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);
1488 * Must be called with the hugetlb lock held
1490 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1492 lockdep_assert_held(&hugetlb_lock);
1494 h->nr_huge_pages_node[nid]++;
1497 static void __prep_new_huge_page(struct page *page)
1499 INIT_LIST_HEAD(&page->lru);
1500 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1501 hugetlb_set_page_subpool(page, NULL);
1502 set_hugetlb_cgroup(page, NULL);
1503 set_hugetlb_cgroup_rsvd(page, NULL);
1506 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1508 __prep_new_huge_page(page);
1509 spin_lock_irq(&hugetlb_lock);
1510 __prep_account_new_huge_page(h, nid);
1511 spin_unlock_irq(&hugetlb_lock);
1514 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1517 int nr_pages = 1 << order;
1518 struct page *p = page + 1;
1520 /* we rely on prep_new_huge_page to set the destructor */
1521 set_compound_order(page, order);
1522 __ClearPageReserved(page);
1523 __SetPageHead(page);
1524 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1526 * For gigantic hugepages allocated through bootmem at
1527 * boot, it's safer to be consistent with the not-gigantic
1528 * hugepages and clear the PG_reserved bit from all tail pages
1529 * too. Otherwise drivers using get_user_pages() to access tail
1530 * pages may get the reference counting wrong if they see
1531 * PG_reserved set on a tail page (despite the head page not
1532 * having PG_reserved set). Enforcing this consistency between
1533 * head and tail pages allows drivers to optimize away a check
1534 * on the head page when they need know if put_page() is needed
1535 * after get_user_pages().
1537 __ClearPageReserved(p);
1538 set_page_count(p, 0);
1539 set_compound_head(p, page);
1541 atomic_set(compound_mapcount_ptr(page), -1);
1542 atomic_set(compound_pincount_ptr(page), 0);
1546 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1547 * transparent huge pages. See the PageTransHuge() documentation for more
1550 int PageHuge(struct page *page)
1552 if (!PageCompound(page))
1555 page = compound_head(page);
1556 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1558 EXPORT_SYMBOL_GPL(PageHuge);
1561 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1562 * normal or transparent huge pages.
1564 int PageHeadHuge(struct page *page_head)
1566 if (!PageHead(page_head))
1569 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1573 * Find and lock address space (mapping) in write mode.
1575 * Upon entry, the page is locked which means that page_mapping() is
1576 * stable. Due to locking order, we can only trylock_write. If we can
1577 * not get the lock, simply return NULL to caller.
1579 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1581 struct address_space *mapping = page_mapping(hpage);
1586 if (i_mmap_trylock_write(mapping))
1592 pgoff_t __basepage_index(struct page *page)
1594 struct page *page_head = compound_head(page);
1595 pgoff_t index = page_index(page_head);
1596 unsigned long compound_idx;
1598 if (!PageHuge(page_head))
1599 return page_index(page);
1601 if (compound_order(page_head) >= MAX_ORDER)
1602 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1604 compound_idx = page - page_head;
1606 return (index << compound_order(page_head)) + compound_idx;
1609 static struct page *alloc_buddy_huge_page(struct hstate *h,
1610 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1611 nodemask_t *node_alloc_noretry)
1613 int order = huge_page_order(h);
1615 bool alloc_try_hard = true;
1618 * By default we always try hard to allocate the page with
1619 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1620 * a loop (to adjust global huge page counts) and previous allocation
1621 * failed, do not continue to try hard on the same node. Use the
1622 * node_alloc_noretry bitmap to manage this state information.
1624 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1625 alloc_try_hard = false;
1626 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1628 gfp_mask |= __GFP_RETRY_MAYFAIL;
1629 if (nid == NUMA_NO_NODE)
1630 nid = numa_mem_id();
1631 page = __alloc_pages(gfp_mask, order, nid, nmask);
1633 __count_vm_event(HTLB_BUDDY_PGALLOC);
1635 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1638 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1639 * indicates an overall state change. Clear bit so that we resume
1640 * normal 'try hard' allocations.
1642 if (node_alloc_noretry && page && !alloc_try_hard)
1643 node_clear(nid, *node_alloc_noretry);
1646 * If we tried hard to get a page but failed, set bit so that
1647 * subsequent attempts will not try as hard until there is an
1648 * overall state change.
1650 if (node_alloc_noretry && !page && alloc_try_hard)
1651 node_set(nid, *node_alloc_noretry);
1657 * Common helper to allocate a fresh hugetlb page. All specific allocators
1658 * should use this function to get new hugetlb pages
1660 static struct page *alloc_fresh_huge_page(struct hstate *h,
1661 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1662 nodemask_t *node_alloc_noretry)
1666 if (hstate_is_gigantic(h))
1667 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1669 page = alloc_buddy_huge_page(h, gfp_mask,
1670 nid, nmask, node_alloc_noretry);
1674 if (hstate_is_gigantic(h))
1675 prep_compound_gigantic_page(page, huge_page_order(h));
1676 prep_new_huge_page(h, page, page_to_nid(page));
1682 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1685 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1686 nodemask_t *node_alloc_noretry)
1690 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1692 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1693 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1694 node_alloc_noretry);
1702 put_page(page); /* free it into the hugepage allocator */
1708 * Remove huge page from pool from next node to free. Attempt to keep
1709 * persistent huge pages more or less balanced over allowed nodes.
1710 * This routine only 'removes' the hugetlb page. The caller must make
1711 * an additional call to free the page to low level allocators.
1712 * Called with hugetlb_lock locked.
1714 static struct page *remove_pool_huge_page(struct hstate *h,
1715 nodemask_t *nodes_allowed,
1719 struct page *page = NULL;
1721 lockdep_assert_held(&hugetlb_lock);
1722 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1724 * If we're returning unused surplus pages, only examine
1725 * nodes with surplus pages.
1727 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1728 !list_empty(&h->hugepage_freelists[node])) {
1729 page = list_entry(h->hugepage_freelists[node].next,
1731 remove_hugetlb_page(h, page, acct_surplus);
1740 * Dissolve a given free hugepage into free buddy pages. This function does
1741 * nothing for in-use hugepages and non-hugepages.
1742 * This function returns values like below:
1744 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1745 * (allocated or reserved.)
1746 * 0: successfully dissolved free hugepages or the page is not a
1747 * hugepage (considered as already dissolved)
1749 int dissolve_free_huge_page(struct page *page)
1754 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1755 if (!PageHuge(page))
1758 spin_lock_irq(&hugetlb_lock);
1759 if (!PageHuge(page)) {
1764 if (!page_count(page)) {
1765 struct page *head = compound_head(page);
1766 struct hstate *h = page_hstate(head);
1767 if (h->free_huge_pages - h->resv_huge_pages == 0)
1771 * We should make sure that the page is already on the free list
1772 * when it is dissolved.
1774 if (unlikely(!HPageFreed(head))) {
1775 spin_unlock_irq(&hugetlb_lock);
1779 * Theoretically, we should return -EBUSY when we
1780 * encounter this race. In fact, we have a chance
1781 * to successfully dissolve the page if we do a
1782 * retry. Because the race window is quite small.
1783 * If we seize this opportunity, it is an optimization
1784 * for increasing the success rate of dissolving page.
1790 * Move PageHWPoison flag from head page to the raw error page,
1791 * which makes any subpages rather than the error page reusable.
1793 if (PageHWPoison(head) && page != head) {
1794 SetPageHWPoison(page);
1795 ClearPageHWPoison(head);
1797 remove_hugetlb_page(h, page, false);
1798 h->max_huge_pages--;
1799 spin_unlock_irq(&hugetlb_lock);
1800 update_and_free_page(h, head);
1804 spin_unlock_irq(&hugetlb_lock);
1809 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1810 * make specified memory blocks removable from the system.
1811 * Note that this will dissolve a free gigantic hugepage completely, if any
1812 * part of it lies within the given range.
1813 * Also note that if dissolve_free_huge_page() returns with an error, all
1814 * free hugepages that were dissolved before that error are lost.
1816 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1822 if (!hugepages_supported())
1825 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1826 page = pfn_to_page(pfn);
1827 rc = dissolve_free_huge_page(page);
1836 * Allocates a fresh surplus page from the page allocator.
1838 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1839 int nid, nodemask_t *nmask)
1841 struct page *page = NULL;
1843 if (hstate_is_gigantic(h))
1846 spin_lock_irq(&hugetlb_lock);
1847 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1849 spin_unlock_irq(&hugetlb_lock);
1851 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1855 spin_lock_irq(&hugetlb_lock);
1857 * We could have raced with the pool size change.
1858 * Double check that and simply deallocate the new page
1859 * if we would end up overcommiting the surpluses. Abuse
1860 * temporary page to workaround the nasty free_huge_page
1863 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1864 SetHPageTemporary(page);
1865 spin_unlock_irq(&hugetlb_lock);
1869 h->surplus_huge_pages++;
1870 h->surplus_huge_pages_node[page_to_nid(page)]++;
1874 spin_unlock_irq(&hugetlb_lock);
1879 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1880 int nid, nodemask_t *nmask)
1884 if (hstate_is_gigantic(h))
1887 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1892 * We do not account these pages as surplus because they are only
1893 * temporary and will be released properly on the last reference
1895 SetHPageTemporary(page);
1901 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1904 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1905 struct vm_area_struct *vma, unsigned long addr)
1908 struct mempolicy *mpol;
1909 gfp_t gfp_mask = htlb_alloc_mask(h);
1911 nodemask_t *nodemask;
1913 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1914 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1915 mpol_cond_put(mpol);
1920 /* page migration callback function */
1921 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1922 nodemask_t *nmask, gfp_t gfp_mask)
1924 spin_lock_irq(&hugetlb_lock);
1925 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1928 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1930 spin_unlock_irq(&hugetlb_lock);
1934 spin_unlock_irq(&hugetlb_lock);
1936 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1939 /* mempolicy aware migration callback */
1940 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1941 unsigned long address)
1943 struct mempolicy *mpol;
1944 nodemask_t *nodemask;
1949 gfp_mask = htlb_alloc_mask(h);
1950 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1951 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
1952 mpol_cond_put(mpol);
1958 * Increase the hugetlb pool such that it can accommodate a reservation
1961 static int gather_surplus_pages(struct hstate *h, long delta)
1962 __must_hold(&hugetlb_lock)
1964 struct list_head surplus_list;
1965 struct page *page, *tmp;
1968 long needed, allocated;
1969 bool alloc_ok = true;
1971 lockdep_assert_held(&hugetlb_lock);
1972 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1974 h->resv_huge_pages += delta;
1979 INIT_LIST_HEAD(&surplus_list);
1983 spin_unlock_irq(&hugetlb_lock);
1984 for (i = 0; i < needed; i++) {
1985 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1986 NUMA_NO_NODE, NULL);
1991 list_add(&page->lru, &surplus_list);
1997 * After retaking hugetlb_lock, we need to recalculate 'needed'
1998 * because either resv_huge_pages or free_huge_pages may have changed.
2000 spin_lock_irq(&hugetlb_lock);
2001 needed = (h->resv_huge_pages + delta) -
2002 (h->free_huge_pages + allocated);
2007 * We were not able to allocate enough pages to
2008 * satisfy the entire reservation so we free what
2009 * we've allocated so far.
2014 * The surplus_list now contains _at_least_ the number of extra pages
2015 * needed to accommodate the reservation. Add the appropriate number
2016 * of pages to the hugetlb pool and free the extras back to the buddy
2017 * allocator. Commit the entire reservation here to prevent another
2018 * process from stealing the pages as they are added to the pool but
2019 * before they are reserved.
2021 needed += allocated;
2022 h->resv_huge_pages += delta;
2025 /* Free the needed pages to the hugetlb pool */
2026 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2032 * This page is now managed by the hugetlb allocator and has
2033 * no users -- drop the buddy allocator's reference.
2035 zeroed = put_page_testzero(page);
2036 VM_BUG_ON_PAGE(!zeroed, page);
2037 enqueue_huge_page(h, page);
2040 spin_unlock_irq(&hugetlb_lock);
2042 /* Free unnecessary surplus pages to the buddy allocator */
2043 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2045 spin_lock_irq(&hugetlb_lock);
2051 * This routine has two main purposes:
2052 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2053 * in unused_resv_pages. This corresponds to the prior adjustments made
2054 * to the associated reservation map.
2055 * 2) Free any unused surplus pages that may have been allocated to satisfy
2056 * the reservation. As many as unused_resv_pages may be freed.
2058 static void return_unused_surplus_pages(struct hstate *h,
2059 unsigned long unused_resv_pages)
2061 unsigned long nr_pages;
2063 LIST_HEAD(page_list);
2065 lockdep_assert_held(&hugetlb_lock);
2066 /* Uncommit the reservation */
2067 h->resv_huge_pages -= unused_resv_pages;
2069 /* Cannot return gigantic pages currently */
2070 if (hstate_is_gigantic(h))
2074 * Part (or even all) of the reservation could have been backed
2075 * by pre-allocated pages. Only free surplus pages.
2077 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2080 * We want to release as many surplus pages as possible, spread
2081 * evenly across all nodes with memory. Iterate across these nodes
2082 * until we can no longer free unreserved surplus pages. This occurs
2083 * when the nodes with surplus pages have no free pages.
2084 * remove_pool_huge_page() will balance the freed pages across the
2085 * on-line nodes with memory and will handle the hstate accounting.
2087 while (nr_pages--) {
2088 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2092 list_add(&page->lru, &page_list);
2096 spin_unlock_irq(&hugetlb_lock);
2097 update_and_free_pages_bulk(h, &page_list);
2098 spin_lock_irq(&hugetlb_lock);
2103 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2104 * are used by the huge page allocation routines to manage reservations.
2106 * vma_needs_reservation is called to determine if the huge page at addr
2107 * within the vma has an associated reservation. If a reservation is
2108 * needed, the value 1 is returned. The caller is then responsible for
2109 * managing the global reservation and subpool usage counts. After
2110 * the huge page has been allocated, vma_commit_reservation is called
2111 * to add the page to the reservation map. If the page allocation fails,
2112 * the reservation must be ended instead of committed. vma_end_reservation
2113 * is called in such cases.
2115 * In the normal case, vma_commit_reservation returns the same value
2116 * as the preceding vma_needs_reservation call. The only time this
2117 * is not the case is if a reserve map was changed between calls. It
2118 * is the responsibility of the caller to notice the difference and
2119 * take appropriate action.
2121 * vma_add_reservation is used in error paths where a reservation must
2122 * be restored when a newly allocated huge page must be freed. It is
2123 * to be called after calling vma_needs_reservation to determine if a
2124 * reservation exists.
2126 enum vma_resv_mode {
2132 static long __vma_reservation_common(struct hstate *h,
2133 struct vm_area_struct *vma, unsigned long addr,
2134 enum vma_resv_mode mode)
2136 struct resv_map *resv;
2139 long dummy_out_regions_needed;
2141 resv = vma_resv_map(vma);
2145 idx = vma_hugecache_offset(h, vma, addr);
2147 case VMA_NEEDS_RESV:
2148 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2149 /* We assume that vma_reservation_* routines always operate on
2150 * 1 page, and that adding to resv map a 1 page entry can only
2151 * ever require 1 region.
2153 VM_BUG_ON(dummy_out_regions_needed != 1);
2155 case VMA_COMMIT_RESV:
2156 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2157 /* region_add calls of range 1 should never fail. */
2161 region_abort(resv, idx, idx + 1, 1);
2165 if (vma->vm_flags & VM_MAYSHARE) {
2166 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2167 /* region_add calls of range 1 should never fail. */
2170 region_abort(resv, idx, idx + 1, 1);
2171 ret = region_del(resv, idx, idx + 1);
2178 if (vma->vm_flags & VM_MAYSHARE)
2181 * We know private mapping must have HPAGE_RESV_OWNER set.
2183 * In most cases, reserves always exist for private mappings.
2184 * However, a file associated with mapping could have been
2185 * hole punched or truncated after reserves were consumed.
2186 * As subsequent fault on such a range will not use reserves.
2187 * Subtle - The reserve map for private mappings has the
2188 * opposite meaning than that of shared mappings. If NO
2189 * entry is in the reserve map, it means a reservation exists.
2190 * If an entry exists in the reserve map, it means the
2191 * reservation has already been consumed. As a result, the
2192 * return value of this routine is the opposite of the
2193 * value returned from reserve map manipulation routines above.
2202 static long vma_needs_reservation(struct hstate *h,
2203 struct vm_area_struct *vma, unsigned long addr)
2205 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2208 static long vma_commit_reservation(struct hstate *h,
2209 struct vm_area_struct *vma, unsigned long addr)
2211 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2214 static void vma_end_reservation(struct hstate *h,
2215 struct vm_area_struct *vma, unsigned long addr)
2217 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2220 static long vma_add_reservation(struct hstate *h,
2221 struct vm_area_struct *vma, unsigned long addr)
2223 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2227 * This routine is called to restore a reservation on error paths. In the
2228 * specific error paths, a huge page was allocated (via alloc_huge_page)
2229 * and is about to be freed. If a reservation for the page existed,
2230 * alloc_huge_page would have consumed the reservation and set
2231 * HPageRestoreReserve in the newly allocated page. When the page is freed
2232 * via free_huge_page, the global reservation count will be incremented if
2233 * HPageRestoreReserve is set. However, free_huge_page can not adjust the
2234 * reserve map. Adjust the reserve map here to be consistent with global
2235 * reserve count adjustments to be made by free_huge_page.
2237 static void restore_reserve_on_error(struct hstate *h,
2238 struct vm_area_struct *vma, unsigned long address,
2241 if (unlikely(HPageRestoreReserve(page))) {
2242 long rc = vma_needs_reservation(h, vma, address);
2244 if (unlikely(rc < 0)) {
2246 * Rare out of memory condition in reserve map
2247 * manipulation. Clear HPageRestoreReserve so that
2248 * global reserve count will not be incremented
2249 * by free_huge_page. This will make it appear
2250 * as though the reservation for this page was
2251 * consumed. This may prevent the task from
2252 * faulting in the page at a later time. This
2253 * is better than inconsistent global huge page
2254 * accounting of reserve counts.
2256 ClearHPageRestoreReserve(page);
2258 rc = vma_add_reservation(h, vma, address);
2259 if (unlikely(rc < 0))
2261 * See above comment about rare out of
2264 ClearHPageRestoreReserve(page);
2266 vma_end_reservation(h, vma, address);
2270 struct page *alloc_huge_page(struct vm_area_struct *vma,
2271 unsigned long addr, int avoid_reserve)
2273 struct hugepage_subpool *spool = subpool_vma(vma);
2274 struct hstate *h = hstate_vma(vma);
2276 long map_chg, map_commit;
2279 struct hugetlb_cgroup *h_cg;
2280 bool deferred_reserve;
2282 idx = hstate_index(h);
2284 * Examine the region/reserve map to determine if the process
2285 * has a reservation for the page to be allocated. A return
2286 * code of zero indicates a reservation exists (no change).
2288 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2290 return ERR_PTR(-ENOMEM);
2293 * Processes that did not create the mapping will have no
2294 * reserves as indicated by the region/reserve map. Check
2295 * that the allocation will not exceed the subpool limit.
2296 * Allocations for MAP_NORESERVE mappings also need to be
2297 * checked against any subpool limit.
2299 if (map_chg || avoid_reserve) {
2300 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2302 vma_end_reservation(h, vma, addr);
2303 return ERR_PTR(-ENOSPC);
2307 * Even though there was no reservation in the region/reserve
2308 * map, there could be reservations associated with the
2309 * subpool that can be used. This would be indicated if the
2310 * return value of hugepage_subpool_get_pages() is zero.
2311 * However, if avoid_reserve is specified we still avoid even
2312 * the subpool reservations.
2318 /* If this allocation is not consuming a reservation, charge it now.
2320 deferred_reserve = map_chg || avoid_reserve;
2321 if (deferred_reserve) {
2322 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2323 idx, pages_per_huge_page(h), &h_cg);
2325 goto out_subpool_put;
2328 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2330 goto out_uncharge_cgroup_reservation;
2332 spin_lock_irq(&hugetlb_lock);
2334 * glb_chg is passed to indicate whether or not a page must be taken
2335 * from the global free pool (global change). gbl_chg == 0 indicates
2336 * a reservation exists for the allocation.
2338 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2340 spin_unlock_irq(&hugetlb_lock);
2341 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2343 goto out_uncharge_cgroup;
2344 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2345 SetHPageRestoreReserve(page);
2346 h->resv_huge_pages--;
2348 spin_lock_irq(&hugetlb_lock);
2349 list_add(&page->lru, &h->hugepage_activelist);
2352 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2353 /* If allocation is not consuming a reservation, also store the
2354 * hugetlb_cgroup pointer on the page.
2356 if (deferred_reserve) {
2357 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2361 spin_unlock_irq(&hugetlb_lock);
2363 hugetlb_set_page_subpool(page, spool);
2365 map_commit = vma_commit_reservation(h, vma, addr);
2366 if (unlikely(map_chg > map_commit)) {
2368 * The page was added to the reservation map between
2369 * vma_needs_reservation and vma_commit_reservation.
2370 * This indicates a race with hugetlb_reserve_pages.
2371 * Adjust for the subpool count incremented above AND
2372 * in hugetlb_reserve_pages for the same page. Also,
2373 * the reservation count added in hugetlb_reserve_pages
2374 * no longer applies.
2378 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2379 hugetlb_acct_memory(h, -rsv_adjust);
2380 if (deferred_reserve)
2381 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2382 pages_per_huge_page(h), page);
2386 out_uncharge_cgroup:
2387 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2388 out_uncharge_cgroup_reservation:
2389 if (deferred_reserve)
2390 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2393 if (map_chg || avoid_reserve)
2394 hugepage_subpool_put_pages(spool, 1);
2395 vma_end_reservation(h, vma, addr);
2396 return ERR_PTR(-ENOSPC);
2399 int alloc_bootmem_huge_page(struct hstate *h)
2400 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2401 int __alloc_bootmem_huge_page(struct hstate *h)
2403 struct huge_bootmem_page *m;
2406 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2409 addr = memblock_alloc_try_nid_raw(
2410 huge_page_size(h), huge_page_size(h),
2411 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2414 * Use the beginning of the huge page to store the
2415 * huge_bootmem_page struct (until gather_bootmem
2416 * puts them into the mem_map).
2425 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2426 /* Put them into a private list first because mem_map is not up yet */
2427 INIT_LIST_HEAD(&m->list);
2428 list_add(&m->list, &huge_boot_pages);
2433 static void __init prep_compound_huge_page(struct page *page,
2436 if (unlikely(order > (MAX_ORDER - 1)))
2437 prep_compound_gigantic_page(page, order);
2439 prep_compound_page(page, order);
2442 /* Put bootmem huge pages into the standard lists after mem_map is up */
2443 static void __init gather_bootmem_prealloc(void)
2445 struct huge_bootmem_page *m;
2447 list_for_each_entry(m, &huge_boot_pages, list) {
2448 struct page *page = virt_to_page(m);
2449 struct hstate *h = m->hstate;
2451 WARN_ON(page_count(page) != 1);
2452 prep_compound_huge_page(page, huge_page_order(h));
2453 WARN_ON(PageReserved(page));
2454 prep_new_huge_page(h, page, page_to_nid(page));
2455 put_page(page); /* free it into the hugepage allocator */
2458 * If we had gigantic hugepages allocated at boot time, we need
2459 * to restore the 'stolen' pages to totalram_pages in order to
2460 * fix confusing memory reports from free(1) and another
2461 * side-effects, like CommitLimit going negative.
2463 if (hstate_is_gigantic(h))
2464 adjust_managed_page_count(page, pages_per_huge_page(h));
2469 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2472 nodemask_t *node_alloc_noretry;
2474 if (!hstate_is_gigantic(h)) {
2476 * Bit mask controlling how hard we retry per-node allocations.
2477 * Ignore errors as lower level routines can deal with
2478 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2479 * time, we are likely in bigger trouble.
2481 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2484 /* allocations done at boot time */
2485 node_alloc_noretry = NULL;
2488 /* bit mask controlling how hard we retry per-node allocations */
2489 if (node_alloc_noretry)
2490 nodes_clear(*node_alloc_noretry);
2492 for (i = 0; i < h->max_huge_pages; ++i) {
2493 if (hstate_is_gigantic(h)) {
2494 if (hugetlb_cma_size) {
2495 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2498 if (!alloc_bootmem_huge_page(h))
2500 } else if (!alloc_pool_huge_page(h,
2501 &node_states[N_MEMORY],
2502 node_alloc_noretry))
2506 if (i < h->max_huge_pages) {
2509 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2510 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2511 h->max_huge_pages, buf, i);
2512 h->max_huge_pages = i;
2515 kfree(node_alloc_noretry);
2518 static void __init hugetlb_init_hstates(void)
2522 for_each_hstate(h) {
2523 if (minimum_order > huge_page_order(h))
2524 minimum_order = huge_page_order(h);
2526 /* oversize hugepages were init'ed in early boot */
2527 if (!hstate_is_gigantic(h))
2528 hugetlb_hstate_alloc_pages(h);
2530 VM_BUG_ON(minimum_order == UINT_MAX);
2533 static void __init report_hugepages(void)
2537 for_each_hstate(h) {
2540 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2541 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2542 buf, h->free_huge_pages);
2546 #ifdef CONFIG_HIGHMEM
2547 static void try_to_free_low(struct hstate *h, unsigned long count,
2548 nodemask_t *nodes_allowed)
2551 LIST_HEAD(page_list);
2553 lockdep_assert_held(&hugetlb_lock);
2554 if (hstate_is_gigantic(h))
2558 * Collect pages to be freed on a list, and free after dropping lock
2560 for_each_node_mask(i, *nodes_allowed) {
2561 struct page *page, *next;
2562 struct list_head *freel = &h->hugepage_freelists[i];
2563 list_for_each_entry_safe(page, next, freel, lru) {
2564 if (count >= h->nr_huge_pages)
2566 if (PageHighMem(page))
2568 remove_hugetlb_page(h, page, false);
2569 list_add(&page->lru, &page_list);
2574 spin_unlock_irq(&hugetlb_lock);
2575 update_and_free_pages_bulk(h, &page_list);
2576 spin_lock_irq(&hugetlb_lock);
2579 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2580 nodemask_t *nodes_allowed)
2586 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2587 * balanced by operating on them in a round-robin fashion.
2588 * Returns 1 if an adjustment was made.
2590 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2595 lockdep_assert_held(&hugetlb_lock);
2596 VM_BUG_ON(delta != -1 && delta != 1);
2599 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2600 if (h->surplus_huge_pages_node[node])
2604 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2605 if (h->surplus_huge_pages_node[node] <
2606 h->nr_huge_pages_node[node])
2613 h->surplus_huge_pages += delta;
2614 h->surplus_huge_pages_node[node] += delta;
2618 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2619 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2620 nodemask_t *nodes_allowed)
2622 unsigned long min_count, ret;
2624 LIST_HEAD(page_list);
2625 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2628 * Bit mask controlling how hard we retry per-node allocations.
2629 * If we can not allocate the bit mask, do not attempt to allocate
2630 * the requested huge pages.
2632 if (node_alloc_noretry)
2633 nodes_clear(*node_alloc_noretry);
2638 * resize_lock mutex prevents concurrent adjustments to number of
2639 * pages in hstate via the proc/sysfs interfaces.
2641 mutex_lock(&h->resize_lock);
2642 spin_lock_irq(&hugetlb_lock);
2645 * Check for a node specific request.
2646 * Changing node specific huge page count may require a corresponding
2647 * change to the global count. In any case, the passed node mask
2648 * (nodes_allowed) will restrict alloc/free to the specified node.
2650 if (nid != NUMA_NO_NODE) {
2651 unsigned long old_count = count;
2653 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2655 * User may have specified a large count value which caused the
2656 * above calculation to overflow. In this case, they wanted
2657 * to allocate as many huge pages as possible. Set count to
2658 * largest possible value to align with their intention.
2660 if (count < old_count)
2665 * Gigantic pages runtime allocation depend on the capability for large
2666 * page range allocation.
2667 * If the system does not provide this feature, return an error when
2668 * the user tries to allocate gigantic pages but let the user free the
2669 * boottime allocated gigantic pages.
2671 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2672 if (count > persistent_huge_pages(h)) {
2673 spin_unlock_irq(&hugetlb_lock);
2674 mutex_unlock(&h->resize_lock);
2675 NODEMASK_FREE(node_alloc_noretry);
2678 /* Fall through to decrease pool */
2682 * Increase the pool size
2683 * First take pages out of surplus state. Then make up the
2684 * remaining difference by allocating fresh huge pages.
2686 * We might race with alloc_surplus_huge_page() here and be unable
2687 * to convert a surplus huge page to a normal huge page. That is
2688 * not critical, though, it just means the overall size of the
2689 * pool might be one hugepage larger than it needs to be, but
2690 * within all the constraints specified by the sysctls.
2692 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2693 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2697 while (count > persistent_huge_pages(h)) {
2699 * If this allocation races such that we no longer need the
2700 * page, free_huge_page will handle it by freeing the page
2701 * and reducing the surplus.
2703 spin_unlock_irq(&hugetlb_lock);
2705 /* yield cpu to avoid soft lockup */
2708 ret = alloc_pool_huge_page(h, nodes_allowed,
2709 node_alloc_noretry);
2710 spin_lock_irq(&hugetlb_lock);
2714 /* Bail for signals. Probably ctrl-c from user */
2715 if (signal_pending(current))
2720 * Decrease the pool size
2721 * First return free pages to the buddy allocator (being careful
2722 * to keep enough around to satisfy reservations). Then place
2723 * pages into surplus state as needed so the pool will shrink
2724 * to the desired size as pages become free.
2726 * By placing pages into the surplus state independent of the
2727 * overcommit value, we are allowing the surplus pool size to
2728 * exceed overcommit. There are few sane options here. Since
2729 * alloc_surplus_huge_page() is checking the global counter,
2730 * though, we'll note that we're not allowed to exceed surplus
2731 * and won't grow the pool anywhere else. Not until one of the
2732 * sysctls are changed, or the surplus pages go out of use.
2734 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2735 min_count = max(count, min_count);
2736 try_to_free_low(h, min_count, nodes_allowed);
2739 * Collect pages to be removed on list without dropping lock
2741 while (min_count < persistent_huge_pages(h)) {
2742 page = remove_pool_huge_page(h, nodes_allowed, 0);
2746 list_add(&page->lru, &page_list);
2748 /* free the pages after dropping lock */
2749 spin_unlock_irq(&hugetlb_lock);
2750 update_and_free_pages_bulk(h, &page_list);
2751 spin_lock_irq(&hugetlb_lock);
2753 while (count < persistent_huge_pages(h)) {
2754 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2758 h->max_huge_pages = persistent_huge_pages(h);
2759 spin_unlock_irq(&hugetlb_lock);
2760 mutex_unlock(&h->resize_lock);
2762 NODEMASK_FREE(node_alloc_noretry);
2767 #define HSTATE_ATTR_RO(_name) \
2768 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2770 #define HSTATE_ATTR(_name) \
2771 static struct kobj_attribute _name##_attr = \
2772 __ATTR(_name, 0644, _name##_show, _name##_store)
2774 static struct kobject *hugepages_kobj;
2775 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2777 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2779 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2783 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2784 if (hstate_kobjs[i] == kobj) {
2786 *nidp = NUMA_NO_NODE;
2790 return kobj_to_node_hstate(kobj, nidp);
2793 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2794 struct kobj_attribute *attr, char *buf)
2797 unsigned long nr_huge_pages;
2800 h = kobj_to_hstate(kobj, &nid);
2801 if (nid == NUMA_NO_NODE)
2802 nr_huge_pages = h->nr_huge_pages;
2804 nr_huge_pages = h->nr_huge_pages_node[nid];
2806 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
2809 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2810 struct hstate *h, int nid,
2811 unsigned long count, size_t len)
2814 nodemask_t nodes_allowed, *n_mask;
2816 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2819 if (nid == NUMA_NO_NODE) {
2821 * global hstate attribute
2823 if (!(obey_mempolicy &&
2824 init_nodemask_of_mempolicy(&nodes_allowed)))
2825 n_mask = &node_states[N_MEMORY];
2827 n_mask = &nodes_allowed;
2830 * Node specific request. count adjustment happens in
2831 * set_max_huge_pages() after acquiring hugetlb_lock.
2833 init_nodemask_of_node(&nodes_allowed, nid);
2834 n_mask = &nodes_allowed;
2837 err = set_max_huge_pages(h, count, nid, n_mask);
2839 return err ? err : len;
2842 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2843 struct kobject *kobj, const char *buf,
2847 unsigned long count;
2851 err = kstrtoul(buf, 10, &count);
2855 h = kobj_to_hstate(kobj, &nid);
2856 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2859 static ssize_t nr_hugepages_show(struct kobject *kobj,
2860 struct kobj_attribute *attr, char *buf)
2862 return nr_hugepages_show_common(kobj, attr, buf);
2865 static ssize_t nr_hugepages_store(struct kobject *kobj,
2866 struct kobj_attribute *attr, const char *buf, size_t len)
2868 return nr_hugepages_store_common(false, kobj, buf, len);
2870 HSTATE_ATTR(nr_hugepages);
2875 * hstate attribute for optionally mempolicy-based constraint on persistent
2876 * huge page alloc/free.
2878 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2879 struct kobj_attribute *attr,
2882 return nr_hugepages_show_common(kobj, attr, buf);
2885 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2886 struct kobj_attribute *attr, const char *buf, size_t len)
2888 return nr_hugepages_store_common(true, kobj, buf, len);
2890 HSTATE_ATTR(nr_hugepages_mempolicy);
2894 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2895 struct kobj_attribute *attr, char *buf)
2897 struct hstate *h = kobj_to_hstate(kobj, NULL);
2898 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
2901 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2902 struct kobj_attribute *attr, const char *buf, size_t count)
2905 unsigned long input;
2906 struct hstate *h = kobj_to_hstate(kobj, NULL);
2908 if (hstate_is_gigantic(h))
2911 err = kstrtoul(buf, 10, &input);
2915 spin_lock_irq(&hugetlb_lock);
2916 h->nr_overcommit_huge_pages = input;
2917 spin_unlock_irq(&hugetlb_lock);
2921 HSTATE_ATTR(nr_overcommit_hugepages);
2923 static ssize_t free_hugepages_show(struct kobject *kobj,
2924 struct kobj_attribute *attr, char *buf)
2927 unsigned long free_huge_pages;
2930 h = kobj_to_hstate(kobj, &nid);
2931 if (nid == NUMA_NO_NODE)
2932 free_huge_pages = h->free_huge_pages;
2934 free_huge_pages = h->free_huge_pages_node[nid];
2936 return sysfs_emit(buf, "%lu\n", free_huge_pages);
2938 HSTATE_ATTR_RO(free_hugepages);
2940 static ssize_t resv_hugepages_show(struct kobject *kobj,
2941 struct kobj_attribute *attr, char *buf)
2943 struct hstate *h = kobj_to_hstate(kobj, NULL);
2944 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
2946 HSTATE_ATTR_RO(resv_hugepages);
2948 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2949 struct kobj_attribute *attr, char *buf)
2952 unsigned long surplus_huge_pages;
2955 h = kobj_to_hstate(kobj, &nid);
2956 if (nid == NUMA_NO_NODE)
2957 surplus_huge_pages = h->surplus_huge_pages;
2959 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2961 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
2963 HSTATE_ATTR_RO(surplus_hugepages);
2965 static struct attribute *hstate_attrs[] = {
2966 &nr_hugepages_attr.attr,
2967 &nr_overcommit_hugepages_attr.attr,
2968 &free_hugepages_attr.attr,
2969 &resv_hugepages_attr.attr,
2970 &surplus_hugepages_attr.attr,
2972 &nr_hugepages_mempolicy_attr.attr,
2977 static const struct attribute_group hstate_attr_group = {
2978 .attrs = hstate_attrs,
2981 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2982 struct kobject **hstate_kobjs,
2983 const struct attribute_group *hstate_attr_group)
2986 int hi = hstate_index(h);
2988 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2989 if (!hstate_kobjs[hi])
2992 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2994 kobject_put(hstate_kobjs[hi]);
2995 hstate_kobjs[hi] = NULL;
3001 static void __init hugetlb_sysfs_init(void)
3006 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3007 if (!hugepages_kobj)
3010 for_each_hstate(h) {
3011 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3012 hstate_kobjs, &hstate_attr_group);
3014 pr_err("HugeTLB: Unable to add hstate %s", h->name);
3021 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3022 * with node devices in node_devices[] using a parallel array. The array
3023 * index of a node device or _hstate == node id.
3024 * This is here to avoid any static dependency of the node device driver, in
3025 * the base kernel, on the hugetlb module.
3027 struct node_hstate {
3028 struct kobject *hugepages_kobj;
3029 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3031 static struct node_hstate node_hstates[MAX_NUMNODES];
3034 * A subset of global hstate attributes for node devices
3036 static struct attribute *per_node_hstate_attrs[] = {
3037 &nr_hugepages_attr.attr,
3038 &free_hugepages_attr.attr,
3039 &surplus_hugepages_attr.attr,
3043 static const struct attribute_group per_node_hstate_attr_group = {
3044 .attrs = per_node_hstate_attrs,
3048 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3049 * Returns node id via non-NULL nidp.
3051 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3055 for (nid = 0; nid < nr_node_ids; nid++) {
3056 struct node_hstate *nhs = &node_hstates[nid];
3058 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3059 if (nhs->hstate_kobjs[i] == kobj) {
3071 * Unregister hstate attributes from a single node device.
3072 * No-op if no hstate attributes attached.
3074 static void hugetlb_unregister_node(struct node *node)
3077 struct node_hstate *nhs = &node_hstates[node->dev.id];
3079 if (!nhs->hugepages_kobj)
3080 return; /* no hstate attributes */
3082 for_each_hstate(h) {
3083 int idx = hstate_index(h);
3084 if (nhs->hstate_kobjs[idx]) {
3085 kobject_put(nhs->hstate_kobjs[idx]);
3086 nhs->hstate_kobjs[idx] = NULL;
3090 kobject_put(nhs->hugepages_kobj);
3091 nhs->hugepages_kobj = NULL;
3096 * Register hstate attributes for a single node device.
3097 * No-op if attributes already registered.
3099 static void hugetlb_register_node(struct node *node)
3102 struct node_hstate *nhs = &node_hstates[node->dev.id];
3105 if (nhs->hugepages_kobj)
3106 return; /* already allocated */
3108 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3110 if (!nhs->hugepages_kobj)
3113 for_each_hstate(h) {
3114 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3116 &per_node_hstate_attr_group);
3118 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3119 h->name, node->dev.id);
3120 hugetlb_unregister_node(node);
3127 * hugetlb init time: register hstate attributes for all registered node
3128 * devices of nodes that have memory. All on-line nodes should have
3129 * registered their associated device by this time.
3131 static void __init hugetlb_register_all_nodes(void)
3135 for_each_node_state(nid, N_MEMORY) {
3136 struct node *node = node_devices[nid];
3137 if (node->dev.id == nid)
3138 hugetlb_register_node(node);
3142 * Let the node device driver know we're here so it can
3143 * [un]register hstate attributes on node hotplug.
3145 register_hugetlbfs_with_node(hugetlb_register_node,
3146 hugetlb_unregister_node);
3148 #else /* !CONFIG_NUMA */
3150 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3158 static void hugetlb_register_all_nodes(void) { }
3162 static int __init hugetlb_init(void)
3166 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
3169 if (!hugepages_supported()) {
3170 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3171 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3176 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3177 * architectures depend on setup being done here.
3179 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3180 if (!parsed_default_hugepagesz) {
3182 * If we did not parse a default huge page size, set
3183 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3184 * number of huge pages for this default size was implicitly
3185 * specified, set that here as well.
3186 * Note that the implicit setting will overwrite an explicit
3187 * setting. A warning will be printed in this case.
3189 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3190 if (default_hstate_max_huge_pages) {
3191 if (default_hstate.max_huge_pages) {
3194 string_get_size(huge_page_size(&default_hstate),
3195 1, STRING_UNITS_2, buf, 32);
3196 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3197 default_hstate.max_huge_pages, buf);
3198 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3199 default_hstate_max_huge_pages);
3201 default_hstate.max_huge_pages =
3202 default_hstate_max_huge_pages;
3206 hugetlb_cma_check();
3207 hugetlb_init_hstates();
3208 gather_bootmem_prealloc();
3211 hugetlb_sysfs_init();
3212 hugetlb_register_all_nodes();
3213 hugetlb_cgroup_file_init();
3216 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3218 num_fault_mutexes = 1;
3220 hugetlb_fault_mutex_table =
3221 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3223 BUG_ON(!hugetlb_fault_mutex_table);
3225 for (i = 0; i < num_fault_mutexes; i++)
3226 mutex_init(&hugetlb_fault_mutex_table[i]);
3229 subsys_initcall(hugetlb_init);
3231 /* Overwritten by architectures with more huge page sizes */
3232 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3234 return size == HPAGE_SIZE;
3237 void __init hugetlb_add_hstate(unsigned int order)
3242 if (size_to_hstate(PAGE_SIZE << order)) {
3245 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3247 h = &hstates[hugetlb_max_hstate++];
3248 mutex_init(&h->resize_lock);
3250 h->mask = ~(huge_page_size(h) - 1);
3251 for (i = 0; i < MAX_NUMNODES; ++i)
3252 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3253 INIT_LIST_HEAD(&h->hugepage_activelist);
3254 h->next_nid_to_alloc = first_memory_node;
3255 h->next_nid_to_free = first_memory_node;
3256 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3257 huge_page_size(h)/1024);
3263 * hugepages command line processing
3264 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3265 * specification. If not, ignore the hugepages value. hugepages can also
3266 * be the first huge page command line option in which case it implicitly
3267 * specifies the number of huge pages for the default size.
3269 static int __init hugepages_setup(char *s)
3272 static unsigned long *last_mhp;
3274 if (!parsed_valid_hugepagesz) {
3275 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3276 parsed_valid_hugepagesz = true;
3281 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3282 * yet, so this hugepages= parameter goes to the "default hstate".
3283 * Otherwise, it goes with the previously parsed hugepagesz or
3284 * default_hugepagesz.
3286 else if (!hugetlb_max_hstate)
3287 mhp = &default_hstate_max_huge_pages;
3289 mhp = &parsed_hstate->max_huge_pages;
3291 if (mhp == last_mhp) {
3292 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3296 if (sscanf(s, "%lu", mhp) <= 0)
3300 * Global state is always initialized later in hugetlb_init.
3301 * But we need to allocate gigantic hstates here early to still
3302 * use the bootmem allocator.
3304 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
3305 hugetlb_hstate_alloc_pages(parsed_hstate);
3311 __setup("hugepages=", hugepages_setup);
3314 * hugepagesz command line processing
3315 * A specific huge page size can only be specified once with hugepagesz.
3316 * hugepagesz is followed by hugepages on the command line. The global
3317 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3318 * hugepagesz argument was valid.
3320 static int __init hugepagesz_setup(char *s)
3325 parsed_valid_hugepagesz = false;
3326 size = (unsigned long)memparse(s, NULL);
3328 if (!arch_hugetlb_valid_size(size)) {
3329 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3333 h = size_to_hstate(size);
3336 * hstate for this size already exists. This is normally
3337 * an error, but is allowed if the existing hstate is the
3338 * default hstate. More specifically, it is only allowed if
3339 * the number of huge pages for the default hstate was not
3340 * previously specified.
3342 if (!parsed_default_hugepagesz || h != &default_hstate ||
3343 default_hstate.max_huge_pages) {
3344 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3349 * No need to call hugetlb_add_hstate() as hstate already
3350 * exists. But, do set parsed_hstate so that a following
3351 * hugepages= parameter will be applied to this hstate.
3354 parsed_valid_hugepagesz = true;
3358 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3359 parsed_valid_hugepagesz = true;
3362 __setup("hugepagesz=", hugepagesz_setup);
3365 * default_hugepagesz command line input
3366 * Only one instance of default_hugepagesz allowed on command line.
3368 static int __init default_hugepagesz_setup(char *s)
3372 parsed_valid_hugepagesz = false;
3373 if (parsed_default_hugepagesz) {
3374 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3378 size = (unsigned long)memparse(s, NULL);
3380 if (!arch_hugetlb_valid_size(size)) {
3381 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3385 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3386 parsed_valid_hugepagesz = true;
3387 parsed_default_hugepagesz = true;
3388 default_hstate_idx = hstate_index(size_to_hstate(size));
3391 * The number of default huge pages (for this size) could have been
3392 * specified as the first hugetlb parameter: hugepages=X. If so,
3393 * then default_hstate_max_huge_pages is set. If the default huge
3394 * page size is gigantic (>= MAX_ORDER), then the pages must be
3395 * allocated here from bootmem allocator.
3397 if (default_hstate_max_huge_pages) {
3398 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3399 if (hstate_is_gigantic(&default_hstate))
3400 hugetlb_hstate_alloc_pages(&default_hstate);
3401 default_hstate_max_huge_pages = 0;
3406 __setup("default_hugepagesz=", default_hugepagesz_setup);
3408 static unsigned int allowed_mems_nr(struct hstate *h)
3411 unsigned int nr = 0;
3412 nodemask_t *mpol_allowed;
3413 unsigned int *array = h->free_huge_pages_node;
3414 gfp_t gfp_mask = htlb_alloc_mask(h);
3416 mpol_allowed = policy_nodemask_current(gfp_mask);
3418 for_each_node_mask(node, cpuset_current_mems_allowed) {
3419 if (!mpol_allowed || node_isset(node, *mpol_allowed))
3426 #ifdef CONFIG_SYSCTL
3427 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3428 void *buffer, size_t *length,
3429 loff_t *ppos, unsigned long *out)
3431 struct ctl_table dup_table;
3434 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3435 * can duplicate the @table and alter the duplicate of it.
3438 dup_table.data = out;
3440 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3443 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3444 struct ctl_table *table, int write,
3445 void *buffer, size_t *length, loff_t *ppos)
3447 struct hstate *h = &default_hstate;
3448 unsigned long tmp = h->max_huge_pages;
3451 if (!hugepages_supported())
3454 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3460 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3461 NUMA_NO_NODE, tmp, *length);
3466 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3467 void *buffer, size_t *length, loff_t *ppos)
3470 return hugetlb_sysctl_handler_common(false, table, write,
3471 buffer, length, ppos);
3475 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3476 void *buffer, size_t *length, loff_t *ppos)
3478 return hugetlb_sysctl_handler_common(true, table, write,
3479 buffer, length, ppos);
3481 #endif /* CONFIG_NUMA */
3483 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3484 void *buffer, size_t *length, loff_t *ppos)
3486 struct hstate *h = &default_hstate;
3490 if (!hugepages_supported())
3493 tmp = h->nr_overcommit_huge_pages;
3495 if (write && hstate_is_gigantic(h))
3498 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3504 spin_lock_irq(&hugetlb_lock);
3505 h->nr_overcommit_huge_pages = tmp;
3506 spin_unlock_irq(&hugetlb_lock);
3512 #endif /* CONFIG_SYSCTL */
3514 void hugetlb_report_meminfo(struct seq_file *m)
3517 unsigned long total = 0;
3519 if (!hugepages_supported())
3522 for_each_hstate(h) {
3523 unsigned long count = h->nr_huge_pages;
3525 total += huge_page_size(h) * count;
3527 if (h == &default_hstate)
3529 "HugePages_Total: %5lu\n"
3530 "HugePages_Free: %5lu\n"
3531 "HugePages_Rsvd: %5lu\n"
3532 "HugePages_Surp: %5lu\n"
3533 "Hugepagesize: %8lu kB\n",
3537 h->surplus_huge_pages,
3538 huge_page_size(h) / SZ_1K);
3541 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
3544 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
3546 struct hstate *h = &default_hstate;
3548 if (!hugepages_supported())
3551 return sysfs_emit_at(buf, len,
3552 "Node %d HugePages_Total: %5u\n"
3553 "Node %d HugePages_Free: %5u\n"
3554 "Node %d HugePages_Surp: %5u\n",
3555 nid, h->nr_huge_pages_node[nid],
3556 nid, h->free_huge_pages_node[nid],
3557 nid, h->surplus_huge_pages_node[nid]);
3560 void hugetlb_show_meminfo(void)
3565 if (!hugepages_supported())
3568 for_each_node_state(nid, N_MEMORY)
3570 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3572 h->nr_huge_pages_node[nid],
3573 h->free_huge_pages_node[nid],
3574 h->surplus_huge_pages_node[nid],
3575 huge_page_size(h) / SZ_1K);
3578 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3580 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3581 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3584 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3585 unsigned long hugetlb_total_pages(void)
3588 unsigned long nr_total_pages = 0;
3591 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3592 return nr_total_pages;
3595 static int hugetlb_acct_memory(struct hstate *h, long delta)
3602 spin_lock_irq(&hugetlb_lock);
3604 * When cpuset is configured, it breaks the strict hugetlb page
3605 * reservation as the accounting is done on a global variable. Such
3606 * reservation is completely rubbish in the presence of cpuset because
3607 * the reservation is not checked against page availability for the
3608 * current cpuset. Application can still potentially OOM'ed by kernel
3609 * with lack of free htlb page in cpuset that the task is in.
3610 * Attempt to enforce strict accounting with cpuset is almost
3611 * impossible (or too ugly) because cpuset is too fluid that
3612 * task or memory node can be dynamically moved between cpusets.
3614 * The change of semantics for shared hugetlb mapping with cpuset is
3615 * undesirable. However, in order to preserve some of the semantics,
3616 * we fall back to check against current free page availability as
3617 * a best attempt and hopefully to minimize the impact of changing
3618 * semantics that cpuset has.
3620 * Apart from cpuset, we also have memory policy mechanism that
3621 * also determines from which node the kernel will allocate memory
3622 * in a NUMA system. So similar to cpuset, we also should consider
3623 * the memory policy of the current task. Similar to the description
3627 if (gather_surplus_pages(h, delta) < 0)
3630 if (delta > allowed_mems_nr(h)) {
3631 return_unused_surplus_pages(h, delta);
3638 return_unused_surplus_pages(h, (unsigned long) -delta);
3641 spin_unlock_irq(&hugetlb_lock);
3645 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3647 struct resv_map *resv = vma_resv_map(vma);
3650 * This new VMA should share its siblings reservation map if present.
3651 * The VMA will only ever have a valid reservation map pointer where
3652 * it is being copied for another still existing VMA. As that VMA
3653 * has a reference to the reservation map it cannot disappear until
3654 * after this open call completes. It is therefore safe to take a
3655 * new reference here without additional locking.
3657 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3658 kref_get(&resv->refs);
3661 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3663 struct hstate *h = hstate_vma(vma);
3664 struct resv_map *resv = vma_resv_map(vma);
3665 struct hugepage_subpool *spool = subpool_vma(vma);
3666 unsigned long reserve, start, end;
3669 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3672 start = vma_hugecache_offset(h, vma, vma->vm_start);
3673 end = vma_hugecache_offset(h, vma, vma->vm_end);
3675 reserve = (end - start) - region_count(resv, start, end);
3676 hugetlb_cgroup_uncharge_counter(resv, start, end);
3679 * Decrement reserve counts. The global reserve count may be
3680 * adjusted if the subpool has a minimum size.
3682 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3683 hugetlb_acct_memory(h, -gbl_reserve);
3686 kref_put(&resv->refs, resv_map_release);
3689 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3691 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3696 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3698 return huge_page_size(hstate_vma(vma));
3702 * We cannot handle pagefaults against hugetlb pages at all. They cause
3703 * handle_mm_fault() to try to instantiate regular-sized pages in the
3704 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3707 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3714 * When a new function is introduced to vm_operations_struct and added
3715 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3716 * This is because under System V memory model, mappings created via
3717 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3718 * their original vm_ops are overwritten with shm_vm_ops.
3720 const struct vm_operations_struct hugetlb_vm_ops = {
3721 .fault = hugetlb_vm_op_fault,
3722 .open = hugetlb_vm_op_open,
3723 .close = hugetlb_vm_op_close,
3724 .may_split = hugetlb_vm_op_split,
3725 .pagesize = hugetlb_vm_op_pagesize,
3728 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3734 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3735 vma->vm_page_prot)));
3737 entry = huge_pte_wrprotect(mk_huge_pte(page,
3738 vma->vm_page_prot));
3740 entry = pte_mkyoung(entry);
3741 entry = pte_mkhuge(entry);
3742 entry = arch_make_huge_pte(entry, vma, page, writable);
3747 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3748 unsigned long address, pte_t *ptep)
3752 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3753 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3754 update_mmu_cache(vma, address, ptep);
3757 bool is_hugetlb_entry_migration(pte_t pte)
3761 if (huge_pte_none(pte) || pte_present(pte))
3763 swp = pte_to_swp_entry(pte);
3764 if (is_migration_entry(swp))
3770 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
3774 if (huge_pte_none(pte) || pte_present(pte))
3776 swp = pte_to_swp_entry(pte);
3777 if (is_hwpoison_entry(swp))
3784 hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
3785 struct page *new_page)
3787 __SetPageUptodate(new_page);
3788 set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
3789 hugepage_add_new_anon_rmap(new_page, vma, addr);
3790 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
3791 ClearHPageRestoreReserve(new_page);
3792 SetHPageMigratable(new_page);
3795 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3796 struct vm_area_struct *vma)
3798 pte_t *src_pte, *dst_pte, entry, dst_entry;
3799 struct page *ptepage;
3801 bool cow = is_cow_mapping(vma->vm_flags);
3802 struct hstate *h = hstate_vma(vma);
3803 unsigned long sz = huge_page_size(h);
3804 unsigned long npages = pages_per_huge_page(h);
3805 struct address_space *mapping = vma->vm_file->f_mapping;
3806 struct mmu_notifier_range range;
3810 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3813 mmu_notifier_invalidate_range_start(&range);
3816 * For shared mappings i_mmap_rwsem must be held to call
3817 * huge_pte_alloc, otherwise the returned ptep could go
3818 * away if part of a shared pmd and another thread calls
3821 i_mmap_lock_read(mapping);
3824 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3825 spinlock_t *src_ptl, *dst_ptl;
3826 src_pte = huge_pte_offset(src, addr, sz);
3829 dst_pte = huge_pte_alloc(dst, vma, addr, sz);
3836 * If the pagetables are shared don't copy or take references.
3837 * dst_pte == src_pte is the common case of src/dest sharing.
3839 * However, src could have 'unshared' and dst shares with
3840 * another vma. If dst_pte !none, this implies sharing.
3841 * Check here before taking page table lock, and once again
3842 * after taking the lock below.
3844 dst_entry = huge_ptep_get(dst_pte);
3845 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3848 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3849 src_ptl = huge_pte_lockptr(h, src, src_pte);
3850 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3851 entry = huge_ptep_get(src_pte);
3852 dst_entry = huge_ptep_get(dst_pte);
3854 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3856 * Skip if src entry none. Also, skip in the
3857 * unlikely case dst entry !none as this implies
3858 * sharing with another vma.
3861 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3862 is_hugetlb_entry_hwpoisoned(entry))) {
3863 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3865 if (is_write_migration_entry(swp_entry) && cow) {
3867 * COW mappings require pages in both
3868 * parent and child to be set to read.
3870 make_migration_entry_read(&swp_entry);
3871 entry = swp_entry_to_pte(swp_entry);
3872 set_huge_swap_pte_at(src, addr, src_pte,
3875 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3877 entry = huge_ptep_get(src_pte);
3878 ptepage = pte_page(entry);
3882 * This is a rare case where we see pinned hugetlb
3883 * pages while they're prone to COW. We need to do the
3884 * COW earlier during fork.
3886 * When pre-allocating the page or copying data, we
3887 * need to be without the pgtable locks since we could
3888 * sleep during the process.
3890 if (unlikely(page_needs_cow_for_dma(vma, ptepage))) {
3891 pte_t src_pte_old = entry;
3894 spin_unlock(src_ptl);
3895 spin_unlock(dst_ptl);
3896 /* Do not use reserve as it's private owned */
3897 new = alloc_huge_page(vma, addr, 1);
3903 copy_user_huge_page(new, ptepage, addr, vma,
3907 /* Install the new huge page if src pte stable */
3908 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3909 src_ptl = huge_pte_lockptr(h, src, src_pte);
3910 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3911 entry = huge_ptep_get(src_pte);
3912 if (!pte_same(src_pte_old, entry)) {
3914 /* dst_entry won't change as in child */
3917 hugetlb_install_page(vma, dst_pte, addr, new);
3918 spin_unlock(src_ptl);
3919 spin_unlock(dst_ptl);
3925 * No need to notify as we are downgrading page
3926 * table protection not changing it to point
3929 * See Documentation/vm/mmu_notifier.rst
3931 huge_ptep_set_wrprotect(src, addr, src_pte);
3934 page_dup_rmap(ptepage, true);
3935 set_huge_pte_at(dst, addr, dst_pte, entry);
3936 hugetlb_count_add(npages, dst);
3938 spin_unlock(src_ptl);
3939 spin_unlock(dst_ptl);
3943 mmu_notifier_invalidate_range_end(&range);
3945 i_mmap_unlock_read(mapping);
3950 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3951 unsigned long start, unsigned long end,
3952 struct page *ref_page)
3954 struct mm_struct *mm = vma->vm_mm;
3955 unsigned long address;
3960 struct hstate *h = hstate_vma(vma);
3961 unsigned long sz = huge_page_size(h);
3962 struct mmu_notifier_range range;
3964 WARN_ON(!is_vm_hugetlb_page(vma));
3965 BUG_ON(start & ~huge_page_mask(h));
3966 BUG_ON(end & ~huge_page_mask(h));
3969 * This is a hugetlb vma, all the pte entries should point
3972 tlb_change_page_size(tlb, sz);
3973 tlb_start_vma(tlb, vma);
3976 * If sharing possible, alert mmu notifiers of worst case.
3978 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3980 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3981 mmu_notifier_invalidate_range_start(&range);
3983 for (; address < end; address += sz) {
3984 ptep = huge_pte_offset(mm, address, sz);
3988 ptl = huge_pte_lock(h, mm, ptep);
3989 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
3992 * We just unmapped a page of PMDs by clearing a PUD.
3993 * The caller's TLB flush range should cover this area.
3998 pte = huge_ptep_get(ptep);
3999 if (huge_pte_none(pte)) {
4005 * Migrating hugepage or HWPoisoned hugepage is already
4006 * unmapped and its refcount is dropped, so just clear pte here.
4008 if (unlikely(!pte_present(pte))) {
4009 huge_pte_clear(mm, address, ptep, sz);
4014 page = pte_page(pte);
4016 * If a reference page is supplied, it is because a specific
4017 * page is being unmapped, not a range. Ensure the page we
4018 * are about to unmap is the actual page of interest.
4021 if (page != ref_page) {
4026 * Mark the VMA as having unmapped its page so that
4027 * future faults in this VMA will fail rather than
4028 * looking like data was lost
4030 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
4033 pte = huge_ptep_get_and_clear(mm, address, ptep);
4034 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
4035 if (huge_pte_dirty(pte))
4036 set_page_dirty(page);
4038 hugetlb_count_sub(pages_per_huge_page(h), mm);
4039 page_remove_rmap(page, true);
4042 tlb_remove_page_size(tlb, page, huge_page_size(h));
4044 * Bail out after unmapping reference page if supplied
4049 mmu_notifier_invalidate_range_end(&range);
4050 tlb_end_vma(tlb, vma);
4053 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4054 struct vm_area_struct *vma, unsigned long start,
4055 unsigned long end, struct page *ref_page)
4057 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4060 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4061 * test will fail on a vma being torn down, and not grab a page table
4062 * on its way out. We're lucky that the flag has such an appropriate
4063 * name, and can in fact be safely cleared here. We could clear it
4064 * before the __unmap_hugepage_range above, but all that's necessary
4065 * is to clear it before releasing the i_mmap_rwsem. This works
4066 * because in the context this is called, the VMA is about to be
4067 * destroyed and the i_mmap_rwsem is held.
4069 vma->vm_flags &= ~VM_MAYSHARE;
4072 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
4073 unsigned long end, struct page *ref_page)
4075 struct mmu_gather tlb;
4077 tlb_gather_mmu(&tlb, vma->vm_mm);
4078 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
4079 tlb_finish_mmu(&tlb);
4083 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4084 * mapping it owns the reserve page for. The intention is to unmap the page
4085 * from other VMAs and let the children be SIGKILLed if they are faulting the
4088 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4089 struct page *page, unsigned long address)
4091 struct hstate *h = hstate_vma(vma);
4092 struct vm_area_struct *iter_vma;
4093 struct address_space *mapping;
4097 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4098 * from page cache lookup which is in HPAGE_SIZE units.
4100 address = address & huge_page_mask(h);
4101 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4103 mapping = vma->vm_file->f_mapping;
4106 * Take the mapping lock for the duration of the table walk. As
4107 * this mapping should be shared between all the VMAs,
4108 * __unmap_hugepage_range() is called as the lock is already held
4110 i_mmap_lock_write(mapping);
4111 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
4112 /* Do not unmap the current VMA */
4113 if (iter_vma == vma)
4117 * Shared VMAs have their own reserves and do not affect
4118 * MAP_PRIVATE accounting but it is possible that a shared
4119 * VMA is using the same page so check and skip such VMAs.
4121 if (iter_vma->vm_flags & VM_MAYSHARE)
4125 * Unmap the page from other VMAs without their own reserves.
4126 * They get marked to be SIGKILLed if they fault in these
4127 * areas. This is because a future no-page fault on this VMA
4128 * could insert a zeroed page instead of the data existing
4129 * from the time of fork. This would look like data corruption
4131 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
4132 unmap_hugepage_range(iter_vma, address,
4133 address + huge_page_size(h), page);
4135 i_mmap_unlock_write(mapping);
4139 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4140 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4141 * cannot race with other handlers or page migration.
4142 * Keep the pte_same checks anyway to make transition from the mutex easier.
4144 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
4145 unsigned long address, pte_t *ptep,
4146 struct page *pagecache_page, spinlock_t *ptl)
4149 struct hstate *h = hstate_vma(vma);
4150 struct page *old_page, *new_page;
4151 int outside_reserve = 0;
4153 unsigned long haddr = address & huge_page_mask(h);
4154 struct mmu_notifier_range range;
4156 pte = huge_ptep_get(ptep);
4157 old_page = pte_page(pte);
4160 /* If no-one else is actually using this page, avoid the copy
4161 * and just make the page writable */
4162 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
4163 page_move_anon_rmap(old_page, vma);
4164 set_huge_ptep_writable(vma, haddr, ptep);
4169 * If the process that created a MAP_PRIVATE mapping is about to
4170 * perform a COW due to a shared page count, attempt to satisfy
4171 * the allocation without using the existing reserves. The pagecache
4172 * page is used to determine if the reserve at this address was
4173 * consumed or not. If reserves were used, a partial faulted mapping
4174 * at the time of fork() could consume its reserves on COW instead
4175 * of the full address range.
4177 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4178 old_page != pagecache_page)
4179 outside_reserve = 1;
4184 * Drop page table lock as buddy allocator may be called. It will
4185 * be acquired again before returning to the caller, as expected.
4188 new_page = alloc_huge_page(vma, haddr, outside_reserve);
4190 if (IS_ERR(new_page)) {
4192 * If a process owning a MAP_PRIVATE mapping fails to COW,
4193 * it is due to references held by a child and an insufficient
4194 * huge page pool. To guarantee the original mappers
4195 * reliability, unmap the page from child processes. The child
4196 * may get SIGKILLed if it later faults.
4198 if (outside_reserve) {
4199 struct address_space *mapping = vma->vm_file->f_mapping;
4204 BUG_ON(huge_pte_none(pte));
4206 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4207 * unmapping. unmapping needs to hold i_mmap_rwsem
4208 * in write mode. Dropping i_mmap_rwsem in read mode
4209 * here is OK as COW mappings do not interact with
4212 * Reacquire both after unmap operation.
4214 idx = vma_hugecache_offset(h, vma, haddr);
4215 hash = hugetlb_fault_mutex_hash(mapping, idx);
4216 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4217 i_mmap_unlock_read(mapping);
4219 unmap_ref_private(mm, vma, old_page, haddr);
4221 i_mmap_lock_read(mapping);
4222 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4224 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4226 pte_same(huge_ptep_get(ptep), pte)))
4227 goto retry_avoidcopy;
4229 * race occurs while re-acquiring page table
4230 * lock, and our job is done.
4235 ret = vmf_error(PTR_ERR(new_page));
4236 goto out_release_old;
4240 * When the original hugepage is shared one, it does not have
4241 * anon_vma prepared.
4243 if (unlikely(anon_vma_prepare(vma))) {
4245 goto out_release_all;
4248 copy_user_huge_page(new_page, old_page, address, vma,
4249 pages_per_huge_page(h));
4250 __SetPageUptodate(new_page);
4252 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4253 haddr + huge_page_size(h));
4254 mmu_notifier_invalidate_range_start(&range);
4257 * Retake the page table lock to check for racing updates
4258 * before the page tables are altered
4261 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4262 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4263 ClearHPageRestoreReserve(new_page);
4266 huge_ptep_clear_flush(vma, haddr, ptep);
4267 mmu_notifier_invalidate_range(mm, range.start, range.end);
4268 set_huge_pte_at(mm, haddr, ptep,
4269 make_huge_pte(vma, new_page, 1));
4270 page_remove_rmap(old_page, true);
4271 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4272 SetHPageMigratable(new_page);
4273 /* Make the old page be freed below */
4274 new_page = old_page;
4277 mmu_notifier_invalidate_range_end(&range);
4279 restore_reserve_on_error(h, vma, haddr, new_page);
4284 spin_lock(ptl); /* Caller expects lock to be held */
4288 /* Return the pagecache page at a given address within a VMA */
4289 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4290 struct vm_area_struct *vma, unsigned long address)
4292 struct address_space *mapping;
4295 mapping = vma->vm_file->f_mapping;
4296 idx = vma_hugecache_offset(h, vma, address);
4298 return find_lock_page(mapping, idx);
4302 * Return whether there is a pagecache page to back given address within VMA.
4303 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4305 static bool hugetlbfs_pagecache_present(struct hstate *h,
4306 struct vm_area_struct *vma, unsigned long address)
4308 struct address_space *mapping;
4312 mapping = vma->vm_file->f_mapping;
4313 idx = vma_hugecache_offset(h, vma, address);
4315 page = find_get_page(mapping, idx);
4318 return page != NULL;
4321 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4324 struct inode *inode = mapping->host;
4325 struct hstate *h = hstate_inode(inode);
4326 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4330 ClearHPageRestoreReserve(page);
4333 * set page dirty so that it will not be removed from cache/file
4334 * by non-hugetlbfs specific code paths.
4336 set_page_dirty(page);
4338 spin_lock(&inode->i_lock);
4339 inode->i_blocks += blocks_per_huge_page(h);
4340 spin_unlock(&inode->i_lock);
4344 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4345 struct vm_area_struct *vma,
4346 struct address_space *mapping, pgoff_t idx,
4347 unsigned long address, pte_t *ptep, unsigned int flags)
4349 struct hstate *h = hstate_vma(vma);
4350 vm_fault_t ret = VM_FAULT_SIGBUS;
4356 unsigned long haddr = address & huge_page_mask(h);
4357 bool new_page = false;
4360 * Currently, we are forced to kill the process in the event the
4361 * original mapper has unmapped pages from the child due to a failed
4362 * COW. Warn that such a situation has occurred as it may not be obvious
4364 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4365 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4371 * We can not race with truncation due to holding i_mmap_rwsem.
4372 * i_size is modified when holding i_mmap_rwsem, so check here
4373 * once for faults beyond end of file.
4375 size = i_size_read(mapping->host) >> huge_page_shift(h);
4380 page = find_lock_page(mapping, idx);
4383 * Check for page in userfault range
4385 if (userfaultfd_missing(vma)) {
4387 struct vm_fault vmf = {
4392 * Hard to debug if it ends up being
4393 * used by a callee that assumes
4394 * something about the other
4395 * uninitialized fields... same as in
4401 * hugetlb_fault_mutex and i_mmap_rwsem must be
4402 * dropped before handling userfault. Reacquire
4403 * after handling fault to make calling code simpler.
4405 hash = hugetlb_fault_mutex_hash(mapping, idx);
4406 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4407 i_mmap_unlock_read(mapping);
4408 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
4409 i_mmap_lock_read(mapping);
4410 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4414 page = alloc_huge_page(vma, haddr, 0);
4417 * Returning error will result in faulting task being
4418 * sent SIGBUS. The hugetlb fault mutex prevents two
4419 * tasks from racing to fault in the same page which
4420 * could result in false unable to allocate errors.
4421 * Page migration does not take the fault mutex, but
4422 * does a clear then write of pte's under page table
4423 * lock. Page fault code could race with migration,
4424 * notice the clear pte and try to allocate a page
4425 * here. Before returning error, get ptl and make
4426 * sure there really is no pte entry.
4428 ptl = huge_pte_lock(h, mm, ptep);
4430 if (huge_pte_none(huge_ptep_get(ptep)))
4431 ret = vmf_error(PTR_ERR(page));
4435 clear_huge_page(page, address, pages_per_huge_page(h));
4436 __SetPageUptodate(page);
4439 if (vma->vm_flags & VM_MAYSHARE) {
4440 int err = huge_add_to_page_cache(page, mapping, idx);
4449 if (unlikely(anon_vma_prepare(vma))) {
4451 goto backout_unlocked;
4457 * If memory error occurs between mmap() and fault, some process
4458 * don't have hwpoisoned swap entry for errored virtual address.
4459 * So we need to block hugepage fault by PG_hwpoison bit check.
4461 if (unlikely(PageHWPoison(page))) {
4462 ret = VM_FAULT_HWPOISON_LARGE |
4463 VM_FAULT_SET_HINDEX(hstate_index(h));
4464 goto backout_unlocked;
4469 * If we are going to COW a private mapping later, we examine the
4470 * pending reservations for this page now. This will ensure that
4471 * any allocations necessary to record that reservation occur outside
4474 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4475 if (vma_needs_reservation(h, vma, haddr) < 0) {
4477 goto backout_unlocked;
4479 /* Just decrements count, does not deallocate */
4480 vma_end_reservation(h, vma, haddr);
4483 ptl = huge_pte_lock(h, mm, ptep);
4485 if (!huge_pte_none(huge_ptep_get(ptep)))
4489 ClearHPageRestoreReserve(page);
4490 hugepage_add_new_anon_rmap(page, vma, haddr);
4492 page_dup_rmap(page, true);
4493 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4494 && (vma->vm_flags & VM_SHARED)));
4495 set_huge_pte_at(mm, haddr, ptep, new_pte);
4497 hugetlb_count_add(pages_per_huge_page(h), mm);
4498 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4499 /* Optimization, do the COW without a second fault */
4500 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4506 * Only set HPageMigratable in newly allocated pages. Existing pages
4507 * found in the pagecache may not have HPageMigratableset if they have
4508 * been isolated for migration.
4511 SetHPageMigratable(page);
4521 restore_reserve_on_error(h, vma, haddr, page);
4527 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4529 unsigned long key[2];
4532 key[0] = (unsigned long) mapping;
4535 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4537 return hash & (num_fault_mutexes - 1);
4541 * For uniprocessor systems we always use a single mutex, so just
4542 * return 0 and avoid the hashing overhead.
4544 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4550 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4551 unsigned long address, unsigned int flags)
4558 struct page *page = NULL;
4559 struct page *pagecache_page = NULL;
4560 struct hstate *h = hstate_vma(vma);
4561 struct address_space *mapping;
4562 int need_wait_lock = 0;
4563 unsigned long haddr = address & huge_page_mask(h);
4565 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4568 * Since we hold no locks, ptep could be stale. That is
4569 * OK as we are only making decisions based on content and
4570 * not actually modifying content here.
4572 entry = huge_ptep_get(ptep);
4573 if (unlikely(is_hugetlb_entry_migration(entry))) {
4574 migration_entry_wait_huge(vma, mm, ptep);
4576 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4577 return VM_FAULT_HWPOISON_LARGE |
4578 VM_FAULT_SET_HINDEX(hstate_index(h));
4582 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4583 * until finished with ptep. This serves two purposes:
4584 * 1) It prevents huge_pmd_unshare from being called elsewhere
4585 * and making the ptep no longer valid.
4586 * 2) It synchronizes us with i_size modifications during truncation.
4588 * ptep could have already be assigned via huge_pte_offset. That
4589 * is OK, as huge_pte_alloc will return the same value unless
4590 * something has changed.
4592 mapping = vma->vm_file->f_mapping;
4593 i_mmap_lock_read(mapping);
4594 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
4596 i_mmap_unlock_read(mapping);
4597 return VM_FAULT_OOM;
4601 * Serialize hugepage allocation and instantiation, so that we don't
4602 * get spurious allocation failures if two CPUs race to instantiate
4603 * the same page in the page cache.
4605 idx = vma_hugecache_offset(h, vma, haddr);
4606 hash = hugetlb_fault_mutex_hash(mapping, idx);
4607 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4609 entry = huge_ptep_get(ptep);
4610 if (huge_pte_none(entry)) {
4611 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4618 * entry could be a migration/hwpoison entry at this point, so this
4619 * check prevents the kernel from going below assuming that we have
4620 * an active hugepage in pagecache. This goto expects the 2nd page
4621 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4622 * properly handle it.
4624 if (!pte_present(entry))
4628 * If we are going to COW the mapping later, we examine the pending
4629 * reservations for this page now. This will ensure that any
4630 * allocations necessary to record that reservation occur outside the
4631 * spinlock. For private mappings, we also lookup the pagecache
4632 * page now as it is used to determine if a reservation has been
4635 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4636 if (vma_needs_reservation(h, vma, haddr) < 0) {
4640 /* Just decrements count, does not deallocate */
4641 vma_end_reservation(h, vma, haddr);
4643 if (!(vma->vm_flags & VM_MAYSHARE))
4644 pagecache_page = hugetlbfs_pagecache_page(h,
4648 ptl = huge_pte_lock(h, mm, ptep);
4650 /* Check for a racing update before calling hugetlb_cow */
4651 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4655 * hugetlb_cow() requires page locks of pte_page(entry) and
4656 * pagecache_page, so here we need take the former one
4657 * when page != pagecache_page or !pagecache_page.
4659 page = pte_page(entry);
4660 if (page != pagecache_page)
4661 if (!trylock_page(page)) {
4668 if (flags & FAULT_FLAG_WRITE) {
4669 if (!huge_pte_write(entry)) {
4670 ret = hugetlb_cow(mm, vma, address, ptep,
4671 pagecache_page, ptl);
4674 entry = huge_pte_mkdirty(entry);
4676 entry = pte_mkyoung(entry);
4677 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4678 flags & FAULT_FLAG_WRITE))
4679 update_mmu_cache(vma, haddr, ptep);
4681 if (page != pagecache_page)
4687 if (pagecache_page) {
4688 unlock_page(pagecache_page);
4689 put_page(pagecache_page);
4692 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4693 i_mmap_unlock_read(mapping);
4695 * Generally it's safe to hold refcount during waiting page lock. But
4696 * here we just wait to defer the next page fault to avoid busy loop and
4697 * the page is not used after unlocked before returning from the current
4698 * page fault. So we are safe from accessing freed page, even if we wait
4699 * here without taking refcount.
4702 wait_on_page_locked(page);
4707 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4708 * modifications for huge pages.
4710 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4712 struct vm_area_struct *dst_vma,
4713 unsigned long dst_addr,
4714 unsigned long src_addr,
4715 struct page **pagep)
4717 struct address_space *mapping;
4720 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4721 struct hstate *h = hstate_vma(dst_vma);
4729 page = alloc_huge_page(dst_vma, dst_addr, 0);
4733 ret = copy_huge_page_from_user(page,
4734 (const void __user *) src_addr,
4735 pages_per_huge_page(h), false);
4737 /* fallback to copy_from_user outside mmap_lock */
4738 if (unlikely(ret)) {
4741 /* don't free the page */
4750 * The memory barrier inside __SetPageUptodate makes sure that
4751 * preceding stores to the page contents become visible before
4752 * the set_pte_at() write.
4754 __SetPageUptodate(page);
4756 mapping = dst_vma->vm_file->f_mapping;
4757 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4760 * If shared, add to page cache
4763 size = i_size_read(mapping->host) >> huge_page_shift(h);
4766 goto out_release_nounlock;
4769 * Serialization between remove_inode_hugepages() and
4770 * huge_add_to_page_cache() below happens through the
4771 * hugetlb_fault_mutex_table that here must be hold by
4774 ret = huge_add_to_page_cache(page, mapping, idx);
4776 goto out_release_nounlock;
4779 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4783 * Recheck the i_size after holding PT lock to make sure not
4784 * to leave any page mapped (as page_mapped()) beyond the end
4785 * of the i_size (remove_inode_hugepages() is strict about
4786 * enforcing that). If we bail out here, we'll also leave a
4787 * page in the radix tree in the vm_shared case beyond the end
4788 * of the i_size, but remove_inode_hugepages() will take care
4789 * of it as soon as we drop the hugetlb_fault_mutex_table.
4791 size = i_size_read(mapping->host) >> huge_page_shift(h);
4794 goto out_release_unlock;
4797 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4798 goto out_release_unlock;
4801 page_dup_rmap(page, true);
4803 ClearHPageRestoreReserve(page);
4804 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4807 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4808 if (dst_vma->vm_flags & VM_WRITE)
4809 _dst_pte = huge_pte_mkdirty(_dst_pte);
4810 _dst_pte = pte_mkyoung(_dst_pte);
4812 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4814 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4815 dst_vma->vm_flags & VM_WRITE);
4816 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4818 /* No need to invalidate - it was non-present before */
4819 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4822 SetHPageMigratable(page);
4832 out_release_nounlock:
4837 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
4838 int refs, struct page **pages,
4839 struct vm_area_struct **vmas)
4843 for (nr = 0; nr < refs; nr++) {
4845 pages[nr] = mem_map_offset(page, nr);
4851 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4852 struct page **pages, struct vm_area_struct **vmas,
4853 unsigned long *position, unsigned long *nr_pages,
4854 long i, unsigned int flags, int *locked)
4856 unsigned long pfn_offset;
4857 unsigned long vaddr = *position;
4858 unsigned long remainder = *nr_pages;
4859 struct hstate *h = hstate_vma(vma);
4860 int err = -EFAULT, refs;
4862 while (vaddr < vma->vm_end && remainder) {
4864 spinlock_t *ptl = NULL;
4869 * If we have a pending SIGKILL, don't keep faulting pages and
4870 * potentially allocating memory.
4872 if (fatal_signal_pending(current)) {
4878 * Some archs (sparc64, sh*) have multiple pte_ts to
4879 * each hugepage. We have to make sure we get the
4880 * first, for the page indexing below to work.
4882 * Note that page table lock is not held when pte is null.
4884 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4887 ptl = huge_pte_lock(h, mm, pte);
4888 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4891 * When coredumping, it suits get_dump_page if we just return
4892 * an error where there's an empty slot with no huge pagecache
4893 * to back it. This way, we avoid allocating a hugepage, and
4894 * the sparse dumpfile avoids allocating disk blocks, but its
4895 * huge holes still show up with zeroes where they need to be.
4897 if (absent && (flags & FOLL_DUMP) &&
4898 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4906 * We need call hugetlb_fault for both hugepages under migration
4907 * (in which case hugetlb_fault waits for the migration,) and
4908 * hwpoisoned hugepages (in which case we need to prevent the
4909 * caller from accessing to them.) In order to do this, we use
4910 * here is_swap_pte instead of is_hugetlb_entry_migration and
4911 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4912 * both cases, and because we can't follow correct pages
4913 * directly from any kind of swap entries.
4915 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4916 ((flags & FOLL_WRITE) &&
4917 !huge_pte_write(huge_ptep_get(pte)))) {
4919 unsigned int fault_flags = 0;
4923 if (flags & FOLL_WRITE)
4924 fault_flags |= FAULT_FLAG_WRITE;
4926 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4927 FAULT_FLAG_KILLABLE;
4928 if (flags & FOLL_NOWAIT)
4929 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4930 FAULT_FLAG_RETRY_NOWAIT;
4931 if (flags & FOLL_TRIED) {
4933 * Note: FAULT_FLAG_ALLOW_RETRY and
4934 * FAULT_FLAG_TRIED can co-exist
4936 fault_flags |= FAULT_FLAG_TRIED;
4938 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4939 if (ret & VM_FAULT_ERROR) {
4940 err = vm_fault_to_errno(ret, flags);
4944 if (ret & VM_FAULT_RETRY) {
4946 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4950 * VM_FAULT_RETRY must not return an
4951 * error, it will return zero
4954 * No need to update "position" as the
4955 * caller will not check it after
4956 * *nr_pages is set to 0.
4963 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4964 page = pte_page(huge_ptep_get(pte));
4967 * If subpage information not requested, update counters
4968 * and skip the same_page loop below.
4970 if (!pages && !vmas && !pfn_offset &&
4971 (vaddr + huge_page_size(h) < vma->vm_end) &&
4972 (remainder >= pages_per_huge_page(h))) {
4973 vaddr += huge_page_size(h);
4974 remainder -= pages_per_huge_page(h);
4975 i += pages_per_huge_page(h);
4980 refs = min3(pages_per_huge_page(h) - pfn_offset,
4981 (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder);
4984 record_subpages_vmas(mem_map_offset(page, pfn_offset),
4986 likely(pages) ? pages + i : NULL,
4987 vmas ? vmas + i : NULL);
4991 * try_grab_compound_head() should always succeed here,
4992 * because: a) we hold the ptl lock, and b) we've just
4993 * checked that the huge page is present in the page
4994 * tables. If the huge page is present, then the tail
4995 * pages must also be present. The ptl prevents the
4996 * head page and tail pages from being rearranged in
4997 * any way. So this page must be available at this
4998 * point, unless the page refcount overflowed:
5000 if (WARN_ON_ONCE(!try_grab_compound_head(pages[i],
5010 vaddr += (refs << PAGE_SHIFT);
5016 *nr_pages = remainder;
5018 * setting position is actually required only if remainder is
5019 * not zero but it's faster not to add a "if (remainder)"
5027 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
5028 unsigned long address, unsigned long end, pgprot_t newprot)
5030 struct mm_struct *mm = vma->vm_mm;
5031 unsigned long start = address;
5034 struct hstate *h = hstate_vma(vma);
5035 unsigned long pages = 0;
5036 bool shared_pmd = false;
5037 struct mmu_notifier_range range;
5040 * In the case of shared PMDs, the area to flush could be beyond
5041 * start/end. Set range.start/range.end to cover the maximum possible
5042 * range if PMD sharing is possible.
5044 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5045 0, vma, mm, start, end);
5046 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5048 BUG_ON(address >= end);
5049 flush_cache_range(vma, range.start, range.end);
5051 mmu_notifier_invalidate_range_start(&range);
5052 i_mmap_lock_write(vma->vm_file->f_mapping);
5053 for (; address < end; address += huge_page_size(h)) {
5055 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5058 ptl = huge_pte_lock(h, mm, ptep);
5059 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
5065 pte = huge_ptep_get(ptep);
5066 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5070 if (unlikely(is_hugetlb_entry_migration(pte))) {
5071 swp_entry_t entry = pte_to_swp_entry(pte);
5073 if (is_write_migration_entry(entry)) {
5076 make_migration_entry_read(&entry);
5077 newpte = swp_entry_to_pte(entry);
5078 set_huge_swap_pte_at(mm, address, ptep,
5079 newpte, huge_page_size(h));
5085 if (!huge_pte_none(pte)) {
5088 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5089 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
5090 pte = arch_make_huge_pte(pte, vma, NULL, 0);
5091 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
5097 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5098 * may have cleared our pud entry and done put_page on the page table:
5099 * once we release i_mmap_rwsem, another task can do the final put_page
5100 * and that page table be reused and filled with junk. If we actually
5101 * did unshare a page of pmds, flush the range corresponding to the pud.
5104 flush_hugetlb_tlb_range(vma, range.start, range.end);
5106 flush_hugetlb_tlb_range(vma, start, end);
5108 * No need to call mmu_notifier_invalidate_range() we are downgrading
5109 * page table protection not changing it to point to a new page.
5111 * See Documentation/vm/mmu_notifier.rst
5113 i_mmap_unlock_write(vma->vm_file->f_mapping);
5114 mmu_notifier_invalidate_range_end(&range);
5116 return pages << h->order;
5119 /* Return true if reservation was successful, false otherwise. */
5120 bool hugetlb_reserve_pages(struct inode *inode,
5122 struct vm_area_struct *vma,
5123 vm_flags_t vm_flags)
5126 struct hstate *h = hstate_inode(inode);
5127 struct hugepage_subpool *spool = subpool_inode(inode);
5128 struct resv_map *resv_map;
5129 struct hugetlb_cgroup *h_cg = NULL;
5130 long gbl_reserve, regions_needed = 0;
5132 /* This should never happen */
5134 VM_WARN(1, "%s called with a negative range\n", __func__);
5139 * Only apply hugepage reservation if asked. At fault time, an
5140 * attempt will be made for VM_NORESERVE to allocate a page
5141 * without using reserves
5143 if (vm_flags & VM_NORESERVE)
5147 * Shared mappings base their reservation on the number of pages that
5148 * are already allocated on behalf of the file. Private mappings need
5149 * to reserve the full area even if read-only as mprotect() may be
5150 * called to make the mapping read-write. Assume !vma is a shm mapping
5152 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5154 * resv_map can not be NULL as hugetlb_reserve_pages is only
5155 * called for inodes for which resv_maps were created (see
5156 * hugetlbfs_get_inode).
5158 resv_map = inode_resv_map(inode);
5160 chg = region_chg(resv_map, from, to, ®ions_needed);
5163 /* Private mapping. */
5164 resv_map = resv_map_alloc();
5170 set_vma_resv_map(vma, resv_map);
5171 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5177 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
5178 chg * pages_per_huge_page(h), &h_cg) < 0)
5181 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5182 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5185 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5189 * There must be enough pages in the subpool for the mapping. If
5190 * the subpool has a minimum size, there may be some global
5191 * reservations already in place (gbl_reserve).
5193 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5194 if (gbl_reserve < 0)
5195 goto out_uncharge_cgroup;
5198 * Check enough hugepages are available for the reservation.
5199 * Hand the pages back to the subpool if there are not
5201 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
5205 * Account for the reservations made. Shared mappings record regions
5206 * that have reservations as they are shared by multiple VMAs.
5207 * When the last VMA disappears, the region map says how much
5208 * the reservation was and the page cache tells how much of
5209 * the reservation was consumed. Private mappings are per-VMA and
5210 * only the consumed reservations are tracked. When the VMA
5211 * disappears, the original reservation is the VMA size and the
5212 * consumed reservations are stored in the map. Hence, nothing
5213 * else has to be done for private mappings here
5215 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5216 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5218 if (unlikely(add < 0)) {
5219 hugetlb_acct_memory(h, -gbl_reserve);
5221 } else if (unlikely(chg > add)) {
5223 * pages in this range were added to the reserve
5224 * map between region_chg and region_add. This
5225 * indicates a race with alloc_huge_page. Adjust
5226 * the subpool and reserve counts modified above
5227 * based on the difference.
5232 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
5233 * reference to h_cg->css. See comment below for detail.
5235 hugetlb_cgroup_uncharge_cgroup_rsvd(
5237 (chg - add) * pages_per_huge_page(h), h_cg);
5239 rsv_adjust = hugepage_subpool_put_pages(spool,
5241 hugetlb_acct_memory(h, -rsv_adjust);
5244 * The file_regions will hold their own reference to
5245 * h_cg->css. So we should release the reference held
5246 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
5249 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
5255 /* put back original number of pages, chg */
5256 (void)hugepage_subpool_put_pages(spool, chg);
5257 out_uncharge_cgroup:
5258 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5259 chg * pages_per_huge_page(h), h_cg);
5261 if (!vma || vma->vm_flags & VM_MAYSHARE)
5262 /* Only call region_abort if the region_chg succeeded but the
5263 * region_add failed or didn't run.
5265 if (chg >= 0 && add < 0)
5266 region_abort(resv_map, from, to, regions_needed);
5267 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5268 kref_put(&resv_map->refs, resv_map_release);
5272 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5275 struct hstate *h = hstate_inode(inode);
5276 struct resv_map *resv_map = inode_resv_map(inode);
5278 struct hugepage_subpool *spool = subpool_inode(inode);
5282 * Since this routine can be called in the evict inode path for all
5283 * hugetlbfs inodes, resv_map could be NULL.
5286 chg = region_del(resv_map, start, end);
5288 * region_del() can fail in the rare case where a region
5289 * must be split and another region descriptor can not be
5290 * allocated. If end == LONG_MAX, it will not fail.
5296 spin_lock(&inode->i_lock);
5297 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5298 spin_unlock(&inode->i_lock);
5301 * If the subpool has a minimum size, the number of global
5302 * reservations to be released may be adjusted.
5304 * Note that !resv_map implies freed == 0. So (chg - freed)
5305 * won't go negative.
5307 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5308 hugetlb_acct_memory(h, -gbl_reserve);
5313 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5314 static unsigned long page_table_shareable(struct vm_area_struct *svma,
5315 struct vm_area_struct *vma,
5316 unsigned long addr, pgoff_t idx)
5318 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5320 unsigned long sbase = saddr & PUD_MASK;
5321 unsigned long s_end = sbase + PUD_SIZE;
5323 /* Allow segments to share if only one is marked locked */
5324 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5325 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5328 * match the virtual addresses, permission and the alignment of the
5331 if (pmd_index(addr) != pmd_index(saddr) ||
5332 vm_flags != svm_flags ||
5333 !range_in_vma(svma, sbase, s_end))
5339 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5341 unsigned long base = addr & PUD_MASK;
5342 unsigned long end = base + PUD_SIZE;
5345 * check on proper vm_flags and page table alignment
5347 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5352 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5354 #ifdef CONFIG_USERFAULTFD
5355 if (uffd_disable_huge_pmd_share(vma))
5358 return vma_shareable(vma, addr);
5362 * Determine if start,end range within vma could be mapped by shared pmd.
5363 * If yes, adjust start and end to cover range associated with possible
5364 * shared pmd mappings.
5366 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5367 unsigned long *start, unsigned long *end)
5369 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
5370 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
5373 * vma need span at least one aligned PUD size and the start,end range
5374 * must at least partialy within it.
5376 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
5377 (*end <= v_start) || (*start >= v_end))
5380 /* Extend the range to be PUD aligned for a worst case scenario */
5381 if (*start > v_start)
5382 *start = ALIGN_DOWN(*start, PUD_SIZE);
5385 *end = ALIGN(*end, PUD_SIZE);
5389 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5390 * and returns the corresponding pte. While this is not necessary for the
5391 * !shared pmd case because we can allocate the pmd later as well, it makes the
5392 * code much cleaner.
5394 * This routine must be called with i_mmap_rwsem held in at least read mode if
5395 * sharing is possible. For hugetlbfs, this prevents removal of any page
5396 * table entries associated with the address space. This is important as we
5397 * are setting up sharing based on existing page table entries (mappings).
5399 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5400 * huge_pte_alloc know that sharing is not possible and do not take
5401 * i_mmap_rwsem as a performance optimization. This is handled by the
5402 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5403 * only required for subsequent processing.
5405 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5406 unsigned long addr, pud_t *pud)
5408 struct address_space *mapping = vma->vm_file->f_mapping;
5409 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5411 struct vm_area_struct *svma;
5412 unsigned long saddr;
5417 i_mmap_assert_locked(mapping);
5418 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5422 saddr = page_table_shareable(svma, vma, addr, idx);
5424 spte = huge_pte_offset(svma->vm_mm, saddr,
5425 vma_mmu_pagesize(svma));
5427 get_page(virt_to_page(spte));
5436 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
5437 if (pud_none(*pud)) {
5438 pud_populate(mm, pud,
5439 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5442 put_page(virt_to_page(spte));
5446 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5451 * unmap huge page backed by shared pte.
5453 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5454 * indicated by page_count > 1, unmap is achieved by clearing pud and
5455 * decrementing the ref count. If count == 1, the pte page is not shared.
5457 * Called with page table lock held and i_mmap_rwsem held in write mode.
5459 * returns: 1 successfully unmapped a shared pte page
5460 * 0 the underlying pte page is not shared, or it is the last user
5462 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5463 unsigned long *addr, pte_t *ptep)
5465 pgd_t *pgd = pgd_offset(mm, *addr);
5466 p4d_t *p4d = p4d_offset(pgd, *addr);
5467 pud_t *pud = pud_offset(p4d, *addr);
5469 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
5470 BUG_ON(page_count(virt_to_page(ptep)) == 0);
5471 if (page_count(virt_to_page(ptep)) == 1)
5475 put_page(virt_to_page(ptep));
5477 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5481 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5482 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5483 unsigned long addr, pud_t *pud)
5488 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5489 unsigned long *addr, pte_t *ptep)
5494 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5495 unsigned long *start, unsigned long *end)
5499 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5503 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5505 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5506 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
5507 unsigned long addr, unsigned long sz)
5514 pgd = pgd_offset(mm, addr);
5515 p4d = p4d_alloc(mm, pgd, addr);
5518 pud = pud_alloc(mm, p4d, addr);
5520 if (sz == PUD_SIZE) {
5523 BUG_ON(sz != PMD_SIZE);
5524 if (want_pmd_share(vma, addr) && pud_none(*pud))
5525 pte = huge_pmd_share(mm, vma, addr, pud);
5527 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5530 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
5536 * huge_pte_offset() - Walk the page table to resolve the hugepage
5537 * entry at address @addr
5539 * Return: Pointer to page table entry (PUD or PMD) for
5540 * address @addr, or NULL if a !p*d_present() entry is encountered and the
5541 * size @sz doesn't match the hugepage size at this level of the page
5544 pte_t *huge_pte_offset(struct mm_struct *mm,
5545 unsigned long addr, unsigned long sz)
5552 pgd = pgd_offset(mm, addr);
5553 if (!pgd_present(*pgd))
5555 p4d = p4d_offset(pgd, addr);
5556 if (!p4d_present(*p4d))
5559 pud = pud_offset(p4d, addr);
5561 /* must be pud huge, non-present or none */
5562 return (pte_t *)pud;
5563 if (!pud_present(*pud))
5565 /* must have a valid entry and size to go further */
5567 pmd = pmd_offset(pud, addr);
5568 /* must be pmd huge, non-present or none */
5569 return (pte_t *)pmd;
5572 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5575 * These functions are overwritable if your architecture needs its own
5578 struct page * __weak
5579 follow_huge_addr(struct mm_struct *mm, unsigned long address,
5582 return ERR_PTR(-EINVAL);
5585 struct page * __weak
5586 follow_huge_pd(struct vm_area_struct *vma,
5587 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5589 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5593 struct page * __weak
5594 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
5595 pmd_t *pmd, int flags)
5597 struct page *page = NULL;
5601 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5602 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
5603 (FOLL_PIN | FOLL_GET)))
5607 ptl = pmd_lockptr(mm, pmd);
5610 * make sure that the address range covered by this pmd is not
5611 * unmapped from other threads.
5613 if (!pmd_huge(*pmd))
5615 pte = huge_ptep_get((pte_t *)pmd);
5616 if (pte_present(pte)) {
5617 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
5619 * try_grab_page() should always succeed here, because: a) we
5620 * hold the pmd (ptl) lock, and b) we've just checked that the
5621 * huge pmd (head) page is present in the page tables. The ptl
5622 * prevents the head page and tail pages from being rearranged
5623 * in any way. So this page must be available at this point,
5624 * unless the page refcount overflowed:
5626 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5631 if (is_hugetlb_entry_migration(pte)) {
5633 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5637 * hwpoisoned entry is treated as no_page_table in
5638 * follow_page_mask().
5646 struct page * __weak
5647 follow_huge_pud(struct mm_struct *mm, unsigned long address,
5648 pud_t *pud, int flags)
5650 if (flags & (FOLL_GET | FOLL_PIN))
5653 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5656 struct page * __weak
5657 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5659 if (flags & (FOLL_GET | FOLL_PIN))
5662 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5665 bool isolate_huge_page(struct page *page, struct list_head *list)
5669 spin_lock_irq(&hugetlb_lock);
5670 if (!PageHeadHuge(page) ||
5671 !HPageMigratable(page) ||
5672 !get_page_unless_zero(page)) {
5676 ClearHPageMigratable(page);
5677 list_move_tail(&page->lru, list);
5679 spin_unlock_irq(&hugetlb_lock);
5683 void putback_active_hugepage(struct page *page)
5685 spin_lock_irq(&hugetlb_lock);
5686 SetHPageMigratable(page);
5687 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5688 spin_unlock_irq(&hugetlb_lock);
5692 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5694 struct hstate *h = page_hstate(oldpage);
5696 hugetlb_cgroup_migrate(oldpage, newpage);
5697 set_page_owner_migrate_reason(newpage, reason);
5700 * transfer temporary state of the new huge page. This is
5701 * reverse to other transitions because the newpage is going to
5702 * be final while the old one will be freed so it takes over
5703 * the temporary status.
5705 * Also note that we have to transfer the per-node surplus state
5706 * here as well otherwise the global surplus count will not match
5709 if (HPageTemporary(newpage)) {
5710 int old_nid = page_to_nid(oldpage);
5711 int new_nid = page_to_nid(newpage);
5713 SetHPageTemporary(oldpage);
5714 ClearHPageTemporary(newpage);
5717 * There is no need to transfer the per-node surplus state
5718 * when we do not cross the node.
5720 if (new_nid == old_nid)
5722 spin_lock_irq(&hugetlb_lock);
5723 if (h->surplus_huge_pages_node[old_nid]) {
5724 h->surplus_huge_pages_node[old_nid]--;
5725 h->surplus_huge_pages_node[new_nid]++;
5727 spin_unlock_irq(&hugetlb_lock);
5732 * This function will unconditionally remove all the shared pmd pgtable entries
5733 * within the specific vma for a hugetlbfs memory range.
5735 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
5737 struct hstate *h = hstate_vma(vma);
5738 unsigned long sz = huge_page_size(h);
5739 struct mm_struct *mm = vma->vm_mm;
5740 struct mmu_notifier_range range;
5741 unsigned long address, start, end;
5745 if (!(vma->vm_flags & VM_MAYSHARE))
5748 start = ALIGN(vma->vm_start, PUD_SIZE);
5749 end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
5755 * No need to call adjust_range_if_pmd_sharing_possible(), because
5756 * we have already done the PUD_SIZE alignment.
5758 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
5760 mmu_notifier_invalidate_range_start(&range);
5761 i_mmap_lock_write(vma->vm_file->f_mapping);
5762 for (address = start; address < end; address += PUD_SIZE) {
5763 unsigned long tmp = address;
5765 ptep = huge_pte_offset(mm, address, sz);
5768 ptl = huge_pte_lock(h, mm, ptep);
5769 /* We don't want 'address' to be changed */
5770 huge_pmd_unshare(mm, vma, &tmp, ptep);
5773 flush_hugetlb_tlb_range(vma, start, end);
5774 i_mmap_unlock_write(vma->vm_file->f_mapping);
5776 * No need to call mmu_notifier_invalidate_range(), see
5777 * Documentation/vm/mmu_notifier.rst.
5779 mmu_notifier_invalidate_range_end(&range);
5783 static bool cma_reserve_called __initdata;
5785 static int __init cmdline_parse_hugetlb_cma(char *p)
5787 hugetlb_cma_size = memparse(p, &p);
5791 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
5793 void __init hugetlb_cma_reserve(int order)
5795 unsigned long size, reserved, per_node;
5798 cma_reserve_called = true;
5800 if (!hugetlb_cma_size)
5803 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
5804 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
5805 (PAGE_SIZE << order) / SZ_1M);
5810 * If 3 GB area is requested on a machine with 4 numa nodes,
5811 * let's allocate 1 GB on first three nodes and ignore the last one.
5813 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
5814 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
5815 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
5818 for_each_node_state(nid, N_ONLINE) {
5820 char name[CMA_MAX_NAME];
5822 size = min(per_node, hugetlb_cma_size - reserved);
5823 size = round_up(size, PAGE_SIZE << order);
5825 snprintf(name, sizeof(name), "hugetlb%d", nid);
5826 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
5828 &hugetlb_cma[nid], nid);
5830 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
5836 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
5839 if (reserved >= hugetlb_cma_size)
5844 void __init hugetlb_cma_check(void)
5846 if (!hugetlb_cma_size || cma_reserve_called)
5849 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
5852 #endif /* CONFIG_CMA */