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;
69 __initdata LIST_HEAD(huge_boot_pages);
71 /* for command line parsing */
72 static struct hstate * __initdata parsed_hstate;
73 static unsigned long __initdata default_hstate_max_huge_pages;
74 static bool __initdata parsed_valid_hugepagesz = true;
75 static bool __initdata parsed_default_hugepagesz;
76 static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata;
79 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
80 * free_huge_pages, and surplus_huge_pages.
82 DEFINE_SPINLOCK(hugetlb_lock);
85 * Serializes faults on the same logical page. This is used to
86 * prevent spurious OOMs when the hugepage pool is fully utilized.
88 static int num_fault_mutexes;
89 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
91 /* Forward declaration */
92 static int hugetlb_acct_memory(struct hstate *h, long delta);
94 static inline bool subpool_is_free(struct hugepage_subpool *spool)
98 if (spool->max_hpages != -1)
99 return spool->used_hpages == 0;
100 if (spool->min_hpages != -1)
101 return spool->rsv_hpages == spool->min_hpages;
106 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
107 unsigned long irq_flags)
109 spin_unlock_irqrestore(&spool->lock, irq_flags);
111 /* If no pages are used, and no other handles to the subpool
112 * remain, give up any reservations based on minimum size and
113 * free the subpool */
114 if (subpool_is_free(spool)) {
115 if (spool->min_hpages != -1)
116 hugetlb_acct_memory(spool->hstate,
122 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
125 struct hugepage_subpool *spool;
127 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
131 spin_lock_init(&spool->lock);
133 spool->max_hpages = max_hpages;
135 spool->min_hpages = min_hpages;
137 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
141 spool->rsv_hpages = min_hpages;
146 void hugepage_put_subpool(struct hugepage_subpool *spool)
150 spin_lock_irqsave(&spool->lock, flags);
151 BUG_ON(!spool->count);
153 unlock_or_release_subpool(spool, flags);
157 * Subpool accounting for allocating and reserving pages.
158 * Return -ENOMEM if there are not enough resources to satisfy the
159 * request. Otherwise, return the number of pages by which the
160 * global pools must be adjusted (upward). The returned value may
161 * only be different than the passed value (delta) in the case where
162 * a subpool minimum size must be maintained.
164 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
172 spin_lock_irq(&spool->lock);
174 if (spool->max_hpages != -1) { /* maximum size accounting */
175 if ((spool->used_hpages + delta) <= spool->max_hpages)
176 spool->used_hpages += delta;
183 /* minimum size accounting */
184 if (spool->min_hpages != -1 && spool->rsv_hpages) {
185 if (delta > spool->rsv_hpages) {
187 * Asking for more reserves than those already taken on
188 * behalf of subpool. Return difference.
190 ret = delta - spool->rsv_hpages;
191 spool->rsv_hpages = 0;
193 ret = 0; /* reserves already accounted for */
194 spool->rsv_hpages -= delta;
199 spin_unlock_irq(&spool->lock);
204 * Subpool accounting for freeing and unreserving pages.
205 * Return the number of global page reservations that must be dropped.
206 * The return value may only be different than the passed value (delta)
207 * in the case where a subpool minimum size must be maintained.
209 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
218 spin_lock_irqsave(&spool->lock, flags);
220 if (spool->max_hpages != -1) /* maximum size accounting */
221 spool->used_hpages -= delta;
223 /* minimum size accounting */
224 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
225 if (spool->rsv_hpages + delta <= spool->min_hpages)
228 ret = spool->rsv_hpages + delta - spool->min_hpages;
230 spool->rsv_hpages += delta;
231 if (spool->rsv_hpages > spool->min_hpages)
232 spool->rsv_hpages = spool->min_hpages;
236 * If hugetlbfs_put_super couldn't free spool due to an outstanding
237 * quota reference, free it now.
239 unlock_or_release_subpool(spool, flags);
244 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
246 return HUGETLBFS_SB(inode->i_sb)->spool;
249 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
251 return subpool_inode(file_inode(vma->vm_file));
254 /* Helper that removes a struct file_region from the resv_map cache and returns
257 static struct file_region *
258 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
260 struct file_region *nrg = NULL;
262 VM_BUG_ON(resv->region_cache_count <= 0);
264 resv->region_cache_count--;
265 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
266 list_del(&nrg->link);
274 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
275 struct file_region *rg)
277 #ifdef CONFIG_CGROUP_HUGETLB
278 nrg->reservation_counter = rg->reservation_counter;
285 /* Helper that records hugetlb_cgroup uncharge info. */
286 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
288 struct resv_map *resv,
289 struct file_region *nrg)
291 #ifdef CONFIG_CGROUP_HUGETLB
293 nrg->reservation_counter =
294 &h_cg->rsvd_hugepage[hstate_index(h)];
295 nrg->css = &h_cg->css;
297 * The caller will hold exactly one h_cg->css reference for the
298 * whole contiguous reservation region. But this area might be
299 * scattered when there are already some file_regions reside in
300 * it. As a result, many file_regions may share only one css
301 * reference. In order to ensure that one file_region must hold
302 * exactly one h_cg->css reference, we should do css_get for
303 * each file_region and leave the reference held by caller
307 if (!resv->pages_per_hpage)
308 resv->pages_per_hpage = pages_per_huge_page(h);
309 /* pages_per_hpage should be the same for all entries in
312 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
314 nrg->reservation_counter = NULL;
320 static void put_uncharge_info(struct file_region *rg)
322 #ifdef CONFIG_CGROUP_HUGETLB
328 static bool has_same_uncharge_info(struct file_region *rg,
329 struct file_region *org)
331 #ifdef CONFIG_CGROUP_HUGETLB
332 return rg->reservation_counter == org->reservation_counter &&
340 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
342 struct file_region *nrg = NULL, *prg = NULL;
344 prg = list_prev_entry(rg, link);
345 if (&prg->link != &resv->regions && prg->to == rg->from &&
346 has_same_uncharge_info(prg, rg)) {
350 put_uncharge_info(rg);
356 nrg = list_next_entry(rg, link);
357 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
358 has_same_uncharge_info(nrg, rg)) {
359 nrg->from = rg->from;
362 put_uncharge_info(rg);
368 hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from,
369 long to, struct hstate *h, struct hugetlb_cgroup *cg,
370 long *regions_needed)
372 struct file_region *nrg;
374 if (!regions_needed) {
375 nrg = get_file_region_entry_from_cache(map, from, to);
376 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
377 list_add(&nrg->link, rg);
378 coalesce_file_region(map, nrg);
380 *regions_needed += 1;
386 * Must be called with resv->lock held.
388 * Calling this with regions_needed != NULL will count the number of pages
389 * to be added but will not modify the linked list. And regions_needed will
390 * indicate the number of file_regions needed in the cache to carry out to add
391 * the regions for this range.
393 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
394 struct hugetlb_cgroup *h_cg,
395 struct hstate *h, long *regions_needed)
398 struct list_head *head = &resv->regions;
399 long last_accounted_offset = f;
400 struct file_region *iter, *trg = NULL;
401 struct list_head *rg = NULL;
406 /* In this loop, we essentially handle an entry for the range
407 * [last_accounted_offset, iter->from), at every iteration, with some
410 list_for_each_entry_safe(iter, trg, head, link) {
411 /* Skip irrelevant regions that start before our range. */
412 if (iter->from < f) {
413 /* If this region ends after the last accounted offset,
414 * then we need to update last_accounted_offset.
416 if (iter->to > last_accounted_offset)
417 last_accounted_offset = iter->to;
421 /* When we find a region that starts beyond our range, we've
424 if (iter->from >= t) {
425 rg = iter->link.prev;
429 /* Add an entry for last_accounted_offset -> iter->from, and
430 * update last_accounted_offset.
432 if (iter->from > last_accounted_offset)
433 add += hugetlb_resv_map_add(resv, iter->link.prev,
434 last_accounted_offset,
438 last_accounted_offset = iter->to;
441 /* Handle the case where our range extends beyond
442 * last_accounted_offset.
446 if (last_accounted_offset < t)
447 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
448 t, h, h_cg, regions_needed);
453 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
455 static int allocate_file_region_entries(struct resv_map *resv,
457 __must_hold(&resv->lock)
459 struct list_head allocated_regions;
460 int to_allocate = 0, i = 0;
461 struct file_region *trg = NULL, *rg = NULL;
463 VM_BUG_ON(regions_needed < 0);
465 INIT_LIST_HEAD(&allocated_regions);
468 * Check for sufficient descriptors in the cache to accommodate
469 * the number of in progress add operations plus regions_needed.
471 * This is a while loop because when we drop the lock, some other call
472 * to region_add or region_del may have consumed some region_entries,
473 * so we keep looping here until we finally have enough entries for
474 * (adds_in_progress + regions_needed).
476 while (resv->region_cache_count <
477 (resv->adds_in_progress + regions_needed)) {
478 to_allocate = resv->adds_in_progress + regions_needed -
479 resv->region_cache_count;
481 /* At this point, we should have enough entries in the cache
482 * for all the existing adds_in_progress. We should only be
483 * needing to allocate for regions_needed.
485 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
487 spin_unlock(&resv->lock);
488 for (i = 0; i < to_allocate; i++) {
489 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
492 list_add(&trg->link, &allocated_regions);
495 spin_lock(&resv->lock);
497 list_splice(&allocated_regions, &resv->region_cache);
498 resv->region_cache_count += to_allocate;
504 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
512 * Add the huge page range represented by [f, t) to the reserve
513 * map. Regions will be taken from the cache to fill in this range.
514 * Sufficient regions should exist in the cache due to the previous
515 * call to region_chg with the same range, but in some cases the cache will not
516 * have sufficient entries due to races with other code doing region_add or
517 * region_del. The extra needed entries will be allocated.
519 * regions_needed is the out value provided by a previous call to region_chg.
521 * Return the number of new huge pages added to the map. This number is greater
522 * than or equal to zero. If file_region entries needed to be allocated for
523 * this operation and we were not able to allocate, it returns -ENOMEM.
524 * region_add of regions of length 1 never allocate file_regions and cannot
525 * fail; region_chg will always allocate at least 1 entry and a region_add for
526 * 1 page will only require at most 1 entry.
528 static long region_add(struct resv_map *resv, long f, long t,
529 long in_regions_needed, struct hstate *h,
530 struct hugetlb_cgroup *h_cg)
532 long add = 0, actual_regions_needed = 0;
534 spin_lock(&resv->lock);
537 /* Count how many regions are actually needed to execute this add. */
538 add_reservation_in_range(resv, f, t, NULL, NULL,
539 &actual_regions_needed);
542 * Check for sufficient descriptors in the cache to accommodate
543 * this add operation. Note that actual_regions_needed may be greater
544 * than in_regions_needed, as the resv_map may have been modified since
545 * the region_chg call. In this case, we need to make sure that we
546 * allocate extra entries, such that we have enough for all the
547 * existing adds_in_progress, plus the excess needed for this
550 if (actual_regions_needed > in_regions_needed &&
551 resv->region_cache_count <
552 resv->adds_in_progress +
553 (actual_regions_needed - in_regions_needed)) {
554 /* region_add operation of range 1 should never need to
555 * allocate file_region entries.
557 VM_BUG_ON(t - f <= 1);
559 if (allocate_file_region_entries(
560 resv, actual_regions_needed - in_regions_needed)) {
567 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
569 resv->adds_in_progress -= in_regions_needed;
571 spin_unlock(&resv->lock);
576 * Examine the existing reserve map and determine how many
577 * huge pages in the specified range [f, t) are NOT currently
578 * represented. This routine is called before a subsequent
579 * call to region_add that will actually modify the reserve
580 * map to add the specified range [f, t). region_chg does
581 * not change the number of huge pages represented by the
582 * map. A number of new file_region structures is added to the cache as a
583 * placeholder, for the subsequent region_add call to use. At least 1
584 * file_region structure is added.
586 * out_regions_needed is the number of regions added to the
587 * resv->adds_in_progress. This value needs to be provided to a follow up call
588 * to region_add or region_abort for proper accounting.
590 * Returns the number of huge pages that need to be added to the existing
591 * reservation map for the range [f, t). This number is greater or equal to
592 * zero. -ENOMEM is returned if a new file_region structure or cache entry
593 * is needed and can not be allocated.
595 static long region_chg(struct resv_map *resv, long f, long t,
596 long *out_regions_needed)
600 spin_lock(&resv->lock);
602 /* Count how many hugepages in this range are NOT represented. */
603 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
606 if (*out_regions_needed == 0)
607 *out_regions_needed = 1;
609 if (allocate_file_region_entries(resv, *out_regions_needed))
612 resv->adds_in_progress += *out_regions_needed;
614 spin_unlock(&resv->lock);
619 * Abort the in progress add operation. The adds_in_progress field
620 * of the resv_map keeps track of the operations in progress between
621 * calls to region_chg and region_add. Operations are sometimes
622 * aborted after the call to region_chg. In such cases, region_abort
623 * is called to decrement the adds_in_progress counter. regions_needed
624 * is the value returned by the region_chg call, it is used to decrement
625 * the adds_in_progress counter.
627 * NOTE: The range arguments [f, t) are not needed or used in this
628 * routine. They are kept to make reading the calling code easier as
629 * arguments will match the associated region_chg call.
631 static void region_abort(struct resv_map *resv, long f, long t,
634 spin_lock(&resv->lock);
635 VM_BUG_ON(!resv->region_cache_count);
636 resv->adds_in_progress -= regions_needed;
637 spin_unlock(&resv->lock);
641 * Delete the specified range [f, t) from the reserve map. If the
642 * t parameter is LONG_MAX, this indicates that ALL regions after f
643 * should be deleted. Locate the regions which intersect [f, t)
644 * and either trim, delete or split the existing regions.
646 * Returns the number of huge pages deleted from the reserve map.
647 * In the normal case, the return value is zero or more. In the
648 * case where a region must be split, a new region descriptor must
649 * be allocated. If the allocation fails, -ENOMEM will be returned.
650 * NOTE: If the parameter t == LONG_MAX, then we will never split
651 * a region and possibly return -ENOMEM. Callers specifying
652 * t == LONG_MAX do not need to check for -ENOMEM error.
654 static long region_del(struct resv_map *resv, long f, long t)
656 struct list_head *head = &resv->regions;
657 struct file_region *rg, *trg;
658 struct file_region *nrg = NULL;
662 spin_lock(&resv->lock);
663 list_for_each_entry_safe(rg, trg, head, link) {
665 * Skip regions before the range to be deleted. file_region
666 * ranges are normally of the form [from, to). However, there
667 * may be a "placeholder" entry in the map which is of the form
668 * (from, to) with from == to. Check for placeholder entries
669 * at the beginning of the range to be deleted.
671 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
677 if (f > rg->from && t < rg->to) { /* Must split region */
679 * Check for an entry in the cache before dropping
680 * lock and attempting allocation.
683 resv->region_cache_count > resv->adds_in_progress) {
684 nrg = list_first_entry(&resv->region_cache,
687 list_del(&nrg->link);
688 resv->region_cache_count--;
692 spin_unlock(&resv->lock);
693 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
700 hugetlb_cgroup_uncharge_file_region(
701 resv, rg, t - f, false);
703 /* New entry for end of split region */
707 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
709 INIT_LIST_HEAD(&nrg->link);
711 /* Original entry is trimmed */
714 list_add(&nrg->link, &rg->link);
719 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
720 del += rg->to - rg->from;
721 hugetlb_cgroup_uncharge_file_region(resv, rg,
722 rg->to - rg->from, true);
728 if (f <= rg->from) { /* Trim beginning of region */
729 hugetlb_cgroup_uncharge_file_region(resv, rg,
730 t - rg->from, false);
734 } else { /* Trim end of region */
735 hugetlb_cgroup_uncharge_file_region(resv, rg,
743 spin_unlock(&resv->lock);
749 * A rare out of memory error was encountered which prevented removal of
750 * the reserve map region for a page. The huge page itself was free'ed
751 * and removed from the page cache. This routine will adjust the subpool
752 * usage count, and the global reserve count if needed. By incrementing
753 * these counts, the reserve map entry which could not be deleted will
754 * appear as a "reserved" entry instead of simply dangling with incorrect
757 void hugetlb_fix_reserve_counts(struct inode *inode)
759 struct hugepage_subpool *spool = subpool_inode(inode);
761 bool reserved = false;
763 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
764 if (rsv_adjust > 0) {
765 struct hstate *h = hstate_inode(inode);
767 if (!hugetlb_acct_memory(h, 1))
769 } else if (!rsv_adjust) {
774 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
778 * Count and return the number of huge pages in the reserve map
779 * that intersect with the range [f, t).
781 static long region_count(struct resv_map *resv, long f, long t)
783 struct list_head *head = &resv->regions;
784 struct file_region *rg;
787 spin_lock(&resv->lock);
788 /* Locate each segment we overlap with, and count that overlap. */
789 list_for_each_entry(rg, head, link) {
798 seg_from = max(rg->from, f);
799 seg_to = min(rg->to, t);
801 chg += seg_to - seg_from;
803 spin_unlock(&resv->lock);
809 * Convert the address within this vma to the page offset within
810 * the mapping, in pagecache page units; huge pages here.
812 static pgoff_t vma_hugecache_offset(struct hstate *h,
813 struct vm_area_struct *vma, unsigned long address)
815 return ((address - vma->vm_start) >> huge_page_shift(h)) +
816 (vma->vm_pgoff >> huge_page_order(h));
819 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
820 unsigned long address)
822 return vma_hugecache_offset(hstate_vma(vma), vma, address);
824 EXPORT_SYMBOL_GPL(linear_hugepage_index);
827 * Return the size of the pages allocated when backing a VMA. In the majority
828 * cases this will be same size as used by the page table entries.
830 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
832 if (vma->vm_ops && vma->vm_ops->pagesize)
833 return vma->vm_ops->pagesize(vma);
836 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
839 * Return the page size being used by the MMU to back a VMA. In the majority
840 * of cases, the page size used by the kernel matches the MMU size. On
841 * architectures where it differs, an architecture-specific 'strong'
842 * version of this symbol is required.
844 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
846 return vma_kernel_pagesize(vma);
850 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
851 * bits of the reservation map pointer, which are always clear due to
854 #define HPAGE_RESV_OWNER (1UL << 0)
855 #define HPAGE_RESV_UNMAPPED (1UL << 1)
856 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
859 * These helpers are used to track how many pages are reserved for
860 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
861 * is guaranteed to have their future faults succeed.
863 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
864 * the reserve counters are updated with the hugetlb_lock held. It is safe
865 * to reset the VMA at fork() time as it is not in use yet and there is no
866 * chance of the global counters getting corrupted as a result of the values.
868 * The private mapping reservation is represented in a subtly different
869 * manner to a shared mapping. A shared mapping has a region map associated
870 * with the underlying file, this region map represents the backing file
871 * pages which have ever had a reservation assigned which this persists even
872 * after the page is instantiated. A private mapping has a region map
873 * associated with the original mmap which is attached to all VMAs which
874 * reference it, this region map represents those offsets which have consumed
875 * reservation ie. where pages have been instantiated.
877 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
879 return (unsigned long)vma->vm_private_data;
882 static void set_vma_private_data(struct vm_area_struct *vma,
885 vma->vm_private_data = (void *)value;
889 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
890 struct hugetlb_cgroup *h_cg,
893 #ifdef CONFIG_CGROUP_HUGETLB
895 resv_map->reservation_counter = NULL;
896 resv_map->pages_per_hpage = 0;
897 resv_map->css = NULL;
899 resv_map->reservation_counter =
900 &h_cg->rsvd_hugepage[hstate_index(h)];
901 resv_map->pages_per_hpage = pages_per_huge_page(h);
902 resv_map->css = &h_cg->css;
907 struct resv_map *resv_map_alloc(void)
909 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
910 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
912 if (!resv_map || !rg) {
918 kref_init(&resv_map->refs);
919 spin_lock_init(&resv_map->lock);
920 INIT_LIST_HEAD(&resv_map->regions);
922 resv_map->adds_in_progress = 0;
924 * Initialize these to 0. On shared mappings, 0's here indicate these
925 * fields don't do cgroup accounting. On private mappings, these will be
926 * re-initialized to the proper values, to indicate that hugetlb cgroup
927 * reservations are to be un-charged from here.
929 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
931 INIT_LIST_HEAD(&resv_map->region_cache);
932 list_add(&rg->link, &resv_map->region_cache);
933 resv_map->region_cache_count = 1;
938 void resv_map_release(struct kref *ref)
940 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
941 struct list_head *head = &resv_map->region_cache;
942 struct file_region *rg, *trg;
944 /* Clear out any active regions before we release the map. */
945 region_del(resv_map, 0, LONG_MAX);
947 /* ... and any entries left in the cache */
948 list_for_each_entry_safe(rg, trg, head, link) {
953 VM_BUG_ON(resv_map->adds_in_progress);
958 static inline struct resv_map *inode_resv_map(struct inode *inode)
961 * At inode evict time, i_mapping may not point to the original
962 * address space within the inode. This original address space
963 * contains the pointer to the resv_map. So, always use the
964 * address space embedded within the inode.
965 * The VERY common case is inode->mapping == &inode->i_data but,
966 * this may not be true for device special inodes.
968 return (struct resv_map *)(&inode->i_data)->private_data;
971 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
973 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
974 if (vma->vm_flags & VM_MAYSHARE) {
975 struct address_space *mapping = vma->vm_file->f_mapping;
976 struct inode *inode = mapping->host;
978 return inode_resv_map(inode);
981 return (struct resv_map *)(get_vma_private_data(vma) &
986 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
988 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
989 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
991 set_vma_private_data(vma, (get_vma_private_data(vma) &
992 HPAGE_RESV_MASK) | (unsigned long)map);
995 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
997 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
998 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1000 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
1003 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
1005 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1007 return (get_vma_private_data(vma) & flag) != 0;
1010 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
1011 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
1013 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1014 if (!(vma->vm_flags & VM_MAYSHARE))
1015 vma->vm_private_data = (void *)0;
1019 * Reset and decrement one ref on hugepage private reservation.
1020 * Called with mm->mmap_sem writer semaphore held.
1021 * This function should be only used by move_vma() and operate on
1022 * same sized vma. It should never come here with last ref on the
1025 void clear_vma_resv_huge_pages(struct vm_area_struct *vma)
1028 * Clear the old hugetlb private page reservation.
1029 * It has already been transferred to new_vma.
1031 * During a mremap() operation of a hugetlb vma we call move_vma()
1032 * which copies vma into new_vma and unmaps vma. After the copy
1033 * operation both new_vma and vma share a reference to the resv_map
1034 * struct, and at that point vma is about to be unmapped. We don't
1035 * want to return the reservation to the pool at unmap of vma because
1036 * the reservation still lives on in new_vma, so simply decrement the
1037 * ref here and remove the resv_map reference from this vma.
1039 struct resv_map *reservations = vma_resv_map(vma);
1041 if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1042 resv_map_put_hugetlb_cgroup_uncharge_info(reservations);
1043 kref_put(&reservations->refs, resv_map_release);
1046 reset_vma_resv_huge_pages(vma);
1049 /* Returns true if the VMA has associated reserve pages */
1050 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1052 if (vma->vm_flags & VM_NORESERVE) {
1054 * This address is already reserved by other process(chg == 0),
1055 * so, we should decrement reserved count. Without decrementing,
1056 * reserve count remains after releasing inode, because this
1057 * allocated page will go into page cache and is regarded as
1058 * coming from reserved pool in releasing step. Currently, we
1059 * don't have any other solution to deal with this situation
1060 * properly, so add work-around here.
1062 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1068 /* Shared mappings always use reserves */
1069 if (vma->vm_flags & VM_MAYSHARE) {
1071 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1072 * be a region map for all pages. The only situation where
1073 * there is no region map is if a hole was punched via
1074 * fallocate. In this case, there really are no reserves to
1075 * use. This situation is indicated if chg != 0.
1084 * Only the process that called mmap() has reserves for
1087 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1089 * Like the shared case above, a hole punch or truncate
1090 * could have been performed on the private mapping.
1091 * Examine the value of chg to determine if reserves
1092 * actually exist or were previously consumed.
1093 * Very Subtle - The value of chg comes from a previous
1094 * call to vma_needs_reserves(). The reserve map for
1095 * private mappings has different (opposite) semantics
1096 * than that of shared mappings. vma_needs_reserves()
1097 * has already taken this difference in semantics into
1098 * account. Therefore, the meaning of chg is the same
1099 * as in the shared case above. Code could easily be
1100 * combined, but keeping it separate draws attention to
1101 * subtle differences.
1112 static void enqueue_huge_page(struct hstate *h, struct page *page)
1114 int nid = page_to_nid(page);
1116 lockdep_assert_held(&hugetlb_lock);
1117 VM_BUG_ON_PAGE(page_count(page), page);
1119 list_move(&page->lru, &h->hugepage_freelists[nid]);
1120 h->free_huge_pages++;
1121 h->free_huge_pages_node[nid]++;
1122 SetHPageFreed(page);
1125 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1128 bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1130 lockdep_assert_held(&hugetlb_lock);
1131 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1132 if (pin && !is_longterm_pinnable_page(page))
1135 if (PageHWPoison(page))
1138 list_move(&page->lru, &h->hugepage_activelist);
1139 set_page_refcounted(page);
1140 ClearHPageFreed(page);
1141 h->free_huge_pages--;
1142 h->free_huge_pages_node[nid]--;
1149 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1152 unsigned int cpuset_mems_cookie;
1153 struct zonelist *zonelist;
1156 int node = NUMA_NO_NODE;
1158 zonelist = node_zonelist(nid, gfp_mask);
1161 cpuset_mems_cookie = read_mems_allowed_begin();
1162 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1165 if (!cpuset_zone_allowed(zone, gfp_mask))
1168 * no need to ask again on the same node. Pool is node rather than
1171 if (zone_to_nid(zone) == node)
1173 node = zone_to_nid(zone);
1175 page = dequeue_huge_page_node_exact(h, node);
1179 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1185 static struct page *dequeue_huge_page_vma(struct hstate *h,
1186 struct vm_area_struct *vma,
1187 unsigned long address, int avoid_reserve,
1190 struct page *page = NULL;
1191 struct mempolicy *mpol;
1193 nodemask_t *nodemask;
1197 * A child process with MAP_PRIVATE mappings created by their parent
1198 * have no page reserves. This check ensures that reservations are
1199 * not "stolen". The child may still get SIGKILLed
1201 if (!vma_has_reserves(vma, chg) &&
1202 h->free_huge_pages - h->resv_huge_pages == 0)
1205 /* If reserves cannot be used, ensure enough pages are in the pool */
1206 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1209 gfp_mask = htlb_alloc_mask(h);
1210 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1212 if (mpol_is_preferred_many(mpol)) {
1213 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1215 /* Fallback to all nodes if page==NULL */
1220 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1222 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1223 SetHPageRestoreReserve(page);
1224 h->resv_huge_pages--;
1227 mpol_cond_put(mpol);
1235 * common helper functions for hstate_next_node_to_{alloc|free}.
1236 * We may have allocated or freed a huge page based on a different
1237 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1238 * be outside of *nodes_allowed. Ensure that we use an allowed
1239 * node for alloc or free.
1241 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1243 nid = next_node_in(nid, *nodes_allowed);
1244 VM_BUG_ON(nid >= MAX_NUMNODES);
1249 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1251 if (!node_isset(nid, *nodes_allowed))
1252 nid = next_node_allowed(nid, nodes_allowed);
1257 * returns the previously saved node ["this node"] from which to
1258 * allocate a persistent huge page for the pool and advance the
1259 * next node from which to allocate, handling wrap at end of node
1262 static int hstate_next_node_to_alloc(struct hstate *h,
1263 nodemask_t *nodes_allowed)
1267 VM_BUG_ON(!nodes_allowed);
1269 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1270 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1276 * helper for remove_pool_huge_page() - return the previously saved
1277 * node ["this node"] from which to free a huge page. Advance the
1278 * next node id whether or not we find a free huge page to free so
1279 * that the next attempt to free addresses the next node.
1281 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1285 VM_BUG_ON(!nodes_allowed);
1287 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1288 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1293 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1294 for (nr_nodes = nodes_weight(*mask); \
1296 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1299 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1300 for (nr_nodes = nodes_weight(*mask); \
1302 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1305 /* used to demote non-gigantic_huge pages as well */
1306 static void __destroy_compound_gigantic_page(struct page *page,
1307 unsigned int order, bool demote)
1310 int nr_pages = 1 << order;
1311 struct page *p = page + 1;
1313 atomic_set(compound_mapcount_ptr(page), 0);
1314 atomic_set(compound_pincount_ptr(page), 0);
1316 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1318 clear_compound_head(p);
1320 set_page_refcounted(p);
1323 set_compound_order(page, 0);
1325 page[1].compound_nr = 0;
1327 __ClearPageHead(page);
1330 static void destroy_compound_hugetlb_page_for_demote(struct page *page,
1333 __destroy_compound_gigantic_page(page, order, true);
1336 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1337 static void destroy_compound_gigantic_page(struct page *page,
1340 __destroy_compound_gigantic_page(page, order, false);
1343 static void free_gigantic_page(struct page *page, unsigned int order)
1346 * If the page isn't allocated using the cma allocator,
1347 * cma_release() returns false.
1350 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1354 free_contig_range(page_to_pfn(page), 1 << order);
1357 #ifdef CONFIG_CONTIG_ALLOC
1358 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1359 int nid, nodemask_t *nodemask)
1361 unsigned long nr_pages = pages_per_huge_page(h);
1362 if (nid == NUMA_NO_NODE)
1363 nid = numa_mem_id();
1370 if (hugetlb_cma[nid]) {
1371 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1372 huge_page_order(h), true);
1377 if (!(gfp_mask & __GFP_THISNODE)) {
1378 for_each_node_mask(node, *nodemask) {
1379 if (node == nid || !hugetlb_cma[node])
1382 page = cma_alloc(hugetlb_cma[node], nr_pages,
1383 huge_page_order(h), true);
1391 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1394 #else /* !CONFIG_CONTIG_ALLOC */
1395 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1396 int nid, nodemask_t *nodemask)
1400 #endif /* CONFIG_CONTIG_ALLOC */
1402 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1403 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1404 int nid, nodemask_t *nodemask)
1408 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1409 static inline void destroy_compound_gigantic_page(struct page *page,
1410 unsigned int order) { }
1414 * Remove hugetlb page from lists, and update dtor so that page appears
1415 * as just a compound page.
1417 * A reference is held on the page, except in the case of demote.
1419 * Must be called with hugetlb lock held.
1421 static void __remove_hugetlb_page(struct hstate *h, struct page *page,
1422 bool adjust_surplus,
1425 int nid = page_to_nid(page);
1427 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1428 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1430 lockdep_assert_held(&hugetlb_lock);
1431 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1434 list_del(&page->lru);
1436 if (HPageFreed(page)) {
1437 h->free_huge_pages--;
1438 h->free_huge_pages_node[nid]--;
1440 if (adjust_surplus) {
1441 h->surplus_huge_pages--;
1442 h->surplus_huge_pages_node[nid]--;
1448 * For non-gigantic pages set the destructor to the normal compound
1449 * page dtor. This is needed in case someone takes an additional
1450 * temporary ref to the page, and freeing is delayed until they drop
1453 * For gigantic pages set the destructor to the null dtor. This
1454 * destructor will never be called. Before freeing the gigantic
1455 * page destroy_compound_gigantic_page will turn the compound page
1456 * into a simple group of pages. After this the destructor does not
1459 * This handles the case where more than one ref is held when and
1460 * after update_and_free_page is called.
1462 * In the case of demote we do not ref count the page as it will soon
1463 * be turned into a page of smaller size.
1466 set_page_refcounted(page);
1467 if (hstate_is_gigantic(h))
1468 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1470 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
1473 h->nr_huge_pages_node[nid]--;
1476 static void remove_hugetlb_page(struct hstate *h, struct page *page,
1477 bool adjust_surplus)
1479 __remove_hugetlb_page(h, page, adjust_surplus, false);
1482 static void remove_hugetlb_page_for_demote(struct hstate *h, struct page *page,
1483 bool adjust_surplus)
1485 __remove_hugetlb_page(h, page, adjust_surplus, true);
1488 static void add_hugetlb_page(struct hstate *h, struct page *page,
1489 bool adjust_surplus)
1492 int nid = page_to_nid(page);
1494 VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page);
1496 lockdep_assert_held(&hugetlb_lock);
1498 INIT_LIST_HEAD(&page->lru);
1500 h->nr_huge_pages_node[nid]++;
1502 if (adjust_surplus) {
1503 h->surplus_huge_pages++;
1504 h->surplus_huge_pages_node[nid]++;
1507 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1508 set_page_private(page, 0);
1509 SetHPageVmemmapOptimized(page);
1512 * This page is about to be managed by the hugetlb allocator and
1513 * should have no users. Drop our reference, and check for others
1516 zeroed = put_page_testzero(page);
1519 * It is VERY unlikely soneone else has taken a ref on
1520 * the page. In this case, we simply return as the
1521 * hugetlb destructor (free_huge_page) will be called
1522 * when this other ref is dropped.
1526 arch_clear_hugepage_flags(page);
1527 enqueue_huge_page(h, page);
1530 static void __update_and_free_page(struct hstate *h, struct page *page)
1533 struct page *subpage = page;
1535 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1538 if (hugetlb_vmemmap_alloc(h, page)) {
1539 spin_lock_irq(&hugetlb_lock);
1541 * If we cannot allocate vmemmap pages, just refuse to free the
1542 * page and put the page back on the hugetlb free list and treat
1543 * as a surplus page.
1545 add_hugetlb_page(h, page, true);
1546 spin_unlock_irq(&hugetlb_lock);
1550 for (i = 0; i < pages_per_huge_page(h);
1551 i++, subpage = mem_map_next(subpage, page, i)) {
1552 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1553 1 << PG_referenced | 1 << PG_dirty |
1554 1 << PG_active | 1 << PG_private |
1559 * Non-gigantic pages demoted from CMA allocated gigantic pages
1560 * need to be given back to CMA in free_gigantic_page.
1562 if (hstate_is_gigantic(h) ||
1563 hugetlb_cma_page(page, huge_page_order(h))) {
1564 destroy_compound_gigantic_page(page, huge_page_order(h));
1565 free_gigantic_page(page, huge_page_order(h));
1567 __free_pages(page, huge_page_order(h));
1572 * As update_and_free_page() can be called under any context, so we cannot
1573 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1574 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1575 * the vmemmap pages.
1577 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1578 * freed and frees them one-by-one. As the page->mapping pointer is going
1579 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1580 * structure of a lockless linked list of huge pages to be freed.
1582 static LLIST_HEAD(hpage_freelist);
1584 static void free_hpage_workfn(struct work_struct *work)
1586 struct llist_node *node;
1588 node = llist_del_all(&hpage_freelist);
1594 page = container_of((struct address_space **)node,
1595 struct page, mapping);
1597 page->mapping = NULL;
1599 * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate()
1600 * is going to trigger because a previous call to
1601 * remove_hugetlb_page() will set_compound_page_dtor(page,
1602 * NULL_COMPOUND_DTOR), so do not use page_hstate() directly.
1604 h = size_to_hstate(page_size(page));
1606 __update_and_free_page(h, page);
1611 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1613 static inline void flush_free_hpage_work(struct hstate *h)
1615 if (hugetlb_optimize_vmemmap_pages(h))
1616 flush_work(&free_hpage_work);
1619 static void update_and_free_page(struct hstate *h, struct page *page,
1622 if (!HPageVmemmapOptimized(page) || !atomic) {
1623 __update_and_free_page(h, page);
1628 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1630 * Only call schedule_work() if hpage_freelist is previously
1631 * empty. Otherwise, schedule_work() had been called but the workfn
1632 * hasn't retrieved the list yet.
1634 if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist))
1635 schedule_work(&free_hpage_work);
1638 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1640 struct page *page, *t_page;
1642 list_for_each_entry_safe(page, t_page, list, lru) {
1643 update_and_free_page(h, page, false);
1648 struct hstate *size_to_hstate(unsigned long size)
1652 for_each_hstate(h) {
1653 if (huge_page_size(h) == size)
1659 void free_huge_page(struct page *page)
1662 * Can't pass hstate in here because it is called from the
1663 * compound page destructor.
1665 struct hstate *h = page_hstate(page);
1666 int nid = page_to_nid(page);
1667 struct hugepage_subpool *spool = hugetlb_page_subpool(page);
1668 bool restore_reserve;
1669 unsigned long flags;
1671 VM_BUG_ON_PAGE(page_count(page), page);
1672 VM_BUG_ON_PAGE(page_mapcount(page), page);
1674 hugetlb_set_page_subpool(page, NULL);
1676 __ClearPageAnonExclusive(page);
1677 page->mapping = NULL;
1678 restore_reserve = HPageRestoreReserve(page);
1679 ClearHPageRestoreReserve(page);
1682 * If HPageRestoreReserve was set on page, page allocation consumed a
1683 * reservation. If the page was associated with a subpool, there
1684 * would have been a page reserved in the subpool before allocation
1685 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1686 * reservation, do not call hugepage_subpool_put_pages() as this will
1687 * remove the reserved page from the subpool.
1689 if (!restore_reserve) {
1691 * A return code of zero implies that the subpool will be
1692 * under its minimum size if the reservation is not restored
1693 * after page is free. Therefore, force restore_reserve
1696 if (hugepage_subpool_put_pages(spool, 1) == 0)
1697 restore_reserve = true;
1700 spin_lock_irqsave(&hugetlb_lock, flags);
1701 ClearHPageMigratable(page);
1702 hugetlb_cgroup_uncharge_page(hstate_index(h),
1703 pages_per_huge_page(h), page);
1704 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1705 pages_per_huge_page(h), page);
1706 if (restore_reserve)
1707 h->resv_huge_pages++;
1709 if (HPageTemporary(page)) {
1710 remove_hugetlb_page(h, page, false);
1711 spin_unlock_irqrestore(&hugetlb_lock, flags);
1712 update_and_free_page(h, page, true);
1713 } else if (h->surplus_huge_pages_node[nid]) {
1714 /* remove the page from active list */
1715 remove_hugetlb_page(h, page, true);
1716 spin_unlock_irqrestore(&hugetlb_lock, flags);
1717 update_and_free_page(h, page, true);
1719 arch_clear_hugepage_flags(page);
1720 enqueue_huge_page(h, page);
1721 spin_unlock_irqrestore(&hugetlb_lock, flags);
1726 * Must be called with the hugetlb lock held
1728 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1730 lockdep_assert_held(&hugetlb_lock);
1732 h->nr_huge_pages_node[nid]++;
1735 static void __prep_new_huge_page(struct hstate *h, struct page *page)
1737 hugetlb_vmemmap_free(h, page);
1738 INIT_LIST_HEAD(&page->lru);
1739 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1740 hugetlb_set_page_subpool(page, NULL);
1741 set_hugetlb_cgroup(page, NULL);
1742 set_hugetlb_cgroup_rsvd(page, NULL);
1745 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1747 __prep_new_huge_page(h, page);
1748 spin_lock_irq(&hugetlb_lock);
1749 __prep_account_new_huge_page(h, nid);
1750 spin_unlock_irq(&hugetlb_lock);
1753 static bool __prep_compound_gigantic_page(struct page *page, unsigned int order,
1757 int nr_pages = 1 << order;
1758 struct page *p = page + 1;
1760 /* we rely on prep_new_huge_page to set the destructor */
1761 set_compound_order(page, order);
1762 __ClearPageReserved(page);
1763 __SetPageHead(page);
1764 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1766 * For gigantic hugepages allocated through bootmem at
1767 * boot, it's safer to be consistent with the not-gigantic
1768 * hugepages and clear the PG_reserved bit from all tail pages
1769 * too. Otherwise drivers using get_user_pages() to access tail
1770 * pages may get the reference counting wrong if they see
1771 * PG_reserved set on a tail page (despite the head page not
1772 * having PG_reserved set). Enforcing this consistency between
1773 * head and tail pages allows drivers to optimize away a check
1774 * on the head page when they need know if put_page() is needed
1775 * after get_user_pages().
1777 __ClearPageReserved(p);
1779 * Subtle and very unlikely
1781 * Gigantic 'page allocators' such as memblock or cma will
1782 * return a set of pages with each page ref counted. We need
1783 * to turn this set of pages into a compound page with tail
1784 * page ref counts set to zero. Code such as speculative page
1785 * cache adding could take a ref on a 'to be' tail page.
1786 * We need to respect any increased ref count, and only set
1787 * the ref count to zero if count is currently 1. If count
1788 * is not 1, we return an error. An error return indicates
1789 * the set of pages can not be converted to a gigantic page.
1790 * The caller who allocated the pages should then discard the
1791 * pages using the appropriate free interface.
1793 * In the case of demote, the ref count will be zero.
1796 if (!page_ref_freeze(p, 1)) {
1797 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
1801 VM_BUG_ON_PAGE(page_count(p), p);
1803 set_compound_head(p, page);
1805 atomic_set(compound_mapcount_ptr(page), -1);
1806 atomic_set(compound_pincount_ptr(page), 0);
1810 /* undo tail page modifications made above */
1812 for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) {
1813 clear_compound_head(p);
1814 set_page_refcounted(p);
1816 /* need to clear PG_reserved on remaining tail pages */
1817 for (; j < nr_pages; j++, p = mem_map_next(p, page, j))
1818 __ClearPageReserved(p);
1819 set_compound_order(page, 0);
1821 page[1].compound_nr = 0;
1823 __ClearPageHead(page);
1827 static bool prep_compound_gigantic_page(struct page *page, unsigned int order)
1829 return __prep_compound_gigantic_page(page, order, false);
1832 static bool prep_compound_gigantic_page_for_demote(struct page *page,
1835 return __prep_compound_gigantic_page(page, order, true);
1839 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1840 * transparent huge pages. See the PageTransHuge() documentation for more
1843 int PageHuge(struct page *page)
1845 if (!PageCompound(page))
1848 page = compound_head(page);
1849 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1851 EXPORT_SYMBOL_GPL(PageHuge);
1854 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1855 * normal or transparent huge pages.
1857 int PageHeadHuge(struct page *page_head)
1859 if (!PageHead(page_head))
1862 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1864 EXPORT_SYMBOL_GPL(PageHeadHuge);
1867 * Find and lock address space (mapping) in write mode.
1869 * Upon entry, the page is locked which means that page_mapping() is
1870 * stable. Due to locking order, we can only trylock_write. If we can
1871 * not get the lock, simply return NULL to caller.
1873 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1875 struct address_space *mapping = page_mapping(hpage);
1880 if (i_mmap_trylock_write(mapping))
1886 pgoff_t hugetlb_basepage_index(struct page *page)
1888 struct page *page_head = compound_head(page);
1889 pgoff_t index = page_index(page_head);
1890 unsigned long compound_idx;
1892 if (compound_order(page_head) >= MAX_ORDER)
1893 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1895 compound_idx = page - page_head;
1897 return (index << compound_order(page_head)) + compound_idx;
1900 static struct page *alloc_buddy_huge_page(struct hstate *h,
1901 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1902 nodemask_t *node_alloc_noretry)
1904 int order = huge_page_order(h);
1906 bool alloc_try_hard = true;
1909 * By default we always try hard to allocate the page with
1910 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1911 * a loop (to adjust global huge page counts) and previous allocation
1912 * failed, do not continue to try hard on the same node. Use the
1913 * node_alloc_noretry bitmap to manage this state information.
1915 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1916 alloc_try_hard = false;
1917 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1919 gfp_mask |= __GFP_RETRY_MAYFAIL;
1920 if (nid == NUMA_NO_NODE)
1921 nid = numa_mem_id();
1922 page = __alloc_pages(gfp_mask, order, nid, nmask);
1924 __count_vm_event(HTLB_BUDDY_PGALLOC);
1926 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1929 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1930 * indicates an overall state change. Clear bit so that we resume
1931 * normal 'try hard' allocations.
1933 if (node_alloc_noretry && page && !alloc_try_hard)
1934 node_clear(nid, *node_alloc_noretry);
1937 * If we tried hard to get a page but failed, set bit so that
1938 * subsequent attempts will not try as hard until there is an
1939 * overall state change.
1941 if (node_alloc_noretry && !page && alloc_try_hard)
1942 node_set(nid, *node_alloc_noretry);
1948 * Common helper to allocate a fresh hugetlb page. All specific allocators
1949 * should use this function to get new hugetlb pages
1951 static struct page *alloc_fresh_huge_page(struct hstate *h,
1952 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1953 nodemask_t *node_alloc_noretry)
1959 if (hstate_is_gigantic(h))
1960 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1962 page = alloc_buddy_huge_page(h, gfp_mask,
1963 nid, nmask, node_alloc_noretry);
1967 if (hstate_is_gigantic(h)) {
1968 if (!prep_compound_gigantic_page(page, huge_page_order(h))) {
1970 * Rare failure to convert pages to compound page.
1971 * Free pages and try again - ONCE!
1973 free_gigantic_page(page, huge_page_order(h));
1981 prep_new_huge_page(h, page, page_to_nid(page));
1987 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1990 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1991 nodemask_t *node_alloc_noretry)
1995 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1997 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1998 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1999 node_alloc_noretry);
2007 put_page(page); /* free it into the hugepage allocator */
2013 * Remove huge page from pool from next node to free. Attempt to keep
2014 * persistent huge pages more or less balanced over allowed nodes.
2015 * This routine only 'removes' the hugetlb page. The caller must make
2016 * an additional call to free the page to low level allocators.
2017 * Called with hugetlb_lock locked.
2019 static struct page *remove_pool_huge_page(struct hstate *h,
2020 nodemask_t *nodes_allowed,
2024 struct page *page = NULL;
2026 lockdep_assert_held(&hugetlb_lock);
2027 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2029 * If we're returning unused surplus pages, only examine
2030 * nodes with surplus pages.
2032 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
2033 !list_empty(&h->hugepage_freelists[node])) {
2034 page = list_entry(h->hugepage_freelists[node].next,
2036 remove_hugetlb_page(h, page, acct_surplus);
2045 * Dissolve a given free hugepage into free buddy pages. This function does
2046 * nothing for in-use hugepages and non-hugepages.
2047 * This function returns values like below:
2049 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
2050 * when the system is under memory pressure and the feature of
2051 * freeing unused vmemmap pages associated with each hugetlb page
2053 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
2054 * (allocated or reserved.)
2055 * 0: successfully dissolved free hugepages or the page is not a
2056 * hugepage (considered as already dissolved)
2058 int dissolve_free_huge_page(struct page *page)
2063 /* Not to disrupt normal path by vainly holding hugetlb_lock */
2064 if (!PageHuge(page))
2067 spin_lock_irq(&hugetlb_lock);
2068 if (!PageHuge(page)) {
2073 if (!page_count(page)) {
2074 struct page *head = compound_head(page);
2075 struct hstate *h = page_hstate(head);
2076 if (h->free_huge_pages - h->resv_huge_pages == 0)
2080 * We should make sure that the page is already on the free list
2081 * when it is dissolved.
2083 if (unlikely(!HPageFreed(head))) {
2084 spin_unlock_irq(&hugetlb_lock);
2088 * Theoretically, we should return -EBUSY when we
2089 * encounter this race. In fact, we have a chance
2090 * to successfully dissolve the page if we do a
2091 * retry. Because the race window is quite small.
2092 * If we seize this opportunity, it is an optimization
2093 * for increasing the success rate of dissolving page.
2098 remove_hugetlb_page(h, head, false);
2099 h->max_huge_pages--;
2100 spin_unlock_irq(&hugetlb_lock);
2103 * Normally update_and_free_page will allocate required vmemmmap
2104 * before freeing the page. update_and_free_page will fail to
2105 * free the page if it can not allocate required vmemmap. We
2106 * need to adjust max_huge_pages if the page is not freed.
2107 * Attempt to allocate vmemmmap here so that we can take
2108 * appropriate action on failure.
2110 rc = hugetlb_vmemmap_alloc(h, head);
2113 * Move PageHWPoison flag from head page to the raw
2114 * error page, which makes any subpages rather than
2115 * the error page reusable.
2117 if (PageHWPoison(head) && page != head) {
2118 SetPageHWPoison(page);
2119 ClearPageHWPoison(head);
2121 update_and_free_page(h, head, false);
2123 spin_lock_irq(&hugetlb_lock);
2124 add_hugetlb_page(h, head, false);
2125 h->max_huge_pages++;
2126 spin_unlock_irq(&hugetlb_lock);
2132 spin_unlock_irq(&hugetlb_lock);
2137 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2138 * make specified memory blocks removable from the system.
2139 * Note that this will dissolve a free gigantic hugepage completely, if any
2140 * part of it lies within the given range.
2141 * Also note that if dissolve_free_huge_page() returns with an error, all
2142 * free hugepages that were dissolved before that error are lost.
2144 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2152 if (!hugepages_supported())
2155 order = huge_page_order(&default_hstate);
2157 order = min(order, huge_page_order(h));
2159 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << 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 /* oversize hugepages were init'ed in early boot */
3152 if (!hstate_is_gigantic(h))
3153 hugetlb_hstate_alloc_pages(h);
3156 * Set demote order for each hstate. Note that
3157 * h->demote_order is initially 0.
3158 * - We can not demote gigantic pages if runtime freeing
3159 * is not supported, so skip this.
3160 * - If CMA allocation is possible, we can not demote
3161 * HUGETLB_PAGE_ORDER or smaller size pages.
3163 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3165 if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER)
3167 for_each_hstate(h2) {
3170 if (h2->order < h->order &&
3171 h2->order > h->demote_order)
3172 h->demote_order = h2->order;
3177 static void __init report_hugepages(void)
3181 for_each_hstate(h) {
3184 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3185 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
3186 buf, h->free_huge_pages);
3190 #ifdef CONFIG_HIGHMEM
3191 static void try_to_free_low(struct hstate *h, unsigned long count,
3192 nodemask_t *nodes_allowed)
3195 LIST_HEAD(page_list);
3197 lockdep_assert_held(&hugetlb_lock);
3198 if (hstate_is_gigantic(h))
3202 * Collect pages to be freed on a list, and free after dropping lock
3204 for_each_node_mask(i, *nodes_allowed) {
3205 struct page *page, *next;
3206 struct list_head *freel = &h->hugepage_freelists[i];
3207 list_for_each_entry_safe(page, next, freel, lru) {
3208 if (count >= h->nr_huge_pages)
3210 if (PageHighMem(page))
3212 remove_hugetlb_page(h, page, false);
3213 list_add(&page->lru, &page_list);
3218 spin_unlock_irq(&hugetlb_lock);
3219 update_and_free_pages_bulk(h, &page_list);
3220 spin_lock_irq(&hugetlb_lock);
3223 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3224 nodemask_t *nodes_allowed)
3230 * Increment or decrement surplus_huge_pages. Keep node-specific counters
3231 * balanced by operating on them in a round-robin fashion.
3232 * Returns 1 if an adjustment was made.
3234 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3239 lockdep_assert_held(&hugetlb_lock);
3240 VM_BUG_ON(delta != -1 && delta != 1);
3243 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3244 if (h->surplus_huge_pages_node[node])
3248 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3249 if (h->surplus_huge_pages_node[node] <
3250 h->nr_huge_pages_node[node])
3257 h->surplus_huge_pages += delta;
3258 h->surplus_huge_pages_node[node] += delta;
3262 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3263 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3264 nodemask_t *nodes_allowed)
3266 unsigned long min_count, ret;
3268 LIST_HEAD(page_list);
3269 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3272 * Bit mask controlling how hard we retry per-node allocations.
3273 * If we can not allocate the bit mask, do not attempt to allocate
3274 * the requested huge pages.
3276 if (node_alloc_noretry)
3277 nodes_clear(*node_alloc_noretry);
3282 * resize_lock mutex prevents concurrent adjustments to number of
3283 * pages in hstate via the proc/sysfs interfaces.
3285 mutex_lock(&h->resize_lock);
3286 flush_free_hpage_work(h);
3287 spin_lock_irq(&hugetlb_lock);
3290 * Check for a node specific request.
3291 * Changing node specific huge page count may require a corresponding
3292 * change to the global count. In any case, the passed node mask
3293 * (nodes_allowed) will restrict alloc/free to the specified node.
3295 if (nid != NUMA_NO_NODE) {
3296 unsigned long old_count = count;
3298 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3300 * User may have specified a large count value which caused the
3301 * above calculation to overflow. In this case, they wanted
3302 * to allocate as many huge pages as possible. Set count to
3303 * largest possible value to align with their intention.
3305 if (count < old_count)
3310 * Gigantic pages runtime allocation depend on the capability for large
3311 * page range allocation.
3312 * If the system does not provide this feature, return an error when
3313 * the user tries to allocate gigantic pages but let the user free the
3314 * boottime allocated gigantic pages.
3316 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3317 if (count > persistent_huge_pages(h)) {
3318 spin_unlock_irq(&hugetlb_lock);
3319 mutex_unlock(&h->resize_lock);
3320 NODEMASK_FREE(node_alloc_noretry);
3323 /* Fall through to decrease pool */
3327 * Increase the pool size
3328 * First take pages out of surplus state. Then make up the
3329 * remaining difference by allocating fresh huge pages.
3331 * We might race with alloc_surplus_huge_page() here and be unable
3332 * to convert a surplus huge page to a normal huge page. That is
3333 * not critical, though, it just means the overall size of the
3334 * pool might be one hugepage larger than it needs to be, but
3335 * within all the constraints specified by the sysctls.
3337 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3338 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3342 while (count > persistent_huge_pages(h)) {
3344 * If this allocation races such that we no longer need the
3345 * page, free_huge_page will handle it by freeing the page
3346 * and reducing the surplus.
3348 spin_unlock_irq(&hugetlb_lock);
3350 /* yield cpu to avoid soft lockup */
3353 ret = alloc_pool_huge_page(h, nodes_allowed,
3354 node_alloc_noretry);
3355 spin_lock_irq(&hugetlb_lock);
3359 /* Bail for signals. Probably ctrl-c from user */
3360 if (signal_pending(current))
3365 * Decrease the pool size
3366 * First return free pages to the buddy allocator (being careful
3367 * to keep enough around to satisfy reservations). Then place
3368 * pages into surplus state as needed so the pool will shrink
3369 * to the desired size as pages become free.
3371 * By placing pages into the surplus state independent of the
3372 * overcommit value, we are allowing the surplus pool size to
3373 * exceed overcommit. There are few sane options here. Since
3374 * alloc_surplus_huge_page() is checking the global counter,
3375 * though, we'll note that we're not allowed to exceed surplus
3376 * and won't grow the pool anywhere else. Not until one of the
3377 * sysctls are changed, or the surplus pages go out of use.
3379 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3380 min_count = max(count, min_count);
3381 try_to_free_low(h, min_count, nodes_allowed);
3384 * Collect pages to be removed on list without dropping lock
3386 while (min_count < persistent_huge_pages(h)) {
3387 page = remove_pool_huge_page(h, nodes_allowed, 0);
3391 list_add(&page->lru, &page_list);
3393 /* free the pages after dropping lock */
3394 spin_unlock_irq(&hugetlb_lock);
3395 update_and_free_pages_bulk(h, &page_list);
3396 flush_free_hpage_work(h);
3397 spin_lock_irq(&hugetlb_lock);
3399 while (count < persistent_huge_pages(h)) {
3400 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3404 h->max_huge_pages = persistent_huge_pages(h);
3405 spin_unlock_irq(&hugetlb_lock);
3406 mutex_unlock(&h->resize_lock);
3408 NODEMASK_FREE(node_alloc_noretry);
3413 static int demote_free_huge_page(struct hstate *h, struct page *page)
3415 int i, nid = page_to_nid(page);
3416 struct hstate *target_hstate;
3419 target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order);
3421 remove_hugetlb_page_for_demote(h, page, false);
3422 spin_unlock_irq(&hugetlb_lock);
3424 rc = hugetlb_vmemmap_alloc(h, page);
3426 /* Allocation of vmemmmap failed, we can not demote page */
3427 spin_lock_irq(&hugetlb_lock);
3428 set_page_refcounted(page);
3429 add_hugetlb_page(h, page, false);
3434 * Use destroy_compound_hugetlb_page_for_demote for all huge page
3435 * sizes as it will not ref count pages.
3437 destroy_compound_hugetlb_page_for_demote(page, huge_page_order(h));
3440 * Taking target hstate mutex synchronizes with set_max_huge_pages.
3441 * Without the mutex, pages added to target hstate could be marked
3444 * Note that we already hold h->resize_lock. To prevent deadlock,
3445 * use the convention of always taking larger size hstate mutex first.
3447 mutex_lock(&target_hstate->resize_lock);
3448 for (i = 0; i < pages_per_huge_page(h);
3449 i += pages_per_huge_page(target_hstate)) {
3450 if (hstate_is_gigantic(target_hstate))
3451 prep_compound_gigantic_page_for_demote(page + i,
3452 target_hstate->order);
3454 prep_compound_page(page + i, target_hstate->order);
3455 set_page_private(page + i, 0);
3456 set_page_refcounted(page + i);
3457 prep_new_huge_page(target_hstate, page + i, nid);
3460 mutex_unlock(&target_hstate->resize_lock);
3462 spin_lock_irq(&hugetlb_lock);
3465 * Not absolutely necessary, but for consistency update max_huge_pages
3466 * based on pool changes for the demoted page.
3468 h->max_huge_pages--;
3469 target_hstate->max_huge_pages += pages_per_huge_page(h);
3474 static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
3475 __must_hold(&hugetlb_lock)
3480 lockdep_assert_held(&hugetlb_lock);
3482 /* We should never get here if no demote order */
3483 if (!h->demote_order) {
3484 pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n");
3485 return -EINVAL; /* internal error */
3488 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3489 list_for_each_entry(page, &h->hugepage_freelists[node], lru) {
3490 if (PageHWPoison(page))
3493 return demote_free_huge_page(h, page);
3498 * Only way to get here is if all pages on free lists are poisoned.
3499 * Return -EBUSY so that caller will not retry.
3504 #define HSTATE_ATTR_RO(_name) \
3505 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3507 #define HSTATE_ATTR_WO(_name) \
3508 static struct kobj_attribute _name##_attr = __ATTR_WO(_name)
3510 #define HSTATE_ATTR(_name) \
3511 static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
3513 static struct kobject *hugepages_kobj;
3514 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3516 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3518 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3522 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3523 if (hstate_kobjs[i] == kobj) {
3525 *nidp = NUMA_NO_NODE;
3529 return kobj_to_node_hstate(kobj, nidp);
3532 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3533 struct kobj_attribute *attr, char *buf)
3536 unsigned long nr_huge_pages;
3539 h = kobj_to_hstate(kobj, &nid);
3540 if (nid == NUMA_NO_NODE)
3541 nr_huge_pages = h->nr_huge_pages;
3543 nr_huge_pages = h->nr_huge_pages_node[nid];
3545 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3548 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3549 struct hstate *h, int nid,
3550 unsigned long count, size_t len)
3553 nodemask_t nodes_allowed, *n_mask;
3555 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3558 if (nid == NUMA_NO_NODE) {
3560 * global hstate attribute
3562 if (!(obey_mempolicy &&
3563 init_nodemask_of_mempolicy(&nodes_allowed)))
3564 n_mask = &node_states[N_MEMORY];
3566 n_mask = &nodes_allowed;
3569 * Node specific request. count adjustment happens in
3570 * set_max_huge_pages() after acquiring hugetlb_lock.
3572 init_nodemask_of_node(&nodes_allowed, nid);
3573 n_mask = &nodes_allowed;
3576 err = set_max_huge_pages(h, count, nid, n_mask);
3578 return err ? err : len;
3581 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3582 struct kobject *kobj, const char *buf,
3586 unsigned long count;
3590 err = kstrtoul(buf, 10, &count);
3594 h = kobj_to_hstate(kobj, &nid);
3595 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3598 static ssize_t nr_hugepages_show(struct kobject *kobj,
3599 struct kobj_attribute *attr, char *buf)
3601 return nr_hugepages_show_common(kobj, attr, buf);
3604 static ssize_t nr_hugepages_store(struct kobject *kobj,
3605 struct kobj_attribute *attr, const char *buf, size_t len)
3607 return nr_hugepages_store_common(false, kobj, buf, len);
3609 HSTATE_ATTR(nr_hugepages);
3614 * hstate attribute for optionally mempolicy-based constraint on persistent
3615 * huge page alloc/free.
3617 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3618 struct kobj_attribute *attr,
3621 return nr_hugepages_show_common(kobj, attr, buf);
3624 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3625 struct kobj_attribute *attr, const char *buf, size_t len)
3627 return nr_hugepages_store_common(true, kobj, buf, len);
3629 HSTATE_ATTR(nr_hugepages_mempolicy);
3633 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3634 struct kobj_attribute *attr, char *buf)
3636 struct hstate *h = kobj_to_hstate(kobj, NULL);
3637 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3640 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3641 struct kobj_attribute *attr, const char *buf, size_t count)
3644 unsigned long input;
3645 struct hstate *h = kobj_to_hstate(kobj, NULL);
3647 if (hstate_is_gigantic(h))
3650 err = kstrtoul(buf, 10, &input);
3654 spin_lock_irq(&hugetlb_lock);
3655 h->nr_overcommit_huge_pages = input;
3656 spin_unlock_irq(&hugetlb_lock);
3660 HSTATE_ATTR(nr_overcommit_hugepages);
3662 static ssize_t free_hugepages_show(struct kobject *kobj,
3663 struct kobj_attribute *attr, char *buf)
3666 unsigned long free_huge_pages;
3669 h = kobj_to_hstate(kobj, &nid);
3670 if (nid == NUMA_NO_NODE)
3671 free_huge_pages = h->free_huge_pages;
3673 free_huge_pages = h->free_huge_pages_node[nid];
3675 return sysfs_emit(buf, "%lu\n", free_huge_pages);
3677 HSTATE_ATTR_RO(free_hugepages);
3679 static ssize_t resv_hugepages_show(struct kobject *kobj,
3680 struct kobj_attribute *attr, char *buf)
3682 struct hstate *h = kobj_to_hstate(kobj, NULL);
3683 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3685 HSTATE_ATTR_RO(resv_hugepages);
3687 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3688 struct kobj_attribute *attr, char *buf)
3691 unsigned long surplus_huge_pages;
3694 h = kobj_to_hstate(kobj, &nid);
3695 if (nid == NUMA_NO_NODE)
3696 surplus_huge_pages = h->surplus_huge_pages;
3698 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3700 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3702 HSTATE_ATTR_RO(surplus_hugepages);
3704 static ssize_t demote_store(struct kobject *kobj,
3705 struct kobj_attribute *attr, const char *buf, size_t len)
3707 unsigned long nr_demote;
3708 unsigned long nr_available;
3709 nodemask_t nodes_allowed, *n_mask;
3714 err = kstrtoul(buf, 10, &nr_demote);
3717 h = kobj_to_hstate(kobj, &nid);
3719 if (nid != NUMA_NO_NODE) {
3720 init_nodemask_of_node(&nodes_allowed, nid);
3721 n_mask = &nodes_allowed;
3723 n_mask = &node_states[N_MEMORY];
3726 /* Synchronize with other sysfs operations modifying huge pages */
3727 mutex_lock(&h->resize_lock);
3728 spin_lock_irq(&hugetlb_lock);
3732 * Check for available pages to demote each time thorough the
3733 * loop as demote_pool_huge_page will drop hugetlb_lock.
3735 if (nid != NUMA_NO_NODE)
3736 nr_available = h->free_huge_pages_node[nid];
3738 nr_available = h->free_huge_pages;
3739 nr_available -= h->resv_huge_pages;
3743 err = demote_pool_huge_page(h, n_mask);
3750 spin_unlock_irq(&hugetlb_lock);
3751 mutex_unlock(&h->resize_lock);
3757 HSTATE_ATTR_WO(demote);
3759 static ssize_t demote_size_show(struct kobject *kobj,
3760 struct kobj_attribute *attr, char *buf)
3763 struct hstate *h = kobj_to_hstate(kobj, &nid);
3764 unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K;
3766 return sysfs_emit(buf, "%lukB\n", demote_size);
3769 static ssize_t demote_size_store(struct kobject *kobj,
3770 struct kobj_attribute *attr,
3771 const char *buf, size_t count)
3773 struct hstate *h, *demote_hstate;
3774 unsigned long demote_size;
3775 unsigned int demote_order;
3778 demote_size = (unsigned long)memparse(buf, NULL);
3780 demote_hstate = size_to_hstate(demote_size);
3783 demote_order = demote_hstate->order;
3784 if (demote_order < HUGETLB_PAGE_ORDER)
3787 /* demote order must be smaller than hstate order */
3788 h = kobj_to_hstate(kobj, &nid);
3789 if (demote_order >= h->order)
3792 /* resize_lock synchronizes access to demote size and writes */
3793 mutex_lock(&h->resize_lock);
3794 h->demote_order = demote_order;
3795 mutex_unlock(&h->resize_lock);
3799 HSTATE_ATTR(demote_size);
3801 static struct attribute *hstate_attrs[] = {
3802 &nr_hugepages_attr.attr,
3803 &nr_overcommit_hugepages_attr.attr,
3804 &free_hugepages_attr.attr,
3805 &resv_hugepages_attr.attr,
3806 &surplus_hugepages_attr.attr,
3808 &nr_hugepages_mempolicy_attr.attr,
3813 static const struct attribute_group hstate_attr_group = {
3814 .attrs = hstate_attrs,
3817 static struct attribute *hstate_demote_attrs[] = {
3818 &demote_size_attr.attr,
3823 static const struct attribute_group hstate_demote_attr_group = {
3824 .attrs = hstate_demote_attrs,
3827 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3828 struct kobject **hstate_kobjs,
3829 const struct attribute_group *hstate_attr_group)
3832 int hi = hstate_index(h);
3834 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3835 if (!hstate_kobjs[hi])
3838 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3840 kobject_put(hstate_kobjs[hi]);
3841 hstate_kobjs[hi] = NULL;
3844 if (h->demote_order) {
3845 if (sysfs_create_group(hstate_kobjs[hi],
3846 &hstate_demote_attr_group))
3847 pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name);
3853 static void __init hugetlb_sysfs_init(void)
3858 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3859 if (!hugepages_kobj)
3862 for_each_hstate(h) {
3863 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3864 hstate_kobjs, &hstate_attr_group);
3866 pr_err("HugeTLB: Unable to add hstate %s", h->name);
3873 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3874 * with node devices in node_devices[] using a parallel array. The array
3875 * index of a node device or _hstate == node id.
3876 * This is here to avoid any static dependency of the node device driver, in
3877 * the base kernel, on the hugetlb module.
3879 struct node_hstate {
3880 struct kobject *hugepages_kobj;
3881 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3883 static struct node_hstate node_hstates[MAX_NUMNODES];
3886 * A subset of global hstate attributes for node devices
3888 static struct attribute *per_node_hstate_attrs[] = {
3889 &nr_hugepages_attr.attr,
3890 &free_hugepages_attr.attr,
3891 &surplus_hugepages_attr.attr,
3895 static const struct attribute_group per_node_hstate_attr_group = {
3896 .attrs = per_node_hstate_attrs,
3900 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3901 * Returns node id via non-NULL nidp.
3903 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3907 for (nid = 0; nid < nr_node_ids; nid++) {
3908 struct node_hstate *nhs = &node_hstates[nid];
3910 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3911 if (nhs->hstate_kobjs[i] == kobj) {
3923 * Unregister hstate attributes from a single node device.
3924 * No-op if no hstate attributes attached.
3926 static void hugetlb_unregister_node(struct node *node)
3929 struct node_hstate *nhs = &node_hstates[node->dev.id];
3931 if (!nhs->hugepages_kobj)
3932 return; /* no hstate attributes */
3934 for_each_hstate(h) {
3935 int idx = hstate_index(h);
3936 if (nhs->hstate_kobjs[idx]) {
3937 kobject_put(nhs->hstate_kobjs[idx]);
3938 nhs->hstate_kobjs[idx] = NULL;
3942 kobject_put(nhs->hugepages_kobj);
3943 nhs->hugepages_kobj = NULL;
3948 * Register hstate attributes for a single node device.
3949 * No-op if attributes already registered.
3951 static void hugetlb_register_node(struct node *node)
3954 struct node_hstate *nhs = &node_hstates[node->dev.id];
3957 if (nhs->hugepages_kobj)
3958 return; /* already allocated */
3960 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3962 if (!nhs->hugepages_kobj)
3965 for_each_hstate(h) {
3966 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3968 &per_node_hstate_attr_group);
3970 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3971 h->name, node->dev.id);
3972 hugetlb_unregister_node(node);
3979 * hugetlb init time: register hstate attributes for all registered node
3980 * devices of nodes that have memory. All on-line nodes should have
3981 * registered their associated device by this time.
3983 static void __init hugetlb_register_all_nodes(void)
3987 for_each_node_state(nid, N_MEMORY) {
3988 struct node *node = node_devices[nid];
3989 if (node->dev.id == nid)
3990 hugetlb_register_node(node);
3994 * Let the node device driver know we're here so it can
3995 * [un]register hstate attributes on node hotplug.
3997 register_hugetlbfs_with_node(hugetlb_register_node,
3998 hugetlb_unregister_node);
4000 #else /* !CONFIG_NUMA */
4002 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4010 static void hugetlb_register_all_nodes(void) { }
4014 static int __init hugetlb_init(void)
4018 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
4021 if (!hugepages_supported()) {
4022 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
4023 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
4028 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
4029 * architectures depend on setup being done here.
4031 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
4032 if (!parsed_default_hugepagesz) {
4034 * If we did not parse a default huge page size, set
4035 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
4036 * number of huge pages for this default size was implicitly
4037 * specified, set that here as well.
4038 * Note that the implicit setting will overwrite an explicit
4039 * setting. A warning will be printed in this case.
4041 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
4042 if (default_hstate_max_huge_pages) {
4043 if (default_hstate.max_huge_pages) {
4046 string_get_size(huge_page_size(&default_hstate),
4047 1, STRING_UNITS_2, buf, 32);
4048 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
4049 default_hstate.max_huge_pages, buf);
4050 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
4051 default_hstate_max_huge_pages);
4053 default_hstate.max_huge_pages =
4054 default_hstate_max_huge_pages;
4056 for_each_online_node(i)
4057 default_hstate.max_huge_pages_node[i] =
4058 default_hugepages_in_node[i];
4062 hugetlb_cma_check();
4063 hugetlb_init_hstates();
4064 gather_bootmem_prealloc();
4067 hugetlb_sysfs_init();
4068 hugetlb_register_all_nodes();
4069 hugetlb_cgroup_file_init();
4072 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
4074 num_fault_mutexes = 1;
4076 hugetlb_fault_mutex_table =
4077 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
4079 BUG_ON(!hugetlb_fault_mutex_table);
4081 for (i = 0; i < num_fault_mutexes; i++)
4082 mutex_init(&hugetlb_fault_mutex_table[i]);
4085 subsys_initcall(hugetlb_init);
4087 /* Overwritten by architectures with more huge page sizes */
4088 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
4090 return size == HPAGE_SIZE;
4093 void __init hugetlb_add_hstate(unsigned int order)
4098 if (size_to_hstate(PAGE_SIZE << order)) {
4101 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
4103 h = &hstates[hugetlb_max_hstate++];
4104 mutex_init(&h->resize_lock);
4106 h->mask = ~(huge_page_size(h) - 1);
4107 for (i = 0; i < MAX_NUMNODES; ++i)
4108 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
4109 INIT_LIST_HEAD(&h->hugepage_activelist);
4110 h->next_nid_to_alloc = first_memory_node;
4111 h->next_nid_to_free = first_memory_node;
4112 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
4113 huge_page_size(h)/1024);
4114 hugetlb_vmemmap_init(h);
4119 bool __init __weak hugetlb_node_alloc_supported(void)
4124 static void __init hugepages_clear_pages_in_node(void)
4126 if (!hugetlb_max_hstate) {
4127 default_hstate_max_huge_pages = 0;
4128 memset(default_hugepages_in_node, 0,
4129 MAX_NUMNODES * sizeof(unsigned int));
4131 parsed_hstate->max_huge_pages = 0;
4132 memset(parsed_hstate->max_huge_pages_node, 0,
4133 MAX_NUMNODES * sizeof(unsigned int));
4138 * hugepages command line processing
4139 * hugepages normally follows a valid hugepagsz or default_hugepagsz
4140 * specification. If not, ignore the hugepages value. hugepages can also
4141 * be the first huge page command line option in which case it implicitly
4142 * specifies the number of huge pages for the default size.
4144 static int __init hugepages_setup(char *s)
4147 static unsigned long *last_mhp;
4148 int node = NUMA_NO_NODE;
4153 if (!parsed_valid_hugepagesz) {
4154 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
4155 parsed_valid_hugepagesz = true;
4160 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
4161 * yet, so this hugepages= parameter goes to the "default hstate".
4162 * Otherwise, it goes with the previously parsed hugepagesz or
4163 * default_hugepagesz.
4165 else if (!hugetlb_max_hstate)
4166 mhp = &default_hstate_max_huge_pages;
4168 mhp = &parsed_hstate->max_huge_pages;
4170 if (mhp == last_mhp) {
4171 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
4177 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4179 /* Parameter is node format */
4180 if (p[count] == ':') {
4181 if (!hugetlb_node_alloc_supported()) {
4182 pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n");
4185 if (tmp >= MAX_NUMNODES || !node_online(tmp))
4187 node = array_index_nospec(tmp, MAX_NUMNODES);
4189 /* Parse hugepages */
4190 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4192 if (!hugetlb_max_hstate)
4193 default_hugepages_in_node[node] = tmp;
4195 parsed_hstate->max_huge_pages_node[node] = tmp;
4197 /* Go to parse next node*/
4198 if (p[count] == ',')
4211 * Global state is always initialized later in hugetlb_init.
4212 * But we need to allocate gigantic hstates here early to still
4213 * use the bootmem allocator.
4215 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
4216 hugetlb_hstate_alloc_pages(parsed_hstate);
4223 pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p);
4224 hugepages_clear_pages_in_node();
4227 __setup("hugepages=", hugepages_setup);
4230 * hugepagesz command line processing
4231 * A specific huge page size can only be specified once with hugepagesz.
4232 * hugepagesz is followed by hugepages on the command line. The global
4233 * variable 'parsed_valid_hugepagesz' is used to determine if prior
4234 * hugepagesz argument was valid.
4236 static int __init hugepagesz_setup(char *s)
4241 parsed_valid_hugepagesz = false;
4242 size = (unsigned long)memparse(s, NULL);
4244 if (!arch_hugetlb_valid_size(size)) {
4245 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
4249 h = size_to_hstate(size);
4252 * hstate for this size already exists. This is normally
4253 * an error, but is allowed if the existing hstate is the
4254 * default hstate. More specifically, it is only allowed if
4255 * the number of huge pages for the default hstate was not
4256 * previously specified.
4258 if (!parsed_default_hugepagesz || h != &default_hstate ||
4259 default_hstate.max_huge_pages) {
4260 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
4265 * No need to call hugetlb_add_hstate() as hstate already
4266 * exists. But, do set parsed_hstate so that a following
4267 * hugepages= parameter will be applied to this hstate.
4270 parsed_valid_hugepagesz = true;
4274 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4275 parsed_valid_hugepagesz = true;
4278 __setup("hugepagesz=", hugepagesz_setup);
4281 * default_hugepagesz command line input
4282 * Only one instance of default_hugepagesz allowed on command line.
4284 static int __init default_hugepagesz_setup(char *s)
4289 parsed_valid_hugepagesz = false;
4290 if (parsed_default_hugepagesz) {
4291 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
4295 size = (unsigned long)memparse(s, NULL);
4297 if (!arch_hugetlb_valid_size(size)) {
4298 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
4302 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4303 parsed_valid_hugepagesz = true;
4304 parsed_default_hugepagesz = true;
4305 default_hstate_idx = hstate_index(size_to_hstate(size));
4308 * The number of default huge pages (for this size) could have been
4309 * specified as the first hugetlb parameter: hugepages=X. If so,
4310 * then default_hstate_max_huge_pages is set. If the default huge
4311 * page size is gigantic (>= MAX_ORDER), then the pages must be
4312 * allocated here from bootmem allocator.
4314 if (default_hstate_max_huge_pages) {
4315 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
4316 for_each_online_node(i)
4317 default_hstate.max_huge_pages_node[i] =
4318 default_hugepages_in_node[i];
4319 if (hstate_is_gigantic(&default_hstate))
4320 hugetlb_hstate_alloc_pages(&default_hstate);
4321 default_hstate_max_huge_pages = 0;
4326 __setup("default_hugepagesz=", default_hugepagesz_setup);
4328 static unsigned int allowed_mems_nr(struct hstate *h)
4331 unsigned int nr = 0;
4332 nodemask_t *mpol_allowed;
4333 unsigned int *array = h->free_huge_pages_node;
4334 gfp_t gfp_mask = htlb_alloc_mask(h);
4336 mpol_allowed = policy_nodemask_current(gfp_mask);
4338 for_each_node_mask(node, cpuset_current_mems_allowed) {
4339 if (!mpol_allowed || node_isset(node, *mpol_allowed))
4346 #ifdef CONFIG_SYSCTL
4347 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
4348 void *buffer, size_t *length,
4349 loff_t *ppos, unsigned long *out)
4351 struct ctl_table dup_table;
4354 * In order to avoid races with __do_proc_doulongvec_minmax(), we
4355 * can duplicate the @table and alter the duplicate of it.
4358 dup_table.data = out;
4360 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
4363 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
4364 struct ctl_table *table, int write,
4365 void *buffer, size_t *length, loff_t *ppos)
4367 struct hstate *h = &default_hstate;
4368 unsigned long tmp = h->max_huge_pages;
4371 if (!hugepages_supported())
4374 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4380 ret = __nr_hugepages_store_common(obey_mempolicy, h,
4381 NUMA_NO_NODE, tmp, *length);
4386 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
4387 void *buffer, size_t *length, loff_t *ppos)
4390 return hugetlb_sysctl_handler_common(false, table, write,
4391 buffer, length, ppos);
4395 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
4396 void *buffer, size_t *length, loff_t *ppos)
4398 return hugetlb_sysctl_handler_common(true, table, write,
4399 buffer, length, ppos);
4401 #endif /* CONFIG_NUMA */
4403 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
4404 void *buffer, size_t *length, loff_t *ppos)
4406 struct hstate *h = &default_hstate;
4410 if (!hugepages_supported())
4413 tmp = h->nr_overcommit_huge_pages;
4415 if (write && hstate_is_gigantic(h))
4418 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4424 spin_lock_irq(&hugetlb_lock);
4425 h->nr_overcommit_huge_pages = tmp;
4426 spin_unlock_irq(&hugetlb_lock);
4432 #endif /* CONFIG_SYSCTL */
4434 void hugetlb_report_meminfo(struct seq_file *m)
4437 unsigned long total = 0;
4439 if (!hugepages_supported())
4442 for_each_hstate(h) {
4443 unsigned long count = h->nr_huge_pages;
4445 total += huge_page_size(h) * count;
4447 if (h == &default_hstate)
4449 "HugePages_Total: %5lu\n"
4450 "HugePages_Free: %5lu\n"
4451 "HugePages_Rsvd: %5lu\n"
4452 "HugePages_Surp: %5lu\n"
4453 "Hugepagesize: %8lu kB\n",
4457 h->surplus_huge_pages,
4458 huge_page_size(h) / SZ_1K);
4461 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
4464 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
4466 struct hstate *h = &default_hstate;
4468 if (!hugepages_supported())
4471 return sysfs_emit_at(buf, len,
4472 "Node %d HugePages_Total: %5u\n"
4473 "Node %d HugePages_Free: %5u\n"
4474 "Node %d HugePages_Surp: %5u\n",
4475 nid, h->nr_huge_pages_node[nid],
4476 nid, h->free_huge_pages_node[nid],
4477 nid, h->surplus_huge_pages_node[nid]);
4480 void hugetlb_show_meminfo(void)
4485 if (!hugepages_supported())
4488 for_each_node_state(nid, N_MEMORY)
4490 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
4492 h->nr_huge_pages_node[nid],
4493 h->free_huge_pages_node[nid],
4494 h->surplus_huge_pages_node[nid],
4495 huge_page_size(h) / SZ_1K);
4498 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
4500 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
4501 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
4504 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
4505 unsigned long hugetlb_total_pages(void)
4508 unsigned long nr_total_pages = 0;
4511 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
4512 return nr_total_pages;
4515 static int hugetlb_acct_memory(struct hstate *h, long delta)
4522 spin_lock_irq(&hugetlb_lock);
4524 * When cpuset is configured, it breaks the strict hugetlb page
4525 * reservation as the accounting is done on a global variable. Such
4526 * reservation is completely rubbish in the presence of cpuset because
4527 * the reservation is not checked against page availability for the
4528 * current cpuset. Application can still potentially OOM'ed by kernel
4529 * with lack of free htlb page in cpuset that the task is in.
4530 * Attempt to enforce strict accounting with cpuset is almost
4531 * impossible (or too ugly) because cpuset is too fluid that
4532 * task or memory node can be dynamically moved between cpusets.
4534 * The change of semantics for shared hugetlb mapping with cpuset is
4535 * undesirable. However, in order to preserve some of the semantics,
4536 * we fall back to check against current free page availability as
4537 * a best attempt and hopefully to minimize the impact of changing
4538 * semantics that cpuset has.
4540 * Apart from cpuset, we also have memory policy mechanism that
4541 * also determines from which node the kernel will allocate memory
4542 * in a NUMA system. So similar to cpuset, we also should consider
4543 * the memory policy of the current task. Similar to the description
4547 if (gather_surplus_pages(h, delta) < 0)
4550 if (delta > allowed_mems_nr(h)) {
4551 return_unused_surplus_pages(h, delta);
4558 return_unused_surplus_pages(h, (unsigned long) -delta);
4561 spin_unlock_irq(&hugetlb_lock);
4565 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4567 struct resv_map *resv = vma_resv_map(vma);
4570 * This new VMA should share its siblings reservation map if present.
4571 * The VMA will only ever have a valid reservation map pointer where
4572 * it is being copied for another still existing VMA. As that VMA
4573 * has a reference to the reservation map it cannot disappear until
4574 * after this open call completes. It is therefore safe to take a
4575 * new reference here without additional locking.
4577 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
4578 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
4579 kref_get(&resv->refs);
4583 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4585 struct hstate *h = hstate_vma(vma);
4586 struct resv_map *resv = vma_resv_map(vma);
4587 struct hugepage_subpool *spool = subpool_vma(vma);
4588 unsigned long reserve, start, end;
4591 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4594 start = vma_hugecache_offset(h, vma, vma->vm_start);
4595 end = vma_hugecache_offset(h, vma, vma->vm_end);
4597 reserve = (end - start) - region_count(resv, start, end);
4598 hugetlb_cgroup_uncharge_counter(resv, start, end);
4601 * Decrement reserve counts. The global reserve count may be
4602 * adjusted if the subpool has a minimum size.
4604 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4605 hugetlb_acct_memory(h, -gbl_reserve);
4608 kref_put(&resv->refs, resv_map_release);
4611 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4613 if (addr & ~(huge_page_mask(hstate_vma(vma))))
4618 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4620 return huge_page_size(hstate_vma(vma));
4624 * We cannot handle pagefaults against hugetlb pages at all. They cause
4625 * handle_mm_fault() to try to instantiate regular-sized pages in the
4626 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
4629 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4636 * When a new function is introduced to vm_operations_struct and added
4637 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4638 * This is because under System V memory model, mappings created via
4639 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4640 * their original vm_ops are overwritten with shm_vm_ops.
4642 const struct vm_operations_struct hugetlb_vm_ops = {
4643 .fault = hugetlb_vm_op_fault,
4644 .open = hugetlb_vm_op_open,
4645 .close = hugetlb_vm_op_close,
4646 .may_split = hugetlb_vm_op_split,
4647 .pagesize = hugetlb_vm_op_pagesize,
4650 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4654 unsigned int shift = huge_page_shift(hstate_vma(vma));
4657 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4658 vma->vm_page_prot)));
4660 entry = huge_pte_wrprotect(mk_huge_pte(page,
4661 vma->vm_page_prot));
4663 entry = pte_mkyoung(entry);
4664 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4669 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4670 unsigned long address, pte_t *ptep)
4674 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4675 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4676 update_mmu_cache(vma, address, ptep);
4679 bool is_hugetlb_entry_migration(pte_t pte)
4683 if (huge_pte_none(pte) || pte_present(pte))
4685 swp = pte_to_swp_entry(pte);
4686 if (is_migration_entry(swp))
4692 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4696 if (huge_pte_none(pte) || pte_present(pte))
4698 swp = pte_to_swp_entry(pte);
4699 if (is_hwpoison_entry(swp))
4706 hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
4707 struct page *new_page)
4709 __SetPageUptodate(new_page);
4710 hugepage_add_new_anon_rmap(new_page, vma, addr);
4711 set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
4712 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
4713 ClearHPageRestoreReserve(new_page);
4714 SetHPageMigratable(new_page);
4717 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
4718 struct vm_area_struct *dst_vma,
4719 struct vm_area_struct *src_vma)
4721 pte_t *src_pte, *dst_pte, entry, dst_entry;
4722 struct page *ptepage;
4724 bool cow = is_cow_mapping(src_vma->vm_flags);
4725 struct hstate *h = hstate_vma(src_vma);
4726 unsigned long sz = huge_page_size(h);
4727 unsigned long npages = pages_per_huge_page(h);
4728 struct address_space *mapping = src_vma->vm_file->f_mapping;
4729 struct mmu_notifier_range range;
4733 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src_vma, src,
4736 mmu_notifier_invalidate_range_start(&range);
4737 mmap_assert_write_locked(src);
4738 raw_write_seqcount_begin(&src->write_protect_seq);
4741 * For shared mappings i_mmap_rwsem must be held to call
4742 * huge_pte_alloc, otherwise the returned ptep could go
4743 * away if part of a shared pmd and another thread calls
4746 i_mmap_lock_read(mapping);
4749 for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) {
4750 spinlock_t *src_ptl, *dst_ptl;
4751 src_pte = huge_pte_offset(src, addr, sz);
4754 dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz);
4761 * If the pagetables are shared don't copy or take references.
4762 * dst_pte == src_pte is the common case of src/dest sharing.
4764 * However, src could have 'unshared' and dst shares with
4765 * another vma. If dst_pte !none, this implies sharing.
4766 * Check here before taking page table lock, and once again
4767 * after taking the lock below.
4769 dst_entry = huge_ptep_get(dst_pte);
4770 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
4773 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4774 src_ptl = huge_pte_lockptr(h, src, src_pte);
4775 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4776 entry = huge_ptep_get(src_pte);
4777 dst_entry = huge_ptep_get(dst_pte);
4779 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
4781 * Skip if src entry none. Also, skip in the
4782 * unlikely case dst entry !none as this implies
4783 * sharing with another vma.
4786 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
4787 is_hugetlb_entry_hwpoisoned(entry))) {
4788 swp_entry_t swp_entry = pte_to_swp_entry(entry);
4789 bool uffd_wp = huge_pte_uffd_wp(entry);
4791 if (!is_readable_migration_entry(swp_entry) && cow) {
4793 * COW mappings require pages in both
4794 * parent and child to be set to read.
4796 swp_entry = make_readable_migration_entry(
4797 swp_offset(swp_entry));
4798 entry = swp_entry_to_pte(swp_entry);
4799 if (userfaultfd_wp(src_vma) && uffd_wp)
4800 entry = huge_pte_mkuffd_wp(entry);
4801 set_huge_pte_at(src, addr, src_pte, entry);
4803 if (!userfaultfd_wp(dst_vma) && uffd_wp)
4804 entry = huge_pte_clear_uffd_wp(entry);
4805 set_huge_pte_at(dst, addr, dst_pte, entry);
4806 } else if (unlikely(is_pte_marker(entry))) {
4808 * We copy the pte marker only if the dst vma has
4811 if (userfaultfd_wp(dst_vma))
4812 set_huge_pte_at(dst, addr, dst_pte, entry);
4814 entry = huge_ptep_get(src_pte);
4815 ptepage = pte_page(entry);
4819 * Failing to duplicate the anon rmap is a rare case
4820 * where we see pinned hugetlb pages while they're
4821 * prone to COW. We need to do the COW earlier during
4824 * When pre-allocating the page or copying data, we
4825 * need to be without the pgtable locks since we could
4826 * sleep during the process.
4828 if (!PageAnon(ptepage)) {
4829 page_dup_file_rmap(ptepage, true);
4830 } else if (page_try_dup_anon_rmap(ptepage, true,
4832 pte_t src_pte_old = entry;
4835 spin_unlock(src_ptl);
4836 spin_unlock(dst_ptl);
4837 /* Do not use reserve as it's private owned */
4838 new = alloc_huge_page(dst_vma, addr, 1);
4844 copy_user_huge_page(new, ptepage, addr, dst_vma,
4848 /* Install the new huge page if src pte stable */
4849 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4850 src_ptl = huge_pte_lockptr(h, src, src_pte);
4851 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4852 entry = huge_ptep_get(src_pte);
4853 if (!pte_same(src_pte_old, entry)) {
4854 restore_reserve_on_error(h, dst_vma, addr,
4857 /* dst_entry won't change as in child */
4860 hugetlb_install_page(dst_vma, dst_pte, addr, new);
4861 spin_unlock(src_ptl);
4862 spin_unlock(dst_ptl);
4868 * No need to notify as we are downgrading page
4869 * table protection not changing it to point
4872 * See Documentation/mm/mmu_notifier.rst
4874 huge_ptep_set_wrprotect(src, addr, src_pte);
4875 entry = huge_pte_wrprotect(entry);
4878 set_huge_pte_at(dst, addr, dst_pte, entry);
4879 hugetlb_count_add(npages, dst);
4881 spin_unlock(src_ptl);
4882 spin_unlock(dst_ptl);
4886 raw_write_seqcount_end(&src->write_protect_seq);
4887 mmu_notifier_invalidate_range_end(&range);
4889 i_mmap_unlock_read(mapping);
4895 static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
4896 unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte)
4898 struct hstate *h = hstate_vma(vma);
4899 struct mm_struct *mm = vma->vm_mm;
4900 spinlock_t *src_ptl, *dst_ptl;
4903 dst_ptl = huge_pte_lock(h, mm, dst_pte);
4904 src_ptl = huge_pte_lockptr(h, mm, src_pte);
4907 * We don't have to worry about the ordering of src and dst ptlocks
4908 * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock.
4910 if (src_ptl != dst_ptl)
4911 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4913 pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);
4914 set_huge_pte_at(mm, new_addr, dst_pte, pte);
4916 if (src_ptl != dst_ptl)
4917 spin_unlock(src_ptl);
4918 spin_unlock(dst_ptl);
4921 int move_hugetlb_page_tables(struct vm_area_struct *vma,
4922 struct vm_area_struct *new_vma,
4923 unsigned long old_addr, unsigned long new_addr,
4926 struct hstate *h = hstate_vma(vma);
4927 struct address_space *mapping = vma->vm_file->f_mapping;
4928 unsigned long sz = huge_page_size(h);
4929 struct mm_struct *mm = vma->vm_mm;
4930 unsigned long old_end = old_addr + len;
4931 unsigned long old_addr_copy;
4932 pte_t *src_pte, *dst_pte;
4933 struct mmu_notifier_range range;
4934 bool shared_pmd = false;
4936 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr,
4938 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4940 * In case of shared PMDs, we should cover the maximum possible
4943 flush_cache_range(vma, range.start, range.end);
4945 mmu_notifier_invalidate_range_start(&range);
4946 /* Prevent race with file truncation */
4947 i_mmap_lock_write(mapping);
4948 for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
4949 src_pte = huge_pte_offset(mm, old_addr, sz);
4952 if (huge_pte_none(huge_ptep_get(src_pte)))
4955 /* old_addr arg to huge_pmd_unshare() is a pointer and so the
4956 * arg may be modified. Pass a copy instead to preserve the
4957 * value in old_addr.
4959 old_addr_copy = old_addr;
4961 if (huge_pmd_unshare(mm, vma, &old_addr_copy, src_pte)) {
4966 dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz);
4970 move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte);
4974 flush_tlb_range(vma, range.start, range.end);
4976 flush_tlb_range(vma, old_end - len, old_end);
4977 mmu_notifier_invalidate_range_end(&range);
4978 i_mmap_unlock_write(mapping);
4980 return len + old_addr - old_end;
4983 static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
4984 unsigned long start, unsigned long end,
4985 struct page *ref_page, zap_flags_t zap_flags)
4987 struct mm_struct *mm = vma->vm_mm;
4988 unsigned long address;
4993 struct hstate *h = hstate_vma(vma);
4994 unsigned long sz = huge_page_size(h);
4995 struct mmu_notifier_range range;
4996 bool force_flush = false;
4998 WARN_ON(!is_vm_hugetlb_page(vma));
4999 BUG_ON(start & ~huge_page_mask(h));
5000 BUG_ON(end & ~huge_page_mask(h));
5003 * This is a hugetlb vma, all the pte entries should point
5006 tlb_change_page_size(tlb, sz);
5007 tlb_start_vma(tlb, vma);
5010 * If sharing possible, alert mmu notifiers of worst case.
5012 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
5014 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5015 mmu_notifier_invalidate_range_start(&range);
5017 for (; address < end; address += sz) {
5018 ptep = huge_pte_offset(mm, address, sz);
5022 ptl = huge_pte_lock(h, mm, ptep);
5023 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
5025 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
5030 pte = huge_ptep_get(ptep);
5031 if (huge_pte_none(pte)) {
5037 * Migrating hugepage or HWPoisoned hugepage is already
5038 * unmapped and its refcount is dropped, so just clear pte here.
5040 if (unlikely(!pte_present(pte))) {
5042 * If the pte was wr-protected by uffd-wp in any of the
5043 * swap forms, meanwhile the caller does not want to
5044 * drop the uffd-wp bit in this zap, then replace the
5045 * pte with a marker.
5047 if (pte_swp_uffd_wp_any(pte) &&
5048 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5049 set_huge_pte_at(mm, address, ptep,
5050 make_pte_marker(PTE_MARKER_UFFD_WP));
5052 huge_pte_clear(mm, address, ptep, sz);
5057 page = pte_page(pte);
5059 * If a reference page is supplied, it is because a specific
5060 * page is being unmapped, not a range. Ensure the page we
5061 * are about to unmap is the actual page of interest.
5064 if (page != ref_page) {
5069 * Mark the VMA as having unmapped its page so that
5070 * future faults in this VMA will fail rather than
5071 * looking like data was lost
5073 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
5076 pte = huge_ptep_get_and_clear(mm, address, ptep);
5077 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
5078 if (huge_pte_dirty(pte))
5079 set_page_dirty(page);
5080 /* Leave a uffd-wp pte marker if needed */
5081 if (huge_pte_uffd_wp(pte) &&
5082 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5083 set_huge_pte_at(mm, address, ptep,
5084 make_pte_marker(PTE_MARKER_UFFD_WP));
5085 hugetlb_count_sub(pages_per_huge_page(h), mm);
5086 page_remove_rmap(page, vma, true);
5089 tlb_remove_page_size(tlb, page, huge_page_size(h));
5091 * Bail out after unmapping reference page if supplied
5096 mmu_notifier_invalidate_range_end(&range);
5097 tlb_end_vma(tlb, vma);
5100 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
5101 * could defer the flush until now, since by holding i_mmap_rwsem we
5102 * guaranteed that the last refernece would not be dropped. But we must
5103 * do the flushing before we return, as otherwise i_mmap_rwsem will be
5104 * dropped and the last reference to the shared PMDs page might be
5107 * In theory we could defer the freeing of the PMD pages as well, but
5108 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
5109 * detect sharing, so we cannot defer the release of the page either.
5110 * Instead, do flush now.
5113 tlb_flush_mmu_tlbonly(tlb);
5116 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
5117 struct vm_area_struct *vma, unsigned long start,
5118 unsigned long end, struct page *ref_page,
5119 zap_flags_t zap_flags)
5121 __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags);
5124 * Clear this flag so that x86's huge_pmd_share page_table_shareable
5125 * test will fail on a vma being torn down, and not grab a page table
5126 * on its way out. We're lucky that the flag has such an appropriate
5127 * name, and can in fact be safely cleared here. We could clear it
5128 * before the __unmap_hugepage_range above, but all that's necessary
5129 * is to clear it before releasing the i_mmap_rwsem. This works
5130 * because in the context this is called, the VMA is about to be
5131 * destroyed and the i_mmap_rwsem is held.
5133 vma->vm_flags &= ~VM_MAYSHARE;
5136 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
5137 unsigned long end, struct page *ref_page,
5138 zap_flags_t zap_flags)
5140 struct mmu_gather tlb;
5142 tlb_gather_mmu(&tlb, vma->vm_mm);
5143 __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags);
5144 tlb_finish_mmu(&tlb);
5148 * This is called when the original mapper is failing to COW a MAP_PRIVATE
5149 * mapping it owns the reserve page for. The intention is to unmap the page
5150 * from other VMAs and let the children be SIGKILLed if they are faulting the
5153 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
5154 struct page *page, unsigned long address)
5156 struct hstate *h = hstate_vma(vma);
5157 struct vm_area_struct *iter_vma;
5158 struct address_space *mapping;
5162 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
5163 * from page cache lookup which is in HPAGE_SIZE units.
5165 address = address & huge_page_mask(h);
5166 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
5168 mapping = vma->vm_file->f_mapping;
5171 * Take the mapping lock for the duration of the table walk. As
5172 * this mapping should be shared between all the VMAs,
5173 * __unmap_hugepage_range() is called as the lock is already held
5175 i_mmap_lock_write(mapping);
5176 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
5177 /* Do not unmap the current VMA */
5178 if (iter_vma == vma)
5182 * Shared VMAs have their own reserves and do not affect
5183 * MAP_PRIVATE accounting but it is possible that a shared
5184 * VMA is using the same page so check and skip such VMAs.
5186 if (iter_vma->vm_flags & VM_MAYSHARE)
5190 * Unmap the page from other VMAs without their own reserves.
5191 * They get marked to be SIGKILLed if they fault in these
5192 * areas. This is because a future no-page fault on this VMA
5193 * could insert a zeroed page instead of the data existing
5194 * from the time of fork. This would look like data corruption
5196 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
5197 unmap_hugepage_range(iter_vma, address,
5198 address + huge_page_size(h), page, 0);
5200 i_mmap_unlock_write(mapping);
5204 * hugetlb_wp() should be called with page lock of the original hugepage held.
5205 * Called with hugetlb_fault_mutex_table held and pte_page locked so we
5206 * cannot race with other handlers or page migration.
5207 * Keep the pte_same checks anyway to make transition from the mutex easier.
5209 static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma,
5210 unsigned long address, pte_t *ptep, unsigned int flags,
5211 struct page *pagecache_page, spinlock_t *ptl)
5213 const bool unshare = flags & FAULT_FLAG_UNSHARE;
5215 struct hstate *h = hstate_vma(vma);
5216 struct page *old_page, *new_page;
5217 int outside_reserve = 0;
5219 unsigned long haddr = address & huge_page_mask(h);
5220 struct mmu_notifier_range range;
5222 VM_BUG_ON(unshare && (flags & FOLL_WRITE));
5223 VM_BUG_ON(!unshare && !(flags & FOLL_WRITE));
5225 pte = huge_ptep_get(ptep);
5226 old_page = pte_page(pte);
5228 delayacct_wpcopy_start();
5232 * If no-one else is actually using this page, we're the exclusive
5233 * owner and can reuse this page.
5235 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
5236 if (!PageAnonExclusive(old_page))
5237 page_move_anon_rmap(old_page, vma);
5238 if (likely(!unshare))
5239 set_huge_ptep_writable(vma, haddr, ptep);
5241 delayacct_wpcopy_end();
5244 VM_BUG_ON_PAGE(PageAnon(old_page) && PageAnonExclusive(old_page),
5248 * If the process that created a MAP_PRIVATE mapping is about to
5249 * perform a COW due to a shared page count, attempt to satisfy
5250 * the allocation without using the existing reserves. The pagecache
5251 * page is used to determine if the reserve at this address was
5252 * consumed or not. If reserves were used, a partial faulted mapping
5253 * at the time of fork() could consume its reserves on COW instead
5254 * of the full address range.
5256 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
5257 old_page != pagecache_page)
5258 outside_reserve = 1;
5263 * Drop page table lock as buddy allocator may be called. It will
5264 * be acquired again before returning to the caller, as expected.
5267 new_page = alloc_huge_page(vma, haddr, outside_reserve);
5269 if (IS_ERR(new_page)) {
5271 * If a process owning a MAP_PRIVATE mapping fails to COW,
5272 * it is due to references held by a child and an insufficient
5273 * huge page pool. To guarantee the original mappers
5274 * reliability, unmap the page from child processes. The child
5275 * may get SIGKILLed if it later faults.
5277 if (outside_reserve) {
5278 struct address_space *mapping = vma->vm_file->f_mapping;
5283 BUG_ON(huge_pte_none(pte));
5285 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
5286 * unmapping. unmapping needs to hold i_mmap_rwsem
5287 * in write mode. Dropping i_mmap_rwsem in read mode
5288 * here is OK as COW mappings do not interact with
5291 * Reacquire both after unmap operation.
5293 idx = vma_hugecache_offset(h, vma, haddr);
5294 hash = hugetlb_fault_mutex_hash(mapping, idx);
5295 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5296 i_mmap_unlock_read(mapping);
5298 unmap_ref_private(mm, vma, old_page, haddr);
5300 i_mmap_lock_read(mapping);
5301 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5303 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5305 pte_same(huge_ptep_get(ptep), pte)))
5306 goto retry_avoidcopy;
5308 * race occurs while re-acquiring page table
5309 * lock, and our job is done.
5311 delayacct_wpcopy_end();
5315 ret = vmf_error(PTR_ERR(new_page));
5316 goto out_release_old;
5320 * When the original hugepage is shared one, it does not have
5321 * anon_vma prepared.
5323 if (unlikely(anon_vma_prepare(vma))) {
5325 goto out_release_all;
5328 copy_user_huge_page(new_page, old_page, address, vma,
5329 pages_per_huge_page(h));
5330 __SetPageUptodate(new_page);
5332 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
5333 haddr + huge_page_size(h));
5334 mmu_notifier_invalidate_range_start(&range);
5337 * Retake the page table lock to check for racing updates
5338 * before the page tables are altered
5341 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5342 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
5343 ClearHPageRestoreReserve(new_page);
5345 /* Break COW or unshare */
5346 huge_ptep_clear_flush(vma, haddr, ptep);
5347 mmu_notifier_invalidate_range(mm, range.start, range.end);
5348 page_remove_rmap(old_page, vma, true);
5349 hugepage_add_new_anon_rmap(new_page, vma, haddr);
5350 set_huge_pte_at(mm, haddr, ptep,
5351 make_huge_pte(vma, new_page, !unshare));
5352 SetHPageMigratable(new_page);
5353 /* Make the old page be freed below */
5354 new_page = old_page;
5357 mmu_notifier_invalidate_range_end(&range);
5360 * No restore in case of successful pagetable update (Break COW or
5363 if (new_page != old_page)
5364 restore_reserve_on_error(h, vma, haddr, new_page);
5369 spin_lock(ptl); /* Caller expects lock to be held */
5371 delayacct_wpcopy_end();
5375 /* Return the pagecache page at a given address within a VMA */
5376 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
5377 struct vm_area_struct *vma, unsigned long address)
5379 struct address_space *mapping;
5382 mapping = vma->vm_file->f_mapping;
5383 idx = vma_hugecache_offset(h, vma, address);
5385 return find_lock_page(mapping, idx);
5389 * Return whether there is a pagecache page to back given address within VMA.
5390 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
5392 static bool hugetlbfs_pagecache_present(struct hstate *h,
5393 struct vm_area_struct *vma, unsigned long address)
5395 struct address_space *mapping;
5399 mapping = vma->vm_file->f_mapping;
5400 idx = vma_hugecache_offset(h, vma, address);
5402 page = find_get_page(mapping, idx);
5405 return page != NULL;
5408 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
5411 struct inode *inode = mapping->host;
5412 struct hstate *h = hstate_inode(inode);
5413 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
5417 ClearHPageRestoreReserve(page);
5420 * set page dirty so that it will not be removed from cache/file
5421 * by non-hugetlbfs specific code paths.
5423 set_page_dirty(page);
5425 spin_lock(&inode->i_lock);
5426 inode->i_blocks += blocks_per_huge_page(h);
5427 spin_unlock(&inode->i_lock);
5431 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
5432 struct address_space *mapping,
5435 unsigned long haddr,
5437 unsigned long reason)
5441 struct vm_fault vmf = {
5444 .real_address = addr,
5448 * Hard to debug if it ends up being
5449 * used by a callee that assumes
5450 * something about the other
5451 * uninitialized fields... same as in
5457 * hugetlb_fault_mutex and i_mmap_rwsem must be
5458 * dropped before handling userfault. Reacquire
5459 * after handling fault to make calling code simpler.
5461 hash = hugetlb_fault_mutex_hash(mapping, idx);
5462 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5463 i_mmap_unlock_read(mapping);
5464 ret = handle_userfault(&vmf, reason);
5465 i_mmap_lock_read(mapping);
5466 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5471 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
5472 struct vm_area_struct *vma,
5473 struct address_space *mapping, pgoff_t idx,
5474 unsigned long address, pte_t *ptep,
5475 pte_t old_pte, unsigned int flags)
5477 struct hstate *h = hstate_vma(vma);
5478 vm_fault_t ret = VM_FAULT_SIGBUS;
5484 unsigned long haddr = address & huge_page_mask(h);
5485 bool new_page, new_pagecache_page = false;
5488 * Currently, we are forced to kill the process in the event the
5489 * original mapper has unmapped pages from the child due to a failed
5490 * COW/unsharing. Warn that such a situation has occurred as it may not
5493 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
5494 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
5500 * We can not race with truncation due to holding i_mmap_rwsem.
5501 * i_size is modified when holding i_mmap_rwsem, so check here
5502 * once for faults beyond end of file.
5504 size = i_size_read(mapping->host) >> huge_page_shift(h);
5510 page = find_lock_page(mapping, idx);
5512 /* Check for page in userfault range */
5513 if (userfaultfd_missing(vma)) {
5514 ret = hugetlb_handle_userfault(vma, mapping, idx,
5515 flags, haddr, address,
5520 page = alloc_huge_page(vma, haddr, 0);
5523 * Returning error will result in faulting task being
5524 * sent SIGBUS. The hugetlb fault mutex prevents two
5525 * tasks from racing to fault in the same page which
5526 * could result in false unable to allocate errors.
5527 * Page migration does not take the fault mutex, but
5528 * does a clear then write of pte's under page table
5529 * lock. Page fault code could race with migration,
5530 * notice the clear pte and try to allocate a page
5531 * here. Before returning error, get ptl and make
5532 * sure there really is no pte entry.
5534 ptl = huge_pte_lock(h, mm, ptep);
5536 if (huge_pte_none(huge_ptep_get(ptep)))
5537 ret = vmf_error(PTR_ERR(page));
5541 clear_huge_page(page, address, pages_per_huge_page(h));
5542 __SetPageUptodate(page);
5545 if (vma->vm_flags & VM_MAYSHARE) {
5546 int err = huge_add_to_page_cache(page, mapping, idx);
5553 new_pagecache_page = true;
5556 if (unlikely(anon_vma_prepare(vma))) {
5558 goto backout_unlocked;
5564 * If memory error occurs between mmap() and fault, some process
5565 * don't have hwpoisoned swap entry for errored virtual address.
5566 * So we need to block hugepage fault by PG_hwpoison bit check.
5568 if (unlikely(PageHWPoison(page))) {
5569 ret = VM_FAULT_HWPOISON_LARGE |
5570 VM_FAULT_SET_HINDEX(hstate_index(h));
5571 goto backout_unlocked;
5574 /* Check for page in userfault range. */
5575 if (userfaultfd_minor(vma)) {
5578 ret = hugetlb_handle_userfault(vma, mapping, idx,
5579 flags, haddr, address,
5586 * If we are going to COW a private mapping later, we examine the
5587 * pending reservations for this page now. This will ensure that
5588 * any allocations necessary to record that reservation occur outside
5591 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5592 if (vma_needs_reservation(h, vma, haddr) < 0) {
5594 goto backout_unlocked;
5596 /* Just decrements count, does not deallocate */
5597 vma_end_reservation(h, vma, haddr);
5600 ptl = huge_pte_lock(h, mm, ptep);
5602 /* If pte changed from under us, retry */
5603 if (!pte_same(huge_ptep_get(ptep), old_pte))
5607 ClearHPageRestoreReserve(page);
5608 hugepage_add_new_anon_rmap(page, vma, haddr);
5610 page_dup_file_rmap(page, true);
5611 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
5612 && (vma->vm_flags & VM_SHARED)));
5614 * If this pte was previously wr-protected, keep it wr-protected even
5617 if (unlikely(pte_marker_uffd_wp(old_pte)))
5618 new_pte = huge_pte_wrprotect(huge_pte_mkuffd_wp(new_pte));
5619 set_huge_pte_at(mm, haddr, ptep, new_pte);
5621 hugetlb_count_add(pages_per_huge_page(h), mm);
5622 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5623 /* Optimization, do the COW without a second fault */
5624 ret = hugetlb_wp(mm, vma, address, ptep, flags, page, ptl);
5630 * Only set HPageMigratable in newly allocated pages. Existing pages
5631 * found in the pagecache may not have HPageMigratableset if they have
5632 * been isolated for migration.
5635 SetHPageMigratable(page);
5645 /* restore reserve for newly allocated pages not in page cache */
5646 if (new_page && !new_pagecache_page)
5647 restore_reserve_on_error(h, vma, haddr, page);
5653 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5655 unsigned long key[2];
5658 key[0] = (unsigned long) mapping;
5661 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
5663 return hash & (num_fault_mutexes - 1);
5667 * For uniprocessor systems we always use a single mutex, so just
5668 * return 0 and avoid the hashing overhead.
5670 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5676 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
5677 unsigned long address, unsigned int flags)
5684 struct page *page = NULL;
5685 struct page *pagecache_page = NULL;
5686 struct hstate *h = hstate_vma(vma);
5687 struct address_space *mapping;
5688 int need_wait_lock = 0;
5689 unsigned long haddr = address & huge_page_mask(h);
5691 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5694 * Since we hold no locks, ptep could be stale. That is
5695 * OK as we are only making decisions based on content and
5696 * not actually modifying content here.
5698 entry = huge_ptep_get(ptep);
5699 if (unlikely(is_hugetlb_entry_migration(entry))) {
5700 migration_entry_wait_huge(vma, ptep);
5702 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
5703 return VM_FAULT_HWPOISON_LARGE |
5704 VM_FAULT_SET_HINDEX(hstate_index(h));
5708 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
5709 * until finished with ptep. This serves two purposes:
5710 * 1) It prevents huge_pmd_unshare from being called elsewhere
5711 * and making the ptep no longer valid.
5712 * 2) It synchronizes us with i_size modifications during truncation.
5714 * ptep could have already be assigned via huge_pte_offset. That
5715 * is OK, as huge_pte_alloc will return the same value unless
5716 * something has changed.
5718 mapping = vma->vm_file->f_mapping;
5719 i_mmap_lock_read(mapping);
5720 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
5722 i_mmap_unlock_read(mapping);
5723 return VM_FAULT_OOM;
5727 * Serialize hugepage allocation and instantiation, so that we don't
5728 * get spurious allocation failures if two CPUs race to instantiate
5729 * the same page in the page cache.
5731 idx = vma_hugecache_offset(h, vma, haddr);
5732 hash = hugetlb_fault_mutex_hash(mapping, idx);
5733 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5735 entry = huge_ptep_get(ptep);
5736 /* PTE markers should be handled the same way as none pte */
5737 if (huge_pte_none_mostly(entry)) {
5738 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep,
5746 * entry could be a migration/hwpoison entry at this point, so this
5747 * check prevents the kernel from going below assuming that we have
5748 * an active hugepage in pagecache. This goto expects the 2nd page
5749 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
5750 * properly handle it.
5752 if (!pte_present(entry))
5756 * If we are going to COW/unshare the mapping later, we examine the
5757 * pending reservations for this page now. This will ensure that any
5758 * allocations necessary to record that reservation occur outside the
5759 * spinlock. For private mappings, we also lookup the pagecache
5760 * page now as it is used to determine if a reservation has been
5763 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
5764 !huge_pte_write(entry)) {
5765 if (vma_needs_reservation(h, vma, haddr) < 0) {
5769 /* Just decrements count, does not deallocate */
5770 vma_end_reservation(h, vma, haddr);
5772 if (!(vma->vm_flags & VM_MAYSHARE))
5773 pagecache_page = hugetlbfs_pagecache_page(h,
5777 ptl = huge_pte_lock(h, mm, ptep);
5779 /* Check for a racing update before calling hugetlb_wp() */
5780 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
5783 /* Handle userfault-wp first, before trying to lock more pages */
5784 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) &&
5785 (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
5786 struct vm_fault vmf = {
5789 .real_address = address,
5794 if (pagecache_page) {
5795 unlock_page(pagecache_page);
5796 put_page(pagecache_page);
5798 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5799 i_mmap_unlock_read(mapping);
5800 return handle_userfault(&vmf, VM_UFFD_WP);
5804 * hugetlb_wp() requires page locks of pte_page(entry) and
5805 * pagecache_page, so here we need take the former one
5806 * when page != pagecache_page or !pagecache_page.
5808 page = pte_page(entry);
5809 if (page != pagecache_page)
5810 if (!trylock_page(page)) {
5817 if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
5818 if (!huge_pte_write(entry)) {
5819 ret = hugetlb_wp(mm, vma, address, ptep, flags,
5820 pagecache_page, ptl);
5822 } else if (likely(flags & FAULT_FLAG_WRITE)) {
5823 entry = huge_pte_mkdirty(entry);
5826 entry = pte_mkyoung(entry);
5827 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
5828 flags & FAULT_FLAG_WRITE))
5829 update_mmu_cache(vma, haddr, ptep);
5831 if (page != pagecache_page)
5837 if (pagecache_page) {
5838 unlock_page(pagecache_page);
5839 put_page(pagecache_page);
5842 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5843 i_mmap_unlock_read(mapping);
5845 * Generally it's safe to hold refcount during waiting page lock. But
5846 * here we just wait to defer the next page fault to avoid busy loop and
5847 * the page is not used after unlocked before returning from the current
5848 * page fault. So we are safe from accessing freed page, even if we wait
5849 * here without taking refcount.
5852 wait_on_page_locked(page);
5856 #ifdef CONFIG_USERFAULTFD
5858 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
5859 * modifications for huge pages.
5861 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
5863 struct vm_area_struct *dst_vma,
5864 unsigned long dst_addr,
5865 unsigned long src_addr,
5866 enum mcopy_atomic_mode mode,
5867 struct page **pagep,
5870 bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE);
5871 struct hstate *h = hstate_vma(dst_vma);
5872 struct address_space *mapping = dst_vma->vm_file->f_mapping;
5873 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
5875 int vm_shared = dst_vma->vm_flags & VM_SHARED;
5881 bool page_in_pagecache = false;
5885 page = find_lock_page(mapping, idx);
5888 page_in_pagecache = true;
5889 } else if (!*pagep) {
5890 /* If a page already exists, then it's UFFDIO_COPY for
5891 * a non-missing case. Return -EEXIST.
5894 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5899 page = alloc_huge_page(dst_vma, dst_addr, 0);
5905 ret = copy_huge_page_from_user(page,
5906 (const void __user *) src_addr,
5907 pages_per_huge_page(h), false);
5909 /* fallback to copy_from_user outside mmap_lock */
5910 if (unlikely(ret)) {
5912 /* Free the allocated page which may have
5913 * consumed a reservation.
5915 restore_reserve_on_error(h, dst_vma, dst_addr, page);
5918 /* Allocate a temporary page to hold the copied
5921 page = alloc_huge_page_vma(h, dst_vma, dst_addr);
5927 /* Set the outparam pagep and return to the caller to
5928 * copy the contents outside the lock. Don't free the
5935 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5942 page = alloc_huge_page(dst_vma, dst_addr, 0);
5948 copy_user_huge_page(page, *pagep, dst_addr, dst_vma,
5949 pages_per_huge_page(h));
5955 * The memory barrier inside __SetPageUptodate makes sure that
5956 * preceding stores to the page contents become visible before
5957 * the set_pte_at() write.
5959 __SetPageUptodate(page);
5961 /* Add shared, newly allocated pages to the page cache. */
5962 if (vm_shared && !is_continue) {
5963 size = i_size_read(mapping->host) >> huge_page_shift(h);
5966 goto out_release_nounlock;
5969 * Serialization between remove_inode_hugepages() and
5970 * huge_add_to_page_cache() below happens through the
5971 * hugetlb_fault_mutex_table that here must be hold by
5974 ret = huge_add_to_page_cache(page, mapping, idx);
5976 goto out_release_nounlock;
5977 page_in_pagecache = true;
5980 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
5984 * Recheck the i_size after holding PT lock to make sure not
5985 * to leave any page mapped (as page_mapped()) beyond the end
5986 * of the i_size (remove_inode_hugepages() is strict about
5987 * enforcing that). If we bail out here, we'll also leave a
5988 * page in the radix tree in the vm_shared case beyond the end
5989 * of the i_size, but remove_inode_hugepages() will take care
5990 * of it as soon as we drop the hugetlb_fault_mutex_table.
5992 size = i_size_read(mapping->host) >> huge_page_shift(h);
5995 goto out_release_unlock;
5999 * We allow to overwrite a pte marker: consider when both MISSING|WP
6000 * registered, we firstly wr-protect a none pte which has no page cache
6001 * page backing it, then access the page.
6003 if (!huge_pte_none_mostly(huge_ptep_get(dst_pte)))
6004 goto out_release_unlock;
6007 page_dup_file_rmap(page, true);
6009 ClearHPageRestoreReserve(page);
6010 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
6014 * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY
6015 * with wp flag set, don't set pte write bit.
6017 if (wp_copy || (is_continue && !vm_shared))
6020 writable = dst_vma->vm_flags & VM_WRITE;
6022 _dst_pte = make_huge_pte(dst_vma, page, writable);
6024 * Always mark UFFDIO_COPY page dirty; note that this may not be
6025 * extremely important for hugetlbfs for now since swapping is not
6026 * supported, but we should still be clear in that this page cannot be
6027 * thrown away at will, even if write bit not set.
6029 _dst_pte = huge_pte_mkdirty(_dst_pte);
6030 _dst_pte = pte_mkyoung(_dst_pte);
6033 _dst_pte = huge_pte_mkuffd_wp(_dst_pte);
6035 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
6037 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
6039 /* No need to invalidate - it was non-present before */
6040 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6044 SetHPageMigratable(page);
6045 if (vm_shared || is_continue)
6052 if (vm_shared || is_continue)
6054 out_release_nounlock:
6055 if (!page_in_pagecache)
6056 restore_reserve_on_error(h, dst_vma, dst_addr, page);
6060 #endif /* CONFIG_USERFAULTFD */
6062 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
6063 int refs, struct page **pages,
6064 struct vm_area_struct **vmas)
6068 for (nr = 0; nr < refs; nr++) {
6070 pages[nr] = mem_map_offset(page, nr);
6076 static inline bool __follow_hugetlb_must_fault(unsigned int flags, pte_t *pte,
6079 pte_t pteval = huge_ptep_get(pte);
6082 if (is_swap_pte(pteval))
6084 if (huge_pte_write(pteval))
6086 if (flags & FOLL_WRITE)
6088 if (gup_must_unshare(flags, pte_page(pteval))) {
6095 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
6096 struct page **pages, struct vm_area_struct **vmas,
6097 unsigned long *position, unsigned long *nr_pages,
6098 long i, unsigned int flags, int *locked)
6100 unsigned long pfn_offset;
6101 unsigned long vaddr = *position;
6102 unsigned long remainder = *nr_pages;
6103 struct hstate *h = hstate_vma(vma);
6104 int err = -EFAULT, refs;
6106 while (vaddr < vma->vm_end && remainder) {
6108 spinlock_t *ptl = NULL;
6109 bool unshare = false;
6114 * If we have a pending SIGKILL, don't keep faulting pages and
6115 * potentially allocating memory.
6117 if (fatal_signal_pending(current)) {
6123 * Some archs (sparc64, sh*) have multiple pte_ts to
6124 * each hugepage. We have to make sure we get the
6125 * first, for the page indexing below to work.
6127 * Note that page table lock is not held when pte is null.
6129 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
6132 ptl = huge_pte_lock(h, mm, pte);
6133 absent = !pte || huge_pte_none(huge_ptep_get(pte));
6136 * When coredumping, it suits get_dump_page if we just return
6137 * an error where there's an empty slot with no huge pagecache
6138 * to back it. This way, we avoid allocating a hugepage, and
6139 * the sparse dumpfile avoids allocating disk blocks, but its
6140 * huge holes still show up with zeroes where they need to be.
6142 if (absent && (flags & FOLL_DUMP) &&
6143 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
6151 * We need call hugetlb_fault for both hugepages under migration
6152 * (in which case hugetlb_fault waits for the migration,) and
6153 * hwpoisoned hugepages (in which case we need to prevent the
6154 * caller from accessing to them.) In order to do this, we use
6155 * here is_swap_pte instead of is_hugetlb_entry_migration and
6156 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
6157 * both cases, and because we can't follow correct pages
6158 * directly from any kind of swap entries.
6161 __follow_hugetlb_must_fault(flags, pte, &unshare)) {
6163 unsigned int fault_flags = 0;
6167 if (flags & FOLL_WRITE)
6168 fault_flags |= FAULT_FLAG_WRITE;
6170 fault_flags |= FAULT_FLAG_UNSHARE;
6172 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
6173 FAULT_FLAG_KILLABLE;
6174 if (flags & FOLL_NOWAIT)
6175 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
6176 FAULT_FLAG_RETRY_NOWAIT;
6177 if (flags & FOLL_TRIED) {
6179 * Note: FAULT_FLAG_ALLOW_RETRY and
6180 * FAULT_FLAG_TRIED can co-exist
6182 fault_flags |= FAULT_FLAG_TRIED;
6184 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
6185 if (ret & VM_FAULT_ERROR) {
6186 err = vm_fault_to_errno(ret, flags);
6190 if (ret & VM_FAULT_RETRY) {
6192 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
6196 * VM_FAULT_RETRY must not return an
6197 * error, it will return zero
6200 * No need to update "position" as the
6201 * caller will not check it after
6202 * *nr_pages is set to 0.
6209 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
6210 page = pte_page(huge_ptep_get(pte));
6212 VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) &&
6213 !PageAnonExclusive(page), page);
6216 * If subpage information not requested, update counters
6217 * and skip the same_page loop below.
6219 if (!pages && !vmas && !pfn_offset &&
6220 (vaddr + huge_page_size(h) < vma->vm_end) &&
6221 (remainder >= pages_per_huge_page(h))) {
6222 vaddr += huge_page_size(h);
6223 remainder -= pages_per_huge_page(h);
6224 i += pages_per_huge_page(h);
6229 /* vaddr may not be aligned to PAGE_SIZE */
6230 refs = min3(pages_per_huge_page(h) - pfn_offset, remainder,
6231 (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT);
6234 record_subpages_vmas(mem_map_offset(page, pfn_offset),
6236 likely(pages) ? pages + i : NULL,
6237 vmas ? vmas + i : NULL);
6241 * try_grab_folio() should always succeed here,
6242 * because: a) we hold the ptl lock, and b) we've just
6243 * checked that the huge page is present in the page
6244 * tables. If the huge page is present, then the tail
6245 * pages must also be present. The ptl prevents the
6246 * head page and tail pages from being rearranged in
6247 * any way. So this page must be available at this
6248 * point, unless the page refcount overflowed:
6250 if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs,
6259 vaddr += (refs << PAGE_SHIFT);
6265 *nr_pages = remainder;
6267 * setting position is actually required only if remainder is
6268 * not zero but it's faster not to add a "if (remainder)"
6276 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
6277 unsigned long address, unsigned long end,
6278 pgprot_t newprot, unsigned long cp_flags)
6280 struct mm_struct *mm = vma->vm_mm;
6281 unsigned long start = address;
6284 struct hstate *h = hstate_vma(vma);
6285 unsigned long pages = 0, psize = huge_page_size(h);
6286 bool shared_pmd = false;
6287 struct mmu_notifier_range range;
6288 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
6289 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
6292 * In the case of shared PMDs, the area to flush could be beyond
6293 * start/end. Set range.start/range.end to cover the maximum possible
6294 * range if PMD sharing is possible.
6296 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
6297 0, vma, mm, start, end);
6298 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
6300 BUG_ON(address >= end);
6301 flush_cache_range(vma, range.start, range.end);
6303 mmu_notifier_invalidate_range_start(&range);
6304 i_mmap_lock_write(vma->vm_file->f_mapping);
6305 for (; address < end; address += psize) {
6307 ptep = huge_pte_offset(mm, address, psize);
6310 ptl = huge_pte_lock(h, mm, ptep);
6311 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
6313 * When uffd-wp is enabled on the vma, unshare
6314 * shouldn't happen at all. Warn about it if it
6315 * happened due to some reason.
6317 WARN_ON_ONCE(uffd_wp || uffd_wp_resolve);
6323 pte = huge_ptep_get(ptep);
6324 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
6328 if (unlikely(is_hugetlb_entry_migration(pte))) {
6329 swp_entry_t entry = pte_to_swp_entry(pte);
6330 struct page *page = pfn_swap_entry_to_page(entry);
6332 if (!is_readable_migration_entry(entry)) {
6336 entry = make_readable_exclusive_migration_entry(
6339 entry = make_readable_migration_entry(
6341 newpte = swp_entry_to_pte(entry);
6343 newpte = pte_swp_mkuffd_wp(newpte);
6344 else if (uffd_wp_resolve)
6345 newpte = pte_swp_clear_uffd_wp(newpte);
6346 set_huge_pte_at(mm, address, ptep, newpte);
6352 if (unlikely(pte_marker_uffd_wp(pte))) {
6354 * This is changing a non-present pte into a none pte,
6355 * no need for huge_ptep_modify_prot_start/commit().
6357 if (uffd_wp_resolve)
6358 huge_pte_clear(mm, address, ptep, psize);
6360 if (!huge_pte_none(pte)) {
6362 unsigned int shift = huge_page_shift(hstate_vma(vma));
6364 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
6365 pte = huge_pte_modify(old_pte, newprot);
6366 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
6368 pte = huge_pte_mkuffd_wp(huge_pte_wrprotect(pte));
6369 else if (uffd_wp_resolve)
6370 pte = huge_pte_clear_uffd_wp(pte);
6371 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
6375 if (unlikely(uffd_wp))
6376 /* Safe to modify directly (none->non-present). */
6377 set_huge_pte_at(mm, address, ptep,
6378 make_pte_marker(PTE_MARKER_UFFD_WP));
6383 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
6384 * may have cleared our pud entry and done put_page on the page table:
6385 * once we release i_mmap_rwsem, another task can do the final put_page
6386 * and that page table be reused and filled with junk. If we actually
6387 * did unshare a page of pmds, flush the range corresponding to the pud.
6390 flush_hugetlb_tlb_range(vma, range.start, range.end);
6392 flush_hugetlb_tlb_range(vma, start, end);
6394 * No need to call mmu_notifier_invalidate_range() we are downgrading
6395 * page table protection not changing it to point to a new page.
6397 * See Documentation/mm/mmu_notifier.rst
6399 i_mmap_unlock_write(vma->vm_file->f_mapping);
6400 mmu_notifier_invalidate_range_end(&range);
6402 return pages << h->order;
6405 /* Return true if reservation was successful, false otherwise. */
6406 bool hugetlb_reserve_pages(struct inode *inode,
6408 struct vm_area_struct *vma,
6409 vm_flags_t vm_flags)
6412 struct hstate *h = hstate_inode(inode);
6413 struct hugepage_subpool *spool = subpool_inode(inode);
6414 struct resv_map *resv_map;
6415 struct hugetlb_cgroup *h_cg = NULL;
6416 long gbl_reserve, regions_needed = 0;
6418 /* This should never happen */
6420 VM_WARN(1, "%s called with a negative range\n", __func__);
6425 * Only apply hugepage reservation if asked. At fault time, an
6426 * attempt will be made for VM_NORESERVE to allocate a page
6427 * without using reserves
6429 if (vm_flags & VM_NORESERVE)
6433 * Shared mappings base their reservation on the number of pages that
6434 * are already allocated on behalf of the file. Private mappings need
6435 * to reserve the full area even if read-only as mprotect() may be
6436 * called to make the mapping read-write. Assume !vma is a shm mapping
6438 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6440 * resv_map can not be NULL as hugetlb_reserve_pages is only
6441 * called for inodes for which resv_maps were created (see
6442 * hugetlbfs_get_inode).
6444 resv_map = inode_resv_map(inode);
6446 chg = region_chg(resv_map, from, to, ®ions_needed);
6449 /* Private mapping. */
6450 resv_map = resv_map_alloc();
6456 set_vma_resv_map(vma, resv_map);
6457 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
6463 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
6464 chg * pages_per_huge_page(h), &h_cg) < 0)
6467 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
6468 /* For private mappings, the hugetlb_cgroup uncharge info hangs
6471 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
6475 * There must be enough pages in the subpool for the mapping. If
6476 * the subpool has a minimum size, there may be some global
6477 * reservations already in place (gbl_reserve).
6479 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
6480 if (gbl_reserve < 0)
6481 goto out_uncharge_cgroup;
6484 * Check enough hugepages are available for the reservation.
6485 * Hand the pages back to the subpool if there are not
6487 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
6491 * Account for the reservations made. Shared mappings record regions
6492 * that have reservations as they are shared by multiple VMAs.
6493 * When the last VMA disappears, the region map says how much
6494 * the reservation was and the page cache tells how much of
6495 * the reservation was consumed. Private mappings are per-VMA and
6496 * only the consumed reservations are tracked. When the VMA
6497 * disappears, the original reservation is the VMA size and the
6498 * consumed reservations are stored in the map. Hence, nothing
6499 * else has to be done for private mappings here
6501 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6502 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
6504 if (unlikely(add < 0)) {
6505 hugetlb_acct_memory(h, -gbl_reserve);
6507 } else if (unlikely(chg > add)) {
6509 * pages in this range were added to the reserve
6510 * map between region_chg and region_add. This
6511 * indicates a race with alloc_huge_page. Adjust
6512 * the subpool and reserve counts modified above
6513 * based on the difference.
6518 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
6519 * reference to h_cg->css. See comment below for detail.
6521 hugetlb_cgroup_uncharge_cgroup_rsvd(
6523 (chg - add) * pages_per_huge_page(h), h_cg);
6525 rsv_adjust = hugepage_subpool_put_pages(spool,
6527 hugetlb_acct_memory(h, -rsv_adjust);
6530 * The file_regions will hold their own reference to
6531 * h_cg->css. So we should release the reference held
6532 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
6535 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
6541 /* put back original number of pages, chg */
6542 (void)hugepage_subpool_put_pages(spool, chg);
6543 out_uncharge_cgroup:
6544 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
6545 chg * pages_per_huge_page(h), h_cg);
6547 if (!vma || vma->vm_flags & VM_MAYSHARE)
6548 /* Only call region_abort if the region_chg succeeded but the
6549 * region_add failed or didn't run.
6551 if (chg >= 0 && add < 0)
6552 region_abort(resv_map, from, to, regions_needed);
6553 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
6554 kref_put(&resv_map->refs, resv_map_release);
6558 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
6561 struct hstate *h = hstate_inode(inode);
6562 struct resv_map *resv_map = inode_resv_map(inode);
6564 struct hugepage_subpool *spool = subpool_inode(inode);
6568 * Since this routine can be called in the evict inode path for all
6569 * hugetlbfs inodes, resv_map could be NULL.
6572 chg = region_del(resv_map, start, end);
6574 * region_del() can fail in the rare case where a region
6575 * must be split and another region descriptor can not be
6576 * allocated. If end == LONG_MAX, it will not fail.
6582 spin_lock(&inode->i_lock);
6583 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
6584 spin_unlock(&inode->i_lock);
6587 * If the subpool has a minimum size, the number of global
6588 * reservations to be released may be adjusted.
6590 * Note that !resv_map implies freed == 0. So (chg - freed)
6591 * won't go negative.
6593 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
6594 hugetlb_acct_memory(h, -gbl_reserve);
6599 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
6600 static unsigned long page_table_shareable(struct vm_area_struct *svma,
6601 struct vm_area_struct *vma,
6602 unsigned long addr, pgoff_t idx)
6604 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
6606 unsigned long sbase = saddr & PUD_MASK;
6607 unsigned long s_end = sbase + PUD_SIZE;
6609 /* Allow segments to share if only one is marked locked */
6610 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
6611 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
6614 * match the virtual addresses, permission and the alignment of the
6617 if (pmd_index(addr) != pmd_index(saddr) ||
6618 vm_flags != svm_flags ||
6619 !range_in_vma(svma, sbase, s_end))
6625 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
6627 unsigned long base = addr & PUD_MASK;
6628 unsigned long end = base + PUD_SIZE;
6631 * check on proper vm_flags and page table alignment
6633 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
6638 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6640 #ifdef CONFIG_USERFAULTFD
6641 if (uffd_disable_huge_pmd_share(vma))
6644 return vma_shareable(vma, addr);
6648 * Determine if start,end range within vma could be mapped by shared pmd.
6649 * If yes, adjust start and end to cover range associated with possible
6650 * shared pmd mappings.
6652 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
6653 unsigned long *start, unsigned long *end)
6655 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
6656 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6659 * vma needs to span at least one aligned PUD size, and the range
6660 * must be at least partially within in.
6662 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
6663 (*end <= v_start) || (*start >= v_end))
6666 /* Extend the range to be PUD aligned for a worst case scenario */
6667 if (*start > v_start)
6668 *start = ALIGN_DOWN(*start, PUD_SIZE);
6671 *end = ALIGN(*end, PUD_SIZE);
6675 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
6676 * and returns the corresponding pte. While this is not necessary for the
6677 * !shared pmd case because we can allocate the pmd later as well, it makes the
6678 * code much cleaner.
6680 * This routine must be called with i_mmap_rwsem held in at least read mode if
6681 * sharing is possible. For hugetlbfs, this prevents removal of any page
6682 * table entries associated with the address space. This is important as we
6683 * are setting up sharing based on existing page table entries (mappings).
6685 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
6686 unsigned long addr, pud_t *pud)
6688 struct address_space *mapping = vma->vm_file->f_mapping;
6689 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
6691 struct vm_area_struct *svma;
6692 unsigned long saddr;
6697 i_mmap_assert_locked(mapping);
6698 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
6702 saddr = page_table_shareable(svma, vma, addr, idx);
6704 spte = huge_pte_offset(svma->vm_mm, saddr,
6705 vma_mmu_pagesize(svma));
6707 get_page(virt_to_page(spte));
6716 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
6717 if (pud_none(*pud)) {
6718 pud_populate(mm, pud,
6719 (pmd_t *)((unsigned long)spte & PAGE_MASK));
6722 put_page(virt_to_page(spte));
6726 pte = (pte_t *)pmd_alloc(mm, pud, addr);
6731 * unmap huge page backed by shared pte.
6733 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
6734 * indicated by page_count > 1, unmap is achieved by clearing pud and
6735 * decrementing the ref count. If count == 1, the pte page is not shared.
6737 * Called with page table lock held and i_mmap_rwsem held in write mode.
6739 * returns: 1 successfully unmapped a shared pte page
6740 * 0 the underlying pte page is not shared, or it is the last user
6742 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
6743 unsigned long *addr, pte_t *ptep)
6745 pgd_t *pgd = pgd_offset(mm, *addr);
6746 p4d_t *p4d = p4d_offset(pgd, *addr);
6747 pud_t *pud = pud_offset(p4d, *addr);
6749 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
6750 BUG_ON(page_count(virt_to_page(ptep)) == 0);
6751 if (page_count(virt_to_page(ptep)) == 1)
6755 put_page(virt_to_page(ptep));
6758 * This update of passed address optimizes loops sequentially
6759 * processing addresses in increments of huge page size (PMD_SIZE
6760 * in this case). By clearing the pud, a PUD_SIZE area is unmapped.
6761 * Update address to the 'last page' in the cleared area so that
6762 * calling loop can move to first page past this area.
6764 *addr |= PUD_SIZE - PMD_SIZE;
6768 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
6769 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
6770 unsigned long addr, pud_t *pud)
6775 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
6776 unsigned long *addr, pte_t *ptep)
6781 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
6782 unsigned long *start, unsigned long *end)
6786 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6790 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
6792 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
6793 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
6794 unsigned long addr, unsigned long sz)
6801 pgd = pgd_offset(mm, addr);
6802 p4d = p4d_alloc(mm, pgd, addr);
6805 pud = pud_alloc(mm, p4d, addr);
6807 if (sz == PUD_SIZE) {
6810 BUG_ON(sz != PMD_SIZE);
6811 if (want_pmd_share(vma, addr) && pud_none(*pud))
6812 pte = huge_pmd_share(mm, vma, addr, pud);
6814 pte = (pte_t *)pmd_alloc(mm, pud, addr);
6817 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
6823 * huge_pte_offset() - Walk the page table to resolve the hugepage
6824 * entry at address @addr
6826 * Return: Pointer to page table entry (PUD or PMD) for
6827 * address @addr, or NULL if a !p*d_present() entry is encountered and the
6828 * size @sz doesn't match the hugepage size at this level of the page
6831 pte_t *huge_pte_offset(struct mm_struct *mm,
6832 unsigned long addr, unsigned long sz)
6839 pgd = pgd_offset(mm, addr);
6840 if (!pgd_present(*pgd))
6842 p4d = p4d_offset(pgd, addr);
6843 if (!p4d_present(*p4d))
6846 pud = pud_offset(p4d, addr);
6848 /* must be pud huge, non-present or none */
6849 return (pte_t *)pud;
6850 if (!pud_present(*pud))
6852 /* must have a valid entry and size to go further */
6854 pmd = pmd_offset(pud, addr);
6855 /* must be pmd huge, non-present or none */
6856 return (pte_t *)pmd;
6859 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
6862 * These functions are overwritable if your architecture needs its own
6865 struct page * __weak
6866 follow_huge_addr(struct mm_struct *mm, unsigned long address,
6869 return ERR_PTR(-EINVAL);
6872 struct page * __weak
6873 follow_huge_pd(struct vm_area_struct *vma,
6874 unsigned long address, hugepd_t hpd, int flags, int pdshift)
6876 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
6880 struct page * __weak
6881 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
6882 pmd_t *pmd, int flags)
6884 struct page *page = NULL;
6889 * FOLL_PIN is not supported for follow_page(). Ordinary GUP goes via
6890 * follow_hugetlb_page().
6892 if (WARN_ON_ONCE(flags & FOLL_PIN))
6896 ptl = pmd_lockptr(mm, pmd);
6899 * make sure that the address range covered by this pmd is not
6900 * unmapped from other threads.
6902 if (!pmd_huge(*pmd))
6904 pte = huge_ptep_get((pte_t *)pmd);
6905 if (pte_present(pte)) {
6906 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
6908 * try_grab_page() should always succeed here, because: a) we
6909 * hold the pmd (ptl) lock, and b) we've just checked that the
6910 * huge pmd (head) page is present in the page tables. The ptl
6911 * prevents the head page and tail pages from being rearranged
6912 * in any way. So this page must be available at this point,
6913 * unless the page refcount overflowed:
6915 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
6920 if (is_hugetlb_entry_migration(pte)) {
6922 __migration_entry_wait_huge((pte_t *)pmd, ptl);
6926 * hwpoisoned entry is treated as no_page_table in
6927 * follow_page_mask().
6935 struct page * __weak
6936 follow_huge_pud(struct mm_struct *mm, unsigned long address,
6937 pud_t *pud, int flags)
6939 if (flags & (FOLL_GET | FOLL_PIN))
6942 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
6945 struct page * __weak
6946 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
6948 if (flags & (FOLL_GET | FOLL_PIN))
6951 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
6954 int isolate_hugetlb(struct page *page, struct list_head *list)
6958 spin_lock_irq(&hugetlb_lock);
6959 if (!PageHeadHuge(page) ||
6960 !HPageMigratable(page) ||
6961 !get_page_unless_zero(page)) {
6965 ClearHPageMigratable(page);
6966 list_move_tail(&page->lru, list);
6968 spin_unlock_irq(&hugetlb_lock);
6972 int get_hwpoison_huge_page(struct page *page, bool *hugetlb)
6977 spin_lock_irq(&hugetlb_lock);
6978 if (PageHeadHuge(page)) {
6980 if (HPageFreed(page))
6982 else if (HPageMigratable(page))
6983 ret = get_page_unless_zero(page);
6987 spin_unlock_irq(&hugetlb_lock);
6991 int get_huge_page_for_hwpoison(unsigned long pfn, int flags)
6995 spin_lock_irq(&hugetlb_lock);
6996 ret = __get_huge_page_for_hwpoison(pfn, flags);
6997 spin_unlock_irq(&hugetlb_lock);
7001 void putback_active_hugepage(struct page *page)
7003 spin_lock_irq(&hugetlb_lock);
7004 SetHPageMigratable(page);
7005 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
7006 spin_unlock_irq(&hugetlb_lock);
7010 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
7012 struct hstate *h = page_hstate(oldpage);
7014 hugetlb_cgroup_migrate(oldpage, newpage);
7015 set_page_owner_migrate_reason(newpage, reason);
7018 * transfer temporary state of the new huge page. This is
7019 * reverse to other transitions because the newpage is going to
7020 * be final while the old one will be freed so it takes over
7021 * the temporary status.
7023 * Also note that we have to transfer the per-node surplus state
7024 * here as well otherwise the global surplus count will not match
7027 if (HPageTemporary(newpage)) {
7028 int old_nid = page_to_nid(oldpage);
7029 int new_nid = page_to_nid(newpage);
7031 SetHPageTemporary(oldpage);
7032 ClearHPageTemporary(newpage);
7035 * There is no need to transfer the per-node surplus state
7036 * when we do not cross the node.
7038 if (new_nid == old_nid)
7040 spin_lock_irq(&hugetlb_lock);
7041 if (h->surplus_huge_pages_node[old_nid]) {
7042 h->surplus_huge_pages_node[old_nid]--;
7043 h->surplus_huge_pages_node[new_nid]++;
7045 spin_unlock_irq(&hugetlb_lock);
7050 * This function will unconditionally remove all the shared pmd pgtable entries
7051 * within the specific vma for a hugetlbfs memory range.
7053 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
7055 struct hstate *h = hstate_vma(vma);
7056 unsigned long sz = huge_page_size(h);
7057 struct mm_struct *mm = vma->vm_mm;
7058 struct mmu_notifier_range range;
7059 unsigned long address, start, end;
7063 if (!(vma->vm_flags & VM_MAYSHARE))
7066 start = ALIGN(vma->vm_start, PUD_SIZE);
7067 end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
7072 flush_cache_range(vma, start, end);
7074 * No need to call adjust_range_if_pmd_sharing_possible(), because
7075 * we have already done the PUD_SIZE alignment.
7077 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
7079 mmu_notifier_invalidate_range_start(&range);
7080 i_mmap_lock_write(vma->vm_file->f_mapping);
7081 for (address = start; address < end; address += PUD_SIZE) {
7082 unsigned long tmp = address;
7084 ptep = huge_pte_offset(mm, address, sz);
7087 ptl = huge_pte_lock(h, mm, ptep);
7088 /* We don't want 'address' to be changed */
7089 huge_pmd_unshare(mm, vma, &tmp, ptep);
7092 flush_hugetlb_tlb_range(vma, start, end);
7093 i_mmap_unlock_write(vma->vm_file->f_mapping);
7095 * No need to call mmu_notifier_invalidate_range(), see
7096 * Documentation/mm/mmu_notifier.rst.
7098 mmu_notifier_invalidate_range_end(&range);
7102 static bool cma_reserve_called __initdata;
7104 static int __init cmdline_parse_hugetlb_cma(char *p)
7111 if (sscanf(s, "%lu%n", &tmp, &count) != 1)
7114 if (s[count] == ':') {
7115 if (tmp >= MAX_NUMNODES)
7117 nid = array_index_nospec(tmp, MAX_NUMNODES);
7120 tmp = memparse(s, &s);
7121 hugetlb_cma_size_in_node[nid] = tmp;
7122 hugetlb_cma_size += tmp;
7125 * Skip the separator if have one, otherwise
7126 * break the parsing.
7133 hugetlb_cma_size = memparse(p, &p);
7141 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
7143 void __init hugetlb_cma_reserve(int order)
7145 unsigned long size, reserved, per_node;
7146 bool node_specific_cma_alloc = false;
7149 cma_reserve_called = true;
7151 if (!hugetlb_cma_size)
7154 for (nid = 0; nid < MAX_NUMNODES; nid++) {
7155 if (hugetlb_cma_size_in_node[nid] == 0)
7158 if (!node_online(nid)) {
7159 pr_warn("hugetlb_cma: invalid node %d specified\n", nid);
7160 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7161 hugetlb_cma_size_in_node[nid] = 0;
7165 if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) {
7166 pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n",
7167 nid, (PAGE_SIZE << order) / SZ_1M);
7168 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7169 hugetlb_cma_size_in_node[nid] = 0;
7171 node_specific_cma_alloc = true;
7175 /* Validate the CMA size again in case some invalid nodes specified. */
7176 if (!hugetlb_cma_size)
7179 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
7180 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
7181 (PAGE_SIZE << order) / SZ_1M);
7182 hugetlb_cma_size = 0;
7186 if (!node_specific_cma_alloc) {
7188 * If 3 GB area is requested on a machine with 4 numa nodes,
7189 * let's allocate 1 GB on first three nodes and ignore the last one.
7191 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
7192 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
7193 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
7197 for_each_online_node(nid) {
7199 char name[CMA_MAX_NAME];
7201 if (node_specific_cma_alloc) {
7202 if (hugetlb_cma_size_in_node[nid] == 0)
7205 size = hugetlb_cma_size_in_node[nid];
7207 size = min(per_node, hugetlb_cma_size - reserved);
7210 size = round_up(size, PAGE_SIZE << order);
7212 snprintf(name, sizeof(name), "hugetlb%d", nid);
7214 * Note that 'order per bit' is based on smallest size that
7215 * may be returned to CMA allocator in the case of
7216 * huge page demotion.
7218 res = cma_declare_contiguous_nid(0, size, 0,
7219 PAGE_SIZE << HUGETLB_PAGE_ORDER,
7221 &hugetlb_cma[nid], nid);
7223 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
7229 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
7232 if (reserved >= hugetlb_cma_size)
7238 * hugetlb_cma_size is used to determine if allocations from
7239 * cma are possible. Set to zero if no cma regions are set up.
7241 hugetlb_cma_size = 0;
7244 void __init hugetlb_cma_check(void)
7246 if (!hugetlb_cma_size || cma_reserve_called)
7249 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
7252 #endif /* CONFIG_CMA */