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>
33 #include <linux/migrate.h>
34 #include <linux/nospec.h>
35 #include <linux/delayacct.h>
38 #include <asm/pgalloc.h>
42 #include <linux/hugetlb.h>
43 #include <linux/hugetlb_cgroup.h>
44 #include <linux/node.h>
45 #include <linux/page_owner.h>
47 #include "hugetlb_vmemmap.h"
49 int hugetlb_max_hstate __read_mostly;
50 unsigned int default_hstate_idx;
51 struct hstate hstates[HUGE_MAX_HSTATE];
54 static struct cma *hugetlb_cma[MAX_NUMNODES];
55 static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata;
56 static bool hugetlb_cma_page(struct page *page, unsigned int order)
58 return cma_pages_valid(hugetlb_cma[page_to_nid(page)], page,
62 static bool hugetlb_cma_page(struct page *page, unsigned int order)
67 static unsigned long hugetlb_cma_size __initdata;
70 * Minimum page order among possible hugepage sizes, set to a proper value
73 static unsigned int minimum_order __read_mostly = UINT_MAX;
75 __initdata LIST_HEAD(huge_boot_pages);
77 /* for command line parsing */
78 static struct hstate * __initdata parsed_hstate;
79 static unsigned long __initdata default_hstate_max_huge_pages;
80 static bool __initdata parsed_valid_hugepagesz = true;
81 static bool __initdata parsed_default_hugepagesz;
82 static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata;
85 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
86 * free_huge_pages, and surplus_huge_pages.
88 DEFINE_SPINLOCK(hugetlb_lock);
91 * Serializes faults on the same logical page. This is used to
92 * prevent spurious OOMs when the hugepage pool is fully utilized.
94 static int num_fault_mutexes;
95 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
97 /* Forward declaration */
98 static int hugetlb_acct_memory(struct hstate *h, long delta);
100 static inline bool subpool_is_free(struct hugepage_subpool *spool)
104 if (spool->max_hpages != -1)
105 return spool->used_hpages == 0;
106 if (spool->min_hpages != -1)
107 return spool->rsv_hpages == spool->min_hpages;
112 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
113 unsigned long irq_flags)
115 spin_unlock_irqrestore(&spool->lock, irq_flags);
117 /* If no pages are used, and no other handles to the subpool
118 * remain, give up any reservations based on minimum size and
119 * free the subpool */
120 if (subpool_is_free(spool)) {
121 if (spool->min_hpages != -1)
122 hugetlb_acct_memory(spool->hstate,
128 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
131 struct hugepage_subpool *spool;
133 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
137 spin_lock_init(&spool->lock);
139 spool->max_hpages = max_hpages;
141 spool->min_hpages = min_hpages;
143 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
147 spool->rsv_hpages = min_hpages;
152 void hugepage_put_subpool(struct hugepage_subpool *spool)
156 spin_lock_irqsave(&spool->lock, flags);
157 BUG_ON(!spool->count);
159 unlock_or_release_subpool(spool, flags);
163 * Subpool accounting for allocating and reserving pages.
164 * Return -ENOMEM if there are not enough resources to satisfy the
165 * request. Otherwise, return the number of pages by which the
166 * global pools must be adjusted (upward). The returned value may
167 * only be different than the passed value (delta) in the case where
168 * a subpool minimum size must be maintained.
170 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
178 spin_lock_irq(&spool->lock);
180 if (spool->max_hpages != -1) { /* maximum size accounting */
181 if ((spool->used_hpages + delta) <= spool->max_hpages)
182 spool->used_hpages += delta;
189 /* minimum size accounting */
190 if (spool->min_hpages != -1 && spool->rsv_hpages) {
191 if (delta > spool->rsv_hpages) {
193 * Asking for more reserves than those already taken on
194 * behalf of subpool. Return difference.
196 ret = delta - spool->rsv_hpages;
197 spool->rsv_hpages = 0;
199 ret = 0; /* reserves already accounted for */
200 spool->rsv_hpages -= delta;
205 spin_unlock_irq(&spool->lock);
210 * Subpool accounting for freeing and unreserving pages.
211 * Return the number of global page reservations that must be dropped.
212 * The return value may only be different than the passed value (delta)
213 * in the case where a subpool minimum size must be maintained.
215 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
224 spin_lock_irqsave(&spool->lock, flags);
226 if (spool->max_hpages != -1) /* maximum size accounting */
227 spool->used_hpages -= delta;
229 /* minimum size accounting */
230 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
231 if (spool->rsv_hpages + delta <= spool->min_hpages)
234 ret = spool->rsv_hpages + delta - spool->min_hpages;
236 spool->rsv_hpages += delta;
237 if (spool->rsv_hpages > spool->min_hpages)
238 spool->rsv_hpages = spool->min_hpages;
242 * If hugetlbfs_put_super couldn't free spool due to an outstanding
243 * quota reference, free it now.
245 unlock_or_release_subpool(spool, flags);
250 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
252 return HUGETLBFS_SB(inode->i_sb)->spool;
255 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
257 return subpool_inode(file_inode(vma->vm_file));
260 /* Helper that removes a struct file_region from the resv_map cache and returns
263 static struct file_region *
264 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
266 struct file_region *nrg = NULL;
268 VM_BUG_ON(resv->region_cache_count <= 0);
270 resv->region_cache_count--;
271 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
272 list_del(&nrg->link);
280 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
281 struct file_region *rg)
283 #ifdef CONFIG_CGROUP_HUGETLB
284 nrg->reservation_counter = rg->reservation_counter;
291 /* Helper that records hugetlb_cgroup uncharge info. */
292 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
294 struct resv_map *resv,
295 struct file_region *nrg)
297 #ifdef CONFIG_CGROUP_HUGETLB
299 nrg->reservation_counter =
300 &h_cg->rsvd_hugepage[hstate_index(h)];
301 nrg->css = &h_cg->css;
303 * The caller will hold exactly one h_cg->css reference for the
304 * whole contiguous reservation region. But this area might be
305 * scattered when there are already some file_regions reside in
306 * it. As a result, many file_regions may share only one css
307 * reference. In order to ensure that one file_region must hold
308 * exactly one h_cg->css reference, we should do css_get for
309 * each file_region and leave the reference held by caller
313 if (!resv->pages_per_hpage)
314 resv->pages_per_hpage = pages_per_huge_page(h);
315 /* pages_per_hpage should be the same for all entries in
318 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
320 nrg->reservation_counter = NULL;
326 static void put_uncharge_info(struct file_region *rg)
328 #ifdef CONFIG_CGROUP_HUGETLB
334 static bool has_same_uncharge_info(struct file_region *rg,
335 struct file_region *org)
337 #ifdef CONFIG_CGROUP_HUGETLB
338 return rg->reservation_counter == org->reservation_counter &&
346 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
348 struct file_region *nrg = NULL, *prg = NULL;
350 prg = list_prev_entry(rg, link);
351 if (&prg->link != &resv->regions && prg->to == rg->from &&
352 has_same_uncharge_info(prg, rg)) {
356 put_uncharge_info(rg);
362 nrg = list_next_entry(rg, link);
363 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
364 has_same_uncharge_info(nrg, rg)) {
365 nrg->from = rg->from;
368 put_uncharge_info(rg);
374 hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from,
375 long to, struct hstate *h, struct hugetlb_cgroup *cg,
376 long *regions_needed)
378 struct file_region *nrg;
380 if (!regions_needed) {
381 nrg = get_file_region_entry_from_cache(map, from, to);
382 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
383 list_add(&nrg->link, rg);
384 coalesce_file_region(map, nrg);
386 *regions_needed += 1;
392 * Must be called with resv->lock held.
394 * Calling this with regions_needed != NULL will count the number of pages
395 * to be added but will not modify the linked list. And regions_needed will
396 * indicate the number of file_regions needed in the cache to carry out to add
397 * the regions for this range.
399 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
400 struct hugetlb_cgroup *h_cg,
401 struct hstate *h, long *regions_needed)
404 struct list_head *head = &resv->regions;
405 long last_accounted_offset = f;
406 struct file_region *iter, *trg = NULL;
407 struct list_head *rg = NULL;
412 /* In this loop, we essentially handle an entry for the range
413 * [last_accounted_offset, iter->from), at every iteration, with some
416 list_for_each_entry_safe(iter, trg, head, link) {
417 /* Skip irrelevant regions that start before our range. */
418 if (iter->from < f) {
419 /* If this region ends after the last accounted offset,
420 * then we need to update last_accounted_offset.
422 if (iter->to > last_accounted_offset)
423 last_accounted_offset = iter->to;
427 /* When we find a region that starts beyond our range, we've
430 if (iter->from >= t) {
431 rg = iter->link.prev;
435 /* Add an entry for last_accounted_offset -> iter->from, and
436 * update last_accounted_offset.
438 if (iter->from > last_accounted_offset)
439 add += hugetlb_resv_map_add(resv, iter->link.prev,
440 last_accounted_offset,
444 last_accounted_offset = iter->to;
447 /* Handle the case where our range extends beyond
448 * last_accounted_offset.
452 if (last_accounted_offset < t)
453 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
454 t, h, h_cg, regions_needed);
459 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
461 static int allocate_file_region_entries(struct resv_map *resv,
463 __must_hold(&resv->lock)
465 struct list_head allocated_regions;
466 int to_allocate = 0, i = 0;
467 struct file_region *trg = NULL, *rg = NULL;
469 VM_BUG_ON(regions_needed < 0);
471 INIT_LIST_HEAD(&allocated_regions);
474 * Check for sufficient descriptors in the cache to accommodate
475 * the number of in progress add operations plus regions_needed.
477 * This is a while loop because when we drop the lock, some other call
478 * to region_add or region_del may have consumed some region_entries,
479 * so we keep looping here until we finally have enough entries for
480 * (adds_in_progress + regions_needed).
482 while (resv->region_cache_count <
483 (resv->adds_in_progress + regions_needed)) {
484 to_allocate = resv->adds_in_progress + regions_needed -
485 resv->region_cache_count;
487 /* At this point, we should have enough entries in the cache
488 * for all the existing adds_in_progress. We should only be
489 * needing to allocate for regions_needed.
491 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
493 spin_unlock(&resv->lock);
494 for (i = 0; i < to_allocate; i++) {
495 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
498 list_add(&trg->link, &allocated_regions);
501 spin_lock(&resv->lock);
503 list_splice(&allocated_regions, &resv->region_cache);
504 resv->region_cache_count += to_allocate;
510 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
518 * Add the huge page range represented by [f, t) to the reserve
519 * map. Regions will be taken from the cache to fill in this range.
520 * Sufficient regions should exist in the cache due to the previous
521 * call to region_chg with the same range, but in some cases the cache will not
522 * have sufficient entries due to races with other code doing region_add or
523 * region_del. The extra needed entries will be allocated.
525 * regions_needed is the out value provided by a previous call to region_chg.
527 * Return the number of new huge pages added to the map. This number is greater
528 * than or equal to zero. If file_region entries needed to be allocated for
529 * this operation and we were not able to allocate, it returns -ENOMEM.
530 * region_add of regions of length 1 never allocate file_regions and cannot
531 * fail; region_chg will always allocate at least 1 entry and a region_add for
532 * 1 page will only require at most 1 entry.
534 static long region_add(struct resv_map *resv, long f, long t,
535 long in_regions_needed, struct hstate *h,
536 struct hugetlb_cgroup *h_cg)
538 long add = 0, actual_regions_needed = 0;
540 spin_lock(&resv->lock);
543 /* Count how many regions are actually needed to execute this add. */
544 add_reservation_in_range(resv, f, t, NULL, NULL,
545 &actual_regions_needed);
548 * Check for sufficient descriptors in the cache to accommodate
549 * this add operation. Note that actual_regions_needed may be greater
550 * than in_regions_needed, as the resv_map may have been modified since
551 * the region_chg call. In this case, we need to make sure that we
552 * allocate extra entries, such that we have enough for all the
553 * existing adds_in_progress, plus the excess needed for this
556 if (actual_regions_needed > in_regions_needed &&
557 resv->region_cache_count <
558 resv->adds_in_progress +
559 (actual_regions_needed - in_regions_needed)) {
560 /* region_add operation of range 1 should never need to
561 * allocate file_region entries.
563 VM_BUG_ON(t - f <= 1);
565 if (allocate_file_region_entries(
566 resv, actual_regions_needed - in_regions_needed)) {
573 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
575 resv->adds_in_progress -= in_regions_needed;
577 spin_unlock(&resv->lock);
582 * Examine the existing reserve map and determine how many
583 * huge pages in the specified range [f, t) are NOT currently
584 * represented. This routine is called before a subsequent
585 * call to region_add that will actually modify the reserve
586 * map to add the specified range [f, t). region_chg does
587 * not change the number of huge pages represented by the
588 * map. A number of new file_region structures is added to the cache as a
589 * placeholder, for the subsequent region_add call to use. At least 1
590 * file_region structure is added.
592 * out_regions_needed is the number of regions added to the
593 * resv->adds_in_progress. This value needs to be provided to a follow up call
594 * to region_add or region_abort for proper accounting.
596 * Returns the number of huge pages that need to be added to the existing
597 * reservation map for the range [f, t). This number is greater or equal to
598 * zero. -ENOMEM is returned if a new file_region structure or cache entry
599 * is needed and can not be allocated.
601 static long region_chg(struct resv_map *resv, long f, long t,
602 long *out_regions_needed)
606 spin_lock(&resv->lock);
608 /* Count how many hugepages in this range are NOT represented. */
609 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
612 if (*out_regions_needed == 0)
613 *out_regions_needed = 1;
615 if (allocate_file_region_entries(resv, *out_regions_needed))
618 resv->adds_in_progress += *out_regions_needed;
620 spin_unlock(&resv->lock);
625 * Abort the in progress add operation. The adds_in_progress field
626 * of the resv_map keeps track of the operations in progress between
627 * calls to region_chg and region_add. Operations are sometimes
628 * aborted after the call to region_chg. In such cases, region_abort
629 * is called to decrement the adds_in_progress counter. regions_needed
630 * is the value returned by the region_chg call, it is used to decrement
631 * the adds_in_progress counter.
633 * NOTE: The range arguments [f, t) are not needed or used in this
634 * routine. They are kept to make reading the calling code easier as
635 * arguments will match the associated region_chg call.
637 static void region_abort(struct resv_map *resv, long f, long t,
640 spin_lock(&resv->lock);
641 VM_BUG_ON(!resv->region_cache_count);
642 resv->adds_in_progress -= regions_needed;
643 spin_unlock(&resv->lock);
647 * Delete the specified range [f, t) from the reserve map. If the
648 * t parameter is LONG_MAX, this indicates that ALL regions after f
649 * should be deleted. Locate the regions which intersect [f, t)
650 * and either trim, delete or split the existing regions.
652 * Returns the number of huge pages deleted from the reserve map.
653 * In the normal case, the return value is zero or more. In the
654 * case where a region must be split, a new region descriptor must
655 * be allocated. If the allocation fails, -ENOMEM will be returned.
656 * NOTE: If the parameter t == LONG_MAX, then we will never split
657 * a region and possibly return -ENOMEM. Callers specifying
658 * t == LONG_MAX do not need to check for -ENOMEM error.
660 static long region_del(struct resv_map *resv, long f, long t)
662 struct list_head *head = &resv->regions;
663 struct file_region *rg, *trg;
664 struct file_region *nrg = NULL;
668 spin_lock(&resv->lock);
669 list_for_each_entry_safe(rg, trg, head, link) {
671 * Skip regions before the range to be deleted. file_region
672 * ranges are normally of the form [from, to). However, there
673 * may be a "placeholder" entry in the map which is of the form
674 * (from, to) with from == to. Check for placeholder entries
675 * at the beginning of the range to be deleted.
677 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
683 if (f > rg->from && t < rg->to) { /* Must split region */
685 * Check for an entry in the cache before dropping
686 * lock and attempting allocation.
689 resv->region_cache_count > resv->adds_in_progress) {
690 nrg = list_first_entry(&resv->region_cache,
693 list_del(&nrg->link);
694 resv->region_cache_count--;
698 spin_unlock(&resv->lock);
699 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
706 hugetlb_cgroup_uncharge_file_region(
707 resv, rg, t - f, false);
709 /* New entry for end of split region */
713 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
715 INIT_LIST_HEAD(&nrg->link);
717 /* Original entry is trimmed */
720 list_add(&nrg->link, &rg->link);
725 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
726 del += rg->to - rg->from;
727 hugetlb_cgroup_uncharge_file_region(resv, rg,
728 rg->to - rg->from, true);
734 if (f <= rg->from) { /* Trim beginning of region */
735 hugetlb_cgroup_uncharge_file_region(resv, rg,
736 t - rg->from, false);
740 } else { /* Trim end of region */
741 hugetlb_cgroup_uncharge_file_region(resv, rg,
749 spin_unlock(&resv->lock);
755 * A rare out of memory error was encountered which prevented removal of
756 * the reserve map region for a page. The huge page itself was free'ed
757 * and removed from the page cache. This routine will adjust the subpool
758 * usage count, and the global reserve count if needed. By incrementing
759 * these counts, the reserve map entry which could not be deleted will
760 * appear as a "reserved" entry instead of simply dangling with incorrect
763 void hugetlb_fix_reserve_counts(struct inode *inode)
765 struct hugepage_subpool *spool = subpool_inode(inode);
767 bool reserved = false;
769 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
770 if (rsv_adjust > 0) {
771 struct hstate *h = hstate_inode(inode);
773 if (!hugetlb_acct_memory(h, 1))
775 } else if (!rsv_adjust) {
780 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
784 * Count and return the number of huge pages in the reserve map
785 * that intersect with the range [f, t).
787 static long region_count(struct resv_map *resv, long f, long t)
789 struct list_head *head = &resv->regions;
790 struct file_region *rg;
793 spin_lock(&resv->lock);
794 /* Locate each segment we overlap with, and count that overlap. */
795 list_for_each_entry(rg, head, link) {
804 seg_from = max(rg->from, f);
805 seg_to = min(rg->to, t);
807 chg += seg_to - seg_from;
809 spin_unlock(&resv->lock);
815 * Convert the address within this vma to the page offset within
816 * the mapping, in pagecache page units; huge pages here.
818 static pgoff_t vma_hugecache_offset(struct hstate *h,
819 struct vm_area_struct *vma, unsigned long address)
821 return ((address - vma->vm_start) >> huge_page_shift(h)) +
822 (vma->vm_pgoff >> huge_page_order(h));
825 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
826 unsigned long address)
828 return vma_hugecache_offset(hstate_vma(vma), vma, address);
830 EXPORT_SYMBOL_GPL(linear_hugepage_index);
833 * Return the size of the pages allocated when backing a VMA. In the majority
834 * cases this will be same size as used by the page table entries.
836 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
838 if (vma->vm_ops && vma->vm_ops->pagesize)
839 return vma->vm_ops->pagesize(vma);
842 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
845 * Return the page size being used by the MMU to back a VMA. In the majority
846 * of cases, the page size used by the kernel matches the MMU size. On
847 * architectures where it differs, an architecture-specific 'strong'
848 * version of this symbol is required.
850 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
852 return vma_kernel_pagesize(vma);
856 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
857 * bits of the reservation map pointer, which are always clear due to
860 #define HPAGE_RESV_OWNER (1UL << 0)
861 #define HPAGE_RESV_UNMAPPED (1UL << 1)
862 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
865 * These helpers are used to track how many pages are reserved for
866 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
867 * is guaranteed to have their future faults succeed.
869 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
870 * the reserve counters are updated with the hugetlb_lock held. It is safe
871 * to reset the VMA at fork() time as it is not in use yet and there is no
872 * chance of the global counters getting corrupted as a result of the values.
874 * The private mapping reservation is represented in a subtly different
875 * manner to a shared mapping. A shared mapping has a region map associated
876 * with the underlying file, this region map represents the backing file
877 * pages which have ever had a reservation assigned which this persists even
878 * after the page is instantiated. A private mapping has a region map
879 * associated with the original mmap which is attached to all VMAs which
880 * reference it, this region map represents those offsets which have consumed
881 * reservation ie. where pages have been instantiated.
883 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
885 return (unsigned long)vma->vm_private_data;
888 static void set_vma_private_data(struct vm_area_struct *vma,
891 vma->vm_private_data = (void *)value;
895 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
896 struct hugetlb_cgroup *h_cg,
899 #ifdef CONFIG_CGROUP_HUGETLB
901 resv_map->reservation_counter = NULL;
902 resv_map->pages_per_hpage = 0;
903 resv_map->css = NULL;
905 resv_map->reservation_counter =
906 &h_cg->rsvd_hugepage[hstate_index(h)];
907 resv_map->pages_per_hpage = pages_per_huge_page(h);
908 resv_map->css = &h_cg->css;
913 struct resv_map *resv_map_alloc(void)
915 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
916 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
918 if (!resv_map || !rg) {
924 kref_init(&resv_map->refs);
925 spin_lock_init(&resv_map->lock);
926 INIT_LIST_HEAD(&resv_map->regions);
928 resv_map->adds_in_progress = 0;
930 * Initialize these to 0. On shared mappings, 0's here indicate these
931 * fields don't do cgroup accounting. On private mappings, these will be
932 * re-initialized to the proper values, to indicate that hugetlb cgroup
933 * reservations are to be un-charged from here.
935 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
937 INIT_LIST_HEAD(&resv_map->region_cache);
938 list_add(&rg->link, &resv_map->region_cache);
939 resv_map->region_cache_count = 1;
944 void resv_map_release(struct kref *ref)
946 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
947 struct list_head *head = &resv_map->region_cache;
948 struct file_region *rg, *trg;
950 /* Clear out any active regions before we release the map. */
951 region_del(resv_map, 0, LONG_MAX);
953 /* ... and any entries left in the cache */
954 list_for_each_entry_safe(rg, trg, head, link) {
959 VM_BUG_ON(resv_map->adds_in_progress);
964 static inline struct resv_map *inode_resv_map(struct inode *inode)
967 * At inode evict time, i_mapping may not point to the original
968 * address space within the inode. This original address space
969 * contains the pointer to the resv_map. So, always use the
970 * address space embedded within the inode.
971 * The VERY common case is inode->mapping == &inode->i_data but,
972 * this may not be true for device special inodes.
974 return (struct resv_map *)(&inode->i_data)->private_data;
977 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
979 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
980 if (vma->vm_flags & VM_MAYSHARE) {
981 struct address_space *mapping = vma->vm_file->f_mapping;
982 struct inode *inode = mapping->host;
984 return inode_resv_map(inode);
987 return (struct resv_map *)(get_vma_private_data(vma) &
992 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
994 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
995 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
997 set_vma_private_data(vma, (get_vma_private_data(vma) &
998 HPAGE_RESV_MASK) | (unsigned long)map);
1001 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
1003 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1004 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1006 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
1009 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
1011 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1013 return (get_vma_private_data(vma) & flag) != 0;
1016 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
1017 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
1019 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1020 if (!(vma->vm_flags & VM_MAYSHARE))
1021 vma->vm_private_data = (void *)0;
1025 * Reset and decrement one ref on hugepage private reservation.
1026 * Called with mm->mmap_sem writer semaphore held.
1027 * This function should be only used by move_vma() and operate on
1028 * same sized vma. It should never come here with last ref on the
1031 void clear_vma_resv_huge_pages(struct vm_area_struct *vma)
1034 * Clear the old hugetlb private page reservation.
1035 * It has already been transferred to new_vma.
1037 * During a mremap() operation of a hugetlb vma we call move_vma()
1038 * which copies vma into new_vma and unmaps vma. After the copy
1039 * operation both new_vma and vma share a reference to the resv_map
1040 * struct, and at that point vma is about to be unmapped. We don't
1041 * want to return the reservation to the pool at unmap of vma because
1042 * the reservation still lives on in new_vma, so simply decrement the
1043 * ref here and remove the resv_map reference from this vma.
1045 struct resv_map *reservations = vma_resv_map(vma);
1047 if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1048 resv_map_put_hugetlb_cgroup_uncharge_info(reservations);
1049 kref_put(&reservations->refs, resv_map_release);
1052 reset_vma_resv_huge_pages(vma);
1055 /* Returns true if the VMA has associated reserve pages */
1056 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1058 if (vma->vm_flags & VM_NORESERVE) {
1060 * This address is already reserved by other process(chg == 0),
1061 * so, we should decrement reserved count. Without decrementing,
1062 * reserve count remains after releasing inode, because this
1063 * allocated page will go into page cache and is regarded as
1064 * coming from reserved pool in releasing step. Currently, we
1065 * don't have any other solution to deal with this situation
1066 * properly, so add work-around here.
1068 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1074 /* Shared mappings always use reserves */
1075 if (vma->vm_flags & VM_MAYSHARE) {
1077 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1078 * be a region map for all pages. The only situation where
1079 * there is no region map is if a hole was punched via
1080 * fallocate. In this case, there really are no reserves to
1081 * use. This situation is indicated if chg != 0.
1090 * Only the process that called mmap() has reserves for
1093 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1095 * Like the shared case above, a hole punch or truncate
1096 * could have been performed on the private mapping.
1097 * Examine the value of chg to determine if reserves
1098 * actually exist or were previously consumed.
1099 * Very Subtle - The value of chg comes from a previous
1100 * call to vma_needs_reserves(). The reserve map for
1101 * private mappings has different (opposite) semantics
1102 * than that of shared mappings. vma_needs_reserves()
1103 * has already taken this difference in semantics into
1104 * account. Therefore, the meaning of chg is the same
1105 * as in the shared case above. Code could easily be
1106 * combined, but keeping it separate draws attention to
1107 * subtle differences.
1118 static void enqueue_huge_page(struct hstate *h, struct page *page)
1120 int nid = page_to_nid(page);
1122 lockdep_assert_held(&hugetlb_lock);
1123 VM_BUG_ON_PAGE(page_count(page), page);
1125 list_move(&page->lru, &h->hugepage_freelists[nid]);
1126 h->free_huge_pages++;
1127 h->free_huge_pages_node[nid]++;
1128 SetHPageFreed(page);
1131 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1134 bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1136 lockdep_assert_held(&hugetlb_lock);
1137 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1138 if (pin && !is_pinnable_page(page))
1141 if (PageHWPoison(page))
1144 list_move(&page->lru, &h->hugepage_activelist);
1145 set_page_refcounted(page);
1146 ClearHPageFreed(page);
1147 h->free_huge_pages--;
1148 h->free_huge_pages_node[nid]--;
1155 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1158 unsigned int cpuset_mems_cookie;
1159 struct zonelist *zonelist;
1162 int node = NUMA_NO_NODE;
1164 zonelist = node_zonelist(nid, gfp_mask);
1167 cpuset_mems_cookie = read_mems_allowed_begin();
1168 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1171 if (!cpuset_zone_allowed(zone, gfp_mask))
1174 * no need to ask again on the same node. Pool is node rather than
1177 if (zone_to_nid(zone) == node)
1179 node = zone_to_nid(zone);
1181 page = dequeue_huge_page_node_exact(h, node);
1185 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1191 static struct page *dequeue_huge_page_vma(struct hstate *h,
1192 struct vm_area_struct *vma,
1193 unsigned long address, int avoid_reserve,
1196 struct page *page = NULL;
1197 struct mempolicy *mpol;
1199 nodemask_t *nodemask;
1203 * A child process with MAP_PRIVATE mappings created by their parent
1204 * have no page reserves. This check ensures that reservations are
1205 * not "stolen". The child may still get SIGKILLed
1207 if (!vma_has_reserves(vma, chg) &&
1208 h->free_huge_pages - h->resv_huge_pages == 0)
1211 /* If reserves cannot be used, ensure enough pages are in the pool */
1212 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1215 gfp_mask = htlb_alloc_mask(h);
1216 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1218 if (mpol_is_preferred_many(mpol)) {
1219 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1221 /* Fallback to all nodes if page==NULL */
1226 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1228 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1229 SetHPageRestoreReserve(page);
1230 h->resv_huge_pages--;
1233 mpol_cond_put(mpol);
1241 * common helper functions for hstate_next_node_to_{alloc|free}.
1242 * We may have allocated or freed a huge page based on a different
1243 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1244 * be outside of *nodes_allowed. Ensure that we use an allowed
1245 * node for alloc or free.
1247 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1249 nid = next_node_in(nid, *nodes_allowed);
1250 VM_BUG_ON(nid >= MAX_NUMNODES);
1255 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1257 if (!node_isset(nid, *nodes_allowed))
1258 nid = next_node_allowed(nid, nodes_allowed);
1263 * returns the previously saved node ["this node"] from which to
1264 * allocate a persistent huge page for the pool and advance the
1265 * next node from which to allocate, handling wrap at end of node
1268 static int hstate_next_node_to_alloc(struct hstate *h,
1269 nodemask_t *nodes_allowed)
1273 VM_BUG_ON(!nodes_allowed);
1275 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1276 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1282 * helper for remove_pool_huge_page() - return the previously saved
1283 * node ["this node"] from which to free a huge page. Advance the
1284 * next node id whether or not we find a free huge page to free so
1285 * that the next attempt to free addresses the next node.
1287 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1291 VM_BUG_ON(!nodes_allowed);
1293 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1294 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1299 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1300 for (nr_nodes = nodes_weight(*mask); \
1302 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1305 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1306 for (nr_nodes = nodes_weight(*mask); \
1308 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1311 /* used to demote non-gigantic_huge pages as well */
1312 static void __destroy_compound_gigantic_page(struct page *page,
1313 unsigned int order, bool demote)
1316 int nr_pages = 1 << order;
1317 struct page *p = page + 1;
1319 atomic_set(compound_mapcount_ptr(page), 0);
1320 atomic_set(compound_pincount_ptr(page), 0);
1322 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1324 clear_compound_head(p);
1326 set_page_refcounted(p);
1329 set_compound_order(page, 0);
1331 page[1].compound_nr = 0;
1333 __ClearPageHead(page);
1336 static void destroy_compound_hugetlb_page_for_demote(struct page *page,
1339 __destroy_compound_gigantic_page(page, order, true);
1342 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1343 static void destroy_compound_gigantic_page(struct page *page,
1346 __destroy_compound_gigantic_page(page, order, false);
1349 static void free_gigantic_page(struct page *page, unsigned int order)
1352 * If the page isn't allocated using the cma allocator,
1353 * cma_release() returns false.
1356 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1360 free_contig_range(page_to_pfn(page), 1 << order);
1363 #ifdef CONFIG_CONTIG_ALLOC
1364 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1365 int nid, nodemask_t *nodemask)
1367 unsigned long nr_pages = pages_per_huge_page(h);
1368 if (nid == NUMA_NO_NODE)
1369 nid = numa_mem_id();
1376 if (hugetlb_cma[nid]) {
1377 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1378 huge_page_order(h), true);
1383 if (!(gfp_mask & __GFP_THISNODE)) {
1384 for_each_node_mask(node, *nodemask) {
1385 if (node == nid || !hugetlb_cma[node])
1388 page = cma_alloc(hugetlb_cma[node], nr_pages,
1389 huge_page_order(h), true);
1397 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1400 #else /* !CONFIG_CONTIG_ALLOC */
1401 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1402 int nid, nodemask_t *nodemask)
1406 #endif /* CONFIG_CONTIG_ALLOC */
1408 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1409 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1410 int nid, nodemask_t *nodemask)
1414 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1415 static inline void destroy_compound_gigantic_page(struct page *page,
1416 unsigned int order) { }
1420 * Remove hugetlb page from lists, and update dtor so that page appears
1421 * as just a compound page.
1423 * A reference is held on the page, except in the case of demote.
1425 * Must be called with hugetlb lock held.
1427 static void __remove_hugetlb_page(struct hstate *h, struct page *page,
1428 bool adjust_surplus,
1431 int nid = page_to_nid(page);
1433 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1434 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1436 lockdep_assert_held(&hugetlb_lock);
1437 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1440 list_del(&page->lru);
1442 if (HPageFreed(page)) {
1443 h->free_huge_pages--;
1444 h->free_huge_pages_node[nid]--;
1446 if (adjust_surplus) {
1447 h->surplus_huge_pages--;
1448 h->surplus_huge_pages_node[nid]--;
1454 * For non-gigantic pages set the destructor to the normal compound
1455 * page dtor. This is needed in case someone takes an additional
1456 * temporary ref to the page, and freeing is delayed until they drop
1459 * For gigantic pages set the destructor to the null dtor. This
1460 * destructor will never be called. Before freeing the gigantic
1461 * page destroy_compound_gigantic_page will turn the compound page
1462 * into a simple group of pages. After this the destructor does not
1465 * This handles the case where more than one ref is held when and
1466 * after update_and_free_page is called.
1468 * In the case of demote we do not ref count the page as it will soon
1469 * be turned into a page of smaller size.
1472 set_page_refcounted(page);
1473 if (hstate_is_gigantic(h))
1474 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1476 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
1479 h->nr_huge_pages_node[nid]--;
1482 static void remove_hugetlb_page(struct hstate *h, struct page *page,
1483 bool adjust_surplus)
1485 __remove_hugetlb_page(h, page, adjust_surplus, false);
1488 static void remove_hugetlb_page_for_demote(struct hstate *h, struct page *page,
1489 bool adjust_surplus)
1491 __remove_hugetlb_page(h, page, adjust_surplus, true);
1494 static void add_hugetlb_page(struct hstate *h, struct page *page,
1495 bool adjust_surplus)
1498 int nid = page_to_nid(page);
1500 VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page);
1502 lockdep_assert_held(&hugetlb_lock);
1504 INIT_LIST_HEAD(&page->lru);
1506 h->nr_huge_pages_node[nid]++;
1508 if (adjust_surplus) {
1509 h->surplus_huge_pages++;
1510 h->surplus_huge_pages_node[nid]++;
1513 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1514 set_page_private(page, 0);
1515 SetHPageVmemmapOptimized(page);
1518 * This page is about to be managed by the hugetlb allocator and
1519 * should have no users. Drop our reference, and check for others
1522 zeroed = put_page_testzero(page);
1525 * It is VERY unlikely soneone else has taken a ref on
1526 * the page. In this case, we simply return as the
1527 * hugetlb destructor (free_huge_page) will be called
1528 * when this other ref is dropped.
1532 arch_clear_hugepage_flags(page);
1533 enqueue_huge_page(h, page);
1536 static void __update_and_free_page(struct hstate *h, struct page *page)
1539 struct page *subpage = page;
1541 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1544 if (hugetlb_vmemmap_alloc(h, page)) {
1545 spin_lock_irq(&hugetlb_lock);
1547 * If we cannot allocate vmemmap pages, just refuse to free the
1548 * page and put the page back on the hugetlb free list and treat
1549 * as a surplus page.
1551 add_hugetlb_page(h, page, true);
1552 spin_unlock_irq(&hugetlb_lock);
1556 for (i = 0; i < pages_per_huge_page(h);
1557 i++, subpage = mem_map_next(subpage, page, i)) {
1558 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1559 1 << PG_referenced | 1 << PG_dirty |
1560 1 << PG_active | 1 << PG_private |
1565 * Non-gigantic pages demoted from CMA allocated gigantic pages
1566 * need to be given back to CMA in free_gigantic_page.
1568 if (hstate_is_gigantic(h) ||
1569 hugetlb_cma_page(page, huge_page_order(h))) {
1570 destroy_compound_gigantic_page(page, huge_page_order(h));
1571 free_gigantic_page(page, huge_page_order(h));
1573 __free_pages(page, huge_page_order(h));
1578 * As update_and_free_page() can be called under any context, so we cannot
1579 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1580 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1581 * the vmemmap pages.
1583 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1584 * freed and frees them one-by-one. As the page->mapping pointer is going
1585 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1586 * structure of a lockless linked list of huge pages to be freed.
1588 static LLIST_HEAD(hpage_freelist);
1590 static void free_hpage_workfn(struct work_struct *work)
1592 struct llist_node *node;
1594 node = llist_del_all(&hpage_freelist);
1600 page = container_of((struct address_space **)node,
1601 struct page, mapping);
1603 page->mapping = NULL;
1605 * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate()
1606 * is going to trigger because a previous call to
1607 * remove_hugetlb_page() will set_compound_page_dtor(page,
1608 * NULL_COMPOUND_DTOR), so do not use page_hstate() directly.
1610 h = size_to_hstate(page_size(page));
1612 __update_and_free_page(h, page);
1617 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1619 static inline void flush_free_hpage_work(struct hstate *h)
1621 if (hugetlb_optimize_vmemmap_pages(h))
1622 flush_work(&free_hpage_work);
1625 static void update_and_free_page(struct hstate *h, struct page *page,
1628 if (!HPageVmemmapOptimized(page) || !atomic) {
1629 __update_and_free_page(h, page);
1634 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1636 * Only call schedule_work() if hpage_freelist is previously
1637 * empty. Otherwise, schedule_work() had been called but the workfn
1638 * hasn't retrieved the list yet.
1640 if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist))
1641 schedule_work(&free_hpage_work);
1644 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1646 struct page *page, *t_page;
1648 list_for_each_entry_safe(page, t_page, list, lru) {
1649 update_and_free_page(h, page, false);
1654 struct hstate *size_to_hstate(unsigned long size)
1658 for_each_hstate(h) {
1659 if (huge_page_size(h) == size)
1665 void free_huge_page(struct page *page)
1668 * Can't pass hstate in here because it is called from the
1669 * compound page destructor.
1671 struct hstate *h = page_hstate(page);
1672 int nid = page_to_nid(page);
1673 struct hugepage_subpool *spool = hugetlb_page_subpool(page);
1674 bool restore_reserve;
1675 unsigned long flags;
1677 VM_BUG_ON_PAGE(page_count(page), page);
1678 VM_BUG_ON_PAGE(page_mapcount(page), page);
1680 hugetlb_set_page_subpool(page, NULL);
1682 __ClearPageAnonExclusive(page);
1683 page->mapping = NULL;
1684 restore_reserve = HPageRestoreReserve(page);
1685 ClearHPageRestoreReserve(page);
1688 * If HPageRestoreReserve was set on page, page allocation consumed a
1689 * reservation. If the page was associated with a subpool, there
1690 * would have been a page reserved in the subpool before allocation
1691 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1692 * reservation, do not call hugepage_subpool_put_pages() as this will
1693 * remove the reserved page from the subpool.
1695 if (!restore_reserve) {
1697 * A return code of zero implies that the subpool will be
1698 * under its minimum size if the reservation is not restored
1699 * after page is free. Therefore, force restore_reserve
1702 if (hugepage_subpool_put_pages(spool, 1) == 0)
1703 restore_reserve = true;
1706 spin_lock_irqsave(&hugetlb_lock, flags);
1707 ClearHPageMigratable(page);
1708 hugetlb_cgroup_uncharge_page(hstate_index(h),
1709 pages_per_huge_page(h), page);
1710 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1711 pages_per_huge_page(h), page);
1712 if (restore_reserve)
1713 h->resv_huge_pages++;
1715 if (HPageTemporary(page)) {
1716 remove_hugetlb_page(h, page, false);
1717 spin_unlock_irqrestore(&hugetlb_lock, flags);
1718 update_and_free_page(h, page, true);
1719 } else if (h->surplus_huge_pages_node[nid]) {
1720 /* remove the page from active list */
1721 remove_hugetlb_page(h, page, true);
1722 spin_unlock_irqrestore(&hugetlb_lock, flags);
1723 update_and_free_page(h, page, true);
1725 arch_clear_hugepage_flags(page);
1726 enqueue_huge_page(h, page);
1727 spin_unlock_irqrestore(&hugetlb_lock, flags);
1732 * Must be called with the hugetlb lock held
1734 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1736 lockdep_assert_held(&hugetlb_lock);
1738 h->nr_huge_pages_node[nid]++;
1741 static void __prep_new_huge_page(struct hstate *h, struct page *page)
1743 hugetlb_vmemmap_free(h, page);
1744 INIT_LIST_HEAD(&page->lru);
1745 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1746 hugetlb_set_page_subpool(page, NULL);
1747 set_hugetlb_cgroup(page, NULL);
1748 set_hugetlb_cgroup_rsvd(page, NULL);
1751 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1753 __prep_new_huge_page(h, page);
1754 spin_lock_irq(&hugetlb_lock);
1755 __prep_account_new_huge_page(h, nid);
1756 spin_unlock_irq(&hugetlb_lock);
1759 static bool __prep_compound_gigantic_page(struct page *page, unsigned int order,
1763 int nr_pages = 1 << order;
1764 struct page *p = page + 1;
1766 /* we rely on prep_new_huge_page to set the destructor */
1767 set_compound_order(page, order);
1768 __ClearPageReserved(page);
1769 __SetPageHead(page);
1770 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1772 * For gigantic hugepages allocated through bootmem at
1773 * boot, it's safer to be consistent with the not-gigantic
1774 * hugepages and clear the PG_reserved bit from all tail pages
1775 * too. Otherwise drivers using get_user_pages() to access tail
1776 * pages may get the reference counting wrong if they see
1777 * PG_reserved set on a tail page (despite the head page not
1778 * having PG_reserved set). Enforcing this consistency between
1779 * head and tail pages allows drivers to optimize away a check
1780 * on the head page when they need know if put_page() is needed
1781 * after get_user_pages().
1783 __ClearPageReserved(p);
1785 * Subtle and very unlikely
1787 * Gigantic 'page allocators' such as memblock or cma will
1788 * return a set of pages with each page ref counted. We need
1789 * to turn this set of pages into a compound page with tail
1790 * page ref counts set to zero. Code such as speculative page
1791 * cache adding could take a ref on a 'to be' tail page.
1792 * We need to respect any increased ref count, and only set
1793 * the ref count to zero if count is currently 1. If count
1794 * is not 1, we return an error. An error return indicates
1795 * the set of pages can not be converted to a gigantic page.
1796 * The caller who allocated the pages should then discard the
1797 * pages using the appropriate free interface.
1799 * In the case of demote, the ref count will be zero.
1802 if (!page_ref_freeze(p, 1)) {
1803 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
1807 VM_BUG_ON_PAGE(page_count(p), p);
1809 set_compound_head(p, page);
1811 atomic_set(compound_mapcount_ptr(page), -1);
1812 atomic_set(compound_pincount_ptr(page), 0);
1816 /* undo tail page modifications made above */
1818 for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) {
1819 clear_compound_head(p);
1820 set_page_refcounted(p);
1822 /* need to clear PG_reserved on remaining tail pages */
1823 for (; j < nr_pages; j++, p = mem_map_next(p, page, j))
1824 __ClearPageReserved(p);
1825 set_compound_order(page, 0);
1827 page[1].compound_nr = 0;
1829 __ClearPageHead(page);
1833 static bool prep_compound_gigantic_page(struct page *page, unsigned int order)
1835 return __prep_compound_gigantic_page(page, order, false);
1838 static bool prep_compound_gigantic_page_for_demote(struct page *page,
1841 return __prep_compound_gigantic_page(page, order, true);
1845 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1846 * transparent huge pages. See the PageTransHuge() documentation for more
1849 int PageHuge(struct page *page)
1851 if (!PageCompound(page))
1854 page = compound_head(page);
1855 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1857 EXPORT_SYMBOL_GPL(PageHuge);
1860 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1861 * normal or transparent huge pages.
1863 int PageHeadHuge(struct page *page_head)
1865 if (!PageHead(page_head))
1868 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1870 EXPORT_SYMBOL_GPL(PageHeadHuge);
1873 * Find and lock address space (mapping) in write mode.
1875 * Upon entry, the page is locked which means that page_mapping() is
1876 * stable. Due to locking order, we can only trylock_write. If we can
1877 * not get the lock, simply return NULL to caller.
1879 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1881 struct address_space *mapping = page_mapping(hpage);
1886 if (i_mmap_trylock_write(mapping))
1892 pgoff_t hugetlb_basepage_index(struct page *page)
1894 struct page *page_head = compound_head(page);
1895 pgoff_t index = page_index(page_head);
1896 unsigned long compound_idx;
1898 if (compound_order(page_head) >= MAX_ORDER)
1899 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1901 compound_idx = page - page_head;
1903 return (index << compound_order(page_head)) + compound_idx;
1906 static struct page *alloc_buddy_huge_page(struct hstate *h,
1907 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1908 nodemask_t *node_alloc_noretry)
1910 int order = huge_page_order(h);
1912 bool alloc_try_hard = true;
1915 * By default we always try hard to allocate the page with
1916 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1917 * a loop (to adjust global huge page counts) and previous allocation
1918 * failed, do not continue to try hard on the same node. Use the
1919 * node_alloc_noretry bitmap to manage this state information.
1921 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1922 alloc_try_hard = false;
1923 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1925 gfp_mask |= __GFP_RETRY_MAYFAIL;
1926 if (nid == NUMA_NO_NODE)
1927 nid = numa_mem_id();
1928 page = __alloc_pages(gfp_mask, order, nid, nmask);
1930 __count_vm_event(HTLB_BUDDY_PGALLOC);
1932 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1935 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1936 * indicates an overall state change. Clear bit so that we resume
1937 * normal 'try hard' allocations.
1939 if (node_alloc_noretry && page && !alloc_try_hard)
1940 node_clear(nid, *node_alloc_noretry);
1943 * If we tried hard to get a page but failed, set bit so that
1944 * subsequent attempts will not try as hard until there is an
1945 * overall state change.
1947 if (node_alloc_noretry && !page && alloc_try_hard)
1948 node_set(nid, *node_alloc_noretry);
1954 * Common helper to allocate a fresh hugetlb page. All specific allocators
1955 * should use this function to get new hugetlb pages
1957 static struct page *alloc_fresh_huge_page(struct hstate *h,
1958 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1959 nodemask_t *node_alloc_noretry)
1965 if (hstate_is_gigantic(h))
1966 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1968 page = alloc_buddy_huge_page(h, gfp_mask,
1969 nid, nmask, node_alloc_noretry);
1973 if (hstate_is_gigantic(h)) {
1974 if (!prep_compound_gigantic_page(page, huge_page_order(h))) {
1976 * Rare failure to convert pages to compound page.
1977 * Free pages and try again - ONCE!
1979 free_gigantic_page(page, huge_page_order(h));
1987 prep_new_huge_page(h, page, page_to_nid(page));
1993 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1996 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1997 nodemask_t *node_alloc_noretry)
2001 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2003 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2004 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
2005 node_alloc_noretry);
2013 put_page(page); /* free it into the hugepage allocator */
2019 * Remove huge page from pool from next node to free. Attempt to keep
2020 * persistent huge pages more or less balanced over allowed nodes.
2021 * This routine only 'removes' the hugetlb page. The caller must make
2022 * an additional call to free the page to low level allocators.
2023 * Called with hugetlb_lock locked.
2025 static struct page *remove_pool_huge_page(struct hstate *h,
2026 nodemask_t *nodes_allowed,
2030 struct page *page = NULL;
2032 lockdep_assert_held(&hugetlb_lock);
2033 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2035 * If we're returning unused surplus pages, only examine
2036 * nodes with surplus pages.
2038 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
2039 !list_empty(&h->hugepage_freelists[node])) {
2040 page = list_entry(h->hugepage_freelists[node].next,
2042 remove_hugetlb_page(h, page, acct_surplus);
2051 * Dissolve a given free hugepage into free buddy pages. This function does
2052 * nothing for in-use hugepages and non-hugepages.
2053 * This function returns values like below:
2055 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
2056 * when the system is under memory pressure and the feature of
2057 * freeing unused vmemmap pages associated with each hugetlb page
2059 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
2060 * (allocated or reserved.)
2061 * 0: successfully dissolved free hugepages or the page is not a
2062 * hugepage (considered as already dissolved)
2064 int dissolve_free_huge_page(struct page *page)
2069 /* Not to disrupt normal path by vainly holding hugetlb_lock */
2070 if (!PageHuge(page))
2073 spin_lock_irq(&hugetlb_lock);
2074 if (!PageHuge(page)) {
2079 if (!page_count(page)) {
2080 struct page *head = compound_head(page);
2081 struct hstate *h = page_hstate(head);
2082 if (h->free_huge_pages - h->resv_huge_pages == 0)
2086 * We should make sure that the page is already on the free list
2087 * when it is dissolved.
2089 if (unlikely(!HPageFreed(head))) {
2090 spin_unlock_irq(&hugetlb_lock);
2094 * Theoretically, we should return -EBUSY when we
2095 * encounter this race. In fact, we have a chance
2096 * to successfully dissolve the page if we do a
2097 * retry. Because the race window is quite small.
2098 * If we seize this opportunity, it is an optimization
2099 * for increasing the success rate of dissolving page.
2104 remove_hugetlb_page(h, head, false);
2105 h->max_huge_pages--;
2106 spin_unlock_irq(&hugetlb_lock);
2109 * Normally update_and_free_page will allocate required vmemmmap
2110 * before freeing the page. update_and_free_page will fail to
2111 * free the page if it can not allocate required vmemmap. We
2112 * need to adjust max_huge_pages if the page is not freed.
2113 * Attempt to allocate vmemmmap here so that we can take
2114 * appropriate action on failure.
2116 rc = hugetlb_vmemmap_alloc(h, head);
2119 * Move PageHWPoison flag from head page to the raw
2120 * error page, which makes any subpages rather than
2121 * the error page reusable.
2123 if (PageHWPoison(head) && page != head) {
2124 SetPageHWPoison(page);
2125 ClearPageHWPoison(head);
2127 update_and_free_page(h, head, false);
2129 spin_lock_irq(&hugetlb_lock);
2130 add_hugetlb_page(h, head, false);
2131 h->max_huge_pages++;
2132 spin_unlock_irq(&hugetlb_lock);
2138 spin_unlock_irq(&hugetlb_lock);
2143 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2144 * make specified memory blocks removable from the system.
2145 * Note that this will dissolve a free gigantic hugepage completely, if any
2146 * part of it lies within the given range.
2147 * Also note that if dissolve_free_huge_page() returns with an error, all
2148 * free hugepages that were dissolved before that error are lost.
2150 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2156 if (!hugepages_supported())
2159 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
2160 page = pfn_to_page(pfn);
2161 rc = dissolve_free_huge_page(page);
2170 * Allocates a fresh surplus page from the page allocator.
2172 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
2173 int nid, nodemask_t *nmask, bool zero_ref)
2175 struct page *page = NULL;
2178 if (hstate_is_gigantic(h))
2181 spin_lock_irq(&hugetlb_lock);
2182 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2184 spin_unlock_irq(&hugetlb_lock);
2187 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2191 spin_lock_irq(&hugetlb_lock);
2193 * We could have raced with the pool size change.
2194 * Double check that and simply deallocate the new page
2195 * if we would end up overcommiting the surpluses. Abuse
2196 * temporary page to workaround the nasty free_huge_page
2199 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2200 SetHPageTemporary(page);
2201 spin_unlock_irq(&hugetlb_lock);
2208 * Caller requires a page with zero ref count.
2209 * We will drop ref count here. If someone else is holding
2210 * a ref, the page will be freed when they drop it. Abuse
2211 * temporary page flag to accomplish this.
2213 SetHPageTemporary(page);
2214 if (!put_page_testzero(page)) {
2216 * Unexpected inflated ref count on freshly allocated
2219 pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n");
2220 spin_unlock_irq(&hugetlb_lock);
2227 ClearHPageTemporary(page);
2230 h->surplus_huge_pages++;
2231 h->surplus_huge_pages_node[page_to_nid(page)]++;
2234 spin_unlock_irq(&hugetlb_lock);
2239 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
2240 int nid, nodemask_t *nmask)
2244 if (hstate_is_gigantic(h))
2247 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2252 * We do not account these pages as surplus because they are only
2253 * temporary and will be released properly on the last reference
2255 SetHPageTemporary(page);
2261 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2264 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
2265 struct vm_area_struct *vma, unsigned long addr)
2267 struct page *page = NULL;
2268 struct mempolicy *mpol;
2269 gfp_t gfp_mask = htlb_alloc_mask(h);
2271 nodemask_t *nodemask;
2273 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2274 if (mpol_is_preferred_many(mpol)) {
2275 gfp_t gfp = gfp_mask | __GFP_NOWARN;
2277 gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2278 page = alloc_surplus_huge_page(h, gfp, nid, nodemask, false);
2280 /* Fallback to all nodes if page==NULL */
2285 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask, false);
2286 mpol_cond_put(mpol);
2290 /* page migration callback function */
2291 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
2292 nodemask_t *nmask, gfp_t gfp_mask)
2294 spin_lock_irq(&hugetlb_lock);
2295 if (h->free_huge_pages - h->resv_huge_pages > 0) {
2298 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
2300 spin_unlock_irq(&hugetlb_lock);
2304 spin_unlock_irq(&hugetlb_lock);
2306 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2309 /* mempolicy aware migration callback */
2310 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2311 unsigned long address)
2313 struct mempolicy *mpol;
2314 nodemask_t *nodemask;
2319 gfp_mask = htlb_alloc_mask(h);
2320 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2321 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
2322 mpol_cond_put(mpol);
2328 * Increase the hugetlb pool such that it can accommodate a reservation
2331 static int gather_surplus_pages(struct hstate *h, long delta)
2332 __must_hold(&hugetlb_lock)
2334 struct list_head surplus_list;
2335 struct page *page, *tmp;
2338 long needed, allocated;
2339 bool alloc_ok = true;
2341 lockdep_assert_held(&hugetlb_lock);
2342 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2344 h->resv_huge_pages += delta;
2349 INIT_LIST_HEAD(&surplus_list);
2353 spin_unlock_irq(&hugetlb_lock);
2354 for (i = 0; i < needed; i++) {
2355 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2356 NUMA_NO_NODE, NULL, true);
2361 list_add(&page->lru, &surplus_list);
2367 * After retaking hugetlb_lock, we need to recalculate 'needed'
2368 * because either resv_huge_pages or free_huge_pages may have changed.
2370 spin_lock_irq(&hugetlb_lock);
2371 needed = (h->resv_huge_pages + delta) -
2372 (h->free_huge_pages + allocated);
2377 * We were not able to allocate enough pages to
2378 * satisfy the entire reservation so we free what
2379 * we've allocated so far.
2384 * The surplus_list now contains _at_least_ the number of extra pages
2385 * needed to accommodate the reservation. Add the appropriate number
2386 * of pages to the hugetlb pool and free the extras back to the buddy
2387 * allocator. Commit the entire reservation here to prevent another
2388 * process from stealing the pages as they are added to the pool but
2389 * before they are reserved.
2391 needed += allocated;
2392 h->resv_huge_pages += delta;
2395 /* Free the needed pages to the hugetlb pool */
2396 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2399 /* Add the page to the hugetlb allocator */
2400 enqueue_huge_page(h, page);
2403 spin_unlock_irq(&hugetlb_lock);
2406 * Free unnecessary surplus pages to the buddy allocator.
2407 * Pages have no ref count, call free_huge_page directly.
2409 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2410 free_huge_page(page);
2411 spin_lock_irq(&hugetlb_lock);
2417 * This routine has two main purposes:
2418 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2419 * in unused_resv_pages. This corresponds to the prior adjustments made
2420 * to the associated reservation map.
2421 * 2) Free any unused surplus pages that may have been allocated to satisfy
2422 * the reservation. As many as unused_resv_pages may be freed.
2424 static void return_unused_surplus_pages(struct hstate *h,
2425 unsigned long unused_resv_pages)
2427 unsigned long nr_pages;
2429 LIST_HEAD(page_list);
2431 lockdep_assert_held(&hugetlb_lock);
2432 /* Uncommit the reservation */
2433 h->resv_huge_pages -= unused_resv_pages;
2435 /* Cannot return gigantic pages currently */
2436 if (hstate_is_gigantic(h))
2440 * Part (or even all) of the reservation could have been backed
2441 * by pre-allocated pages. Only free surplus pages.
2443 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2446 * We want to release as many surplus pages as possible, spread
2447 * evenly across all nodes with memory. Iterate across these nodes
2448 * until we can no longer free unreserved surplus pages. This occurs
2449 * when the nodes with surplus pages have no free pages.
2450 * remove_pool_huge_page() will balance the freed pages across the
2451 * on-line nodes with memory and will handle the hstate accounting.
2453 while (nr_pages--) {
2454 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2458 list_add(&page->lru, &page_list);
2462 spin_unlock_irq(&hugetlb_lock);
2463 update_and_free_pages_bulk(h, &page_list);
2464 spin_lock_irq(&hugetlb_lock);
2469 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2470 * are used by the huge page allocation routines to manage reservations.
2472 * vma_needs_reservation is called to determine if the huge page at addr
2473 * within the vma has an associated reservation. If a reservation is
2474 * needed, the value 1 is returned. The caller is then responsible for
2475 * managing the global reservation and subpool usage counts. After
2476 * the huge page has been allocated, vma_commit_reservation is called
2477 * to add the page to the reservation map. If the page allocation fails,
2478 * the reservation must be ended instead of committed. vma_end_reservation
2479 * is called in such cases.
2481 * In the normal case, vma_commit_reservation returns the same value
2482 * as the preceding vma_needs_reservation call. The only time this
2483 * is not the case is if a reserve map was changed between calls. It
2484 * is the responsibility of the caller to notice the difference and
2485 * take appropriate action.
2487 * vma_add_reservation is used in error paths where a reservation must
2488 * be restored when a newly allocated huge page must be freed. It is
2489 * to be called after calling vma_needs_reservation to determine if a
2490 * reservation exists.
2492 * vma_del_reservation is used in error paths where an entry in the reserve
2493 * map was created during huge page allocation and must be removed. It is to
2494 * be called after calling vma_needs_reservation to determine if a reservation
2497 enum vma_resv_mode {
2504 static long __vma_reservation_common(struct hstate *h,
2505 struct vm_area_struct *vma, unsigned long addr,
2506 enum vma_resv_mode mode)
2508 struct resv_map *resv;
2511 long dummy_out_regions_needed;
2513 resv = vma_resv_map(vma);
2517 idx = vma_hugecache_offset(h, vma, addr);
2519 case VMA_NEEDS_RESV:
2520 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2521 /* We assume that vma_reservation_* routines always operate on
2522 * 1 page, and that adding to resv map a 1 page entry can only
2523 * ever require 1 region.
2525 VM_BUG_ON(dummy_out_regions_needed != 1);
2527 case VMA_COMMIT_RESV:
2528 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2529 /* region_add calls of range 1 should never fail. */
2533 region_abort(resv, idx, idx + 1, 1);
2537 if (vma->vm_flags & VM_MAYSHARE) {
2538 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2539 /* region_add calls of range 1 should never fail. */
2542 region_abort(resv, idx, idx + 1, 1);
2543 ret = region_del(resv, idx, idx + 1);
2547 if (vma->vm_flags & VM_MAYSHARE) {
2548 region_abort(resv, idx, idx + 1, 1);
2549 ret = region_del(resv, idx, idx + 1);
2551 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2552 /* region_add calls of range 1 should never fail. */
2560 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2563 * We know private mapping must have HPAGE_RESV_OWNER set.
2565 * In most cases, reserves always exist for private mappings.
2566 * However, a file associated with mapping could have been
2567 * hole punched or truncated after reserves were consumed.
2568 * As subsequent fault on such a range will not use reserves.
2569 * Subtle - The reserve map for private mappings has the
2570 * opposite meaning than that of shared mappings. If NO
2571 * entry is in the reserve map, it means a reservation exists.
2572 * If an entry exists in the reserve map, it means the
2573 * reservation has already been consumed. As a result, the
2574 * return value of this routine is the opposite of the
2575 * value returned from reserve map manipulation routines above.
2584 static long vma_needs_reservation(struct hstate *h,
2585 struct vm_area_struct *vma, unsigned long addr)
2587 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2590 static long vma_commit_reservation(struct hstate *h,
2591 struct vm_area_struct *vma, unsigned long addr)
2593 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2596 static void vma_end_reservation(struct hstate *h,
2597 struct vm_area_struct *vma, unsigned long addr)
2599 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2602 static long vma_add_reservation(struct hstate *h,
2603 struct vm_area_struct *vma, unsigned long addr)
2605 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2608 static long vma_del_reservation(struct hstate *h,
2609 struct vm_area_struct *vma, unsigned long addr)
2611 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2615 * This routine is called to restore reservation information on error paths.
2616 * It should ONLY be called for pages allocated via alloc_huge_page(), and
2617 * the hugetlb mutex should remain held when calling this routine.
2619 * It handles two specific cases:
2620 * 1) A reservation was in place and the page consumed the reservation.
2621 * HPageRestoreReserve is set in the page.
2622 * 2) No reservation was in place for the page, so HPageRestoreReserve is
2623 * not set. However, alloc_huge_page always updates the reserve map.
2625 * In case 1, free_huge_page later in the error path will increment the
2626 * global reserve count. But, free_huge_page does not have enough context
2627 * to adjust the reservation map. This case deals primarily with private
2628 * mappings. Adjust the reserve map here to be consistent with global
2629 * reserve count adjustments to be made by free_huge_page. Make sure the
2630 * reserve map indicates there is a reservation present.
2632 * In case 2, simply undo reserve map modifications done by alloc_huge_page.
2634 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2635 unsigned long address, struct page *page)
2637 long rc = vma_needs_reservation(h, vma, address);
2639 if (HPageRestoreReserve(page)) {
2640 if (unlikely(rc < 0))
2642 * Rare out of memory condition in reserve map
2643 * manipulation. Clear HPageRestoreReserve so that
2644 * global reserve count will not be incremented
2645 * by free_huge_page. This will make it appear
2646 * as though the reservation for this page was
2647 * consumed. This may prevent the task from
2648 * faulting in the page at a later time. This
2649 * is better than inconsistent global huge page
2650 * accounting of reserve counts.
2652 ClearHPageRestoreReserve(page);
2654 (void)vma_add_reservation(h, vma, address);
2656 vma_end_reservation(h, vma, address);
2660 * This indicates there is an entry in the reserve map
2661 * not added by alloc_huge_page. We know it was added
2662 * before the alloc_huge_page call, otherwise
2663 * HPageRestoreReserve would be set on the page.
2664 * Remove the entry so that a subsequent allocation
2665 * does not consume a reservation.
2667 rc = vma_del_reservation(h, vma, address);
2670 * VERY rare out of memory condition. Since
2671 * we can not delete the entry, set
2672 * HPageRestoreReserve so that the reserve
2673 * count will be incremented when the page
2674 * is freed. This reserve will be consumed
2675 * on a subsequent allocation.
2677 SetHPageRestoreReserve(page);
2678 } else if (rc < 0) {
2680 * Rare out of memory condition from
2681 * vma_needs_reservation call. Memory allocation is
2682 * only attempted if a new entry is needed. Therefore,
2683 * this implies there is not an entry in the
2686 * For shared mappings, no entry in the map indicates
2687 * no reservation. We are done.
2689 if (!(vma->vm_flags & VM_MAYSHARE))
2691 * For private mappings, no entry indicates
2692 * a reservation is present. Since we can
2693 * not add an entry, set SetHPageRestoreReserve
2694 * on the page so reserve count will be
2695 * incremented when freed. This reserve will
2696 * be consumed on a subsequent allocation.
2698 SetHPageRestoreReserve(page);
2701 * No reservation present, do nothing
2703 vma_end_reservation(h, vma, address);
2708 * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one
2709 * @h: struct hstate old page belongs to
2710 * @old_page: Old page to dissolve
2711 * @list: List to isolate the page in case we need to
2712 * Returns 0 on success, otherwise negated error.
2714 static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page,
2715 struct list_head *list)
2717 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2718 int nid = page_to_nid(old_page);
2719 bool alloc_retry = false;
2720 struct page *new_page;
2724 * Before dissolving the page, we need to allocate a new one for the
2725 * pool to remain stable. Here, we allocate the page and 'prep' it
2726 * by doing everything but actually updating counters and adding to
2727 * the pool. This simplifies and let us do most of the processing
2731 new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL);
2735 * If all goes well, this page will be directly added to the free
2736 * list in the pool. For this the ref count needs to be zero.
2737 * Attempt to drop now, and retry once if needed. It is VERY
2738 * unlikely there is another ref on the page.
2740 * If someone else has a reference to the page, it will be freed
2741 * when they drop their ref. Abuse temporary page flag to accomplish
2742 * this. Retry once if there is an inflated ref count.
2744 SetHPageTemporary(new_page);
2745 if (!put_page_testzero(new_page)) {
2752 ClearHPageTemporary(new_page);
2754 __prep_new_huge_page(h, new_page);
2757 spin_lock_irq(&hugetlb_lock);
2758 if (!PageHuge(old_page)) {
2760 * Freed from under us. Drop new_page too.
2763 } else if (page_count(old_page)) {
2765 * Someone has grabbed the page, try to isolate it here.
2766 * Fail with -EBUSY if not possible.
2768 spin_unlock_irq(&hugetlb_lock);
2769 ret = isolate_hugetlb(old_page, list);
2770 spin_lock_irq(&hugetlb_lock);
2772 } else if (!HPageFreed(old_page)) {
2774 * Page's refcount is 0 but it has not been enqueued in the
2775 * freelist yet. Race window is small, so we can succeed here if
2778 spin_unlock_irq(&hugetlb_lock);
2783 * Ok, old_page is still a genuine free hugepage. Remove it from
2784 * the freelist and decrease the counters. These will be
2785 * incremented again when calling __prep_account_new_huge_page()
2786 * and enqueue_huge_page() for new_page. The counters will remain
2787 * stable since this happens under the lock.
2789 remove_hugetlb_page(h, old_page, false);
2792 * Ref count on new page is already zero as it was dropped
2793 * earlier. It can be directly added to the pool free list.
2795 __prep_account_new_huge_page(h, nid);
2796 enqueue_huge_page(h, new_page);
2799 * Pages have been replaced, we can safely free the old one.
2801 spin_unlock_irq(&hugetlb_lock);
2802 update_and_free_page(h, old_page, false);
2808 spin_unlock_irq(&hugetlb_lock);
2809 /* Page has a zero ref count, but needs a ref to be freed */
2810 set_page_refcounted(new_page);
2811 update_and_free_page(h, new_page, false);
2816 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
2823 * The page might have been dissolved from under our feet, so make sure
2824 * to carefully check the state under the lock.
2825 * Return success when racing as if we dissolved the page ourselves.
2827 spin_lock_irq(&hugetlb_lock);
2828 if (PageHuge(page)) {
2829 head = compound_head(page);
2830 h = page_hstate(head);
2832 spin_unlock_irq(&hugetlb_lock);
2835 spin_unlock_irq(&hugetlb_lock);
2838 * Fence off gigantic pages as there is a cyclic dependency between
2839 * alloc_contig_range and them. Return -ENOMEM as this has the effect
2840 * of bailing out right away without further retrying.
2842 if (hstate_is_gigantic(h))
2845 if (page_count(head) && !isolate_hugetlb(head, list))
2847 else if (!page_count(head))
2848 ret = alloc_and_dissolve_huge_page(h, head, list);
2853 struct page *alloc_huge_page(struct vm_area_struct *vma,
2854 unsigned long addr, int avoid_reserve)
2856 struct hugepage_subpool *spool = subpool_vma(vma);
2857 struct hstate *h = hstate_vma(vma);
2859 long map_chg, map_commit;
2862 struct hugetlb_cgroup *h_cg;
2863 bool deferred_reserve;
2865 idx = hstate_index(h);
2867 * Examine the region/reserve map to determine if the process
2868 * has a reservation for the page to be allocated. A return
2869 * code of zero indicates a reservation exists (no change).
2871 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2873 return ERR_PTR(-ENOMEM);
2876 * Processes that did not create the mapping will have no
2877 * reserves as indicated by the region/reserve map. Check
2878 * that the allocation will not exceed the subpool limit.
2879 * Allocations for MAP_NORESERVE mappings also need to be
2880 * checked against any subpool limit.
2882 if (map_chg || avoid_reserve) {
2883 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2885 vma_end_reservation(h, vma, addr);
2886 return ERR_PTR(-ENOSPC);
2890 * Even though there was no reservation in the region/reserve
2891 * map, there could be reservations associated with the
2892 * subpool that can be used. This would be indicated if the
2893 * return value of hugepage_subpool_get_pages() is zero.
2894 * However, if avoid_reserve is specified we still avoid even
2895 * the subpool reservations.
2901 /* If this allocation is not consuming a reservation, charge it now.
2903 deferred_reserve = map_chg || avoid_reserve;
2904 if (deferred_reserve) {
2905 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2906 idx, pages_per_huge_page(h), &h_cg);
2908 goto out_subpool_put;
2911 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2913 goto out_uncharge_cgroup_reservation;
2915 spin_lock_irq(&hugetlb_lock);
2917 * glb_chg is passed to indicate whether or not a page must be taken
2918 * from the global free pool (global change). gbl_chg == 0 indicates
2919 * a reservation exists for the allocation.
2921 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2923 spin_unlock_irq(&hugetlb_lock);
2924 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2926 goto out_uncharge_cgroup;
2927 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2928 SetHPageRestoreReserve(page);
2929 h->resv_huge_pages--;
2931 spin_lock_irq(&hugetlb_lock);
2932 list_add(&page->lru, &h->hugepage_activelist);
2935 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2936 /* If allocation is not consuming a reservation, also store the
2937 * hugetlb_cgroup pointer on the page.
2939 if (deferred_reserve) {
2940 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2944 spin_unlock_irq(&hugetlb_lock);
2946 hugetlb_set_page_subpool(page, spool);
2948 map_commit = vma_commit_reservation(h, vma, addr);
2949 if (unlikely(map_chg > map_commit)) {
2951 * The page was added to the reservation map between
2952 * vma_needs_reservation and vma_commit_reservation.
2953 * This indicates a race with hugetlb_reserve_pages.
2954 * Adjust for the subpool count incremented above AND
2955 * in hugetlb_reserve_pages for the same page. Also,
2956 * the reservation count added in hugetlb_reserve_pages
2957 * no longer applies.
2961 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2962 hugetlb_acct_memory(h, -rsv_adjust);
2963 if (deferred_reserve)
2964 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2965 pages_per_huge_page(h), page);
2969 out_uncharge_cgroup:
2970 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2971 out_uncharge_cgroup_reservation:
2972 if (deferred_reserve)
2973 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2976 if (map_chg || avoid_reserve)
2977 hugepage_subpool_put_pages(spool, 1);
2978 vma_end_reservation(h, vma, addr);
2979 return ERR_PTR(-ENOSPC);
2982 int alloc_bootmem_huge_page(struct hstate *h, int nid)
2983 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2984 int __alloc_bootmem_huge_page(struct hstate *h, int nid)
2986 struct huge_bootmem_page *m = NULL; /* initialize for clang */
2989 /* do node specific alloc */
2990 if (nid != NUMA_NO_NODE) {
2991 m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h),
2992 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
2997 /* allocate from next node when distributing huge pages */
2998 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2999 m = memblock_alloc_try_nid_raw(
3000 huge_page_size(h), huge_page_size(h),
3001 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
3003 * Use the beginning of the huge page to store the
3004 * huge_bootmem_page struct (until gather_bootmem
3005 * puts them into the mem_map).
3013 /* Put them into a private list first because mem_map is not up yet */
3014 INIT_LIST_HEAD(&m->list);
3015 list_add(&m->list, &huge_boot_pages);
3021 * Put bootmem huge pages into the standard lists after mem_map is up.
3022 * Note: This only applies to gigantic (order > MAX_ORDER) pages.
3024 static void __init gather_bootmem_prealloc(void)
3026 struct huge_bootmem_page *m;
3028 list_for_each_entry(m, &huge_boot_pages, list) {
3029 struct page *page = virt_to_page(m);
3030 struct hstate *h = m->hstate;
3032 VM_BUG_ON(!hstate_is_gigantic(h));
3033 WARN_ON(page_count(page) != 1);
3034 if (prep_compound_gigantic_page(page, huge_page_order(h))) {
3035 WARN_ON(PageReserved(page));
3036 prep_new_huge_page(h, page, page_to_nid(page));
3037 put_page(page); /* add to the hugepage allocator */
3039 /* VERY unlikely inflated ref count on a tail page */
3040 free_gigantic_page(page, huge_page_order(h));
3044 * We need to restore the 'stolen' pages to totalram_pages
3045 * in order to fix confusing memory reports from free(1) and
3046 * other side-effects, like CommitLimit going negative.
3048 adjust_managed_page_count(page, pages_per_huge_page(h));
3052 static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)
3057 for (i = 0; i < h->max_huge_pages_node[nid]; ++i) {
3058 if (hstate_is_gigantic(h)) {
3059 if (!alloc_bootmem_huge_page(h, nid))
3063 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3065 page = alloc_fresh_huge_page(h, gfp_mask, nid,
3066 &node_states[N_MEMORY], NULL);
3069 put_page(page); /* free it into the hugepage allocator */
3073 if (i == h->max_huge_pages_node[nid])
3076 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3077 pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n",
3078 h->max_huge_pages_node[nid], buf, nid, i);
3079 h->max_huge_pages -= (h->max_huge_pages_node[nid] - i);
3080 h->max_huge_pages_node[nid] = i;
3083 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
3086 nodemask_t *node_alloc_noretry;
3087 bool node_specific_alloc = false;
3089 /* skip gigantic hugepages allocation if hugetlb_cma enabled */
3090 if (hstate_is_gigantic(h) && hugetlb_cma_size) {
3091 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
3095 /* do node specific alloc */
3096 for_each_online_node(i) {
3097 if (h->max_huge_pages_node[i] > 0) {
3098 hugetlb_hstate_alloc_pages_onenode(h, i);
3099 node_specific_alloc = true;
3103 if (node_specific_alloc)
3106 /* below will do all node balanced alloc */
3107 if (!hstate_is_gigantic(h)) {
3109 * Bit mask controlling how hard we retry per-node allocations.
3110 * Ignore errors as lower level routines can deal with
3111 * node_alloc_noretry == NULL. If this kmalloc fails at boot
3112 * time, we are likely in bigger trouble.
3114 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
3117 /* allocations done at boot time */
3118 node_alloc_noretry = NULL;
3121 /* bit mask controlling how hard we retry per-node allocations */
3122 if (node_alloc_noretry)
3123 nodes_clear(*node_alloc_noretry);
3125 for (i = 0; i < h->max_huge_pages; ++i) {
3126 if (hstate_is_gigantic(h)) {
3127 if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))
3129 } else if (!alloc_pool_huge_page(h,
3130 &node_states[N_MEMORY],
3131 node_alloc_noretry))
3135 if (i < h->max_huge_pages) {
3138 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3139 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
3140 h->max_huge_pages, buf, i);
3141 h->max_huge_pages = i;
3143 kfree(node_alloc_noretry);
3146 static void __init hugetlb_init_hstates(void)
3148 struct hstate *h, *h2;
3150 for_each_hstate(h) {
3151 if (minimum_order > huge_page_order(h))
3152 minimum_order = huge_page_order(h);
3154 /* oversize hugepages were init'ed in early boot */
3155 if (!hstate_is_gigantic(h))
3156 hugetlb_hstate_alloc_pages(h);
3159 * Set demote order for each hstate. Note that
3160 * h->demote_order is initially 0.
3161 * - We can not demote gigantic pages if runtime freeing
3162 * is not supported, so skip this.
3163 * - If CMA allocation is possible, we can not demote
3164 * HUGETLB_PAGE_ORDER or smaller size pages.
3166 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3168 if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER)
3170 for_each_hstate(h2) {
3173 if (h2->order < h->order &&
3174 h2->order > h->demote_order)
3175 h->demote_order = h2->order;
3178 VM_BUG_ON(minimum_order == UINT_MAX);
3181 static void __init report_hugepages(void)
3185 for_each_hstate(h) {
3188 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3189 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
3190 buf, h->free_huge_pages);
3194 #ifdef CONFIG_HIGHMEM
3195 static void try_to_free_low(struct hstate *h, unsigned long count,
3196 nodemask_t *nodes_allowed)
3199 LIST_HEAD(page_list);
3201 lockdep_assert_held(&hugetlb_lock);
3202 if (hstate_is_gigantic(h))
3206 * Collect pages to be freed on a list, and free after dropping lock
3208 for_each_node_mask(i, *nodes_allowed) {
3209 struct page *page, *next;
3210 struct list_head *freel = &h->hugepage_freelists[i];
3211 list_for_each_entry_safe(page, next, freel, lru) {
3212 if (count >= h->nr_huge_pages)
3214 if (PageHighMem(page))
3216 remove_hugetlb_page(h, page, false);
3217 list_add(&page->lru, &page_list);
3222 spin_unlock_irq(&hugetlb_lock);
3223 update_and_free_pages_bulk(h, &page_list);
3224 spin_lock_irq(&hugetlb_lock);
3227 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3228 nodemask_t *nodes_allowed)
3234 * Increment or decrement surplus_huge_pages. Keep node-specific counters
3235 * balanced by operating on them in a round-robin fashion.
3236 * Returns 1 if an adjustment was made.
3238 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3243 lockdep_assert_held(&hugetlb_lock);
3244 VM_BUG_ON(delta != -1 && delta != 1);
3247 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3248 if (h->surplus_huge_pages_node[node])
3252 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3253 if (h->surplus_huge_pages_node[node] <
3254 h->nr_huge_pages_node[node])
3261 h->surplus_huge_pages += delta;
3262 h->surplus_huge_pages_node[node] += delta;
3266 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3267 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3268 nodemask_t *nodes_allowed)
3270 unsigned long min_count, ret;
3272 LIST_HEAD(page_list);
3273 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3276 * Bit mask controlling how hard we retry per-node allocations.
3277 * If we can not allocate the bit mask, do not attempt to allocate
3278 * the requested huge pages.
3280 if (node_alloc_noretry)
3281 nodes_clear(*node_alloc_noretry);
3286 * resize_lock mutex prevents concurrent adjustments to number of
3287 * pages in hstate via the proc/sysfs interfaces.
3289 mutex_lock(&h->resize_lock);
3290 flush_free_hpage_work(h);
3291 spin_lock_irq(&hugetlb_lock);
3294 * Check for a node specific request.
3295 * Changing node specific huge page count may require a corresponding
3296 * change to the global count. In any case, the passed node mask
3297 * (nodes_allowed) will restrict alloc/free to the specified node.
3299 if (nid != NUMA_NO_NODE) {
3300 unsigned long old_count = count;
3302 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3304 * User may have specified a large count value which caused the
3305 * above calculation to overflow. In this case, they wanted
3306 * to allocate as many huge pages as possible. Set count to
3307 * largest possible value to align with their intention.
3309 if (count < old_count)
3314 * Gigantic pages runtime allocation depend on the capability for large
3315 * page range allocation.
3316 * If the system does not provide this feature, return an error when
3317 * the user tries to allocate gigantic pages but let the user free the
3318 * boottime allocated gigantic pages.
3320 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3321 if (count > persistent_huge_pages(h)) {
3322 spin_unlock_irq(&hugetlb_lock);
3323 mutex_unlock(&h->resize_lock);
3324 NODEMASK_FREE(node_alloc_noretry);
3327 /* Fall through to decrease pool */
3331 * Increase the pool size
3332 * First take pages out of surplus state. Then make up the
3333 * remaining difference by allocating fresh huge pages.
3335 * We might race with alloc_surplus_huge_page() here and be unable
3336 * to convert a surplus huge page to a normal huge page. That is
3337 * not critical, though, it just means the overall size of the
3338 * pool might be one hugepage larger than it needs to be, but
3339 * within all the constraints specified by the sysctls.
3341 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3342 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3346 while (count > persistent_huge_pages(h)) {
3348 * If this allocation races such that we no longer need the
3349 * page, free_huge_page will handle it by freeing the page
3350 * and reducing the surplus.
3352 spin_unlock_irq(&hugetlb_lock);
3354 /* yield cpu to avoid soft lockup */
3357 ret = alloc_pool_huge_page(h, nodes_allowed,
3358 node_alloc_noretry);
3359 spin_lock_irq(&hugetlb_lock);
3363 /* Bail for signals. Probably ctrl-c from user */
3364 if (signal_pending(current))
3369 * Decrease the pool size
3370 * First return free pages to the buddy allocator (being careful
3371 * to keep enough around to satisfy reservations). Then place
3372 * pages into surplus state as needed so the pool will shrink
3373 * to the desired size as pages become free.
3375 * By placing pages into the surplus state independent of the
3376 * overcommit value, we are allowing the surplus pool size to
3377 * exceed overcommit. There are few sane options here. Since
3378 * alloc_surplus_huge_page() is checking the global counter,
3379 * though, we'll note that we're not allowed to exceed surplus
3380 * and won't grow the pool anywhere else. Not until one of the
3381 * sysctls are changed, or the surplus pages go out of use.
3383 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3384 min_count = max(count, min_count);
3385 try_to_free_low(h, min_count, nodes_allowed);
3388 * Collect pages to be removed on list without dropping lock
3390 while (min_count < persistent_huge_pages(h)) {
3391 page = remove_pool_huge_page(h, nodes_allowed, 0);
3395 list_add(&page->lru, &page_list);
3397 /* free the pages after dropping lock */
3398 spin_unlock_irq(&hugetlb_lock);
3399 update_and_free_pages_bulk(h, &page_list);
3400 flush_free_hpage_work(h);
3401 spin_lock_irq(&hugetlb_lock);
3403 while (count < persistent_huge_pages(h)) {
3404 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3408 h->max_huge_pages = persistent_huge_pages(h);
3409 spin_unlock_irq(&hugetlb_lock);
3410 mutex_unlock(&h->resize_lock);
3412 NODEMASK_FREE(node_alloc_noretry);
3417 static int demote_free_huge_page(struct hstate *h, struct page *page)
3419 int i, nid = page_to_nid(page);
3420 struct hstate *target_hstate;
3423 target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order);
3425 remove_hugetlb_page_for_demote(h, page, false);
3426 spin_unlock_irq(&hugetlb_lock);
3428 rc = hugetlb_vmemmap_alloc(h, page);
3430 /* Allocation of vmemmmap failed, we can not demote page */
3431 spin_lock_irq(&hugetlb_lock);
3432 set_page_refcounted(page);
3433 add_hugetlb_page(h, page, false);
3438 * Use destroy_compound_hugetlb_page_for_demote for all huge page
3439 * sizes as it will not ref count pages.
3441 destroy_compound_hugetlb_page_for_demote(page, huge_page_order(h));
3444 * Taking target hstate mutex synchronizes with set_max_huge_pages.
3445 * Without the mutex, pages added to target hstate could be marked
3448 * Note that we already hold h->resize_lock. To prevent deadlock,
3449 * use the convention of always taking larger size hstate mutex first.
3451 mutex_lock(&target_hstate->resize_lock);
3452 for (i = 0; i < pages_per_huge_page(h);
3453 i += pages_per_huge_page(target_hstate)) {
3454 if (hstate_is_gigantic(target_hstate))
3455 prep_compound_gigantic_page_for_demote(page + i,
3456 target_hstate->order);
3458 prep_compound_page(page + i, target_hstate->order);
3459 set_page_private(page + i, 0);
3460 set_page_refcounted(page + i);
3461 prep_new_huge_page(target_hstate, page + i, nid);
3464 mutex_unlock(&target_hstate->resize_lock);
3466 spin_lock_irq(&hugetlb_lock);
3469 * Not absolutely necessary, but for consistency update max_huge_pages
3470 * based on pool changes for the demoted page.
3472 h->max_huge_pages--;
3473 target_hstate->max_huge_pages += pages_per_huge_page(h);
3478 static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
3479 __must_hold(&hugetlb_lock)
3484 lockdep_assert_held(&hugetlb_lock);
3486 /* We should never get here if no demote order */
3487 if (!h->demote_order) {
3488 pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n");
3489 return -EINVAL; /* internal error */
3492 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3493 list_for_each_entry(page, &h->hugepage_freelists[node], lru) {
3494 if (PageHWPoison(page))
3497 return demote_free_huge_page(h, page);
3502 * Only way to get here is if all pages on free lists are poisoned.
3503 * Return -EBUSY so that caller will not retry.
3508 #define HSTATE_ATTR_RO(_name) \
3509 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3511 #define HSTATE_ATTR_WO(_name) \
3512 static struct kobj_attribute _name##_attr = __ATTR_WO(_name)
3514 #define HSTATE_ATTR(_name) \
3515 static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
3517 static struct kobject *hugepages_kobj;
3518 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3520 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3522 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3526 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3527 if (hstate_kobjs[i] == kobj) {
3529 *nidp = NUMA_NO_NODE;
3533 return kobj_to_node_hstate(kobj, nidp);
3536 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3537 struct kobj_attribute *attr, char *buf)
3540 unsigned long nr_huge_pages;
3543 h = kobj_to_hstate(kobj, &nid);
3544 if (nid == NUMA_NO_NODE)
3545 nr_huge_pages = h->nr_huge_pages;
3547 nr_huge_pages = h->nr_huge_pages_node[nid];
3549 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3552 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3553 struct hstate *h, int nid,
3554 unsigned long count, size_t len)
3557 nodemask_t nodes_allowed, *n_mask;
3559 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3562 if (nid == NUMA_NO_NODE) {
3564 * global hstate attribute
3566 if (!(obey_mempolicy &&
3567 init_nodemask_of_mempolicy(&nodes_allowed)))
3568 n_mask = &node_states[N_MEMORY];
3570 n_mask = &nodes_allowed;
3573 * Node specific request. count adjustment happens in
3574 * set_max_huge_pages() after acquiring hugetlb_lock.
3576 init_nodemask_of_node(&nodes_allowed, nid);
3577 n_mask = &nodes_allowed;
3580 err = set_max_huge_pages(h, count, nid, n_mask);
3582 return err ? err : len;
3585 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3586 struct kobject *kobj, const char *buf,
3590 unsigned long count;
3594 err = kstrtoul(buf, 10, &count);
3598 h = kobj_to_hstate(kobj, &nid);
3599 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3602 static ssize_t nr_hugepages_show(struct kobject *kobj,
3603 struct kobj_attribute *attr, char *buf)
3605 return nr_hugepages_show_common(kobj, attr, buf);
3608 static ssize_t nr_hugepages_store(struct kobject *kobj,
3609 struct kobj_attribute *attr, const char *buf, size_t len)
3611 return nr_hugepages_store_common(false, kobj, buf, len);
3613 HSTATE_ATTR(nr_hugepages);
3618 * hstate attribute for optionally mempolicy-based constraint on persistent
3619 * huge page alloc/free.
3621 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3622 struct kobj_attribute *attr,
3625 return nr_hugepages_show_common(kobj, attr, buf);
3628 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3629 struct kobj_attribute *attr, const char *buf, size_t len)
3631 return nr_hugepages_store_common(true, kobj, buf, len);
3633 HSTATE_ATTR(nr_hugepages_mempolicy);
3637 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3638 struct kobj_attribute *attr, char *buf)
3640 struct hstate *h = kobj_to_hstate(kobj, NULL);
3641 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3644 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3645 struct kobj_attribute *attr, const char *buf, size_t count)
3648 unsigned long input;
3649 struct hstate *h = kobj_to_hstate(kobj, NULL);
3651 if (hstate_is_gigantic(h))
3654 err = kstrtoul(buf, 10, &input);
3658 spin_lock_irq(&hugetlb_lock);
3659 h->nr_overcommit_huge_pages = input;
3660 spin_unlock_irq(&hugetlb_lock);
3664 HSTATE_ATTR(nr_overcommit_hugepages);
3666 static ssize_t free_hugepages_show(struct kobject *kobj,
3667 struct kobj_attribute *attr, char *buf)
3670 unsigned long free_huge_pages;
3673 h = kobj_to_hstate(kobj, &nid);
3674 if (nid == NUMA_NO_NODE)
3675 free_huge_pages = h->free_huge_pages;
3677 free_huge_pages = h->free_huge_pages_node[nid];
3679 return sysfs_emit(buf, "%lu\n", free_huge_pages);
3681 HSTATE_ATTR_RO(free_hugepages);
3683 static ssize_t resv_hugepages_show(struct kobject *kobj,
3684 struct kobj_attribute *attr, char *buf)
3686 struct hstate *h = kobj_to_hstate(kobj, NULL);
3687 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3689 HSTATE_ATTR_RO(resv_hugepages);
3691 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3692 struct kobj_attribute *attr, char *buf)
3695 unsigned long surplus_huge_pages;
3698 h = kobj_to_hstate(kobj, &nid);
3699 if (nid == NUMA_NO_NODE)
3700 surplus_huge_pages = h->surplus_huge_pages;
3702 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3704 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3706 HSTATE_ATTR_RO(surplus_hugepages);
3708 static ssize_t demote_store(struct kobject *kobj,
3709 struct kobj_attribute *attr, const char *buf, size_t len)
3711 unsigned long nr_demote;
3712 unsigned long nr_available;
3713 nodemask_t nodes_allowed, *n_mask;
3718 err = kstrtoul(buf, 10, &nr_demote);
3721 h = kobj_to_hstate(kobj, &nid);
3723 if (nid != NUMA_NO_NODE) {
3724 init_nodemask_of_node(&nodes_allowed, nid);
3725 n_mask = &nodes_allowed;
3727 n_mask = &node_states[N_MEMORY];
3730 /* Synchronize with other sysfs operations modifying huge pages */
3731 mutex_lock(&h->resize_lock);
3732 spin_lock_irq(&hugetlb_lock);
3736 * Check for available pages to demote each time thorough the
3737 * loop as demote_pool_huge_page will drop hugetlb_lock.
3739 if (nid != NUMA_NO_NODE)
3740 nr_available = h->free_huge_pages_node[nid];
3742 nr_available = h->free_huge_pages;
3743 nr_available -= h->resv_huge_pages;
3747 err = demote_pool_huge_page(h, n_mask);
3754 spin_unlock_irq(&hugetlb_lock);
3755 mutex_unlock(&h->resize_lock);
3761 HSTATE_ATTR_WO(demote);
3763 static ssize_t demote_size_show(struct kobject *kobj,
3764 struct kobj_attribute *attr, char *buf)
3767 struct hstate *h = kobj_to_hstate(kobj, &nid);
3768 unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K;
3770 return sysfs_emit(buf, "%lukB\n", demote_size);
3773 static ssize_t demote_size_store(struct kobject *kobj,
3774 struct kobj_attribute *attr,
3775 const char *buf, size_t count)
3777 struct hstate *h, *demote_hstate;
3778 unsigned long demote_size;
3779 unsigned int demote_order;
3782 demote_size = (unsigned long)memparse(buf, NULL);
3784 demote_hstate = size_to_hstate(demote_size);
3787 demote_order = demote_hstate->order;
3788 if (demote_order < HUGETLB_PAGE_ORDER)
3791 /* demote order must be smaller than hstate order */
3792 h = kobj_to_hstate(kobj, &nid);
3793 if (demote_order >= h->order)
3796 /* resize_lock synchronizes access to demote size and writes */
3797 mutex_lock(&h->resize_lock);
3798 h->demote_order = demote_order;
3799 mutex_unlock(&h->resize_lock);
3803 HSTATE_ATTR(demote_size);
3805 static struct attribute *hstate_attrs[] = {
3806 &nr_hugepages_attr.attr,
3807 &nr_overcommit_hugepages_attr.attr,
3808 &free_hugepages_attr.attr,
3809 &resv_hugepages_attr.attr,
3810 &surplus_hugepages_attr.attr,
3812 &nr_hugepages_mempolicy_attr.attr,
3817 static const struct attribute_group hstate_attr_group = {
3818 .attrs = hstate_attrs,
3821 static struct attribute *hstate_demote_attrs[] = {
3822 &demote_size_attr.attr,
3827 static const struct attribute_group hstate_demote_attr_group = {
3828 .attrs = hstate_demote_attrs,
3831 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3832 struct kobject **hstate_kobjs,
3833 const struct attribute_group *hstate_attr_group)
3836 int hi = hstate_index(h);
3838 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3839 if (!hstate_kobjs[hi])
3842 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3844 kobject_put(hstate_kobjs[hi]);
3845 hstate_kobjs[hi] = NULL;
3848 if (h->demote_order) {
3849 if (sysfs_create_group(hstate_kobjs[hi],
3850 &hstate_demote_attr_group))
3851 pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name);
3857 static void __init hugetlb_sysfs_init(void)
3862 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3863 if (!hugepages_kobj)
3866 for_each_hstate(h) {
3867 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3868 hstate_kobjs, &hstate_attr_group);
3870 pr_err("HugeTLB: Unable to add hstate %s", h->name);
3877 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3878 * with node devices in node_devices[] using a parallel array. The array
3879 * index of a node device or _hstate == node id.
3880 * This is here to avoid any static dependency of the node device driver, in
3881 * the base kernel, on the hugetlb module.
3883 struct node_hstate {
3884 struct kobject *hugepages_kobj;
3885 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3887 static struct node_hstate node_hstates[MAX_NUMNODES];
3890 * A subset of global hstate attributes for node devices
3892 static struct attribute *per_node_hstate_attrs[] = {
3893 &nr_hugepages_attr.attr,
3894 &free_hugepages_attr.attr,
3895 &surplus_hugepages_attr.attr,
3899 static const struct attribute_group per_node_hstate_attr_group = {
3900 .attrs = per_node_hstate_attrs,
3904 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3905 * Returns node id via non-NULL nidp.
3907 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3911 for (nid = 0; nid < nr_node_ids; nid++) {
3912 struct node_hstate *nhs = &node_hstates[nid];
3914 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3915 if (nhs->hstate_kobjs[i] == kobj) {
3927 * Unregister hstate attributes from a single node device.
3928 * No-op if no hstate attributes attached.
3930 static void hugetlb_unregister_node(struct node *node)
3933 struct node_hstate *nhs = &node_hstates[node->dev.id];
3935 if (!nhs->hugepages_kobj)
3936 return; /* no hstate attributes */
3938 for_each_hstate(h) {
3939 int idx = hstate_index(h);
3940 if (nhs->hstate_kobjs[idx]) {
3941 kobject_put(nhs->hstate_kobjs[idx]);
3942 nhs->hstate_kobjs[idx] = NULL;
3946 kobject_put(nhs->hugepages_kobj);
3947 nhs->hugepages_kobj = NULL;
3952 * Register hstate attributes for a single node device.
3953 * No-op if attributes already registered.
3955 static void hugetlb_register_node(struct node *node)
3958 struct node_hstate *nhs = &node_hstates[node->dev.id];
3961 if (nhs->hugepages_kobj)
3962 return; /* already allocated */
3964 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3966 if (!nhs->hugepages_kobj)
3969 for_each_hstate(h) {
3970 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3972 &per_node_hstate_attr_group);
3974 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3975 h->name, node->dev.id);
3976 hugetlb_unregister_node(node);
3983 * hugetlb init time: register hstate attributes for all registered node
3984 * devices of nodes that have memory. All on-line nodes should have
3985 * registered their associated device by this time.
3987 static void __init hugetlb_register_all_nodes(void)
3991 for_each_node_state(nid, N_MEMORY) {
3992 struct node *node = node_devices[nid];
3993 if (node->dev.id == nid)
3994 hugetlb_register_node(node);
3998 * Let the node device driver know we're here so it can
3999 * [un]register hstate attributes on node hotplug.
4001 register_hugetlbfs_with_node(hugetlb_register_node,
4002 hugetlb_unregister_node);
4004 #else /* !CONFIG_NUMA */
4006 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4014 static void hugetlb_register_all_nodes(void) { }
4018 static int __init hugetlb_init(void)
4022 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
4025 if (!hugepages_supported()) {
4026 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
4027 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
4032 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
4033 * architectures depend on setup being done here.
4035 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
4036 if (!parsed_default_hugepagesz) {
4038 * If we did not parse a default huge page size, set
4039 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
4040 * number of huge pages for this default size was implicitly
4041 * specified, set that here as well.
4042 * Note that the implicit setting will overwrite an explicit
4043 * setting. A warning will be printed in this case.
4045 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
4046 if (default_hstate_max_huge_pages) {
4047 if (default_hstate.max_huge_pages) {
4050 string_get_size(huge_page_size(&default_hstate),
4051 1, STRING_UNITS_2, buf, 32);
4052 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
4053 default_hstate.max_huge_pages, buf);
4054 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
4055 default_hstate_max_huge_pages);
4057 default_hstate.max_huge_pages =
4058 default_hstate_max_huge_pages;
4060 for_each_online_node(i)
4061 default_hstate.max_huge_pages_node[i] =
4062 default_hugepages_in_node[i];
4066 hugetlb_cma_check();
4067 hugetlb_init_hstates();
4068 gather_bootmem_prealloc();
4071 hugetlb_sysfs_init();
4072 hugetlb_register_all_nodes();
4073 hugetlb_cgroup_file_init();
4076 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
4078 num_fault_mutexes = 1;
4080 hugetlb_fault_mutex_table =
4081 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
4083 BUG_ON(!hugetlb_fault_mutex_table);
4085 for (i = 0; i < num_fault_mutexes; i++)
4086 mutex_init(&hugetlb_fault_mutex_table[i]);
4089 subsys_initcall(hugetlb_init);
4091 /* Overwritten by architectures with more huge page sizes */
4092 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
4094 return size == HPAGE_SIZE;
4097 void __init hugetlb_add_hstate(unsigned int order)
4102 if (size_to_hstate(PAGE_SIZE << order)) {
4105 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
4107 h = &hstates[hugetlb_max_hstate++];
4108 mutex_init(&h->resize_lock);
4110 h->mask = ~(huge_page_size(h) - 1);
4111 for (i = 0; i < MAX_NUMNODES; ++i)
4112 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
4113 INIT_LIST_HEAD(&h->hugepage_activelist);
4114 h->next_nid_to_alloc = first_memory_node;
4115 h->next_nid_to_free = first_memory_node;
4116 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
4117 huge_page_size(h)/1024);
4118 hugetlb_vmemmap_init(h);
4123 bool __init __weak hugetlb_node_alloc_supported(void)
4128 static void __init hugepages_clear_pages_in_node(void)
4130 if (!hugetlb_max_hstate) {
4131 default_hstate_max_huge_pages = 0;
4132 memset(default_hugepages_in_node, 0,
4133 MAX_NUMNODES * sizeof(unsigned int));
4135 parsed_hstate->max_huge_pages = 0;
4136 memset(parsed_hstate->max_huge_pages_node, 0,
4137 MAX_NUMNODES * sizeof(unsigned int));
4142 * hugepages command line processing
4143 * hugepages normally follows a valid hugepagsz or default_hugepagsz
4144 * specification. If not, ignore the hugepages value. hugepages can also
4145 * be the first huge page command line option in which case it implicitly
4146 * specifies the number of huge pages for the default size.
4148 static int __init hugepages_setup(char *s)
4151 static unsigned long *last_mhp;
4152 int node = NUMA_NO_NODE;
4157 if (!parsed_valid_hugepagesz) {
4158 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
4159 parsed_valid_hugepagesz = true;
4164 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
4165 * yet, so this hugepages= parameter goes to the "default hstate".
4166 * Otherwise, it goes with the previously parsed hugepagesz or
4167 * default_hugepagesz.
4169 else if (!hugetlb_max_hstate)
4170 mhp = &default_hstate_max_huge_pages;
4172 mhp = &parsed_hstate->max_huge_pages;
4174 if (mhp == last_mhp) {
4175 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
4181 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4183 /* Parameter is node format */
4184 if (p[count] == ':') {
4185 if (!hugetlb_node_alloc_supported()) {
4186 pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n");
4189 if (tmp >= MAX_NUMNODES || !node_online(tmp))
4191 node = array_index_nospec(tmp, MAX_NUMNODES);
4193 /* Parse hugepages */
4194 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4196 if (!hugetlb_max_hstate)
4197 default_hugepages_in_node[node] = tmp;
4199 parsed_hstate->max_huge_pages_node[node] = tmp;
4201 /* Go to parse next node*/
4202 if (p[count] == ',')
4215 * Global state is always initialized later in hugetlb_init.
4216 * But we need to allocate gigantic hstates here early to still
4217 * use the bootmem allocator.
4219 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
4220 hugetlb_hstate_alloc_pages(parsed_hstate);
4227 pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p);
4228 hugepages_clear_pages_in_node();
4231 __setup("hugepages=", hugepages_setup);
4234 * hugepagesz command line processing
4235 * A specific huge page size can only be specified once with hugepagesz.
4236 * hugepagesz is followed by hugepages on the command line. The global
4237 * variable 'parsed_valid_hugepagesz' is used to determine if prior
4238 * hugepagesz argument was valid.
4240 static int __init hugepagesz_setup(char *s)
4245 parsed_valid_hugepagesz = false;
4246 size = (unsigned long)memparse(s, NULL);
4248 if (!arch_hugetlb_valid_size(size)) {
4249 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
4253 h = size_to_hstate(size);
4256 * hstate for this size already exists. This is normally
4257 * an error, but is allowed if the existing hstate is the
4258 * default hstate. More specifically, it is only allowed if
4259 * the number of huge pages for the default hstate was not
4260 * previously specified.
4262 if (!parsed_default_hugepagesz || h != &default_hstate ||
4263 default_hstate.max_huge_pages) {
4264 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
4269 * No need to call hugetlb_add_hstate() as hstate already
4270 * exists. But, do set parsed_hstate so that a following
4271 * hugepages= parameter will be applied to this hstate.
4274 parsed_valid_hugepagesz = true;
4278 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4279 parsed_valid_hugepagesz = true;
4282 __setup("hugepagesz=", hugepagesz_setup);
4285 * default_hugepagesz command line input
4286 * Only one instance of default_hugepagesz allowed on command line.
4288 static int __init default_hugepagesz_setup(char *s)
4293 parsed_valid_hugepagesz = false;
4294 if (parsed_default_hugepagesz) {
4295 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
4299 size = (unsigned long)memparse(s, NULL);
4301 if (!arch_hugetlb_valid_size(size)) {
4302 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
4306 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4307 parsed_valid_hugepagesz = true;
4308 parsed_default_hugepagesz = true;
4309 default_hstate_idx = hstate_index(size_to_hstate(size));
4312 * The number of default huge pages (for this size) could have been
4313 * specified as the first hugetlb parameter: hugepages=X. If so,
4314 * then default_hstate_max_huge_pages is set. If the default huge
4315 * page size is gigantic (>= MAX_ORDER), then the pages must be
4316 * allocated here from bootmem allocator.
4318 if (default_hstate_max_huge_pages) {
4319 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
4320 for_each_online_node(i)
4321 default_hstate.max_huge_pages_node[i] =
4322 default_hugepages_in_node[i];
4323 if (hstate_is_gigantic(&default_hstate))
4324 hugetlb_hstate_alloc_pages(&default_hstate);
4325 default_hstate_max_huge_pages = 0;
4330 __setup("default_hugepagesz=", default_hugepagesz_setup);
4332 static unsigned int allowed_mems_nr(struct hstate *h)
4335 unsigned int nr = 0;
4336 nodemask_t *mpol_allowed;
4337 unsigned int *array = h->free_huge_pages_node;
4338 gfp_t gfp_mask = htlb_alloc_mask(h);
4340 mpol_allowed = policy_nodemask_current(gfp_mask);
4342 for_each_node_mask(node, cpuset_current_mems_allowed) {
4343 if (!mpol_allowed || node_isset(node, *mpol_allowed))
4350 #ifdef CONFIG_SYSCTL
4351 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
4352 void *buffer, size_t *length,
4353 loff_t *ppos, unsigned long *out)
4355 struct ctl_table dup_table;
4358 * In order to avoid races with __do_proc_doulongvec_minmax(), we
4359 * can duplicate the @table and alter the duplicate of it.
4362 dup_table.data = out;
4364 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
4367 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
4368 struct ctl_table *table, int write,
4369 void *buffer, size_t *length, loff_t *ppos)
4371 struct hstate *h = &default_hstate;
4372 unsigned long tmp = h->max_huge_pages;
4375 if (!hugepages_supported())
4378 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4384 ret = __nr_hugepages_store_common(obey_mempolicy, h,
4385 NUMA_NO_NODE, tmp, *length);
4390 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
4391 void *buffer, size_t *length, loff_t *ppos)
4394 return hugetlb_sysctl_handler_common(false, table, write,
4395 buffer, length, ppos);
4399 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
4400 void *buffer, size_t *length, loff_t *ppos)
4402 return hugetlb_sysctl_handler_common(true, table, write,
4403 buffer, length, ppos);
4405 #endif /* CONFIG_NUMA */
4407 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
4408 void *buffer, size_t *length, loff_t *ppos)
4410 struct hstate *h = &default_hstate;
4414 if (!hugepages_supported())
4417 tmp = h->nr_overcommit_huge_pages;
4419 if (write && hstate_is_gigantic(h))
4422 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4428 spin_lock_irq(&hugetlb_lock);
4429 h->nr_overcommit_huge_pages = tmp;
4430 spin_unlock_irq(&hugetlb_lock);
4436 #endif /* CONFIG_SYSCTL */
4438 void hugetlb_report_meminfo(struct seq_file *m)
4441 unsigned long total = 0;
4443 if (!hugepages_supported())
4446 for_each_hstate(h) {
4447 unsigned long count = h->nr_huge_pages;
4449 total += huge_page_size(h) * count;
4451 if (h == &default_hstate)
4453 "HugePages_Total: %5lu\n"
4454 "HugePages_Free: %5lu\n"
4455 "HugePages_Rsvd: %5lu\n"
4456 "HugePages_Surp: %5lu\n"
4457 "Hugepagesize: %8lu kB\n",
4461 h->surplus_huge_pages,
4462 huge_page_size(h) / SZ_1K);
4465 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
4468 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
4470 struct hstate *h = &default_hstate;
4472 if (!hugepages_supported())
4475 return sysfs_emit_at(buf, len,
4476 "Node %d HugePages_Total: %5u\n"
4477 "Node %d HugePages_Free: %5u\n"
4478 "Node %d HugePages_Surp: %5u\n",
4479 nid, h->nr_huge_pages_node[nid],
4480 nid, h->free_huge_pages_node[nid],
4481 nid, h->surplus_huge_pages_node[nid]);
4484 void hugetlb_show_meminfo(void)
4489 if (!hugepages_supported())
4492 for_each_node_state(nid, N_MEMORY)
4494 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
4496 h->nr_huge_pages_node[nid],
4497 h->free_huge_pages_node[nid],
4498 h->surplus_huge_pages_node[nid],
4499 huge_page_size(h) / SZ_1K);
4502 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
4504 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
4505 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
4508 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
4509 unsigned long hugetlb_total_pages(void)
4512 unsigned long nr_total_pages = 0;
4515 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
4516 return nr_total_pages;
4519 static int hugetlb_acct_memory(struct hstate *h, long delta)
4526 spin_lock_irq(&hugetlb_lock);
4528 * When cpuset is configured, it breaks the strict hugetlb page
4529 * reservation as the accounting is done on a global variable. Such
4530 * reservation is completely rubbish in the presence of cpuset because
4531 * the reservation is not checked against page availability for the
4532 * current cpuset. Application can still potentially OOM'ed by kernel
4533 * with lack of free htlb page in cpuset that the task is in.
4534 * Attempt to enforce strict accounting with cpuset is almost
4535 * impossible (or too ugly) because cpuset is too fluid that
4536 * task or memory node can be dynamically moved between cpusets.
4538 * The change of semantics for shared hugetlb mapping with cpuset is
4539 * undesirable. However, in order to preserve some of the semantics,
4540 * we fall back to check against current free page availability as
4541 * a best attempt and hopefully to minimize the impact of changing
4542 * semantics that cpuset has.
4544 * Apart from cpuset, we also have memory policy mechanism that
4545 * also determines from which node the kernel will allocate memory
4546 * in a NUMA system. So similar to cpuset, we also should consider
4547 * the memory policy of the current task. Similar to the description
4551 if (gather_surplus_pages(h, delta) < 0)
4554 if (delta > allowed_mems_nr(h)) {
4555 return_unused_surplus_pages(h, delta);
4562 return_unused_surplus_pages(h, (unsigned long) -delta);
4565 spin_unlock_irq(&hugetlb_lock);
4569 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4571 struct resv_map *resv = vma_resv_map(vma);
4574 * This new VMA should share its siblings reservation map if present.
4575 * The VMA will only ever have a valid reservation map pointer where
4576 * it is being copied for another still existing VMA. As that VMA
4577 * has a reference to the reservation map it cannot disappear until
4578 * after this open call completes. It is therefore safe to take a
4579 * new reference here without additional locking.
4581 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
4582 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
4583 kref_get(&resv->refs);
4587 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4589 struct hstate *h = hstate_vma(vma);
4590 struct resv_map *resv = vma_resv_map(vma);
4591 struct hugepage_subpool *spool = subpool_vma(vma);
4592 unsigned long reserve, start, end;
4595 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4598 start = vma_hugecache_offset(h, vma, vma->vm_start);
4599 end = vma_hugecache_offset(h, vma, vma->vm_end);
4601 reserve = (end - start) - region_count(resv, start, end);
4602 hugetlb_cgroup_uncharge_counter(resv, start, end);
4605 * Decrement reserve counts. The global reserve count may be
4606 * adjusted if the subpool has a minimum size.
4608 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4609 hugetlb_acct_memory(h, -gbl_reserve);
4612 kref_put(&resv->refs, resv_map_release);
4615 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4617 if (addr & ~(huge_page_mask(hstate_vma(vma))))
4622 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4624 return huge_page_size(hstate_vma(vma));
4628 * We cannot handle pagefaults against hugetlb pages at all. They cause
4629 * handle_mm_fault() to try to instantiate regular-sized pages in the
4630 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
4633 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4640 * When a new function is introduced to vm_operations_struct and added
4641 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4642 * This is because under System V memory model, mappings created via
4643 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4644 * their original vm_ops are overwritten with shm_vm_ops.
4646 const struct vm_operations_struct hugetlb_vm_ops = {
4647 .fault = hugetlb_vm_op_fault,
4648 .open = hugetlb_vm_op_open,
4649 .close = hugetlb_vm_op_close,
4650 .may_split = hugetlb_vm_op_split,
4651 .pagesize = hugetlb_vm_op_pagesize,
4654 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4658 unsigned int shift = huge_page_shift(hstate_vma(vma));
4661 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4662 vma->vm_page_prot)));
4664 entry = huge_pte_wrprotect(mk_huge_pte(page,
4665 vma->vm_page_prot));
4667 entry = pte_mkyoung(entry);
4668 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4673 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4674 unsigned long address, pte_t *ptep)
4678 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4679 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4680 update_mmu_cache(vma, address, ptep);
4683 bool is_hugetlb_entry_migration(pte_t pte)
4687 if (huge_pte_none(pte) || pte_present(pte))
4689 swp = pte_to_swp_entry(pte);
4690 if (is_migration_entry(swp))
4696 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4700 if (huge_pte_none(pte) || pte_present(pte))
4702 swp = pte_to_swp_entry(pte);
4703 if (is_hwpoison_entry(swp))
4710 hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
4711 struct page *new_page)
4713 __SetPageUptodate(new_page);
4714 hugepage_add_new_anon_rmap(new_page, vma, addr);
4715 set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
4716 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
4717 ClearHPageRestoreReserve(new_page);
4718 SetHPageMigratable(new_page);
4721 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
4722 struct vm_area_struct *dst_vma,
4723 struct vm_area_struct *src_vma)
4725 pte_t *src_pte, *dst_pte, entry, dst_entry;
4726 struct page *ptepage;
4728 bool cow = is_cow_mapping(src_vma->vm_flags);
4729 struct hstate *h = hstate_vma(src_vma);
4730 unsigned long sz = huge_page_size(h);
4731 unsigned long npages = pages_per_huge_page(h);
4732 struct address_space *mapping = src_vma->vm_file->f_mapping;
4733 struct mmu_notifier_range range;
4737 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src_vma, src,
4740 mmu_notifier_invalidate_range_start(&range);
4741 mmap_assert_write_locked(src);
4742 raw_write_seqcount_begin(&src->write_protect_seq);
4745 * For shared mappings i_mmap_rwsem must be held to call
4746 * huge_pte_alloc, otherwise the returned ptep could go
4747 * away if part of a shared pmd and another thread calls
4750 i_mmap_lock_read(mapping);
4753 for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) {
4754 spinlock_t *src_ptl, *dst_ptl;
4755 src_pte = huge_pte_offset(src, addr, sz);
4758 dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz);
4765 * If the pagetables are shared don't copy or take references.
4766 * dst_pte == src_pte is the common case of src/dest sharing.
4768 * However, src could have 'unshared' and dst shares with
4769 * another vma. If dst_pte !none, this implies sharing.
4770 * Check here before taking page table lock, and once again
4771 * after taking the lock below.
4773 dst_entry = huge_ptep_get(dst_pte);
4774 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
4777 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4778 src_ptl = huge_pte_lockptr(h, src, src_pte);
4779 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4780 entry = huge_ptep_get(src_pte);
4781 dst_entry = huge_ptep_get(dst_pte);
4783 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
4785 * Skip if src entry none. Also, skip in the
4786 * unlikely case dst entry !none as this implies
4787 * sharing with another vma.
4790 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
4791 is_hugetlb_entry_hwpoisoned(entry))) {
4792 swp_entry_t swp_entry = pte_to_swp_entry(entry);
4793 bool uffd_wp = huge_pte_uffd_wp(entry);
4795 if (!is_readable_migration_entry(swp_entry) && cow) {
4797 * COW mappings require pages in both
4798 * parent and child to be set to read.
4800 swp_entry = make_readable_migration_entry(
4801 swp_offset(swp_entry));
4802 entry = swp_entry_to_pte(swp_entry);
4803 if (userfaultfd_wp(src_vma) && uffd_wp)
4804 entry = huge_pte_mkuffd_wp(entry);
4805 set_huge_swap_pte_at(src, addr, src_pte,
4808 if (!userfaultfd_wp(dst_vma) && uffd_wp)
4809 entry = huge_pte_clear_uffd_wp(entry);
4810 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
4811 } else if (unlikely(is_pte_marker(entry))) {
4813 * We copy the pte marker only if the dst vma has
4816 if (userfaultfd_wp(dst_vma))
4817 set_huge_pte_at(dst, addr, dst_pte, entry);
4819 entry = huge_ptep_get(src_pte);
4820 ptepage = pte_page(entry);
4824 * Failing to duplicate the anon rmap is a rare case
4825 * where we see pinned hugetlb pages while they're
4826 * prone to COW. We need to do the COW earlier during
4829 * When pre-allocating the page or copying data, we
4830 * need to be without the pgtable locks since we could
4831 * sleep during the process.
4833 if (!PageAnon(ptepage)) {
4834 page_dup_file_rmap(ptepage, true);
4835 } else if (page_try_dup_anon_rmap(ptepage, true,
4837 pte_t src_pte_old = entry;
4840 spin_unlock(src_ptl);
4841 spin_unlock(dst_ptl);
4842 /* Do not use reserve as it's private owned */
4843 new = alloc_huge_page(dst_vma, addr, 1);
4849 copy_user_huge_page(new, ptepage, addr, dst_vma,
4853 /* Install the new huge page if src pte stable */
4854 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4855 src_ptl = huge_pte_lockptr(h, src, src_pte);
4856 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4857 entry = huge_ptep_get(src_pte);
4858 if (!pte_same(src_pte_old, entry)) {
4859 restore_reserve_on_error(h, dst_vma, addr,
4862 /* dst_entry won't change as in child */
4865 hugetlb_install_page(dst_vma, dst_pte, addr, new);
4866 spin_unlock(src_ptl);
4867 spin_unlock(dst_ptl);
4873 * No need to notify as we are downgrading page
4874 * table protection not changing it to point
4877 * See Documentation/mm/mmu_notifier.rst
4879 huge_ptep_set_wrprotect(src, addr, src_pte);
4880 entry = huge_pte_wrprotect(entry);
4883 set_huge_pte_at(dst, addr, dst_pte, entry);
4884 hugetlb_count_add(npages, dst);
4886 spin_unlock(src_ptl);
4887 spin_unlock(dst_ptl);
4891 raw_write_seqcount_end(&src->write_protect_seq);
4892 mmu_notifier_invalidate_range_end(&range);
4894 i_mmap_unlock_read(mapping);
4900 static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
4901 unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte)
4903 struct hstate *h = hstate_vma(vma);
4904 struct mm_struct *mm = vma->vm_mm;
4905 spinlock_t *src_ptl, *dst_ptl;
4908 dst_ptl = huge_pte_lock(h, mm, dst_pte);
4909 src_ptl = huge_pte_lockptr(h, mm, src_pte);
4912 * We don't have to worry about the ordering of src and dst ptlocks
4913 * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock.
4915 if (src_ptl != dst_ptl)
4916 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4918 pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);
4919 set_huge_pte_at(mm, new_addr, dst_pte, pte);
4921 if (src_ptl != dst_ptl)
4922 spin_unlock(src_ptl);
4923 spin_unlock(dst_ptl);
4926 int move_hugetlb_page_tables(struct vm_area_struct *vma,
4927 struct vm_area_struct *new_vma,
4928 unsigned long old_addr, unsigned long new_addr,
4931 struct hstate *h = hstate_vma(vma);
4932 struct address_space *mapping = vma->vm_file->f_mapping;
4933 unsigned long sz = huge_page_size(h);
4934 struct mm_struct *mm = vma->vm_mm;
4935 unsigned long old_end = old_addr + len;
4936 unsigned long old_addr_copy;
4937 pte_t *src_pte, *dst_pte;
4938 struct mmu_notifier_range range;
4939 bool shared_pmd = false;
4941 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr,
4943 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4945 * In case of shared PMDs, we should cover the maximum possible
4948 flush_cache_range(vma, range.start, range.end);
4950 mmu_notifier_invalidate_range_start(&range);
4951 /* Prevent race with file truncation */
4952 i_mmap_lock_write(mapping);
4953 for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
4954 src_pte = huge_pte_offset(mm, old_addr, sz);
4957 if (huge_pte_none(huge_ptep_get(src_pte)))
4960 /* old_addr arg to huge_pmd_unshare() is a pointer and so the
4961 * arg may be modified. Pass a copy instead to preserve the
4962 * value in old_addr.
4964 old_addr_copy = old_addr;
4966 if (huge_pmd_unshare(mm, vma, &old_addr_copy, src_pte)) {
4971 dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz);
4975 move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte);
4979 flush_tlb_range(vma, range.start, range.end);
4981 flush_tlb_range(vma, old_end - len, old_end);
4982 mmu_notifier_invalidate_range_end(&range);
4983 i_mmap_unlock_write(mapping);
4985 return len + old_addr - old_end;
4988 static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
4989 unsigned long start, unsigned long end,
4990 struct page *ref_page, zap_flags_t zap_flags)
4992 struct mm_struct *mm = vma->vm_mm;
4993 unsigned long address;
4998 struct hstate *h = hstate_vma(vma);
4999 unsigned long sz = huge_page_size(h);
5000 struct mmu_notifier_range range;
5001 bool force_flush = false;
5003 WARN_ON(!is_vm_hugetlb_page(vma));
5004 BUG_ON(start & ~huge_page_mask(h));
5005 BUG_ON(end & ~huge_page_mask(h));
5008 * This is a hugetlb vma, all the pte entries should point
5011 tlb_change_page_size(tlb, sz);
5012 tlb_start_vma(tlb, vma);
5015 * If sharing possible, alert mmu notifiers of worst case.
5017 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
5019 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5020 mmu_notifier_invalidate_range_start(&range);
5022 for (; address < end; address += sz) {
5023 ptep = huge_pte_offset(mm, address, sz);
5027 ptl = huge_pte_lock(h, mm, ptep);
5028 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
5030 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
5035 pte = huge_ptep_get(ptep);
5036 if (huge_pte_none(pte)) {
5042 * Migrating hugepage or HWPoisoned hugepage is already
5043 * unmapped and its refcount is dropped, so just clear pte here.
5045 if (unlikely(!pte_present(pte))) {
5047 * If the pte was wr-protected by uffd-wp in any of the
5048 * swap forms, meanwhile the caller does not want to
5049 * drop the uffd-wp bit in this zap, then replace the
5050 * pte with a marker.
5052 if (pte_swp_uffd_wp_any(pte) &&
5053 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5054 set_huge_pte_at(mm, address, ptep,
5055 make_pte_marker(PTE_MARKER_UFFD_WP));
5057 huge_pte_clear(mm, address, ptep, sz);
5062 page = pte_page(pte);
5064 * If a reference page is supplied, it is because a specific
5065 * page is being unmapped, not a range. Ensure the page we
5066 * are about to unmap is the actual page of interest.
5069 if (page != ref_page) {
5074 * Mark the VMA as having unmapped its page so that
5075 * future faults in this VMA will fail rather than
5076 * looking like data was lost
5078 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
5081 pte = huge_ptep_get_and_clear(mm, address, ptep);
5082 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
5083 if (huge_pte_dirty(pte))
5084 set_page_dirty(page);
5085 /* Leave a uffd-wp pte marker if needed */
5086 if (huge_pte_uffd_wp(pte) &&
5087 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5088 set_huge_pte_at(mm, address, ptep,
5089 make_pte_marker(PTE_MARKER_UFFD_WP));
5090 hugetlb_count_sub(pages_per_huge_page(h), mm);
5091 page_remove_rmap(page, vma, true);
5094 tlb_remove_page_size(tlb, page, huge_page_size(h));
5096 * Bail out after unmapping reference page if supplied
5101 mmu_notifier_invalidate_range_end(&range);
5102 tlb_end_vma(tlb, vma);
5105 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
5106 * could defer the flush until now, since by holding i_mmap_rwsem we
5107 * guaranteed that the last refernece would not be dropped. But we must
5108 * do the flushing before we return, as otherwise i_mmap_rwsem will be
5109 * dropped and the last reference to the shared PMDs page might be
5112 * In theory we could defer the freeing of the PMD pages as well, but
5113 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
5114 * detect sharing, so we cannot defer the release of the page either.
5115 * Instead, do flush now.
5118 tlb_flush_mmu_tlbonly(tlb);
5121 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
5122 struct vm_area_struct *vma, unsigned long start,
5123 unsigned long end, struct page *ref_page,
5124 zap_flags_t zap_flags)
5126 __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags);
5129 * Clear this flag so that x86's huge_pmd_share page_table_shareable
5130 * test will fail on a vma being torn down, and not grab a page table
5131 * on its way out. We're lucky that the flag has such an appropriate
5132 * name, and can in fact be safely cleared here. We could clear it
5133 * before the __unmap_hugepage_range above, but all that's necessary
5134 * is to clear it before releasing the i_mmap_rwsem. This works
5135 * because in the context this is called, the VMA is about to be
5136 * destroyed and the i_mmap_rwsem is held.
5138 vma->vm_flags &= ~VM_MAYSHARE;
5141 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
5142 unsigned long end, struct page *ref_page,
5143 zap_flags_t zap_flags)
5145 struct mmu_gather tlb;
5147 tlb_gather_mmu(&tlb, vma->vm_mm);
5148 __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags);
5149 tlb_finish_mmu(&tlb);
5153 * This is called when the original mapper is failing to COW a MAP_PRIVATE
5154 * mapping it owns the reserve page for. The intention is to unmap the page
5155 * from other VMAs and let the children be SIGKILLed if they are faulting the
5158 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
5159 struct page *page, unsigned long address)
5161 struct hstate *h = hstate_vma(vma);
5162 struct vm_area_struct *iter_vma;
5163 struct address_space *mapping;
5167 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
5168 * from page cache lookup which is in HPAGE_SIZE units.
5170 address = address & huge_page_mask(h);
5171 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
5173 mapping = vma->vm_file->f_mapping;
5176 * Take the mapping lock for the duration of the table walk. As
5177 * this mapping should be shared between all the VMAs,
5178 * __unmap_hugepage_range() is called as the lock is already held
5180 i_mmap_lock_write(mapping);
5181 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
5182 /* Do not unmap the current VMA */
5183 if (iter_vma == vma)
5187 * Shared VMAs have their own reserves and do not affect
5188 * MAP_PRIVATE accounting but it is possible that a shared
5189 * VMA is using the same page so check and skip such VMAs.
5191 if (iter_vma->vm_flags & VM_MAYSHARE)
5195 * Unmap the page from other VMAs without their own reserves.
5196 * They get marked to be SIGKILLed if they fault in these
5197 * areas. This is because a future no-page fault on this VMA
5198 * could insert a zeroed page instead of the data existing
5199 * from the time of fork. This would look like data corruption
5201 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
5202 unmap_hugepage_range(iter_vma, address,
5203 address + huge_page_size(h), page, 0);
5205 i_mmap_unlock_write(mapping);
5209 * hugetlb_wp() should be called with page lock of the original hugepage held.
5210 * Called with hugetlb_fault_mutex_table held and pte_page locked so we
5211 * cannot race with other handlers or page migration.
5212 * Keep the pte_same checks anyway to make transition from the mutex easier.
5214 static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma,
5215 unsigned long address, pte_t *ptep, unsigned int flags,
5216 struct page *pagecache_page, spinlock_t *ptl)
5218 const bool unshare = flags & FAULT_FLAG_UNSHARE;
5220 struct hstate *h = hstate_vma(vma);
5221 struct page *old_page, *new_page;
5222 int outside_reserve = 0;
5224 unsigned long haddr = address & huge_page_mask(h);
5225 struct mmu_notifier_range range;
5227 VM_BUG_ON(unshare && (flags & FOLL_WRITE));
5228 VM_BUG_ON(!unshare && !(flags & FOLL_WRITE));
5230 pte = huge_ptep_get(ptep);
5231 old_page = pte_page(pte);
5233 delayacct_wpcopy_start();
5237 * If no-one else is actually using this page, we're the exclusive
5238 * owner and can reuse this page.
5240 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
5241 if (!PageAnonExclusive(old_page))
5242 page_move_anon_rmap(old_page, vma);
5243 if (likely(!unshare))
5244 set_huge_ptep_writable(vma, haddr, ptep);
5246 delayacct_wpcopy_end();
5249 VM_BUG_ON_PAGE(PageAnon(old_page) && PageAnonExclusive(old_page),
5253 * If the process that created a MAP_PRIVATE mapping is about to
5254 * perform a COW due to a shared page count, attempt to satisfy
5255 * the allocation without using the existing reserves. The pagecache
5256 * page is used to determine if the reserve at this address was
5257 * consumed or not. If reserves were used, a partial faulted mapping
5258 * at the time of fork() could consume its reserves on COW instead
5259 * of the full address range.
5261 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
5262 old_page != pagecache_page)
5263 outside_reserve = 1;
5268 * Drop page table lock as buddy allocator may be called. It will
5269 * be acquired again before returning to the caller, as expected.
5272 new_page = alloc_huge_page(vma, haddr, outside_reserve);
5274 if (IS_ERR(new_page)) {
5276 * If a process owning a MAP_PRIVATE mapping fails to COW,
5277 * it is due to references held by a child and an insufficient
5278 * huge page pool. To guarantee the original mappers
5279 * reliability, unmap the page from child processes. The child
5280 * may get SIGKILLed if it later faults.
5282 if (outside_reserve) {
5283 struct address_space *mapping = vma->vm_file->f_mapping;
5288 BUG_ON(huge_pte_none(pte));
5290 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
5291 * unmapping. unmapping needs to hold i_mmap_rwsem
5292 * in write mode. Dropping i_mmap_rwsem in read mode
5293 * here is OK as COW mappings do not interact with
5296 * Reacquire both after unmap operation.
5298 idx = vma_hugecache_offset(h, vma, haddr);
5299 hash = hugetlb_fault_mutex_hash(mapping, idx);
5300 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5301 i_mmap_unlock_read(mapping);
5303 unmap_ref_private(mm, vma, old_page, haddr);
5305 i_mmap_lock_read(mapping);
5306 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5308 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5310 pte_same(huge_ptep_get(ptep), pte)))
5311 goto retry_avoidcopy;
5313 * race occurs while re-acquiring page table
5314 * lock, and our job is done.
5316 delayacct_wpcopy_end();
5320 ret = vmf_error(PTR_ERR(new_page));
5321 goto out_release_old;
5325 * When the original hugepage is shared one, it does not have
5326 * anon_vma prepared.
5328 if (unlikely(anon_vma_prepare(vma))) {
5330 goto out_release_all;
5333 copy_user_huge_page(new_page, old_page, address, vma,
5334 pages_per_huge_page(h));
5335 __SetPageUptodate(new_page);
5337 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
5338 haddr + huge_page_size(h));
5339 mmu_notifier_invalidate_range_start(&range);
5342 * Retake the page table lock to check for racing updates
5343 * before the page tables are altered
5346 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5347 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
5348 ClearHPageRestoreReserve(new_page);
5350 /* Break COW or unshare */
5351 huge_ptep_clear_flush(vma, haddr, ptep);
5352 mmu_notifier_invalidate_range(mm, range.start, range.end);
5353 page_remove_rmap(old_page, vma, true);
5354 hugepage_add_new_anon_rmap(new_page, vma, haddr);
5355 set_huge_pte_at(mm, haddr, ptep,
5356 make_huge_pte(vma, new_page, !unshare));
5357 SetHPageMigratable(new_page);
5358 /* Make the old page be freed below */
5359 new_page = old_page;
5362 mmu_notifier_invalidate_range_end(&range);
5365 * No restore in case of successful pagetable update (Break COW or
5368 if (new_page != old_page)
5369 restore_reserve_on_error(h, vma, haddr, new_page);
5374 spin_lock(ptl); /* Caller expects lock to be held */
5376 delayacct_wpcopy_end();
5380 /* Return the pagecache page at a given address within a VMA */
5381 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
5382 struct vm_area_struct *vma, unsigned long address)
5384 struct address_space *mapping;
5387 mapping = vma->vm_file->f_mapping;
5388 idx = vma_hugecache_offset(h, vma, address);
5390 return find_lock_page(mapping, idx);
5394 * Return whether there is a pagecache page to back given address within VMA.
5395 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
5397 static bool hugetlbfs_pagecache_present(struct hstate *h,
5398 struct vm_area_struct *vma, unsigned long address)
5400 struct address_space *mapping;
5404 mapping = vma->vm_file->f_mapping;
5405 idx = vma_hugecache_offset(h, vma, address);
5407 page = find_get_page(mapping, idx);
5410 return page != NULL;
5413 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
5416 struct inode *inode = mapping->host;
5417 struct hstate *h = hstate_inode(inode);
5418 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
5422 ClearHPageRestoreReserve(page);
5425 * set page dirty so that it will not be removed from cache/file
5426 * by non-hugetlbfs specific code paths.
5428 set_page_dirty(page);
5430 spin_lock(&inode->i_lock);
5431 inode->i_blocks += blocks_per_huge_page(h);
5432 spin_unlock(&inode->i_lock);
5436 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
5437 struct address_space *mapping,
5440 unsigned long haddr,
5442 unsigned long reason)
5446 struct vm_fault vmf = {
5449 .real_address = addr,
5453 * Hard to debug if it ends up being
5454 * used by a callee that assumes
5455 * something about the other
5456 * uninitialized fields... same as in
5462 * hugetlb_fault_mutex and i_mmap_rwsem must be
5463 * dropped before handling userfault. Reacquire
5464 * after handling fault to make calling code simpler.
5466 hash = hugetlb_fault_mutex_hash(mapping, idx);
5467 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5468 i_mmap_unlock_read(mapping);
5469 ret = handle_userfault(&vmf, reason);
5470 i_mmap_lock_read(mapping);
5471 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5476 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
5477 struct vm_area_struct *vma,
5478 struct address_space *mapping, pgoff_t idx,
5479 unsigned long address, pte_t *ptep,
5480 pte_t old_pte, unsigned int flags)
5482 struct hstate *h = hstate_vma(vma);
5483 vm_fault_t ret = VM_FAULT_SIGBUS;
5489 unsigned long haddr = address & huge_page_mask(h);
5490 bool new_page, new_pagecache_page = false;
5493 * Currently, we are forced to kill the process in the event the
5494 * original mapper has unmapped pages from the child due to a failed
5495 * COW/unsharing. Warn that such a situation has occurred as it may not
5498 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
5499 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
5505 * We can not race with truncation due to holding i_mmap_rwsem.
5506 * i_size is modified when holding i_mmap_rwsem, so check here
5507 * once for faults beyond end of file.
5509 size = i_size_read(mapping->host) >> huge_page_shift(h);
5515 page = find_lock_page(mapping, idx);
5517 /* Check for page in userfault range */
5518 if (userfaultfd_missing(vma)) {
5519 ret = hugetlb_handle_userfault(vma, mapping, idx,
5520 flags, haddr, address,
5525 page = alloc_huge_page(vma, haddr, 0);
5528 * Returning error will result in faulting task being
5529 * sent SIGBUS. The hugetlb fault mutex prevents two
5530 * tasks from racing to fault in the same page which
5531 * could result in false unable to allocate errors.
5532 * Page migration does not take the fault mutex, but
5533 * does a clear then write of pte's under page table
5534 * lock. Page fault code could race with migration,
5535 * notice the clear pte and try to allocate a page
5536 * here. Before returning error, get ptl and make
5537 * sure there really is no pte entry.
5539 ptl = huge_pte_lock(h, mm, ptep);
5541 if (huge_pte_none(huge_ptep_get(ptep)))
5542 ret = vmf_error(PTR_ERR(page));
5546 clear_huge_page(page, address, pages_per_huge_page(h));
5547 __SetPageUptodate(page);
5550 if (vma->vm_flags & VM_MAYSHARE) {
5551 int err = huge_add_to_page_cache(page, mapping, idx);
5558 new_pagecache_page = true;
5561 if (unlikely(anon_vma_prepare(vma))) {
5563 goto backout_unlocked;
5569 * If memory error occurs between mmap() and fault, some process
5570 * don't have hwpoisoned swap entry for errored virtual address.
5571 * So we need to block hugepage fault by PG_hwpoison bit check.
5573 if (unlikely(PageHWPoison(page))) {
5574 ret = VM_FAULT_HWPOISON_LARGE |
5575 VM_FAULT_SET_HINDEX(hstate_index(h));
5576 goto backout_unlocked;
5579 /* Check for page in userfault range. */
5580 if (userfaultfd_minor(vma)) {
5583 ret = hugetlb_handle_userfault(vma, mapping, idx,
5584 flags, haddr, address,
5591 * If we are going to COW a private mapping later, we examine the
5592 * pending reservations for this page now. This will ensure that
5593 * any allocations necessary to record that reservation occur outside
5596 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5597 if (vma_needs_reservation(h, vma, haddr) < 0) {
5599 goto backout_unlocked;
5601 /* Just decrements count, does not deallocate */
5602 vma_end_reservation(h, vma, haddr);
5605 ptl = huge_pte_lock(h, mm, ptep);
5607 /* If pte changed from under us, retry */
5608 if (!pte_same(huge_ptep_get(ptep), old_pte))
5612 ClearHPageRestoreReserve(page);
5613 hugepage_add_new_anon_rmap(page, vma, haddr);
5615 page_dup_file_rmap(page, true);
5616 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
5617 && (vma->vm_flags & VM_SHARED)));
5619 * If this pte was previously wr-protected, keep it wr-protected even
5622 if (unlikely(pte_marker_uffd_wp(old_pte)))
5623 new_pte = huge_pte_wrprotect(huge_pte_mkuffd_wp(new_pte));
5624 set_huge_pte_at(mm, haddr, ptep, new_pte);
5626 hugetlb_count_add(pages_per_huge_page(h), mm);
5627 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5628 /* Optimization, do the COW without a second fault */
5629 ret = hugetlb_wp(mm, vma, address, ptep, flags, page, ptl);
5635 * Only set HPageMigratable in newly allocated pages. Existing pages
5636 * found in the pagecache may not have HPageMigratableset if they have
5637 * been isolated for migration.
5640 SetHPageMigratable(page);
5650 /* restore reserve for newly allocated pages not in page cache */
5651 if (new_page && !new_pagecache_page)
5652 restore_reserve_on_error(h, vma, haddr, page);
5658 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5660 unsigned long key[2];
5663 key[0] = (unsigned long) mapping;
5666 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
5668 return hash & (num_fault_mutexes - 1);
5672 * For uniprocessor systems we always use a single mutex, so just
5673 * return 0 and avoid the hashing overhead.
5675 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5681 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
5682 unsigned long address, unsigned int flags)
5689 struct page *page = NULL;
5690 struct page *pagecache_page = NULL;
5691 struct hstate *h = hstate_vma(vma);
5692 struct address_space *mapping;
5693 int need_wait_lock = 0;
5694 unsigned long haddr = address & huge_page_mask(h);
5696 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5699 * Since we hold no locks, ptep could be stale. That is
5700 * OK as we are only making decisions based on content and
5701 * not actually modifying content here.
5703 entry = huge_ptep_get(ptep);
5704 if (unlikely(is_hugetlb_entry_migration(entry))) {
5705 migration_entry_wait_huge(vma, ptep);
5707 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
5708 return VM_FAULT_HWPOISON_LARGE |
5709 VM_FAULT_SET_HINDEX(hstate_index(h));
5713 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
5714 * until finished with ptep. This serves two purposes:
5715 * 1) It prevents huge_pmd_unshare from being called elsewhere
5716 * and making the ptep no longer valid.
5717 * 2) It synchronizes us with i_size modifications during truncation.
5719 * ptep could have already be assigned via huge_pte_offset. That
5720 * is OK, as huge_pte_alloc will return the same value unless
5721 * something has changed.
5723 mapping = vma->vm_file->f_mapping;
5724 i_mmap_lock_read(mapping);
5725 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
5727 i_mmap_unlock_read(mapping);
5728 return VM_FAULT_OOM;
5732 * Serialize hugepage allocation and instantiation, so that we don't
5733 * get spurious allocation failures if two CPUs race to instantiate
5734 * the same page in the page cache.
5736 idx = vma_hugecache_offset(h, vma, haddr);
5737 hash = hugetlb_fault_mutex_hash(mapping, idx);
5738 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5740 entry = huge_ptep_get(ptep);
5741 /* PTE markers should be handled the same way as none pte */
5742 if (huge_pte_none_mostly(entry)) {
5743 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep,
5751 * entry could be a migration/hwpoison entry at this point, so this
5752 * check prevents the kernel from going below assuming that we have
5753 * an active hugepage in pagecache. This goto expects the 2nd page
5754 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
5755 * properly handle it.
5757 if (!pte_present(entry))
5761 * If we are going to COW/unshare the mapping later, we examine the
5762 * pending reservations for this page now. This will ensure that any
5763 * allocations necessary to record that reservation occur outside the
5764 * spinlock. For private mappings, we also lookup the pagecache
5765 * page now as it is used to determine if a reservation has been
5768 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
5769 !huge_pte_write(entry)) {
5770 if (vma_needs_reservation(h, vma, haddr) < 0) {
5774 /* Just decrements count, does not deallocate */
5775 vma_end_reservation(h, vma, haddr);
5777 if (!(vma->vm_flags & VM_MAYSHARE))
5778 pagecache_page = hugetlbfs_pagecache_page(h,
5782 ptl = huge_pte_lock(h, mm, ptep);
5784 /* Check for a racing update before calling hugetlb_wp() */
5785 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
5788 /* Handle userfault-wp first, before trying to lock more pages */
5789 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) &&
5790 (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
5791 struct vm_fault vmf = {
5794 .real_address = address,
5799 if (pagecache_page) {
5800 unlock_page(pagecache_page);
5801 put_page(pagecache_page);
5803 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5804 i_mmap_unlock_read(mapping);
5805 return handle_userfault(&vmf, VM_UFFD_WP);
5809 * hugetlb_wp() requires page locks of pte_page(entry) and
5810 * pagecache_page, so here we need take the former one
5811 * when page != pagecache_page or !pagecache_page.
5813 page = pte_page(entry);
5814 if (page != pagecache_page)
5815 if (!trylock_page(page)) {
5822 if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
5823 if (!huge_pte_write(entry)) {
5824 ret = hugetlb_wp(mm, vma, address, ptep, flags,
5825 pagecache_page, ptl);
5827 } else if (likely(flags & FAULT_FLAG_WRITE)) {
5828 entry = huge_pte_mkdirty(entry);
5831 entry = pte_mkyoung(entry);
5832 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
5833 flags & FAULT_FLAG_WRITE))
5834 update_mmu_cache(vma, haddr, ptep);
5836 if (page != pagecache_page)
5842 if (pagecache_page) {
5843 unlock_page(pagecache_page);
5844 put_page(pagecache_page);
5847 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5848 i_mmap_unlock_read(mapping);
5850 * Generally it's safe to hold refcount during waiting page lock. But
5851 * here we just wait to defer the next page fault to avoid busy loop and
5852 * the page is not used after unlocked before returning from the current
5853 * page fault. So we are safe from accessing freed page, even if we wait
5854 * here without taking refcount.
5857 wait_on_page_locked(page);
5861 #ifdef CONFIG_USERFAULTFD
5863 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
5864 * modifications for huge pages.
5866 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
5868 struct vm_area_struct *dst_vma,
5869 unsigned long dst_addr,
5870 unsigned long src_addr,
5871 enum mcopy_atomic_mode mode,
5872 struct page **pagep,
5875 bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE);
5876 struct hstate *h = hstate_vma(dst_vma);
5877 struct address_space *mapping = dst_vma->vm_file->f_mapping;
5878 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
5880 int vm_shared = dst_vma->vm_flags & VM_SHARED;
5886 bool page_in_pagecache = false;
5890 page = find_lock_page(mapping, idx);
5893 page_in_pagecache = true;
5894 } else if (!*pagep) {
5895 /* If a page already exists, then it's UFFDIO_COPY for
5896 * a non-missing case. Return -EEXIST.
5899 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5904 page = alloc_huge_page(dst_vma, dst_addr, 0);
5910 ret = copy_huge_page_from_user(page,
5911 (const void __user *) src_addr,
5912 pages_per_huge_page(h), false);
5914 /* fallback to copy_from_user outside mmap_lock */
5915 if (unlikely(ret)) {
5917 /* Free the allocated page which may have
5918 * consumed a reservation.
5920 restore_reserve_on_error(h, dst_vma, dst_addr, page);
5923 /* Allocate a temporary page to hold the copied
5926 page = alloc_huge_page_vma(h, dst_vma, dst_addr);
5932 /* Set the outparam pagep and return to the caller to
5933 * copy the contents outside the lock. Don't free the
5940 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5947 page = alloc_huge_page(dst_vma, dst_addr, 0);
5953 copy_user_huge_page(page, *pagep, dst_addr, dst_vma,
5954 pages_per_huge_page(h));
5960 * The memory barrier inside __SetPageUptodate makes sure that
5961 * preceding stores to the page contents become visible before
5962 * the set_pte_at() write.
5964 __SetPageUptodate(page);
5966 /* Add shared, newly allocated pages to the page cache. */
5967 if (vm_shared && !is_continue) {
5968 size = i_size_read(mapping->host) >> huge_page_shift(h);
5971 goto out_release_nounlock;
5974 * Serialization between remove_inode_hugepages() and
5975 * huge_add_to_page_cache() below happens through the
5976 * hugetlb_fault_mutex_table that here must be hold by
5979 ret = huge_add_to_page_cache(page, mapping, idx);
5981 goto out_release_nounlock;
5982 page_in_pagecache = true;
5985 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
5989 * Recheck the i_size after holding PT lock to make sure not
5990 * to leave any page mapped (as page_mapped()) beyond the end
5991 * of the i_size (remove_inode_hugepages() is strict about
5992 * enforcing that). If we bail out here, we'll also leave a
5993 * page in the radix tree in the vm_shared case beyond the end
5994 * of the i_size, but remove_inode_hugepages() will take care
5995 * of it as soon as we drop the hugetlb_fault_mutex_table.
5997 size = i_size_read(mapping->host) >> huge_page_shift(h);
6000 goto out_release_unlock;
6004 * We allow to overwrite a pte marker: consider when both MISSING|WP
6005 * registered, we firstly wr-protect a none pte which has no page cache
6006 * page backing it, then access the page.
6008 if (!huge_pte_none_mostly(huge_ptep_get(dst_pte)))
6009 goto out_release_unlock;
6012 page_dup_file_rmap(page, true);
6014 ClearHPageRestoreReserve(page);
6015 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
6019 * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY
6020 * with wp flag set, don't set pte write bit.
6022 if (wp_copy || (is_continue && !vm_shared))
6025 writable = dst_vma->vm_flags & VM_WRITE;
6027 _dst_pte = make_huge_pte(dst_vma, page, writable);
6029 * Always mark UFFDIO_COPY page dirty; note that this may not be
6030 * extremely important for hugetlbfs for now since swapping is not
6031 * supported, but we should still be clear in that this page cannot be
6032 * thrown away at will, even if write bit not set.
6034 _dst_pte = huge_pte_mkdirty(_dst_pte);
6035 _dst_pte = pte_mkyoung(_dst_pte);
6038 _dst_pte = huge_pte_mkuffd_wp(_dst_pte);
6040 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
6042 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
6043 dst_vma->vm_flags & VM_WRITE);
6044 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
6046 /* No need to invalidate - it was non-present before */
6047 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6051 SetHPageMigratable(page);
6052 if (vm_shared || is_continue)
6059 if (vm_shared || is_continue)
6061 out_release_nounlock:
6062 if (!page_in_pagecache)
6063 restore_reserve_on_error(h, dst_vma, dst_addr, page);
6067 #endif /* CONFIG_USERFAULTFD */
6069 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
6070 int refs, struct page **pages,
6071 struct vm_area_struct **vmas)
6075 for (nr = 0; nr < refs; nr++) {
6077 pages[nr] = mem_map_offset(page, nr);
6083 static inline bool __follow_hugetlb_must_fault(unsigned int flags, pte_t *pte,
6086 pte_t pteval = huge_ptep_get(pte);
6089 if (is_swap_pte(pteval))
6091 if (huge_pte_write(pteval))
6093 if (flags & FOLL_WRITE)
6095 if (gup_must_unshare(flags, pte_page(pteval))) {
6102 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
6103 struct page **pages, struct vm_area_struct **vmas,
6104 unsigned long *position, unsigned long *nr_pages,
6105 long i, unsigned int flags, int *locked)
6107 unsigned long pfn_offset;
6108 unsigned long vaddr = *position;
6109 unsigned long remainder = *nr_pages;
6110 struct hstate *h = hstate_vma(vma);
6111 int err = -EFAULT, refs;
6113 while (vaddr < vma->vm_end && remainder) {
6115 spinlock_t *ptl = NULL;
6116 bool unshare = false;
6121 * If we have a pending SIGKILL, don't keep faulting pages and
6122 * potentially allocating memory.
6124 if (fatal_signal_pending(current)) {
6130 * Some archs (sparc64, sh*) have multiple pte_ts to
6131 * each hugepage. We have to make sure we get the
6132 * first, for the page indexing below to work.
6134 * Note that page table lock is not held when pte is null.
6136 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
6139 ptl = huge_pte_lock(h, mm, pte);
6140 absent = !pte || huge_pte_none(huge_ptep_get(pte));
6143 * When coredumping, it suits get_dump_page if we just return
6144 * an error where there's an empty slot with no huge pagecache
6145 * to back it. This way, we avoid allocating a hugepage, and
6146 * the sparse dumpfile avoids allocating disk blocks, but its
6147 * huge holes still show up with zeroes where they need to be.
6149 if (absent && (flags & FOLL_DUMP) &&
6150 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
6158 * We need call hugetlb_fault for both hugepages under migration
6159 * (in which case hugetlb_fault waits for the migration,) and
6160 * hwpoisoned hugepages (in which case we need to prevent the
6161 * caller from accessing to them.) In order to do this, we use
6162 * here is_swap_pte instead of is_hugetlb_entry_migration and
6163 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
6164 * both cases, and because we can't follow correct pages
6165 * directly from any kind of swap entries.
6168 __follow_hugetlb_must_fault(flags, pte, &unshare)) {
6170 unsigned int fault_flags = 0;
6174 if (flags & FOLL_WRITE)
6175 fault_flags |= FAULT_FLAG_WRITE;
6177 fault_flags |= FAULT_FLAG_UNSHARE;
6179 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
6180 FAULT_FLAG_KILLABLE;
6181 if (flags & FOLL_NOWAIT)
6182 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
6183 FAULT_FLAG_RETRY_NOWAIT;
6184 if (flags & FOLL_TRIED) {
6186 * Note: FAULT_FLAG_ALLOW_RETRY and
6187 * FAULT_FLAG_TRIED can co-exist
6189 fault_flags |= FAULT_FLAG_TRIED;
6191 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
6192 if (ret & VM_FAULT_ERROR) {
6193 err = vm_fault_to_errno(ret, flags);
6197 if (ret & VM_FAULT_RETRY) {
6199 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
6203 * VM_FAULT_RETRY must not return an
6204 * error, it will return zero
6207 * No need to update "position" as the
6208 * caller will not check it after
6209 * *nr_pages is set to 0.
6216 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
6217 page = pte_page(huge_ptep_get(pte));
6219 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
6220 !PageAnonExclusive(page), page);
6223 * If subpage information not requested, update counters
6224 * and skip the same_page loop below.
6226 if (!pages && !vmas && !pfn_offset &&
6227 (vaddr + huge_page_size(h) < vma->vm_end) &&
6228 (remainder >= pages_per_huge_page(h))) {
6229 vaddr += huge_page_size(h);
6230 remainder -= pages_per_huge_page(h);
6231 i += pages_per_huge_page(h);
6236 /* vaddr may not be aligned to PAGE_SIZE */
6237 refs = min3(pages_per_huge_page(h) - pfn_offset, remainder,
6238 (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT);
6241 record_subpages_vmas(mem_map_offset(page, pfn_offset),
6243 likely(pages) ? pages + i : NULL,
6244 vmas ? vmas + i : NULL);
6248 * try_grab_folio() should always succeed here,
6249 * because: a) we hold the ptl lock, and b) we've just
6250 * checked that the huge page is present in the page
6251 * tables. If the huge page is present, then the tail
6252 * pages must also be present. The ptl prevents the
6253 * head page and tail pages from being rearranged in
6254 * any way. So this page must be available at this
6255 * point, unless the page refcount overflowed:
6257 if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs,
6266 vaddr += (refs << PAGE_SHIFT);
6272 *nr_pages = remainder;
6274 * setting position is actually required only if remainder is
6275 * not zero but it's faster not to add a "if (remainder)"
6283 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
6284 unsigned long address, unsigned long end,
6285 pgprot_t newprot, unsigned long cp_flags)
6287 struct mm_struct *mm = vma->vm_mm;
6288 unsigned long start = address;
6291 struct hstate *h = hstate_vma(vma);
6292 unsigned long pages = 0, psize = huge_page_size(h);
6293 bool shared_pmd = false;
6294 struct mmu_notifier_range range;
6295 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
6296 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
6299 * In the case of shared PMDs, the area to flush could be beyond
6300 * start/end. Set range.start/range.end to cover the maximum possible
6301 * range if PMD sharing is possible.
6303 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
6304 0, vma, mm, start, end);
6305 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
6307 BUG_ON(address >= end);
6308 flush_cache_range(vma, range.start, range.end);
6310 mmu_notifier_invalidate_range_start(&range);
6311 i_mmap_lock_write(vma->vm_file->f_mapping);
6312 for (; address < end; address += psize) {
6314 ptep = huge_pte_offset(mm, address, psize);
6317 ptl = huge_pte_lock(h, mm, ptep);
6318 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
6320 * When uffd-wp is enabled on the vma, unshare
6321 * shouldn't happen at all. Warn about it if it
6322 * happened due to some reason.
6324 WARN_ON_ONCE(uffd_wp || uffd_wp_resolve);
6330 pte = huge_ptep_get(ptep);
6331 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
6335 if (unlikely(is_hugetlb_entry_migration(pte))) {
6336 swp_entry_t entry = pte_to_swp_entry(pte);
6337 struct page *page = pfn_swap_entry_to_page(entry);
6339 if (!is_readable_migration_entry(entry)) {
6343 entry = make_readable_exclusive_migration_entry(
6346 entry = make_readable_migration_entry(
6348 newpte = swp_entry_to_pte(entry);
6350 newpte = pte_swp_mkuffd_wp(newpte);
6351 else if (uffd_wp_resolve)
6352 newpte = pte_swp_clear_uffd_wp(newpte);
6353 set_huge_swap_pte_at(mm, address, ptep,
6360 if (unlikely(pte_marker_uffd_wp(pte))) {
6362 * This is changing a non-present pte into a none pte,
6363 * no need for huge_ptep_modify_prot_start/commit().
6365 if (uffd_wp_resolve)
6366 huge_pte_clear(mm, address, ptep, psize);
6368 if (!huge_pte_none(pte)) {
6370 unsigned int shift = huge_page_shift(hstate_vma(vma));
6372 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
6373 pte = huge_pte_modify(old_pte, newprot);
6374 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
6376 pte = huge_pte_mkuffd_wp(huge_pte_wrprotect(pte));
6377 else if (uffd_wp_resolve)
6378 pte = huge_pte_clear_uffd_wp(pte);
6379 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
6383 if (unlikely(uffd_wp))
6384 /* Safe to modify directly (none->non-present). */
6385 set_huge_pte_at(mm, address, ptep,
6386 make_pte_marker(PTE_MARKER_UFFD_WP));
6391 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
6392 * may have cleared our pud entry and done put_page on the page table:
6393 * once we release i_mmap_rwsem, another task can do the final put_page
6394 * and that page table be reused and filled with junk. If we actually
6395 * did unshare a page of pmds, flush the range corresponding to the pud.
6398 flush_hugetlb_tlb_range(vma, range.start, range.end);
6400 flush_hugetlb_tlb_range(vma, start, end);
6402 * No need to call mmu_notifier_invalidate_range() we are downgrading
6403 * page table protection not changing it to point to a new page.
6405 * See Documentation/mm/mmu_notifier.rst
6407 i_mmap_unlock_write(vma->vm_file->f_mapping);
6408 mmu_notifier_invalidate_range_end(&range);
6410 return pages << h->order;
6413 /* Return true if reservation was successful, false otherwise. */
6414 bool hugetlb_reserve_pages(struct inode *inode,
6416 struct vm_area_struct *vma,
6417 vm_flags_t vm_flags)
6420 struct hstate *h = hstate_inode(inode);
6421 struct hugepage_subpool *spool = subpool_inode(inode);
6422 struct resv_map *resv_map;
6423 struct hugetlb_cgroup *h_cg = NULL;
6424 long gbl_reserve, regions_needed = 0;
6426 /* This should never happen */
6428 VM_WARN(1, "%s called with a negative range\n", __func__);
6433 * Only apply hugepage reservation if asked. At fault time, an
6434 * attempt will be made for VM_NORESERVE to allocate a page
6435 * without using reserves
6437 if (vm_flags & VM_NORESERVE)
6441 * Shared mappings base their reservation on the number of pages that
6442 * are already allocated on behalf of the file. Private mappings need
6443 * to reserve the full area even if read-only as mprotect() may be
6444 * called to make the mapping read-write. Assume !vma is a shm mapping
6446 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6448 * resv_map can not be NULL as hugetlb_reserve_pages is only
6449 * called for inodes for which resv_maps were created (see
6450 * hugetlbfs_get_inode).
6452 resv_map = inode_resv_map(inode);
6454 chg = region_chg(resv_map, from, to, ®ions_needed);
6457 /* Private mapping. */
6458 resv_map = resv_map_alloc();
6464 set_vma_resv_map(vma, resv_map);
6465 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
6471 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
6472 chg * pages_per_huge_page(h), &h_cg) < 0)
6475 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
6476 /* For private mappings, the hugetlb_cgroup uncharge info hangs
6479 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
6483 * There must be enough pages in the subpool for the mapping. If
6484 * the subpool has a minimum size, there may be some global
6485 * reservations already in place (gbl_reserve).
6487 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
6488 if (gbl_reserve < 0)
6489 goto out_uncharge_cgroup;
6492 * Check enough hugepages are available for the reservation.
6493 * Hand the pages back to the subpool if there are not
6495 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
6499 * Account for the reservations made. Shared mappings record regions
6500 * that have reservations as they are shared by multiple VMAs.
6501 * When the last VMA disappears, the region map says how much
6502 * the reservation was and the page cache tells how much of
6503 * the reservation was consumed. Private mappings are per-VMA and
6504 * only the consumed reservations are tracked. When the VMA
6505 * disappears, the original reservation is the VMA size and the
6506 * consumed reservations are stored in the map. Hence, nothing
6507 * else has to be done for private mappings here
6509 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6510 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
6512 if (unlikely(add < 0)) {
6513 hugetlb_acct_memory(h, -gbl_reserve);
6515 } else if (unlikely(chg > add)) {
6517 * pages in this range were added to the reserve
6518 * map between region_chg and region_add. This
6519 * indicates a race with alloc_huge_page. Adjust
6520 * the subpool and reserve counts modified above
6521 * based on the difference.
6526 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
6527 * reference to h_cg->css. See comment below for detail.
6529 hugetlb_cgroup_uncharge_cgroup_rsvd(
6531 (chg - add) * pages_per_huge_page(h), h_cg);
6533 rsv_adjust = hugepage_subpool_put_pages(spool,
6535 hugetlb_acct_memory(h, -rsv_adjust);
6538 * The file_regions will hold their own reference to
6539 * h_cg->css. So we should release the reference held
6540 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
6543 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
6549 /* put back original number of pages, chg */
6550 (void)hugepage_subpool_put_pages(spool, chg);
6551 out_uncharge_cgroup:
6552 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
6553 chg * pages_per_huge_page(h), h_cg);
6555 if (!vma || vma->vm_flags & VM_MAYSHARE)
6556 /* Only call region_abort if the region_chg succeeded but the
6557 * region_add failed or didn't run.
6559 if (chg >= 0 && add < 0)
6560 region_abort(resv_map, from, to, regions_needed);
6561 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
6562 kref_put(&resv_map->refs, resv_map_release);
6566 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
6569 struct hstate *h = hstate_inode(inode);
6570 struct resv_map *resv_map = inode_resv_map(inode);
6572 struct hugepage_subpool *spool = subpool_inode(inode);
6576 * Since this routine can be called in the evict inode path for all
6577 * hugetlbfs inodes, resv_map could be NULL.
6580 chg = region_del(resv_map, start, end);
6582 * region_del() can fail in the rare case where a region
6583 * must be split and another region descriptor can not be
6584 * allocated. If end == LONG_MAX, it will not fail.
6590 spin_lock(&inode->i_lock);
6591 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
6592 spin_unlock(&inode->i_lock);
6595 * If the subpool has a minimum size, the number of global
6596 * reservations to be released may be adjusted.
6598 * Note that !resv_map implies freed == 0. So (chg - freed)
6599 * won't go negative.
6601 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
6602 hugetlb_acct_memory(h, -gbl_reserve);
6607 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
6608 static unsigned long page_table_shareable(struct vm_area_struct *svma,
6609 struct vm_area_struct *vma,
6610 unsigned long addr, pgoff_t idx)
6612 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
6614 unsigned long sbase = saddr & PUD_MASK;
6615 unsigned long s_end = sbase + PUD_SIZE;
6617 /* Allow segments to share if only one is marked locked */
6618 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
6619 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
6622 * match the virtual addresses, permission and the alignment of the
6625 if (pmd_index(addr) != pmd_index(saddr) ||
6626 vm_flags != svm_flags ||
6627 !range_in_vma(svma, sbase, s_end))
6633 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
6635 unsigned long base = addr & PUD_MASK;
6636 unsigned long end = base + PUD_SIZE;
6639 * check on proper vm_flags and page table alignment
6641 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
6646 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6648 #ifdef CONFIG_USERFAULTFD
6649 if (uffd_disable_huge_pmd_share(vma))
6652 return vma_shareable(vma, addr);
6656 * Determine if start,end range within vma could be mapped by shared pmd.
6657 * If yes, adjust start and end to cover range associated with possible
6658 * shared pmd mappings.
6660 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
6661 unsigned long *start, unsigned long *end)
6663 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
6664 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6667 * vma needs to span at least one aligned PUD size, and the range
6668 * must be at least partially within in.
6670 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
6671 (*end <= v_start) || (*start >= v_end))
6674 /* Extend the range to be PUD aligned for a worst case scenario */
6675 if (*start > v_start)
6676 *start = ALIGN_DOWN(*start, PUD_SIZE);
6679 *end = ALIGN(*end, PUD_SIZE);
6683 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
6684 * and returns the corresponding pte. While this is not necessary for the
6685 * !shared pmd case because we can allocate the pmd later as well, it makes the
6686 * code much cleaner.
6688 * This routine must be called with i_mmap_rwsem held in at least read mode if
6689 * sharing is possible. For hugetlbfs, this prevents removal of any page
6690 * table entries associated with the address space. This is important as we
6691 * are setting up sharing based on existing page table entries (mappings).
6693 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
6694 unsigned long addr, pud_t *pud)
6696 struct address_space *mapping = vma->vm_file->f_mapping;
6697 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
6699 struct vm_area_struct *svma;
6700 unsigned long saddr;
6705 i_mmap_assert_locked(mapping);
6706 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
6710 saddr = page_table_shareable(svma, vma, addr, idx);
6712 spte = huge_pte_offset(svma->vm_mm, saddr,
6713 vma_mmu_pagesize(svma));
6715 get_page(virt_to_page(spte));
6724 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
6725 if (pud_none(*pud)) {
6726 pud_populate(mm, pud,
6727 (pmd_t *)((unsigned long)spte & PAGE_MASK));
6730 put_page(virt_to_page(spte));
6734 pte = (pte_t *)pmd_alloc(mm, pud, addr);
6739 * unmap huge page backed by shared pte.
6741 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
6742 * indicated by page_count > 1, unmap is achieved by clearing pud and
6743 * decrementing the ref count. If count == 1, the pte page is not shared.
6745 * Called with page table lock held and i_mmap_rwsem held in write mode.
6747 * returns: 1 successfully unmapped a shared pte page
6748 * 0 the underlying pte page is not shared, or it is the last user
6750 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
6751 unsigned long *addr, pte_t *ptep)
6753 pgd_t *pgd = pgd_offset(mm, *addr);
6754 p4d_t *p4d = p4d_offset(pgd, *addr);
6755 pud_t *pud = pud_offset(p4d, *addr);
6757 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
6758 BUG_ON(page_count(virt_to_page(ptep)) == 0);
6759 if (page_count(virt_to_page(ptep)) == 1)
6763 put_page(virt_to_page(ptep));
6766 * This update of passed address optimizes loops sequentially
6767 * processing addresses in increments of huge page size (PMD_SIZE
6768 * in this case). By clearing the pud, a PUD_SIZE area is unmapped.
6769 * Update address to the 'last page' in the cleared area so that
6770 * calling loop can move to first page past this area.
6772 *addr |= PUD_SIZE - PMD_SIZE;
6776 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
6777 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
6778 unsigned long addr, pud_t *pud)
6783 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
6784 unsigned long *addr, pte_t *ptep)
6789 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
6790 unsigned long *start, unsigned long *end)
6794 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6798 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
6800 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
6801 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
6802 unsigned long addr, unsigned long sz)
6809 pgd = pgd_offset(mm, addr);
6810 p4d = p4d_alloc(mm, pgd, addr);
6813 pud = pud_alloc(mm, p4d, addr);
6815 if (sz == PUD_SIZE) {
6818 BUG_ON(sz != PMD_SIZE);
6819 if (want_pmd_share(vma, addr) && pud_none(*pud))
6820 pte = huge_pmd_share(mm, vma, addr, pud);
6822 pte = (pte_t *)pmd_alloc(mm, pud, addr);
6825 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
6831 * huge_pte_offset() - Walk the page table to resolve the hugepage
6832 * entry at address @addr
6834 * Return: Pointer to page table entry (PUD or PMD) for
6835 * address @addr, or NULL if a !p*d_present() entry is encountered and the
6836 * size @sz doesn't match the hugepage size at this level of the page
6839 pte_t *huge_pte_offset(struct mm_struct *mm,
6840 unsigned long addr, unsigned long sz)
6847 pgd = pgd_offset(mm, addr);
6848 if (!pgd_present(*pgd))
6850 p4d = p4d_offset(pgd, addr);
6851 if (!p4d_present(*p4d))
6854 pud = pud_offset(p4d, addr);
6856 /* must be pud huge, non-present or none */
6857 return (pte_t *)pud;
6858 if (!pud_present(*pud))
6860 /* must have a valid entry and size to go further */
6862 pmd = pmd_offset(pud, addr);
6863 /* must be pmd huge, non-present or none */
6864 return (pte_t *)pmd;
6867 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
6870 * These functions are overwritable if your architecture needs its own
6873 struct page * __weak
6874 follow_huge_addr(struct mm_struct *mm, unsigned long address,
6877 return ERR_PTR(-EINVAL);
6880 struct page * __weak
6881 follow_huge_pd(struct vm_area_struct *vma,
6882 unsigned long address, hugepd_t hpd, int flags, int pdshift)
6884 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
6888 struct page * __weak
6889 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
6890 pmd_t *pmd, int flags)
6892 struct page *page = NULL;
6897 * FOLL_PIN is not supported for follow_page(). Ordinary GUP goes via
6898 * follow_hugetlb_page().
6900 if (WARN_ON_ONCE(flags & FOLL_PIN))
6904 ptl = pmd_lockptr(mm, pmd);
6907 * make sure that the address range covered by this pmd is not
6908 * unmapped from other threads.
6910 if (!pmd_huge(*pmd))
6912 pte = huge_ptep_get((pte_t *)pmd);
6913 if (pte_present(pte)) {
6914 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
6916 * try_grab_page() should always succeed here, because: a) we
6917 * hold the pmd (ptl) lock, and b) we've just checked that the
6918 * huge pmd (head) page is present in the page tables. The ptl
6919 * prevents the head page and tail pages from being rearranged
6920 * in any way. So this page must be available at this point,
6921 * unless the page refcount overflowed:
6923 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
6928 if (is_hugetlb_entry_migration(pte)) {
6930 __migration_entry_wait_huge((pte_t *)pmd, ptl);
6934 * hwpoisoned entry is treated as no_page_table in
6935 * follow_page_mask().
6943 struct page * __weak
6944 follow_huge_pud(struct mm_struct *mm, unsigned long address,
6945 pud_t *pud, int flags)
6947 if (flags & (FOLL_GET | FOLL_PIN))
6950 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
6953 struct page * __weak
6954 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
6956 if (flags & (FOLL_GET | FOLL_PIN))
6959 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
6962 int isolate_hugetlb(struct page *page, struct list_head *list)
6966 spin_lock_irq(&hugetlb_lock);
6967 if (!PageHeadHuge(page) ||
6968 !HPageMigratable(page) ||
6969 !get_page_unless_zero(page)) {
6973 ClearHPageMigratable(page);
6974 list_move_tail(&page->lru, list);
6976 spin_unlock_irq(&hugetlb_lock);
6980 int get_hwpoison_huge_page(struct page *page, bool *hugetlb)
6985 spin_lock_irq(&hugetlb_lock);
6986 if (PageHeadHuge(page)) {
6988 if (HPageFreed(page))
6990 else if (HPageMigratable(page))
6991 ret = get_page_unless_zero(page);
6995 spin_unlock_irq(&hugetlb_lock);
6999 int get_huge_page_for_hwpoison(unsigned long pfn, int flags)
7003 spin_lock_irq(&hugetlb_lock);
7004 ret = __get_huge_page_for_hwpoison(pfn, flags);
7005 spin_unlock_irq(&hugetlb_lock);
7009 void putback_active_hugepage(struct page *page)
7011 spin_lock_irq(&hugetlb_lock);
7012 SetHPageMigratable(page);
7013 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
7014 spin_unlock_irq(&hugetlb_lock);
7018 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
7020 struct hstate *h = page_hstate(oldpage);
7022 hugetlb_cgroup_migrate(oldpage, newpage);
7023 set_page_owner_migrate_reason(newpage, reason);
7026 * transfer temporary state of the new huge page. This is
7027 * reverse to other transitions because the newpage is going to
7028 * be final while the old one will be freed so it takes over
7029 * the temporary status.
7031 * Also note that we have to transfer the per-node surplus state
7032 * here as well otherwise the global surplus count will not match
7035 if (HPageTemporary(newpage)) {
7036 int old_nid = page_to_nid(oldpage);
7037 int new_nid = page_to_nid(newpage);
7039 SetHPageTemporary(oldpage);
7040 ClearHPageTemporary(newpage);
7043 * There is no need to transfer the per-node surplus state
7044 * when we do not cross the node.
7046 if (new_nid == old_nid)
7048 spin_lock_irq(&hugetlb_lock);
7049 if (h->surplus_huge_pages_node[old_nid]) {
7050 h->surplus_huge_pages_node[old_nid]--;
7051 h->surplus_huge_pages_node[new_nid]++;
7053 spin_unlock_irq(&hugetlb_lock);
7058 * This function will unconditionally remove all the shared pmd pgtable entries
7059 * within the specific vma for a hugetlbfs memory range.
7061 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
7063 struct hstate *h = hstate_vma(vma);
7064 unsigned long sz = huge_page_size(h);
7065 struct mm_struct *mm = vma->vm_mm;
7066 struct mmu_notifier_range range;
7067 unsigned long address, start, end;
7071 if (!(vma->vm_flags & VM_MAYSHARE))
7074 start = ALIGN(vma->vm_start, PUD_SIZE);
7075 end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
7080 flush_cache_range(vma, start, end);
7082 * No need to call adjust_range_if_pmd_sharing_possible(), because
7083 * we have already done the PUD_SIZE alignment.
7085 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
7087 mmu_notifier_invalidate_range_start(&range);
7088 i_mmap_lock_write(vma->vm_file->f_mapping);
7089 for (address = start; address < end; address += PUD_SIZE) {
7090 unsigned long tmp = address;
7092 ptep = huge_pte_offset(mm, address, sz);
7095 ptl = huge_pte_lock(h, mm, ptep);
7096 /* We don't want 'address' to be changed */
7097 huge_pmd_unshare(mm, vma, &tmp, ptep);
7100 flush_hugetlb_tlb_range(vma, start, end);
7101 i_mmap_unlock_write(vma->vm_file->f_mapping);
7103 * No need to call mmu_notifier_invalidate_range(), see
7104 * Documentation/mm/mmu_notifier.rst.
7106 mmu_notifier_invalidate_range_end(&range);
7110 static bool cma_reserve_called __initdata;
7112 static int __init cmdline_parse_hugetlb_cma(char *p)
7119 if (sscanf(s, "%lu%n", &tmp, &count) != 1)
7122 if (s[count] == ':') {
7123 if (tmp >= MAX_NUMNODES)
7125 nid = array_index_nospec(tmp, MAX_NUMNODES);
7128 tmp = memparse(s, &s);
7129 hugetlb_cma_size_in_node[nid] = tmp;
7130 hugetlb_cma_size += tmp;
7133 * Skip the separator if have one, otherwise
7134 * break the parsing.
7141 hugetlb_cma_size = memparse(p, &p);
7149 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
7151 void __init hugetlb_cma_reserve(int order)
7153 unsigned long size, reserved, per_node;
7154 bool node_specific_cma_alloc = false;
7157 cma_reserve_called = true;
7159 if (!hugetlb_cma_size)
7162 for (nid = 0; nid < MAX_NUMNODES; nid++) {
7163 if (hugetlb_cma_size_in_node[nid] == 0)
7166 if (!node_online(nid)) {
7167 pr_warn("hugetlb_cma: invalid node %d specified\n", nid);
7168 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7169 hugetlb_cma_size_in_node[nid] = 0;
7173 if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) {
7174 pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n",
7175 nid, (PAGE_SIZE << order) / SZ_1M);
7176 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7177 hugetlb_cma_size_in_node[nid] = 0;
7179 node_specific_cma_alloc = true;
7183 /* Validate the CMA size again in case some invalid nodes specified. */
7184 if (!hugetlb_cma_size)
7187 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
7188 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
7189 (PAGE_SIZE << order) / SZ_1M);
7190 hugetlb_cma_size = 0;
7194 if (!node_specific_cma_alloc) {
7196 * If 3 GB area is requested on a machine with 4 numa nodes,
7197 * let's allocate 1 GB on first three nodes and ignore the last one.
7199 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
7200 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
7201 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
7205 for_each_online_node(nid) {
7207 char name[CMA_MAX_NAME];
7209 if (node_specific_cma_alloc) {
7210 if (hugetlb_cma_size_in_node[nid] == 0)
7213 size = hugetlb_cma_size_in_node[nid];
7215 size = min(per_node, hugetlb_cma_size - reserved);
7218 size = round_up(size, PAGE_SIZE << order);
7220 snprintf(name, sizeof(name), "hugetlb%d", nid);
7222 * Note that 'order per bit' is based on smallest size that
7223 * may be returned to CMA allocator in the case of
7224 * huge page demotion.
7226 res = cma_declare_contiguous_nid(0, size, 0,
7227 PAGE_SIZE << HUGETLB_PAGE_ORDER,
7229 &hugetlb_cma[nid], nid);
7231 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
7237 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
7240 if (reserved >= hugetlb_cma_size)
7246 * hugetlb_cma_size is used to determine if allocations from
7247 * cma are possible. Set to zero if no cma regions are set up.
7249 hugetlb_cma_size = 0;
7252 void __init hugetlb_cma_check(void)
7254 if (!hugetlb_cma_size || cma_reserve_called)
7257 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
7260 #endif /* CONFIG_CMA */