
3153 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
3156 if (!hugepages_supported()) {
3157 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3158 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3163 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3164 * architectures depend on setup being done here.
3166 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3167 if (!parsed_default_hugepagesz) {
3169 * If we did not parse a default huge page size, set
3170 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3171 * number of huge pages for this default size was implicitly
3172 * specified, set that here as well.
3173 * Note that the implicit setting will overwrite an explicit
3174 * setting. A warning will be printed in this case.
3176 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3177 if (default_hstate_max_huge_pages) {
3178 if (default_hstate.max_huge_pages) {
3181 string_get_size(huge_page_size(&default_hstate),
3182 1, STRING_UNITS_2, buf, 32);
3183 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3184 default_hstate.max_huge_pages, buf);
3185 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3186 default_hstate_max_huge_pages);
3188 default_hstate.max_huge_pages =
3189 default_hstate_max_huge_pages;
3193 hugetlb_cma_check();
3194 hugetlb_init_hstates();
3195 gather_bootmem_prealloc();
3198 hugetlb_sysfs_init();
3199 hugetlb_register_all_nodes();
3200 hugetlb_cgroup_file_init();
3203 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3205 num_fault_mutexes = 1;
3207 hugetlb_fault_mutex_table =
3208 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3210 BUG_ON(!hugetlb_fault_mutex_table);
3212 for (i = 0; i < num_fault_mutexes; i++)
3213 mutex_init(&hugetlb_fault_mutex_table[i]);
3216 subsys_initcall(hugetlb_init);
3218 /* Overwritten by architectures with more huge page sizes */
3219 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3221 return size == HPAGE_SIZE;
3224 void __init hugetlb_add_hstate(unsigned int order)
3229 if (size_to_hstate(PAGE_SIZE << order)) {
3232 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3234 h = &hstates[hugetlb_max_hstate++];
3235 mutex_init(&h->resize_lock);
3237 h->mask = ~(huge_page_size(h) - 1);
3238 for (i = 0; i < MAX_NUMNODES; ++i)
3239 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3240 INIT_LIST_HEAD(&h->hugepage_activelist);
3241 h->next_nid_to_alloc = first_memory_node;
3242 h->next_nid_to_free = first_memory_node;
3243 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3244 huge_page_size(h)/1024);
3250 * hugepages command line processing
3251 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3252 * specification. If not, ignore the hugepages value. hugepages can also
3253 * be the first huge page command line option in which case it implicitly
3254 * specifies the number of huge pages for the default size.
3256 static int __init hugepages_setup(char *s)
3259 static unsigned long *last_mhp;
3261 if (!parsed_valid_hugepagesz) {
3262 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3263 parsed_valid_hugepagesz = true;
3268 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3269 * yet, so this hugepages= parameter goes to the "default hstate".
3270 * Otherwise, it goes with the previously parsed hugepagesz or
3271 * default_hugepagesz.
3273 else if (!hugetlb_max_hstate)
3274 mhp = &default_hstate_max_huge_pages;
3276 mhp = &parsed_hstate->max_huge_pages;
3278 if (mhp == last_mhp) {
3279 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3283 if (sscanf(s, "%lu", mhp) <= 0)
3287 * Global state is always initialized later in hugetlb_init.
3288 * But we need to allocate gigantic hstates here early to still
3289 * use the bootmem allocator.
3291 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
3292 hugetlb_hstate_alloc_pages(parsed_hstate);
3298 __setup("hugepages=", hugepages_setup);
3301 * hugepagesz command line processing
3302 * A specific huge page size can only be specified once with hugepagesz.
3303 * hugepagesz is followed by hugepages on the command line. The global
3304 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3305 * hugepagesz argument was valid.
3307 static int __init hugepagesz_setup(char *s)
3312 parsed_valid_hugepagesz = false;
3313 size = (unsigned long)memparse(s, NULL);
3315 if (!arch_hugetlb_valid_size(size)) {
3316 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3320 h = size_to_hstate(size);
3323 * hstate for this size already exists. This is normally
3324 * an error, but is allowed if the existing hstate is the
3325 * default hstate. More specifically, it is only allowed if
3326 * the number of huge pages for the default hstate was not
3327 * previously specified.
3329 if (!parsed_default_hugepagesz || h != &default_hstate ||
3330 default_hstate.max_huge_pages) {
3331 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3336 * No need to call hugetlb_add_hstate() as hstate already
3337 * exists. But, do set parsed_hstate so that a following
3338 * hugepages= parameter will be applied to this hstate.
3341 parsed_valid_hugepagesz = true;
3345 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3346 parsed_valid_hugepagesz = true;
3349 __setup("hugepagesz=", hugepagesz_setup);
3352 * default_hugepagesz command line input
3353 * Only one instance of default_hugepagesz allowed on command line.
3355 static int __init default_hugepagesz_setup(char *s)
3359 parsed_valid_hugepagesz = false;
3360 if (parsed_default_hugepagesz) {
3361 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3365 size = (unsigned long)memparse(s, NULL);
3367 if (!arch_hugetlb_valid_size(size)) {
3368 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3372 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3373 parsed_valid_hugepagesz = true;
3374 parsed_default_hugepagesz = true;
3375 default_hstate_idx = hstate_index(size_to_hstate(size));
3378 * The number of default huge pages (for this size) could have been
3379 * specified as the first hugetlb parameter: hugepages=X. If so,
3380 * then default_hstate_max_huge_pages is set. If the default huge
3381 * page size is gigantic (>= MAX_ORDER), then the pages must be
3382 * allocated here from bootmem allocator.
3384 if (default_hstate_max_huge_pages) {
3385 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3386 if (hstate_is_gigantic(&default_hstate))
3387 hugetlb_hstate_alloc_pages(&default_hstate);
3388 default_hstate_max_huge_pages = 0;
3393 __setup("default_hugepagesz=", default_hugepagesz_setup);
3395 static unsigned int allowed_mems_nr(struct hstate *h)
3398 unsigned int nr = 0;
3399 nodemask_t *mpol_allowed;
3400 unsigned int *array = h->free_huge_pages_node;
3401 gfp_t gfp_mask = htlb_alloc_mask(h);
3403 mpol_allowed = policy_nodemask_current(gfp_mask);
3405 for_each_node_mask(node, cpuset_current_mems_allowed) {
3406 if (!mpol_allowed || node_isset(node, *mpol_allowed))
3413 #ifdef CONFIG_SYSCTL
3414 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3415 void *buffer, size_t *length,
3416 loff_t *ppos, unsigned long *out)
3418 struct ctl_table dup_table;
3421 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3422 * can duplicate the @table and alter the duplicate of it.
3425 dup_table.data = out;
3427 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3430 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3431 struct ctl_table *table, int write,
3432 void *buffer, size_t *length, loff_t *ppos)
3434 struct hstate *h = &default_hstate;
3435 unsigned long tmp = h->max_huge_pages;
3438 if (!hugepages_supported())
3441 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3447 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3448 NUMA_NO_NODE, tmp, *length);
3453 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3454 void *buffer, size_t *length, loff_t *ppos)
3457 return hugetlb_sysctl_handler_common(false, table, write,
3458 buffer, length, ppos);
3462 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3463 void *buffer, size_t *length, loff_t *ppos)
3465 return hugetlb_sysctl_handler_common(true, table, write,
3466 buffer, length, ppos);
3468 #endif /* CONFIG_NUMA */
3470 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3471 void *buffer, size_t *length, loff_t *ppos)
3473 struct hstate *h = &default_hstate;
3477 if (!hugepages_supported())
3480 tmp = h->nr_overcommit_huge_pages;
3482 if (write && hstate_is_gigantic(h))
3485 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3491 spin_lock_irq(&hugetlb_lock);
3492 h->nr_overcommit_huge_pages = tmp;
3493 spin_unlock_irq(&hugetlb_lock);
3499 #endif /* CONFIG_SYSCTL */
3501 void hugetlb_report_meminfo(struct seq_file *m)
3504 unsigned long total = 0;
3506 if (!hugepages_supported())
3509 for_each_hstate(h) {
3510 unsigned long count = h->nr_huge_pages;
3512 total += huge_page_size(h) * count;
3514 if (h == &default_hstate)
3516 "HugePages_Total: %5lu\n"
3517 "HugePages_Free: %5lu\n"
3518 "HugePages_Rsvd: %5lu\n"
3519 "HugePages_Surp: %5lu\n"
3520 "Hugepagesize: %8lu kB\n",
3524 h->surplus_huge_pages,
3525 huge_page_size(h) / SZ_1K);
3528 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
3531 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
3533 struct hstate *h = &default_hstate;
3535 if (!hugepages_supported())
3538 return sysfs_emit_at(buf, len,
3539 "Node %d HugePages_Total: %5u\n"
3540 "Node %d HugePages_Free: %5u\n"
3541 "Node %d HugePages_Surp: %5u\n",
3542 nid, h->nr_huge_pages_node[nid],
3543 nid, h->free_huge_pages_node[nid],
3544 nid, h->surplus_huge_pages_node[nid]);
3547 void hugetlb_show_meminfo(void)
3552 if (!hugepages_supported())
3555 for_each_node_state(nid, N_MEMORY)
3557 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3559 h->nr_huge_pages_node[nid],
3560 h->free_huge_pages_node[nid],
3561 h->surplus_huge_pages_node[nid],
3562 huge_page_size(h) / SZ_1K);
3565 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3567 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3568 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3571 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3572 unsigned long hugetlb_total_pages(void)
3575 unsigned long nr_total_pages = 0;
3578 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3579 return nr_total_pages;
3582 static int hugetlb_acct_memory(struct hstate *h, long delta)
3589 spin_lock_irq(&hugetlb_lock);
3591 * When cpuset is configured, it breaks the strict hugetlb page
3592 * reservation as the accounting is done on a global variable. Such
3593 * reservation is completely rubbish in the presence of cpuset because
3594 * the reservation is not checked against page availability for the
3595 * current cpuset. Application can still potentially OOM'ed by kernel
3596 * with lack of free htlb page in cpuset that the task is in.
3597 * Attempt to enforce strict accounting with cpuset is almost
3598 * impossible (or too ugly) because cpuset is too fluid that
3599 * task or memory node can be dynamically moved between cpusets.
3601 * The change of semantics for shared hugetlb mapping with cpuset is
3602 * undesirable. However, in order to preserve some of the semantics,
3603 * we fall back to check against current free page availability as
3604 * a best attempt and hopefully to minimize the impact of changing
3605 * semantics that cpuset has.
3607 * Apart from cpuset, we also have memory policy mechanism that
3608 * also determines from which node the kernel will allocate memory
3609 * in a NUMA system. So similar to cpuset, we also should consider
3610 * the memory policy of the current task. Similar to the description
3614 if (gather_surplus_pages(h, delta) < 0)
3617 if (delta > allowed_mems_nr(h)) {
3618 return_unused_surplus_pages(h, delta);
3625 return_unused_surplus_pages(h, (unsigned long) -delta);
3628 spin_unlock_irq(&hugetlb_lock);
3632 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3634 struct resv_map *resv = vma_resv_map(vma);
3637 * This new VMA should share its siblings reservation map if present.
3638 * The VMA will only ever have a valid reservation map pointer where
3639 * it is being copied for another still existing VMA. As that VMA
3640 * has a reference to the reservation map it cannot disappear until
3641 * after this open call completes. It is therefore safe to take a
3642 * new reference here without additional locking.
3644 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3645 kref_get(&resv->refs);
3648 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3650 struct hstate *h = hstate_vma(vma);
3651 struct resv_map *resv = vma_resv_map(vma);
3652 struct hugepage_subpool *spool = subpool_vma(vma);
3653 unsigned long reserve, start, end;
3656 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3659 start = vma_hugecache_offset(h, vma, vma->vm_start);
3660 end = vma_hugecache_offset(h, vma, vma->vm_end);
3662 reserve = (end - start) - region_count(resv, start, end);
3663 hugetlb_cgroup_uncharge_counter(resv, start, end);
3666 * Decrement reserve counts. The global reserve count may be
3667 * adjusted if the subpool has a minimum size.
3669 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3670 hugetlb_acct_memory(h, -gbl_reserve);
3673 kref_put(&resv->refs, resv_map_release);
3676 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3678 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3683 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3685 return huge_page_size(hstate_vma(vma));
3689 * We cannot handle pagefaults against hugetlb pages at all. They cause
3690 * handle_mm_fault() to try to instantiate regular-sized pages in the
3691 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3694 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3701 * When a new function is introduced to vm_operations_struct and added
3702 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3703 * This is because under System V memory model, mappings created via
3704 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3705 * their original vm_ops are overwritten with shm_vm_ops.
3707 const struct vm_operations_struct hugetlb_vm_ops = {
3708 .fault = hugetlb_vm_op_fault,
3709 .open = hugetlb_vm_op_open,
3710 .close = hugetlb_vm_op_close,
3711 .may_split = hugetlb_vm_op_split,
3712 .pagesize = hugetlb_vm_op_pagesize,
3715 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3721 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3722 vma->vm_page_prot)));
3724 entry = huge_pte_wrprotect(mk_huge_pte(page,
3725 vma->vm_page_prot));
3727 entry = pte_mkyoung(entry);
3728 entry = pte_mkhuge(entry);
3729 entry = arch_make_huge_pte(entry, vma, page, writable);
3734 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3735 unsigned long address, pte_t *ptep)
3739 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3740 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3741 update_mmu_cache(vma, address, ptep);
3744 bool is_hugetlb_entry_migration(pte_t pte)
3748 if (huge_pte_none(pte) || pte_present(pte))
3750 swp = pte_to_swp_entry(pte);
3751 if (is_migration_entry(swp))
3757 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
3761 if (huge_pte_none(pte) || pte_present(pte))
3763 swp = pte_to_swp_entry(pte);
3764 if (is_hwpoison_entry(swp))
3771 hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
3772 struct page *new_page)
3774 __SetPageUptodate(new_page);
3775 set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
3776 hugepage_add_new_anon_rmap(new_page, vma, addr);
3777 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
3778 ClearHPageRestoreReserve(new_page);
3779 SetHPageMigratable(new_page);
3782 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3783 struct vm_area_struct *vma)
3785 pte_t *src_pte, *dst_pte, entry, dst_entry;
3786 struct page *ptepage;
3788 bool cow = is_cow_mapping(vma->vm_flags);
3789 struct hstate *h = hstate_vma(vma);
3790 unsigned long sz = huge_page_size(h);
3791 unsigned long npages = pages_per_huge_page(h);
3792 struct address_space *mapping = vma->vm_file->f_mapping;
3793 struct mmu_notifier_range range;
3797 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3800 mmu_notifier_invalidate_range_start(&range);
3803 * For shared mappings i_mmap_rwsem must be held to call
3804 * huge_pte_alloc, otherwise the returned ptep could go
3805 * away if part of a shared pmd and another thread calls
3808 i_mmap_lock_read(mapping);
3811 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3812 spinlock_t *src_ptl, *dst_ptl;
3813 src_pte = huge_pte_offset(src, addr, sz);
3816 dst_pte = huge_pte_alloc(dst, vma, addr, sz);
3823 * If the pagetables are shared don't copy or take references.
3824 * dst_pte == src_pte is the common case of src/dest sharing.
3826 * However, src could have 'unshared' and dst shares with
3827 * another vma. If dst_pte !none, this implies sharing.
3828 * Check here before taking page table lock, and once again
3829 * after taking the lock below.
3831 dst_entry = huge_ptep_get(dst_pte);
3832 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3835 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3836 src_ptl = huge_pte_lockptr(h, src, src_pte);
3837 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3838 entry = huge_ptep_get(src_pte);
3839 dst_entry = huge_ptep_get(dst_pte);
3841 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3843 * Skip if src entry none. Also, skip in the
3844 * unlikely case dst entry !none as this implies
3845 * sharing with another vma.
3848 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3849 is_hugetlb_entry_hwpoisoned(entry))) {
3850 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3852 if (is_write_migration_entry(swp_entry) && cow) {
3854 * COW mappings require pages in both
3855 * parent and child to be set to read.
3857 make_migration_entry_read(&swp_entry);
3858 entry = swp_entry_to_pte(swp_entry);
3859 set_huge_swap_pte_at(src, addr, src_pte,
3862 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3864 entry = huge_ptep_get(src_pte);
3865 ptepage = pte_page(entry);
3869 * This is a rare case where we see pinned hugetlb
3870 * pages while they're prone to COW. We need to do the
3871 * COW earlier during fork.
3873 * When pre-allocating the page or copying data, we
3874 * need to be without the pgtable locks since we could
3875 * sleep during the process.
3877 if (unlikely(page_needs_cow_for_dma(vma, ptepage))) {
3878 pte_t src_pte_old = entry;
3881 spin_unlock(src_ptl);
3882 spin_unlock(dst_ptl);
3883 /* Do not use reserve as it's private owned */
3884 new = alloc_huge_page(vma, addr, 1);
3890 copy_user_huge_page(new, ptepage, addr, vma,
3894 /* Install the new huge page if src pte stable */
3895 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3896 src_ptl = huge_pte_lockptr(h, src, src_pte);
3897 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3898 entry = huge_ptep_get(src_pte);
3899 if (!pte_same(src_pte_old, entry)) {
3901 /* dst_entry won't change as in child */
3904 hugetlb_install_page(vma, dst_pte, addr, new);
3905 spin_unlock(src_ptl);
3906 spin_unlock(dst_ptl);
3912 * No need to notify as we are downgrading page
3913 * table protection not changing it to point
3916 * See Documentation/vm/mmu_notifier.rst
3918 huge_ptep_set_wrprotect(src, addr, src_pte);
3921 page_dup_rmap(ptepage, true);
3922 set_huge_pte_at(dst, addr, dst_pte, entry);
3923 hugetlb_count_add(npages, dst);
3925 spin_unlock(src_ptl);
3926 spin_unlock(dst_ptl);
3930 mmu_notifier_invalidate_range_end(&range);
3932 i_mmap_unlock_read(mapping);
3937 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3938 unsigned long start, unsigned long end,
3939 struct page *ref_page)
3941 struct mm_struct *mm = vma->vm_mm;
3942 unsigned long address;
3947 struct hstate *h = hstate_vma(vma);
3948 unsigned long sz = huge_page_size(h);
3949 struct mmu_notifier_range range;
3951 WARN_ON(!is_vm_hugetlb_page(vma));
3952 BUG_ON(start & ~huge_page_mask(h));
3953 BUG_ON(end & ~huge_page_mask(h));
3956 * This is a hugetlb vma, all the pte entries should point
3959 tlb_change_page_size(tlb, sz);
3960 tlb_start_vma(tlb, vma);
3963 * If sharing possible, alert mmu notifiers of worst case.
3965 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3967 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3968 mmu_notifier_invalidate_range_start(&range);
3970 for (; address < end; address += sz) {
3971 ptep = huge_pte_offset(mm, address, sz);
3975 ptl = huge_pte_lock(h, mm, ptep);
3976 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
3979 * We just unmapped a page of PMDs by clearing a PUD.
3980 * The caller's TLB flush range should cover this area.
3985 pte = huge_ptep_get(ptep);
3986 if (huge_pte_none(pte)) {
3992 * Migrating hugepage or HWPoisoned hugepage is already
3993 * unmapped and its refcount is dropped, so just clear pte here.
3995 if (unlikely(!pte_present(pte))) {
3996 huge_pte_clear(mm, address, ptep, sz);
4001 page = pte_page(pte);
4003 * If a reference page is supplied, it is because a specific
4004 * page is being unmapped, not a range. Ensure the page we
4005 * are about to unmap is the actual page of interest.
4008 if (page != ref_page) {
4013 * Mark the VMA as having unmapped its page so that
4014 * future faults in this VMA will fail rather than
4015 * looking like data was lost
4017 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
4020 pte = huge_ptep_get_and_clear(mm, address, ptep);
4021 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
4022 if (huge_pte_dirty(pte))
4023 set_page_dirty(page);
4025 hugetlb_count_sub(pages_per_huge_page(h), mm);
4026 page_remove_rmap(page, true);
4029 tlb_remove_page_size(tlb, page, huge_page_size(h));
4031 * Bail out after unmapping reference page if supplied
4036 mmu_notifier_invalidate_range_end(&range);
4037 tlb_end_vma(tlb, vma);
4040 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4041 struct vm_area_struct *vma, unsigned long start,
4042 unsigned long end, struct page *ref_page)
4044 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4047 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4048 * test will fail on a vma being torn down, and not grab a page table
4049 * on its way out. We're lucky that the flag has such an appropriate
4050 * name, and can in fact be safely cleared here. We could clear it
4051 * before the __unmap_hugepage_range above, but all that's necessary
4052 * is to clear it before releasing the i_mmap_rwsem. This works
4053 * because in the context this is called, the VMA is about to be
4054 * destroyed and the i_mmap_rwsem is held.
4056 vma->vm_flags &= ~VM_MAYSHARE;
4059 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
4060 unsigned long end, struct page *ref_page)
4062 struct mmu_gather tlb;
4064 tlb_gather_mmu(&tlb, vma->vm_mm);
4065 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
4066 tlb_finish_mmu(&tlb);
4070 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4071 * mapping it owns the reserve page for. The intention is to unmap the page
4072 * from other VMAs and let the children be SIGKILLed if they are faulting the
4075 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4076 struct page *page, unsigned long address)
4078 struct hstate *h = hstate_vma(vma);
4079 struct vm_area_struct *iter_vma;
4080 struct address_space *mapping;
4084 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4085 * from page cache lookup which is in HPAGE_SIZE units.
4087 address = address & huge_page_mask(h);
4088 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4090 mapping = vma->vm_file->f_mapping;
4093 * Take the mapping lock for the duration of the table walk. As
4094 * this mapping should be shared between all the VMAs,
4095 * __unmap_hugepage_range() is called as the lock is already held
4097 i_mmap_lock_write(mapping);
4098 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
4099 /* Do not unmap the current VMA */
4100 if (iter_vma == vma)
4104 * Shared VMAs have their own reserves and do not affect
4105 * MAP_PRIVATE accounting but it is possible that a shared
4106 * VMA is using the same page so check and skip such VMAs.
4108 if (iter_vma->vm_flags & VM_MAYSHARE)
4112 * Unmap the page from other VMAs without their own reserves.
4113 * They get marked to be SIGKILLed if they fault in these
4114 * areas. This is because a future no-page fault on this VMA
4115 * could insert a zeroed page instead of the data existing
4116 * from the time of fork. This would look like data corruption
4118 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
4119 unmap_hugepage_range(iter_vma, address,
4120 address + huge_page_size(h), page);
4122 i_mmap_unlock_write(mapping);
4126 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4127 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4128 * cannot race with other handlers or page migration.
4129 * Keep the pte_same checks anyway to make transition from the mutex easier.
4131 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
4132 unsigned long address, pte_t *ptep,
4133 struct page *pagecache_page, spinlock_t *ptl)
4136 struct hstate *h = hstate_vma(vma);
4137 struct page *old_page, *new_page;
4138 int outside_reserve = 0;
4140 unsigned long haddr = address & huge_page_mask(h);
4141 struct mmu_notifier_range range;
4143 pte = huge_ptep_get(ptep);
4144 old_page = pte_page(pte);
4147 /* If no-one else is actually using this page, avoid the copy
4148 * and just make the page writable */
4149 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
4150 page_move_anon_rmap(old_page, vma);
4151 set_huge_ptep_writable(vma, haddr, ptep);
4156 * If the process that created a MAP_PRIVATE mapping is about to
4157 * perform a COW due to a shared page count, attempt to satisfy
4158 * the allocation without using the existing reserves. The pagecache
4159 * page is used to determine if the reserve at this address was
4160 * consumed or not. If reserves were used, a partial faulted mapping
4161 * at the time of fork() could consume its reserves on COW instead
4162 * of the full address range.
4164 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4165 old_page != pagecache_page)
4166 outside_reserve = 1;
4171 * Drop page table lock as buddy allocator may be called. It will
4172 * be acquired again before returning to the caller, as expected.
4175 new_page = alloc_huge_page(vma, haddr, outside_reserve);
4177 if (IS_ERR(new_page)) {
4179 * If a process owning a MAP_PRIVATE mapping fails to COW,
4180 * it is due to references held by a child and an insufficient
4181 * huge page pool. To guarantee the original mappers
4182 * reliability, unmap the page from child processes. The child
4183 * may get SIGKILLed if it later faults.
4185 if (outside_reserve) {
4186 struct address_space *mapping = vma->vm_file->f_mapping;
4191 BUG_ON(huge_pte_none(pte));
4193 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4194 * unmapping. unmapping needs to hold i_mmap_rwsem
4195 * in write mode. Dropping i_mmap_rwsem in read mode
4196 * here is OK as COW mappings do not interact with
4199 * Reacquire both after unmap operation.
4201 idx = vma_hugecache_offset(h, vma, haddr);
4202 hash = hugetlb_fault_mutex_hash(mapping, idx);
4203 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4204 i_mmap_unlock_read(mapping);
4206 unmap_ref_private(mm, vma, old_page, haddr);
4208 i_mmap_lock_read(mapping);
4209 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4211 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4213 pte_same(huge_ptep_get(ptep), pte)))
4214 goto retry_avoidcopy;
4216 * race occurs while re-acquiring page table
4217 * lock, and our job is done.
4222 ret = vmf_error(PTR_ERR(new_page));
4223 goto out_release_old;
4227 * When the original hugepage is shared one, it does not have
4228 * anon_vma prepared.
4230 if (unlikely(anon_vma_prepare(vma))) {
4232 goto out_release_all;
4235 copy_user_huge_page(new_page, old_page, address, vma,
4236 pages_per_huge_page(h));
4237 __SetPageUptodate(new_page);
4239 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4240 haddr + huge_page_size(h));
4241 mmu_notifier_invalidate_range_start(&range);
4244 * Retake the page table lock to check for racing updates
4245 * before the page tables are altered
4248 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4249 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4250 ClearHPageRestoreReserve(new_page);
4253 huge_ptep_clear_flush(vma, haddr, ptep);
4254 mmu_notifier_invalidate_range(mm, range.start, range.end);
4255 set_huge_pte_at(mm, haddr, ptep,
4256 make_huge_pte(vma, new_page, 1));
4257 page_remove_rmap(old_page, true);
4258 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4259 SetHPageMigratable(new_page);
4260 /* Make the old page be freed below */
4261 new_page = old_page;
4264 mmu_notifier_invalidate_range_end(&range);
4266 restore_reserve_on_error(h, vma, haddr, new_page);
4271 spin_lock(ptl); /* Caller expects lock to be held */
4275 /* Return the pagecache page at a given address within a VMA */
4276 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4277 struct vm_area_struct *vma, unsigned long address)
4279 struct address_space *mapping;
4282 mapping = vma->vm_file->f_mapping;
4283 idx = vma_hugecache_offset(h, vma, address);
4285 return find_lock_page(mapping, idx);
4289 * Return whether there is a pagecache page to back given address within VMA.
4290 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4292 static bool hugetlbfs_pagecache_present(struct hstate *h,
4293 struct vm_area_struct *vma, unsigned long address)
4295 struct address_space *mapping;
4299 mapping = vma->vm_file->f_mapping;
4300 idx = vma_hugecache_offset(h, vma, address);
4302 page = find_get_page(mapping, idx);
4305 return page != NULL;
4308 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4311 struct inode *inode = mapping->host;
4312 struct hstate *h = hstate_inode(inode);
4313 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4317 ClearHPageRestoreReserve(page);
4320 * set page dirty so that it will not be removed from cache/file
4321 * by non-hugetlbfs specific code paths.
4323 set_page_dirty(page);
4325 spin_lock(&inode->i_lock);
4326 inode->i_blocks += blocks_per_huge_page(h);
4327 spin_unlock(&inode->i_lock);
4331 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4332 struct vm_area_struct *vma,
4333 struct address_space *mapping, pgoff_t idx,
4334 unsigned long address, pte_t *ptep, unsigned int flags)
4336 struct hstate *h = hstate_vma(vma);
4337 vm_fault_t ret = VM_FAULT_SIGBUS;
4343 unsigned long haddr = address & huge_page_mask(h);
4344 bool new_page = false;
4347 * Currently, we are forced to kill the process in the event the
4348 * original mapper has unmapped pages from the child due to a failed
4349 * COW. Warn that such a situation has occurred as it may not be obvious
4351 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4352 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4358 * We can not race with truncation due to holding i_mmap_rwsem.
4359 * i_size is modified when holding i_mmap_rwsem, so check here
4360 * once for faults beyond end of file.
4362 size = i_size_read(mapping->host) >> huge_page_shift(h);
4367 page = find_lock_page(mapping, idx);
4370 * Check for page in userfault range
4372 if (userfaultfd_missing(vma)) {
4374 struct vm_fault vmf = {
4379 * Hard to debug if it ends up being
4380 * used by a callee that assumes
4381 * something about the other
4382 * uninitialized fields... same as in
4388 * hugetlb_fault_mutex and i_mmap_rwsem must be
4389 * dropped before handling userfault. Reacquire
4390 * after handling fault to make calling code simpler.
4392 hash = hugetlb_fault_mutex_hash(mapping, idx);
4393 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4394 i_mmap_unlock_read(mapping);
4395 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
4396 i_mmap_lock_read(mapping);
4397 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4401 page = alloc_huge_page(vma, haddr, 0);
4404 * Returning error will result in faulting task being
4405 * sent SIGBUS. The hugetlb fault mutex prevents two
4406 * tasks from racing to fault in the same page which
4407 * could result in false unable to allocate errors.
4408 * Page migration does not take the fault mutex, but
4409 * does a clear then write of pte's under page table
4410 * lock. Page fault code could race with migration,
4411 * notice the clear pte and try to allocate a page
4412 * here. Before returning error, get ptl and make
4413 * sure there really is no pte entry.
4415 ptl = huge_pte_lock(h, mm, ptep);
4417 if (huge_pte_none(huge_ptep_get(ptep)))
4418 ret = vmf_error(PTR_ERR(page));
4422 clear_huge_page(page, address, pages_per_huge_page(h));
4423 __SetPageUptodate(page);
4426 if (vma->vm_flags & VM_MAYSHARE) {
4427 int err = huge_add_to_page_cache(page, mapping, idx);
4436 if (unlikely(anon_vma_prepare(vma))) {
4438 goto backout_unlocked;
4444 * If memory error occurs between mmap() and fault, some process
4445 * don't have hwpoisoned swap entry for errored virtual address.
4446 * So we need to block hugepage fault by PG_hwpoison bit check.
4448 if (unlikely(PageHWPoison(page))) {
4449 ret = VM_FAULT_HWPOISON_LARGE |
4450 VM_FAULT_SET_HINDEX(hstate_index(h));
4451 goto backout_unlocked;
4456 * If we are going to COW a private mapping later, we examine the
4457 * pending reservations for this page now. This will ensure that
4458 * any allocations necessary to record that reservation occur outside
4461 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4462 if (vma_needs_reservation(h, vma, haddr) < 0) {
4464 goto backout_unlocked;
4466 /* Just decrements count, does not deallocate */
4467 vma_end_reservation(h, vma, haddr);
4470 ptl = huge_pte_lock(h, mm, ptep);
4472 if (!huge_pte_none(huge_ptep_get(ptep)))
4476 ClearHPageRestoreReserve(page);
4477 hugepage_add_new_anon_rmap(page, vma, haddr);
4479 page_dup_rmap(page, true);
4480 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4481 && (vma->vm_flags & VM_SHARED)));
4482 set_huge_pte_at(mm, haddr, ptep, new_pte);
4484 hugetlb_count_add(pages_per_huge_page(h), mm);
4485 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4486 /* Optimization, do the COW without a second fault */
4487 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4493 * Only set HPageMigratable in newly allocated pages. Existing pages
4494 * found in the pagecache may not have HPageMigratableset if they have
4495 * been isolated for migration.
4498 SetHPageMigratable(page);
4508 restore_reserve_on_error(h, vma, haddr, page);
4514 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4516 unsigned long key[2];
4519 key[0] = (unsigned long) mapping;
4522 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4524 return hash & (num_fault_mutexes - 1);
4528 * For uniprocessor systems we always use a single mutex, so just
4529 * return 0 and avoid the hashing overhead.
4531 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4537 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4538 unsigned long address, unsigned int flags)
4545 struct page *page = NULL;
4546 struct page *pagecache_page = NULL;
4547 struct hstate *h = hstate_vma(vma);
4548 struct address_space *mapping;
4549 int need_wait_lock = 0;
4550 unsigned long haddr = address & huge_page_mask(h);
4552 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4555 * Since we hold no locks, ptep could be stale. That is
4556 * OK as we are only making decisions based on content and
4557 * not actually modifying content here.
4559 entry = huge_ptep_get(ptep);
4560 if (unlikely(is_hugetlb_entry_migration(entry))) {
4561 migration_entry_wait_huge(vma, mm, ptep);
4563 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4564 return VM_FAULT_HWPOISON_LARGE |
4565 VM_FAULT_SET_HINDEX(hstate_index(h));
4569 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4570 * until finished with ptep. This serves two purposes:
4571 * 1) It prevents huge_pmd_unshare from being called elsewhere
4572 * and making the ptep no longer valid.
4573 * 2) It synchronizes us with i_size modifications during truncation.
4575 * ptep could have already be assigned via huge_pte_offset. That
4576 * is OK, as huge_pte_alloc will return the same value unless
4577 * something has changed.
4579 mapping = vma->vm_file->f_mapping;
4580 i_mmap_lock_read(mapping);
4581 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
4583 i_mmap_unlock_read(mapping);
4584 return VM_FAULT_OOM;
4588 * Serialize hugepage allocation and instantiation, so that we don't
4589 * get spurious allocation failures if two CPUs race to instantiate
4590 * the same page in the page cache.
4592 idx = vma_hugecache_offset(h, vma, haddr);
4593 hash = hugetlb_fault_mutex_hash(mapping, idx);
4594 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4596 entry = huge_ptep_get(ptep);
4597 if (huge_pte_none(entry)) {
4598 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4605 * entry could be a migration/hwpoison entry at this point, so this
4606 * check prevents the kernel from going below assuming that we have
4607 * an active hugepage in pagecache. This goto expects the 2nd page
4608 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4609 * properly handle it.
4611 if (!pte_present(entry))
4615 * If we are going to COW the mapping later, we examine the pending
4616 * reservations for this page now. This will ensure that any
4617 * allocations necessary to record that reservation occur outside the
4618 * spinlock. For private mappings, we also lookup the pagecache
4619 * page now as it is used to determine if a reservation has been
4622 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4623 if (vma_needs_reservation(h, vma, haddr) < 0) {
4627 /* Just decrements count, does not deallocate */
4628 vma_end_reservation(h, vma, haddr);
4630 if (!(vma->vm_flags & VM_MAYSHARE))
4631 pagecache_page = hugetlbfs_pagecache_page(h,
4635 ptl = huge_pte_lock(h, mm, ptep);
4637 /* Check for a racing update before calling hugetlb_cow */
4638 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4642 * hugetlb_cow() requires page locks of pte_page(entry) and
4643 * pagecache_page, so here we need take the former one
4644 * when page != pagecache_page or !pagecache_page.
4646 page = pte_page(entry);
4647 if (page != pagecache_page)
4648 if (!trylock_page(page)) {
4655 if (flags & FAULT_FLAG_WRITE) {
4656 if (!huge_pte_write(entry)) {
4657 ret = hugetlb_cow(mm, vma, address, ptep,
4658 pagecache_page, ptl);
4661 entry = huge_pte_mkdirty(entry);
4663 entry = pte_mkyoung(entry);
4664 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4665 flags & FAULT_FLAG_WRITE))
4666 update_mmu_cache(vma, haddr, ptep);
4668 if (page != pagecache_page)
4674 if (pagecache_page) {
4675 unlock_page(pagecache_page);
4676 put_page(pagecache_page);
4679 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4680 i_mmap_unlock_read(mapping);
4682 * Generally it's safe to hold refcount during waiting page lock. But
4683 * here we just wait to defer the next page fault to avoid busy loop and
4684 * the page is not used after unlocked before returning from the current
4685 * page fault. So we are safe from accessing freed page, even if we wait
4686 * here without taking refcount.
4689 wait_on_page_locked(page);
4694 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4695 * modifications for huge pages.
4697 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4699 struct vm_area_struct *dst_vma,
4700 unsigned long dst_addr,
4701 unsigned long src_addr,
4702 struct page **pagep)
4704 struct address_space *mapping;
4707 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4708 struct hstate *h = hstate_vma(dst_vma);
4716 page = alloc_huge_page(dst_vma, dst_addr, 0);
4720 ret = copy_huge_page_from_user(page,
4721 (const void __user *) src_addr,
4722 pages_per_huge_page(h), false);
4724 /* fallback to copy_from_user outside mmap_lock */
4725 if (unlikely(ret)) {
4728 /* don't free the page */
4737 * The memory barrier inside __SetPageUptodate makes sure that
4738 * preceding stores to the page contents become visible before
4739 * the set_pte_at() write.
4741 __SetPageUptodate(page);
4743 mapping = dst_vma->vm_file->f_mapping;
4744 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4747 * If shared, add to page cache
4750 size = i_size_read(mapping->host) >> huge_page_shift(h);
4753 goto out_release_nounlock;
4756 * Serialization between remove_inode_hugepages() and
4757 * huge_add_to_page_cache() below happens through the
4758 * hugetlb_fault_mutex_table that here must be hold by
4761 ret = huge_add_to_page_cache(page, mapping, idx);
4763 goto out_release_nounlock;
4766 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4770 * Recheck the i_size after holding PT lock to make sure not
4771 * to leave any page mapped (as page_mapped()) beyond the end
4772 * of the i_size (remove_inode_hugepages() is strict about
4773 * enforcing that). If we bail out here, we'll also leave a
4774 * page in the radix tree in the vm_shared case beyond the end
4775 * of the i_size, but remove_inode_hugepages() will take care
4776 * of it as soon as we drop the hugetlb_fault_mutex_table.
4778 size = i_size_read(mapping->host) >> huge_page_shift(h);
4781 goto out_release_unlock;
4784 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4785 goto out_release_unlock;
4788 page_dup_rmap(page, true);
4790 ClearHPageRestoreReserve(page);
4791 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4794 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4795 if (dst_vma->vm_flags & VM_WRITE)
4796 _dst_pte = huge_pte_mkdirty(_dst_pte);
4797 _dst_pte = pte_mkyoung(_dst_pte);
4799 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4801 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4802 dst_vma->vm_flags & VM_WRITE);
4803 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4805 /* No need to invalidate - it was non-present before */
4806 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4809 SetHPageMigratable(page);
4819 out_release_nounlock:
4824 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
4825 int refs, struct page **pages,
4826 struct vm_area_struct **vmas)
4830 for (nr = 0; nr < refs; nr++) {
4832 pages[nr] = mem_map_offset(page, nr);
4838 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4839 struct page **pages, struct vm_area_struct **vmas,
4840 unsigned long *position, unsigned long *nr_pages,
4841 long i, unsigned int flags, int *locked)
4843 unsigned long pfn_offset;
4844 unsigned long vaddr = *position;
4845 unsigned long remainder = *nr_pages;
4846 struct hstate *h = hstate_vma(vma);
4847 int err = -EFAULT, refs;
4849 while (vaddr < vma->vm_end && remainder) {
4851 spinlock_t *ptl = NULL;
4856 * If we have a pending SIGKILL, don't keep faulting pages and
4857 * potentially allocating memory.
4859 if (fatal_signal_pending(current)) {
4865 * Some archs (sparc64, sh*) have multiple pte_ts to
4866 * each hugepage. We have to make sure we get the
4867 * first, for the page indexing below to work.
4869 * Note that page table lock is not held when pte is null.
4871 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4874 ptl = huge_pte_lock(h, mm, pte);
4875 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4878 * When coredumping, it suits get_dump_page if we just return
4879 * an error where there's an empty slot with no huge pagecache
4880 * to back it. This way, we avoid allocating a hugepage, and
4881 * the sparse dumpfile avoids allocating disk blocks, but its
4882 * huge holes still show up with zeroes where they need to be.
4884 if (absent && (flags & FOLL_DUMP) &&
4885 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4893 * We need call hugetlb_fault for both hugepages under migration
4894 * (in which case hugetlb_fault waits for the migration,) and
4895 * hwpoisoned hugepages (in which case we need to prevent the
4896 * caller from accessing to them.) In order to do this, we use
4897 * here is_swap_pte instead of is_hugetlb_entry_migration and
4898 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4899 * both cases, and because we can't follow correct pages
4900 * directly from any kind of swap entries.
4902 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4903 ((flags & FOLL_WRITE) &&
4904 !huge_pte_write(huge_ptep_get(pte)))) {
4906 unsigned int fault_flags = 0;
4910 if (flags & FOLL_WRITE)
4911 fault_flags |= FAULT_FLAG_WRITE;
4913 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4914 FAULT_FLAG_KILLABLE;
4915 if (flags & FOLL_NOWAIT)
4916 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4917 FAULT_FLAG_RETRY_NOWAIT;
4918 if (flags & FOLL_TRIED) {
4920 * Note: FAULT_FLAG_ALLOW_RETRY and
4921 * FAULT_FLAG_TRIED can co-exist
4923 fault_flags |= FAULT_FLAG_TRIED;
4925 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4926 if (ret & VM_FAULT_ERROR) {
4927 err = vm_fault_to_errno(ret, flags);
4931 if (ret & VM_FAULT_RETRY) {
4933 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4937 * VM_FAULT_RETRY must not return an
4938 * error, it will return zero
4941 * No need to update "position" as the
4942 * caller will not check it after
4943 * *nr_pages is set to 0.
4950 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4951 page = pte_page(huge_ptep_get(pte));
4954 * If subpage information not requested, update counters
4955 * and skip the same_page loop below.
4957 if (!pages && !vmas && !pfn_offset &&
4958 (vaddr + huge_page_size(h) < vma->vm_end) &&
4959 (remainder >= pages_per_huge_page(h))) {
4960 vaddr += huge_page_size(h);
4961 remainder -= pages_per_huge_page(h);
4962 i += pages_per_huge_page(h);
4967 refs = min3(pages_per_huge_page(h) - pfn_offset,
4968 (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder);
4971 record_subpages_vmas(mem_map_offset(page, pfn_offset),
4973 likely(pages) ? pages + i : NULL,
4974 vmas ? vmas + i : NULL);
4978 * try_grab_compound_head() should always succeed here,
4979 * because: a) we hold the ptl lock, and b) we've just
4980 * checked that the huge page is present in the page
4981 * tables. If the huge page is present, then the tail
4982 * pages must also be present. The ptl prevents the
4983 * head page and tail pages from being rearranged in
4984 * any way. So this page must be available at this
4985 * point, unless the page refcount overflowed:
4987 if (WARN_ON_ONCE(!try_grab_compound_head(pages[i],
4997 vaddr += (refs << PAGE_SHIFT);
5003 *nr_pages = remainder;
5005 * setting position is actually required only if remainder is
5006 * not zero but it's faster not to add a "if (remainder)"
5014 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
5015 unsigned long address, unsigned long end, pgprot_t newprot)
5017 struct mm_struct *mm = vma->vm_mm;
5018 unsigned long start = address;
5021 struct hstate *h = hstate_vma(vma);
5022 unsigned long pages = 0;
5023 bool shared_pmd = false;
5024 struct mmu_notifier_range range;
5027 * In the case of shared PMDs, the area to flush could be beyond
5028 * start/end. Set range.start/range.end to cover the maximum possible
5029 * range if PMD sharing is possible.
5031 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5032 0, vma, mm, start, end);
5033 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5035 BUG_ON(address >= end);
5036 flush_cache_range(vma, range.start, range.end);
5038 mmu_notifier_invalidate_range_start(&range);
5039 i_mmap_lock_write(vma->vm_file->f_mapping);
5040 for (; address < end; address += huge_page_size(h)) {
5042 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5045 ptl = huge_pte_lock(h, mm, ptep);
5046 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
5052 pte = huge_ptep_get(ptep);
5053 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5057 if (unlikely(is_hugetlb_entry_migration(pte))) {
5058 swp_entry_t entry = pte_to_swp_entry(pte);
5060 if (is_write_migration_entry(entry)) {
5063 make_migration_entry_read(&entry);
5064 newpte = swp_entry_to_pte(entry);
5065 set_huge_swap_pte_at(mm, address, ptep,
5066 newpte, huge_page_size(h));
5072 if (!huge_pte_none(pte)) {
5075 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5076 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
5077 pte = arch_make_huge_pte(pte, vma, NULL, 0);
5078 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
5084 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5085 * may have cleared our pud entry and done put_page on the page table:
5086 * once we release i_mmap_rwsem, another task can do the final put_page
5087 * and that page table be reused and filled with junk. If we actually
5088 * did unshare a page of pmds, flush the range corresponding to the pud.
5091 flush_hugetlb_tlb_range(vma, range.start, range.end);
5093 flush_hugetlb_tlb_range(vma, start, end);
5095 * No need to call mmu_notifier_invalidate_range() we are downgrading
5096 * page table protection not changing it to point to a new page.
5098 * See Documentation/vm/mmu_notifier.rst
5100 i_mmap_unlock_write(vma->vm_file->f_mapping);
5101 mmu_notifier_invalidate_range_end(&range);
5103 return pages << h->order;
5106 /* Return true if reservation was successful, false otherwise. */
5107 bool hugetlb_reserve_pages(struct inode *inode,
5109 struct vm_area_struct *vma,
5110 vm_flags_t vm_flags)
5113 struct hstate *h = hstate_inode(inode);
5114 struct hugepage_subpool *spool = subpool_inode(inode);
5115 struct resv_map *resv_map;
5116 struct hugetlb_cgroup *h_cg = NULL;
5117 long gbl_reserve, regions_needed = 0;
5119 /* This should never happen */
5121 VM_WARN(1, "%s called with a negative range\n", __func__);
5126 * Only apply hugepage reservation if asked. At fault time, an
5127 * attempt will be made for VM_NORESERVE to allocate a page
5128 * without using reserves
5130 if (vm_flags & VM_NORESERVE)
5134 * Shared mappings base their reservation on the number of pages that
5135 * are already allocated on behalf of the file. Private mappings need
5136 * to reserve the full area even if read-only as mprotect() may be
5137 * called to make the mapping read-write. Assume !vma is a shm mapping
5139 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5141 * resv_map can not be NULL as hugetlb_reserve_pages is only
5142 * called for inodes for which resv_maps were created (see
5143 * hugetlbfs_get_inode).
5145 resv_map = inode_resv_map(inode);
5147 chg = region_chg(resv_map, from, to, ®ions_needed);
5150 /* Private mapping. */
5151 resv_map = resv_map_alloc();
5157 set_vma_resv_map(vma, resv_map);
5158 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5164 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
5165 chg * pages_per_huge_page(h), &h_cg) < 0)
5168 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5169 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5172 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5176 * There must be enough pages in the subpool for the mapping. If
5177 * the subpool has a minimum size, there may be some global
5178 * reservations already in place (gbl_reserve).
5180 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5181 if (gbl_reserve < 0)
5182 goto out_uncharge_cgroup;
5185 * Check enough hugepages are available for the reservation.
5186 * Hand the pages back to the subpool if there are not
5188 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
5192 * Account for the reservations made. Shared mappings record regions
5193 * that have reservations as they are shared by multiple VMAs.
5194 * When the last VMA disappears, the region map says how much
5195 * the reservation was and the page cache tells how much of
5196 * the reservation was consumed. Private mappings are per-VMA and
5197 * only the consumed reservations are tracked. When the VMA
5198 * disappears, the original reservation is the VMA size and the
5199 * consumed reservations are stored in the map. Hence, nothing
5200 * else has to be done for private mappings here
5202 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5203 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5205 if (unlikely(add < 0)) {
5206 hugetlb_acct_memory(h, -gbl_reserve);
5208 } else if (unlikely(chg > add)) {
5210 * pages in this range were added to the reserve
5211 * map between region_chg and region_add. This
5212 * indicates a race with alloc_huge_page. Adjust
5213 * the subpool and reserve counts modified above
5214 * based on the difference.
5219 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
5220 * reference to h_cg->css. See comment below for detail.
5222 hugetlb_cgroup_uncharge_cgroup_rsvd(
5224 (chg - add) * pages_per_huge_page(h), h_cg);
5226 rsv_adjust = hugepage_subpool_put_pages(spool,
5228 hugetlb_acct_memory(h, -rsv_adjust);
5231 * The file_regions will hold their own reference to
5232 * h_cg->css. So we should release the reference held
5233 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
5236 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
5242 /* put back original number of pages, chg */
5243 (void)hugepage_subpool_put_pages(spool, chg);
5244 out_uncharge_cgroup:
5245 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5246 chg * pages_per_huge_page(h), h_cg);
5248 if (!vma || vma->vm_flags & VM_MAYSHARE)
5249 /* Only call region_abort if the region_chg succeeded but the
5250 * region_add failed or didn't run.
5252 if (chg >= 0 && add < 0)
5253 region_abort(resv_map, from, to, regions_needed);
5254 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5255 kref_put(&resv_map->refs, resv_map_release);
5259 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5262 struct hstate *h = hstate_inode(inode);
5263 struct resv_map *resv_map = inode_resv_map(inode);
5265 struct hugepage_subpool *spool = subpool_inode(inode);
5269 * Since this routine can be called in the evict inode path for all
5270 * hugetlbfs inodes, resv_map could be NULL.
5273 chg = region_del(resv_map, start, end);
5275 * region_del() can fail in the rare case where a region
5276 * must be split and another region descriptor can not be
5277 * allocated. If end == LONG_MAX, it will not fail.
5283 spin_lock(&inode->i_lock);
5284 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5285 spin_unlock(&inode->i_lock);
5288 * If the subpool has a minimum size, the number of global
5289 * reservations to be released may be adjusted.
5291 * Note that !resv_map implies freed == 0. So (chg - freed)
5292 * won't go negative.
5294 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5295 hugetlb_acct_memory(h, -gbl_reserve);
5300 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5301 static unsigned long page_table_shareable(struct vm_area_struct *svma,
5302 struct vm_area_struct *vma,
5303 unsigned long addr, pgoff_t idx)
5305 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5307 unsigned long sbase = saddr & PUD_MASK;
5308 unsigned long s_end = sbase + PUD_SIZE;
5310 /* Allow segments to share if only one is marked locked */
5311 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5312 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5315 * match the virtual addresses, permission and the alignment of the
5318 if (pmd_index(addr) != pmd_index(saddr) ||
5319 vm_flags != svm_flags ||
5320 !range_in_vma(svma, sbase, s_end))
5326 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5328 unsigned long base = addr & PUD_MASK;
5329 unsigned long end = base + PUD_SIZE;
5332 * check on proper vm_flags and page table alignment
5334 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5339 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5341 #ifdef CONFIG_USERFAULTFD
5342 if (uffd_disable_huge_pmd_share(vma))
5345 return vma_shareable(vma, addr);
5349 * Determine if start,end range within vma could be mapped by shared pmd.
5350 * If yes, adjust start and end to cover range associated with possible
5351 * shared pmd mappings.
5353 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5354 unsigned long *start, unsigned long *end)
5356 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
5357 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
5360 * vma need span at least one aligned PUD size and the start,end range
5361 * must at least partialy within it.
5363 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
5364 (*end <= v_start) || (*start >= v_end))
5367 /* Extend the range to be PUD aligned for a worst case scenario */
5368 if (*start > v_start)
5369 *start = ALIGN_DOWN(*start, PUD_SIZE);
5372 *end = ALIGN(*end, PUD_SIZE);
5376 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5377 * and returns the corresponding pte. While this is not necessary for the
5378 * !shared pmd case because we can allocate the pmd later as well, it makes the
5379 * code much cleaner.
5381 * This routine must be called with i_mmap_rwsem held in at least read mode if
5382 * sharing is possible. For hugetlbfs, this prevents removal of any page
5383 * table entries associated with the address space. This is important as we
5384 * are setting up sharing based on existing page table entries (mappings).
5386 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5387 * huge_pte_alloc know that sharing is not possible and do not take
5388 * i_mmap_rwsem as a performance optimization. This is handled by the
5389 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5390 * only required for subsequent processing.
5392 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5393 unsigned long addr, pud_t *pud)
5395 struct address_space *mapping = vma->vm_file->f_mapping;
5396 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5398 struct vm_area_struct *svma;
5399 unsigned long saddr;
5404 i_mmap_assert_locked(mapping);
5405 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5409 saddr = page_table_shareable(svma, vma, addr, idx);
5411 spte = huge_pte_offset(svma->vm_mm, saddr,
5412 vma_mmu_pagesize(svma));
5414 get_page(virt_to_page(spte));
5423 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
5424 if (pud_none(*pud)) {
5425 pud_populate(mm, pud,
5426 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5429 put_page(virt_to_page(spte));
5433 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5438 * unmap huge page backed by shared pte.
5440 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5441 * indicated by page_count > 1, unmap is achieved by clearing pud and
5442 * decrementing the ref count. If count == 1, the pte page is not shared.
5444 * Called with page table lock held and i_mmap_rwsem held in write mode.
5446 * returns: 1 successfully unmapped a shared pte page
5447 * 0 the underlying pte page is not shared, or it is the last user
5449 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5450 unsigned long *addr, pte_t *ptep)
5452 pgd_t *pgd = pgd_offset(mm, *addr);
5453 p4d_t *p4d = p4d_offset(pgd, *addr);
5454 pud_t *pud = pud_offset(p4d, *addr);
5456 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
5457 BUG_ON(page_count(virt_to_page(ptep)) == 0);
5458 if (page_count(virt_to_page(ptep)) == 1)
5462 put_page(virt_to_page(ptep));
5464 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5468 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5469 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5470 unsigned long addr, pud_t *pud)
5475 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5476 unsigned long *addr, pte_t *ptep)
5481 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5482 unsigned long *start, unsigned long *end)
5486 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5490 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5492 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5493 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
5494 unsigned long addr, unsigned long sz)
5501 pgd = pgd_offset(mm, addr);
5502 p4d = p4d_alloc(mm, pgd, addr);
5505 pud = pud_alloc(mm, p4d, addr);
5507 if (sz == PUD_SIZE) {
5510 BUG_ON(sz != PMD_SIZE);
5511 if (want_pmd_share(vma, addr) && pud_none(*pud))
5512 pte = huge_pmd_share(mm, vma, addr, pud);
5514 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5517 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
5523 * huge_pte_offset() - Walk the page table to resolve the hugepage
5524 * entry at address @addr
5526 * Return: Pointer to page table entry (PUD or PMD) for
5527 * address @addr, or NULL if a !p*d_present() entry is encountered and the
5528 * size @sz doesn't match the hugepage size at this level of the page
5531 pte_t *huge_pte_offset(struct mm_struct *mm,
5532 unsigned long addr, unsigned long sz)
5539 pgd = pgd_offset(mm, addr);
5540 if (!pgd_present(*pgd))
5542 p4d = p4d_offset(pgd, addr);
5543 if (!p4d_present(*p4d))
5546 pud = pud_offset(p4d, addr);
5548 /* must be pud huge, non-present or none */
5549 return (pte_t *)pud;
5550 if (!pud_present(*pud))
5552 /* must have a valid entry and size to go further */
5554 pmd = pmd_offset(pud, addr);
5555 /* must be pmd huge, non-present or none */
5556 return (pte_t *)pmd;
5559 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5562 * These functions are overwritable if your architecture needs its own
5565 struct page * __weak
5566 follow_huge_addr(struct mm_struct *mm, unsigned long address,
5569 return ERR_PTR(-EINVAL);
5572 struct page * __weak
5573 follow_huge_pd(struct vm_area_struct *vma,
5574 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5576 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5580 struct page * __weak
5581 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
5582 pmd_t *pmd, int flags)
5584 struct page *page = NULL;
5588 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5589 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
5590 (FOLL_PIN | FOLL_GET)))
5594 ptl = pmd_lockptr(mm, pmd);
5597 * make sure that the address range covered by this pmd is not
5598 * unmapped from other threads.
5600 if (!pmd_huge(*pmd))
5602 pte = huge_ptep_get((pte_t *)pmd);
5603 if (pte_present(pte)) {
5604 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
5606 * try_grab_page() should always succeed here, because: a) we
5607 * hold the pmd (ptl) lock, and b) we've just checked that the
5608 * huge pmd (head) page is present in the page tables. The ptl
5609 * prevents the head page and tail pages from being rearranged
5610 * in any way. So this page must be available at this point,
5611 * unless the page refcount overflowed:
5613 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5618 if (is_hugetlb_entry_migration(pte)) {
5620 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5624 * hwpoisoned entry is treated as no_page_table in
5625 * follow_page_mask().
5633 struct page * __weak
5634 follow_huge_pud(struct mm_struct *mm, unsigned long address,
5635 pud_t *pud, int flags)
5637 if (flags & (FOLL_GET | FOLL_PIN))
5640 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5643 struct page * __weak
5644 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5646 if (flags & (FOLL_GET | FOLL_PIN))
5649 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5652 bool isolate_huge_page(struct page *page, struct list_head *list)
5656 spin_lock_irq(&hugetlb_lock);
5657 if (!PageHeadHuge(page) ||
5658 !HPageMigratable(page) ||
5659 !get_page_unless_zero(page)) {
5663 ClearHPageMigratable(page);
5664 list_move_tail(&page->lru, list);
5666 spin_unlock_irq(&hugetlb_lock);
5670 void putback_active_hugepage(struct page *page)
5672 spin_lock_irq(&hugetlb_lock);
5673 SetHPageMigratable(page);
5674 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5675 spin_unlock_irq(&hugetlb_lock);
5679 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5681 struct hstate *h = page_hstate(oldpage);
5683 hugetlb_cgroup_migrate(oldpage, newpage);
5684 set_page_owner_migrate_reason(newpage, reason);
5687 * transfer temporary state of the new huge page. This is
5688 * reverse to other transitions because the newpage is going to
5689 * be final while the old one will be freed so it takes over
5690 * the temporary status.
5692 * Also note that we have to transfer the per-node surplus state
5693 * here as well otherwise the global surplus count will not match
5696 if (HPageTemporary(newpage)) {
5697 int old_nid = page_to_nid(oldpage);
5698 int new_nid = page_to_nid(newpage);
5700 SetHPageTemporary(oldpage);
5701 ClearHPageTemporary(newpage);
5704 * There is no need to transfer the per-node surplus state
5705 * when we do not cross the node.
5707 if (new_nid == old_nid)
5709 spin_lock_irq(&hugetlb_lock);
5710 if (h->surplus_huge_pages_node[old_nid]) {
5711 h->surplus_huge_pages_node[old_nid]--;
5712 h->surplus_huge_pages_node[new_nid]++;
5714 spin_unlock_irq(&hugetlb_lock);
5719 * This function will unconditionally remove all the shared pmd pgtable entries
5720 * within the specific vma for a hugetlbfs memory range.
5722 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
5724 struct hstate *h = hstate_vma(vma);
5725 unsigned long sz = huge_page_size(h);
5726 struct mm_struct *mm = vma->vm_mm;
5727 struct mmu_notifier_range range;
5728 unsigned long address, start, end;
5732 if (!(vma->vm_flags & VM_MAYSHARE))
5735 start = ALIGN(vma->vm_start, PUD_SIZE);
5736 end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
5742 * No need to call adjust_range_if_pmd_sharing_possible(), because
5743 * we have already done the PUD_SIZE alignment.
5745 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
5747 mmu_notifier_invalidate_range_start(&range);
5748 i_mmap_lock_write(vma->vm_file->f_mapping);
5749 for (address = start; address < end; address += PUD_SIZE) {
5750 unsigned long tmp = address;
5752 ptep = huge_pte_offset(mm, address, sz);
5755 ptl = huge_pte_lock(h, mm, ptep);
5756 /* We don't want 'address' to be changed */
5757 huge_pmd_unshare(mm, vma, &tmp, ptep);
5760 flush_hugetlb_tlb_range(vma, start, end);
5761 i_mmap_unlock_write(vma->vm_file->f_mapping);
5763 * No need to call mmu_notifier_invalidate_range(), see
5764 * Documentation/vm/mmu_notifier.rst.
5766 mmu_notifier_invalidate_range_end(&range);
5770 static bool cma_reserve_called __initdata;
5772 static int __init cmdline_parse_hugetlb_cma(char *p)
5774 hugetlb_cma_size = memparse(p, &p);
5778 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
5780 void __init hugetlb_cma_reserve(int order)
5782 unsigned long size, reserved, per_node;
5785 cma_reserve_called = true;
5787 if (!hugetlb_cma_size)
5790 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
5791 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
5792 (PAGE_SIZE << order) / SZ_1M);
5797 * If 3 GB area is requested on a machine with 4 numa nodes,
5798 * let's allocate 1 GB on first three nodes and ignore the last one.
5800 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
5801 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
5802 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
5805 for_each_node_state(nid, N_ONLINE) {
5807 char name[CMA_MAX_NAME];
5809 size = min(per_node, hugetlb_cma_size - reserved);
5810 size = round_up(size, PAGE_SIZE << order);
5812 snprintf(name, sizeof(name), "hugetlb%d", nid);
5813 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
5815 &hugetlb_cma[nid], nid);
5817 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
5823 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
5826 if (reserved >= hugetlb_cma_size)
5831 void __init hugetlb_cma_check(void)
5833 if (!hugetlb_cma_size || cma_reserve_called)
5836 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
5839 #endif /* CONFIG_CMA */