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>
37 #include <asm/pgalloc.h>
41 #include <linux/hugetlb.h>
42 #include <linux/hugetlb_cgroup.h>
43 #include <linux/node.h>
44 #include <linux/page_owner.h>
46 #include "hugetlb_vmemmap.h"
48 int hugetlb_max_hstate __read_mostly;
49 unsigned int default_hstate_idx;
50 struct hstate hstates[HUGE_MAX_HSTATE];
53 static struct cma *hugetlb_cma[MAX_NUMNODES];
54 static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata;
55 static bool hugetlb_cma_page(struct page *page, unsigned int order)
57 return cma_pages_valid(hugetlb_cma[page_to_nid(page)], page,
61 static bool hugetlb_cma_page(struct page *page, unsigned int order)
66 static unsigned long hugetlb_cma_size __initdata;
69 * Minimum page order among possible hugepage sizes, set to a proper value
72 static unsigned int minimum_order __read_mostly = UINT_MAX;
74 __initdata LIST_HEAD(huge_boot_pages);
76 /* for command line parsing */
77 static struct hstate * __initdata parsed_hstate;
78 static unsigned long __initdata default_hstate_max_huge_pages;
79 static bool __initdata parsed_valid_hugepagesz = true;
80 static bool __initdata parsed_default_hugepagesz;
81 static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata;
84 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
85 * free_huge_pages, and surplus_huge_pages.
87 DEFINE_SPINLOCK(hugetlb_lock);
90 * Serializes faults on the same logical page. This is used to
91 * prevent spurious OOMs when the hugepage pool is fully utilized.
93 static int num_fault_mutexes;
94 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
96 /* Forward declaration */
97 static int hugetlb_acct_memory(struct hstate *h, long delta);
99 static inline bool subpool_is_free(struct hugepage_subpool *spool)
103 if (spool->max_hpages != -1)
104 return spool->used_hpages == 0;
105 if (spool->min_hpages != -1)
106 return spool->rsv_hpages == spool->min_hpages;
111 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
112 unsigned long irq_flags)
114 spin_unlock_irqrestore(&spool->lock, irq_flags);
116 /* If no pages are used, and no other handles to the subpool
117 * remain, give up any reservations based on minimum size and
118 * free the subpool */
119 if (subpool_is_free(spool)) {
120 if (spool->min_hpages != -1)
121 hugetlb_acct_memory(spool->hstate,
127 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
130 struct hugepage_subpool *spool;
132 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
136 spin_lock_init(&spool->lock);
138 spool->max_hpages = max_hpages;
140 spool->min_hpages = min_hpages;
142 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
146 spool->rsv_hpages = min_hpages;
151 void hugepage_put_subpool(struct hugepage_subpool *spool)
155 spin_lock_irqsave(&spool->lock, flags);
156 BUG_ON(!spool->count);
158 unlock_or_release_subpool(spool, flags);
162 * Subpool accounting for allocating and reserving pages.
163 * Return -ENOMEM if there are not enough resources to satisfy the
164 * request. Otherwise, return the number of pages by which the
165 * global pools must be adjusted (upward). The returned value may
166 * only be different than the passed value (delta) in the case where
167 * a subpool minimum size must be maintained.
169 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
177 spin_lock_irq(&spool->lock);
179 if (spool->max_hpages != -1) { /* maximum size accounting */
180 if ((spool->used_hpages + delta) <= spool->max_hpages)
181 spool->used_hpages += delta;
188 /* minimum size accounting */
189 if (spool->min_hpages != -1 && spool->rsv_hpages) {
190 if (delta > spool->rsv_hpages) {
192 * Asking for more reserves than those already taken on
193 * behalf of subpool. Return difference.
195 ret = delta - spool->rsv_hpages;
196 spool->rsv_hpages = 0;
198 ret = 0; /* reserves already accounted for */
199 spool->rsv_hpages -= delta;
204 spin_unlock_irq(&spool->lock);
209 * Subpool accounting for freeing and unreserving pages.
210 * Return the number of global page reservations that must be dropped.
211 * The return value may only be different than the passed value (delta)
212 * in the case where a subpool minimum size must be maintained.
214 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
223 spin_lock_irqsave(&spool->lock, flags);
225 if (spool->max_hpages != -1) /* maximum size accounting */
226 spool->used_hpages -= delta;
228 /* minimum size accounting */
229 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
230 if (spool->rsv_hpages + delta <= spool->min_hpages)
233 ret = spool->rsv_hpages + delta - spool->min_hpages;
235 spool->rsv_hpages += delta;
236 if (spool->rsv_hpages > spool->min_hpages)
237 spool->rsv_hpages = spool->min_hpages;
241 * If hugetlbfs_put_super couldn't free spool due to an outstanding
242 * quota reference, free it now.
244 unlock_or_release_subpool(spool, flags);
249 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
251 return HUGETLBFS_SB(inode->i_sb)->spool;
254 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
256 return subpool_inode(file_inode(vma->vm_file));
259 /* Helper that removes a struct file_region from the resv_map cache and returns
262 static struct file_region *
263 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
265 struct file_region *nrg = NULL;
267 VM_BUG_ON(resv->region_cache_count <= 0);
269 resv->region_cache_count--;
270 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
271 list_del(&nrg->link);
279 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
280 struct file_region *rg)
282 #ifdef CONFIG_CGROUP_HUGETLB
283 nrg->reservation_counter = rg->reservation_counter;
290 /* Helper that records hugetlb_cgroup uncharge info. */
291 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
293 struct resv_map *resv,
294 struct file_region *nrg)
296 #ifdef CONFIG_CGROUP_HUGETLB
298 nrg->reservation_counter =
299 &h_cg->rsvd_hugepage[hstate_index(h)];
300 nrg->css = &h_cg->css;
302 * The caller will hold exactly one h_cg->css reference for the
303 * whole contiguous reservation region. But this area might be
304 * scattered when there are already some file_regions reside in
305 * it. As a result, many file_regions may share only one css
306 * reference. In order to ensure that one file_region must hold
307 * exactly one h_cg->css reference, we should do css_get for
308 * each file_region and leave the reference held by caller
312 if (!resv->pages_per_hpage)
313 resv->pages_per_hpage = pages_per_huge_page(h);
314 /* pages_per_hpage should be the same for all entries in
317 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
319 nrg->reservation_counter = NULL;
325 static void put_uncharge_info(struct file_region *rg)
327 #ifdef CONFIG_CGROUP_HUGETLB
333 static bool has_same_uncharge_info(struct file_region *rg,
334 struct file_region *org)
336 #ifdef CONFIG_CGROUP_HUGETLB
337 return rg->reservation_counter == org->reservation_counter &&
345 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
347 struct file_region *nrg = NULL, *prg = NULL;
349 prg = list_prev_entry(rg, link);
350 if (&prg->link != &resv->regions && prg->to == rg->from &&
351 has_same_uncharge_info(prg, rg)) {
355 put_uncharge_info(rg);
361 nrg = list_next_entry(rg, link);
362 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
363 has_same_uncharge_info(nrg, rg)) {
364 nrg->from = rg->from;
367 put_uncharge_info(rg);
373 hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from,
374 long to, struct hstate *h, struct hugetlb_cgroup *cg,
375 long *regions_needed)
377 struct file_region *nrg;
379 if (!regions_needed) {
380 nrg = get_file_region_entry_from_cache(map, from, to);
381 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
382 list_add(&nrg->link, rg->link.prev);
383 coalesce_file_region(map, nrg);
385 *regions_needed += 1;
391 * Must be called with resv->lock held.
393 * Calling this with regions_needed != NULL will count the number of pages
394 * to be added but will not modify the linked list. And regions_needed will
395 * indicate the number of file_regions needed in the cache to carry out to add
396 * the regions for this range.
398 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
399 struct hugetlb_cgroup *h_cg,
400 struct hstate *h, long *regions_needed)
403 struct list_head *head = &resv->regions;
404 long last_accounted_offset = f;
405 struct file_region *rg = NULL, *trg = NULL;
410 /* In this loop, we essentially handle an entry for the range
411 * [last_accounted_offset, rg->from), at every iteration, with some
414 list_for_each_entry_safe(rg, trg, head, link) {
415 /* Skip irrelevant regions that start before our range. */
417 /* If this region ends after the last accounted offset,
418 * then we need to update last_accounted_offset.
420 if (rg->to > last_accounted_offset)
421 last_accounted_offset = rg->to;
425 /* When we find a region that starts beyond our range, we've
431 /* Add an entry for last_accounted_offset -> rg->from, and
432 * update last_accounted_offset.
434 if (rg->from > last_accounted_offset)
435 add += hugetlb_resv_map_add(resv, rg,
436 last_accounted_offset,
440 last_accounted_offset = rg->to;
443 /* Handle the case where our range extends beyond
444 * 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_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);
1324 page[1].compound_nr = 0;
1325 __ClearPageHead(page);
1328 static void destroy_compound_hugetlb_page_for_demote(struct page *page,
1331 __destroy_compound_gigantic_page(page, order, true);
1334 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1335 static void destroy_compound_gigantic_page(struct page *page,
1338 __destroy_compound_gigantic_page(page, order, false);
1341 static void free_gigantic_page(struct page *page, unsigned int order)
1344 * If the page isn't allocated using the cma allocator,
1345 * cma_release() returns false.
1348 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1352 free_contig_range(page_to_pfn(page), 1 << order);
1355 #ifdef CONFIG_CONTIG_ALLOC
1356 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1357 int nid, nodemask_t *nodemask)
1359 unsigned long nr_pages = pages_per_huge_page(h);
1360 if (nid == NUMA_NO_NODE)
1361 nid = numa_mem_id();
1368 if (hugetlb_cma[nid]) {
1369 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1370 huge_page_order(h), true);
1375 if (!(gfp_mask & __GFP_THISNODE)) {
1376 for_each_node_mask(node, *nodemask) {
1377 if (node == nid || !hugetlb_cma[node])
1380 page = cma_alloc(hugetlb_cma[node], nr_pages,
1381 huge_page_order(h), true);
1389 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1392 #else /* !CONFIG_CONTIG_ALLOC */
1393 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1394 int nid, nodemask_t *nodemask)
1398 #endif /* CONFIG_CONTIG_ALLOC */
1400 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1401 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1402 int nid, nodemask_t *nodemask)
1406 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1407 static inline void destroy_compound_gigantic_page(struct page *page,
1408 unsigned int order) { }
1412 * Remove hugetlb page from lists, and update dtor so that page appears
1413 * as just a compound page.
1415 * A reference is held on the page, except in the case of demote.
1417 * Must be called with hugetlb lock held.
1419 static void __remove_hugetlb_page(struct hstate *h, struct page *page,
1420 bool adjust_surplus,
1423 int nid = page_to_nid(page);
1425 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1426 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1428 lockdep_assert_held(&hugetlb_lock);
1429 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1432 list_del(&page->lru);
1434 if (HPageFreed(page)) {
1435 h->free_huge_pages--;
1436 h->free_huge_pages_node[nid]--;
1438 if (adjust_surplus) {
1439 h->surplus_huge_pages--;
1440 h->surplus_huge_pages_node[nid]--;
1446 * For non-gigantic pages set the destructor to the normal compound
1447 * page dtor. This is needed in case someone takes an additional
1448 * temporary ref to the page, and freeing is delayed until they drop
1451 * For gigantic pages set the destructor to the null dtor. This
1452 * destructor will never be called. Before freeing the gigantic
1453 * page destroy_compound_gigantic_page will turn the compound page
1454 * into a simple group of pages. After this the destructor does not
1457 * This handles the case where more than one ref is held when and
1458 * after update_and_free_page is called.
1460 * In the case of demote we do not ref count the page as it will soon
1461 * be turned into a page of smaller size.
1464 set_page_refcounted(page);
1465 if (hstate_is_gigantic(h))
1466 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1468 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
1471 h->nr_huge_pages_node[nid]--;
1474 static void remove_hugetlb_page(struct hstate *h, struct page *page,
1475 bool adjust_surplus)
1477 __remove_hugetlb_page(h, page, adjust_surplus, false);
1480 static void remove_hugetlb_page_for_demote(struct hstate *h, struct page *page,
1481 bool adjust_surplus)
1483 __remove_hugetlb_page(h, page, adjust_surplus, true);
1486 static void add_hugetlb_page(struct hstate *h, struct page *page,
1487 bool adjust_surplus)
1490 int nid = page_to_nid(page);
1492 VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page);
1494 lockdep_assert_held(&hugetlb_lock);
1496 INIT_LIST_HEAD(&page->lru);
1498 h->nr_huge_pages_node[nid]++;
1500 if (adjust_surplus) {
1501 h->surplus_huge_pages++;
1502 h->surplus_huge_pages_node[nid]++;
1505 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1506 set_page_private(page, 0);
1507 SetHPageVmemmapOptimized(page);
1510 * This page is about to be managed by the hugetlb allocator and
1511 * should have no users. Drop our reference, and check for others
1514 zeroed = put_page_testzero(page);
1517 * It is VERY unlikely soneone else has taken a ref on
1518 * the page. In this case, we simply return as the
1519 * hugetlb destructor (free_huge_page) will be called
1520 * when this other ref is dropped.
1524 arch_clear_hugepage_flags(page);
1525 enqueue_huge_page(h, page);
1528 static void __update_and_free_page(struct hstate *h, struct page *page)
1531 struct page *subpage = page;
1533 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1536 if (alloc_huge_page_vmemmap(h, page)) {
1537 spin_lock_irq(&hugetlb_lock);
1539 * If we cannot allocate vmemmap pages, just refuse to free the
1540 * page and put the page back on the hugetlb free list and treat
1541 * as a surplus page.
1543 add_hugetlb_page(h, page, true);
1544 spin_unlock_irq(&hugetlb_lock);
1548 for (i = 0; i < pages_per_huge_page(h);
1549 i++, subpage = mem_map_next(subpage, page, i)) {
1550 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1551 1 << PG_referenced | 1 << PG_dirty |
1552 1 << PG_active | 1 << PG_private |
1557 * Non-gigantic pages demoted from CMA allocated gigantic pages
1558 * need to be given back to CMA in free_gigantic_page.
1560 if (hstate_is_gigantic(h) ||
1561 hugetlb_cma_page(page, huge_page_order(h))) {
1562 destroy_compound_gigantic_page(page, huge_page_order(h));
1563 free_gigantic_page(page, huge_page_order(h));
1565 __free_pages(page, huge_page_order(h));
1570 * As update_and_free_page() can be called under any context, so we cannot
1571 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1572 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1573 * the vmemmap pages.
1575 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1576 * freed and frees them one-by-one. As the page->mapping pointer is going
1577 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1578 * structure of a lockless linked list of huge pages to be freed.
1580 static LLIST_HEAD(hpage_freelist);
1582 static void free_hpage_workfn(struct work_struct *work)
1584 struct llist_node *node;
1586 node = llist_del_all(&hpage_freelist);
1592 page = container_of((struct address_space **)node,
1593 struct page, mapping);
1595 page->mapping = NULL;
1597 * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate()
1598 * is going to trigger because a previous call to
1599 * remove_hugetlb_page() will set_compound_page_dtor(page,
1600 * NULL_COMPOUND_DTOR), so do not use page_hstate() directly.
1602 h = size_to_hstate(page_size(page));
1604 __update_and_free_page(h, page);
1609 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1611 static inline void flush_free_hpage_work(struct hstate *h)
1613 if (free_vmemmap_pages_per_hpage(h))
1614 flush_work(&free_hpage_work);
1617 static void update_and_free_page(struct hstate *h, struct page *page,
1620 if (!HPageVmemmapOptimized(page) || !atomic) {
1621 __update_and_free_page(h, page);
1626 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1628 * Only call schedule_work() if hpage_freelist is previously
1629 * empty. Otherwise, schedule_work() had been called but the workfn
1630 * hasn't retrieved the list yet.
1632 if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist))
1633 schedule_work(&free_hpage_work);
1636 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1638 struct page *page, *t_page;
1640 list_for_each_entry_safe(page, t_page, list, lru) {
1641 update_and_free_page(h, page, false);
1646 struct hstate *size_to_hstate(unsigned long size)
1650 for_each_hstate(h) {
1651 if (huge_page_size(h) == size)
1657 void free_huge_page(struct page *page)
1660 * Can't pass hstate in here because it is called from the
1661 * compound page destructor.
1663 struct hstate *h = page_hstate(page);
1664 int nid = page_to_nid(page);
1665 struct hugepage_subpool *spool = hugetlb_page_subpool(page);
1666 bool restore_reserve;
1667 unsigned long flags;
1669 VM_BUG_ON_PAGE(page_count(page), page);
1670 VM_BUG_ON_PAGE(page_mapcount(page), page);
1672 hugetlb_set_page_subpool(page, NULL);
1673 page->mapping = NULL;
1674 restore_reserve = HPageRestoreReserve(page);
1675 ClearHPageRestoreReserve(page);
1678 * If HPageRestoreReserve was set on page, page allocation consumed a
1679 * reservation. If the page was associated with a subpool, there
1680 * would have been a page reserved in the subpool before allocation
1681 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1682 * reservation, do not call hugepage_subpool_put_pages() as this will
1683 * remove the reserved page from the subpool.
1685 if (!restore_reserve) {
1687 * A return code of zero implies that the subpool will be
1688 * under its minimum size if the reservation is not restored
1689 * after page is free. Therefore, force restore_reserve
1692 if (hugepage_subpool_put_pages(spool, 1) == 0)
1693 restore_reserve = true;
1696 spin_lock_irqsave(&hugetlb_lock, flags);
1697 ClearHPageMigratable(page);
1698 hugetlb_cgroup_uncharge_page(hstate_index(h),
1699 pages_per_huge_page(h), page);
1700 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1701 pages_per_huge_page(h), page);
1702 if (restore_reserve)
1703 h->resv_huge_pages++;
1705 if (HPageTemporary(page)) {
1706 remove_hugetlb_page(h, page, false);
1707 spin_unlock_irqrestore(&hugetlb_lock, flags);
1708 update_and_free_page(h, page, true);
1709 } else if (h->surplus_huge_pages_node[nid]) {
1710 /* remove the page from active list */
1711 remove_hugetlb_page(h, page, true);
1712 spin_unlock_irqrestore(&hugetlb_lock, flags);
1713 update_and_free_page(h, page, true);
1715 arch_clear_hugepage_flags(page);
1716 enqueue_huge_page(h, page);
1717 spin_unlock_irqrestore(&hugetlb_lock, flags);
1722 * Must be called with the hugetlb lock held
1724 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1726 lockdep_assert_held(&hugetlb_lock);
1728 h->nr_huge_pages_node[nid]++;
1731 static void __prep_new_huge_page(struct hstate *h, struct page *page)
1733 free_huge_page_vmemmap(h, page);
1734 INIT_LIST_HEAD(&page->lru);
1735 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1736 hugetlb_set_page_subpool(page, NULL);
1737 set_hugetlb_cgroup(page, NULL);
1738 set_hugetlb_cgroup_rsvd(page, NULL);
1741 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1743 __prep_new_huge_page(h, page);
1744 spin_lock_irq(&hugetlb_lock);
1745 __prep_account_new_huge_page(h, nid);
1746 spin_unlock_irq(&hugetlb_lock);
1749 static bool __prep_compound_gigantic_page(struct page *page, unsigned int order,
1753 int nr_pages = 1 << order;
1754 struct page *p = page + 1;
1756 /* we rely on prep_new_huge_page to set the destructor */
1757 set_compound_order(page, order);
1758 __ClearPageReserved(page);
1759 __SetPageHead(page);
1760 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1762 * For gigantic hugepages allocated through bootmem at
1763 * boot, it's safer to be consistent with the not-gigantic
1764 * hugepages and clear the PG_reserved bit from all tail pages
1765 * too. Otherwise drivers using get_user_pages() to access tail
1766 * pages may get the reference counting wrong if they see
1767 * PG_reserved set on a tail page (despite the head page not
1768 * having PG_reserved set). Enforcing this consistency between
1769 * head and tail pages allows drivers to optimize away a check
1770 * on the head page when they need know if put_page() is needed
1771 * after get_user_pages().
1773 __ClearPageReserved(p);
1775 * Subtle and very unlikely
1777 * Gigantic 'page allocators' such as memblock or cma will
1778 * return a set of pages with each page ref counted. We need
1779 * to turn this set of pages into a compound page with tail
1780 * page ref counts set to zero. Code such as speculative page
1781 * cache adding could take a ref on a 'to be' tail page.
1782 * We need to respect any increased ref count, and only set
1783 * the ref count to zero if count is currently 1. If count
1784 * is not 1, we return an error. An error return indicates
1785 * the set of pages can not be converted to a gigantic page.
1786 * The caller who allocated the pages should then discard the
1787 * pages using the appropriate free interface.
1789 * In the case of demote, the ref count will be zero.
1792 if (!page_ref_freeze(p, 1)) {
1793 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
1797 VM_BUG_ON_PAGE(page_count(p), p);
1799 set_compound_head(p, page);
1801 atomic_set(compound_mapcount_ptr(page), -1);
1802 atomic_set(compound_pincount_ptr(page), 0);
1806 /* undo tail page modifications made above */
1808 for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) {
1809 clear_compound_head(p);
1810 set_page_refcounted(p);
1812 /* need to clear PG_reserved on remaining tail pages */
1813 for (; j < nr_pages; j++, p = mem_map_next(p, page, j))
1814 __ClearPageReserved(p);
1815 set_compound_order(page, 0);
1816 page[1].compound_nr = 0;
1817 __ClearPageHead(page);
1821 static bool prep_compound_gigantic_page(struct page *page, unsigned int order)
1823 return __prep_compound_gigantic_page(page, order, false);
1826 static bool prep_compound_gigantic_page_for_demote(struct page *page,
1829 return __prep_compound_gigantic_page(page, order, true);
1833 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1834 * transparent huge pages. See the PageTransHuge() documentation for more
1837 int PageHuge(struct page *page)
1839 if (!PageCompound(page))
1842 page = compound_head(page);
1843 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1845 EXPORT_SYMBOL_GPL(PageHuge);
1848 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1849 * normal or transparent huge pages.
1851 int PageHeadHuge(struct page *page_head)
1853 if (!PageHead(page_head))
1856 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1858 EXPORT_SYMBOL_GPL(PageHeadHuge);
1861 * Find and lock address space (mapping) in write mode.
1863 * Upon entry, the page is locked which means that page_mapping() is
1864 * stable. Due to locking order, we can only trylock_write. If we can
1865 * not get the lock, simply return NULL to caller.
1867 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1869 struct address_space *mapping = page_mapping(hpage);
1874 if (i_mmap_trylock_write(mapping))
1880 pgoff_t hugetlb_basepage_index(struct page *page)
1882 struct page *page_head = compound_head(page);
1883 pgoff_t index = page_index(page_head);
1884 unsigned long compound_idx;
1886 if (compound_order(page_head) >= MAX_ORDER)
1887 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1889 compound_idx = page - page_head;
1891 return (index << compound_order(page_head)) + compound_idx;
1894 static struct page *alloc_buddy_huge_page(struct hstate *h,
1895 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1896 nodemask_t *node_alloc_noretry)
1898 int order = huge_page_order(h);
1900 bool alloc_try_hard = true;
1903 * By default we always try hard to allocate the page with
1904 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1905 * a loop (to adjust global huge page counts) and previous allocation
1906 * failed, do not continue to try hard on the same node. Use the
1907 * node_alloc_noretry bitmap to manage this state information.
1909 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1910 alloc_try_hard = false;
1911 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1913 gfp_mask |= __GFP_RETRY_MAYFAIL;
1914 if (nid == NUMA_NO_NODE)
1915 nid = numa_mem_id();
1916 page = __alloc_pages(gfp_mask, order, nid, nmask);
1918 __count_vm_event(HTLB_BUDDY_PGALLOC);
1920 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1923 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1924 * indicates an overall state change. Clear bit so that we resume
1925 * normal 'try hard' allocations.
1927 if (node_alloc_noretry && page && !alloc_try_hard)
1928 node_clear(nid, *node_alloc_noretry);
1931 * If we tried hard to get a page but failed, set bit so that
1932 * subsequent attempts will not try as hard until there is an
1933 * overall state change.
1935 if (node_alloc_noretry && !page && alloc_try_hard)
1936 node_set(nid, *node_alloc_noretry);
1942 * Common helper to allocate a fresh hugetlb page. All specific allocators
1943 * should use this function to get new hugetlb pages
1945 static struct page *alloc_fresh_huge_page(struct hstate *h,
1946 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1947 nodemask_t *node_alloc_noretry)
1953 if (hstate_is_gigantic(h))
1954 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1956 page = alloc_buddy_huge_page(h, gfp_mask,
1957 nid, nmask, node_alloc_noretry);
1961 if (hstate_is_gigantic(h)) {
1962 if (!prep_compound_gigantic_page(page, huge_page_order(h))) {
1964 * Rare failure to convert pages to compound page.
1965 * Free pages and try again - ONCE!
1967 free_gigantic_page(page, huge_page_order(h));
1975 prep_new_huge_page(h, page, page_to_nid(page));
1981 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1984 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1985 nodemask_t *node_alloc_noretry)
1989 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1991 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1992 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1993 node_alloc_noretry);
2001 put_page(page); /* free it into the hugepage allocator */
2007 * Remove huge page from pool from next node to free. Attempt to keep
2008 * persistent huge pages more or less balanced over allowed nodes.
2009 * This routine only 'removes' the hugetlb page. The caller must make
2010 * an additional call to free the page to low level allocators.
2011 * Called with hugetlb_lock locked.
2013 static struct page *remove_pool_huge_page(struct hstate *h,
2014 nodemask_t *nodes_allowed,
2018 struct page *page = NULL;
2020 lockdep_assert_held(&hugetlb_lock);
2021 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2023 * If we're returning unused surplus pages, only examine
2024 * nodes with surplus pages.
2026 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
2027 !list_empty(&h->hugepage_freelists[node])) {
2028 page = list_entry(h->hugepage_freelists[node].next,
2030 remove_hugetlb_page(h, page, acct_surplus);
2039 * Dissolve a given free hugepage into free buddy pages. This function does
2040 * nothing for in-use hugepages and non-hugepages.
2041 * This function returns values like below:
2043 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
2044 * when the system is under memory pressure and the feature of
2045 * freeing unused vmemmap pages associated with each hugetlb page
2047 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
2048 * (allocated or reserved.)
2049 * 0: successfully dissolved free hugepages or the page is not a
2050 * hugepage (considered as already dissolved)
2052 int dissolve_free_huge_page(struct page *page)
2057 /* Not to disrupt normal path by vainly holding hugetlb_lock */
2058 if (!PageHuge(page))
2061 spin_lock_irq(&hugetlb_lock);
2062 if (!PageHuge(page)) {
2067 if (!page_count(page)) {
2068 struct page *head = compound_head(page);
2069 struct hstate *h = page_hstate(head);
2070 if (h->free_huge_pages - h->resv_huge_pages == 0)
2074 * We should make sure that the page is already on the free list
2075 * when it is dissolved.
2077 if (unlikely(!HPageFreed(head))) {
2078 spin_unlock_irq(&hugetlb_lock);
2082 * Theoretically, we should return -EBUSY when we
2083 * encounter this race. In fact, we have a chance
2084 * to successfully dissolve the page if we do a
2085 * retry. Because the race window is quite small.
2086 * If we seize this opportunity, it is an optimization
2087 * for increasing the success rate of dissolving page.
2092 remove_hugetlb_page(h, head, false);
2093 h->max_huge_pages--;
2094 spin_unlock_irq(&hugetlb_lock);
2097 * Normally update_and_free_page will allocate required vmemmmap
2098 * before freeing the page. update_and_free_page will fail to
2099 * free the page if it can not allocate required vmemmap. We
2100 * need to adjust max_huge_pages if the page is not freed.
2101 * Attempt to allocate vmemmmap here so that we can take
2102 * appropriate action on failure.
2104 rc = alloc_huge_page_vmemmap(h, head);
2107 * Move PageHWPoison flag from head page to the raw
2108 * error page, which makes any subpages rather than
2109 * the error page reusable.
2111 if (PageHWPoison(head) && page != head) {
2112 SetPageHWPoison(page);
2113 ClearPageHWPoison(head);
2115 update_and_free_page(h, head, false);
2117 spin_lock_irq(&hugetlb_lock);
2118 add_hugetlb_page(h, head, false);
2119 h->max_huge_pages++;
2120 spin_unlock_irq(&hugetlb_lock);
2126 spin_unlock_irq(&hugetlb_lock);
2131 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2132 * make specified memory blocks removable from the system.
2133 * Note that this will dissolve a free gigantic hugepage completely, if any
2134 * part of it lies within the given range.
2135 * Also note that if dissolve_free_huge_page() returns with an error, all
2136 * free hugepages that were dissolved before that error are lost.
2138 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2144 if (!hugepages_supported())
2147 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
2148 page = pfn_to_page(pfn);
2149 rc = dissolve_free_huge_page(page);
2158 * Allocates a fresh surplus page from the page allocator.
2160 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
2161 int nid, nodemask_t *nmask, bool zero_ref)
2163 struct page *page = NULL;
2166 if (hstate_is_gigantic(h))
2169 spin_lock_irq(&hugetlb_lock);
2170 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2172 spin_unlock_irq(&hugetlb_lock);
2175 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2179 spin_lock_irq(&hugetlb_lock);
2181 * We could have raced with the pool size change.
2182 * Double check that and simply deallocate the new page
2183 * if we would end up overcommiting the surpluses. Abuse
2184 * temporary page to workaround the nasty free_huge_page
2187 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2188 SetHPageTemporary(page);
2189 spin_unlock_irq(&hugetlb_lock);
2196 * Caller requires a page with zero ref count.
2197 * We will drop ref count here. If someone else is holding
2198 * a ref, the page will be freed when they drop it. Abuse
2199 * temporary page flag to accomplish this.
2201 SetHPageTemporary(page);
2202 if (!put_page_testzero(page)) {
2204 * Unexpected inflated ref count on freshly allocated
2207 pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n");
2208 spin_unlock_irq(&hugetlb_lock);
2215 ClearHPageTemporary(page);
2218 h->surplus_huge_pages++;
2219 h->surplus_huge_pages_node[page_to_nid(page)]++;
2222 spin_unlock_irq(&hugetlb_lock);
2227 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
2228 int nid, nodemask_t *nmask)
2232 if (hstate_is_gigantic(h))
2235 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2240 * We do not account these pages as surplus because they are only
2241 * temporary and will be released properly on the last reference
2243 SetHPageTemporary(page);
2249 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2252 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
2253 struct vm_area_struct *vma, unsigned long addr)
2255 struct page *page = NULL;
2256 struct mempolicy *mpol;
2257 gfp_t gfp_mask = htlb_alloc_mask(h);
2259 nodemask_t *nodemask;
2261 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2262 if (mpol_is_preferred_many(mpol)) {
2263 gfp_t gfp = gfp_mask | __GFP_NOWARN;
2265 gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2266 page = alloc_surplus_huge_page(h, gfp, nid, nodemask, false);
2268 /* Fallback to all nodes if page==NULL */
2273 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask, false);
2274 mpol_cond_put(mpol);
2278 /* page migration callback function */
2279 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
2280 nodemask_t *nmask, gfp_t gfp_mask)
2282 spin_lock_irq(&hugetlb_lock);
2283 if (h->free_huge_pages - h->resv_huge_pages > 0) {
2286 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
2288 spin_unlock_irq(&hugetlb_lock);
2292 spin_unlock_irq(&hugetlb_lock);
2294 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2297 /* mempolicy aware migration callback */
2298 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2299 unsigned long address)
2301 struct mempolicy *mpol;
2302 nodemask_t *nodemask;
2307 gfp_mask = htlb_alloc_mask(h);
2308 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2309 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
2310 mpol_cond_put(mpol);
2316 * Increase the hugetlb pool such that it can accommodate a reservation
2319 static int gather_surplus_pages(struct hstate *h, long delta)
2320 __must_hold(&hugetlb_lock)
2322 struct list_head surplus_list;
2323 struct page *page, *tmp;
2326 long needed, allocated;
2327 bool alloc_ok = true;
2329 lockdep_assert_held(&hugetlb_lock);
2330 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2332 h->resv_huge_pages += delta;
2337 INIT_LIST_HEAD(&surplus_list);
2341 spin_unlock_irq(&hugetlb_lock);
2342 for (i = 0; i < needed; i++) {
2343 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2344 NUMA_NO_NODE, NULL, true);
2349 list_add(&page->lru, &surplus_list);
2355 * After retaking hugetlb_lock, we need to recalculate 'needed'
2356 * because either resv_huge_pages or free_huge_pages may have changed.
2358 spin_lock_irq(&hugetlb_lock);
2359 needed = (h->resv_huge_pages + delta) -
2360 (h->free_huge_pages + allocated);
2365 * We were not able to allocate enough pages to
2366 * satisfy the entire reservation so we free what
2367 * we've allocated so far.
2372 * The surplus_list now contains _at_least_ the number of extra pages
2373 * needed to accommodate the reservation. Add the appropriate number
2374 * of pages to the hugetlb pool and free the extras back to the buddy
2375 * allocator. Commit the entire reservation here to prevent another
2376 * process from stealing the pages as they are added to the pool but
2377 * before they are reserved.
2379 needed += allocated;
2380 h->resv_huge_pages += delta;
2383 /* Free the needed pages to the hugetlb pool */
2384 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2387 /* Add the page to the hugetlb allocator */
2388 enqueue_huge_page(h, page);
2391 spin_unlock_irq(&hugetlb_lock);
2394 * Free unnecessary surplus pages to the buddy allocator.
2395 * Pages have no ref count, call free_huge_page directly.
2397 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2398 free_huge_page(page);
2399 spin_lock_irq(&hugetlb_lock);
2405 * This routine has two main purposes:
2406 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2407 * in unused_resv_pages. This corresponds to the prior adjustments made
2408 * to the associated reservation map.
2409 * 2) Free any unused surplus pages that may have been allocated to satisfy
2410 * the reservation. As many as unused_resv_pages may be freed.
2412 static void return_unused_surplus_pages(struct hstate *h,
2413 unsigned long unused_resv_pages)
2415 unsigned long nr_pages;
2417 LIST_HEAD(page_list);
2419 lockdep_assert_held(&hugetlb_lock);
2420 /* Uncommit the reservation */
2421 h->resv_huge_pages -= unused_resv_pages;
2423 /* Cannot return gigantic pages currently */
2424 if (hstate_is_gigantic(h))
2428 * Part (or even all) of the reservation could have been backed
2429 * by pre-allocated pages. Only free surplus pages.
2431 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2434 * We want to release as many surplus pages as possible, spread
2435 * evenly across all nodes with memory. Iterate across these nodes
2436 * until we can no longer free unreserved surplus pages. This occurs
2437 * when the nodes with surplus pages have no free pages.
2438 * remove_pool_huge_page() will balance the freed pages across the
2439 * on-line nodes with memory and will handle the hstate accounting.
2441 while (nr_pages--) {
2442 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2446 list_add(&page->lru, &page_list);
2450 spin_unlock_irq(&hugetlb_lock);
2451 update_and_free_pages_bulk(h, &page_list);
2452 spin_lock_irq(&hugetlb_lock);
2457 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2458 * are used by the huge page allocation routines to manage reservations.
2460 * vma_needs_reservation is called to determine if the huge page at addr
2461 * within the vma has an associated reservation. If a reservation is
2462 * needed, the value 1 is returned. The caller is then responsible for
2463 * managing the global reservation and subpool usage counts. After
2464 * the huge page has been allocated, vma_commit_reservation is called
2465 * to add the page to the reservation map. If the page allocation fails,
2466 * the reservation must be ended instead of committed. vma_end_reservation
2467 * is called in such cases.
2469 * In the normal case, vma_commit_reservation returns the same value
2470 * as the preceding vma_needs_reservation call. The only time this
2471 * is not the case is if a reserve map was changed between calls. It
2472 * is the responsibility of the caller to notice the difference and
2473 * take appropriate action.
2475 * vma_add_reservation is used in error paths where a reservation must
2476 * be restored when a newly allocated huge page must be freed. It is
2477 * to be called after calling vma_needs_reservation to determine if a
2478 * reservation exists.
2480 * vma_del_reservation is used in error paths where an entry in the reserve
2481 * map was created during huge page allocation and must be removed. It is to
2482 * be called after calling vma_needs_reservation to determine if a reservation
2485 enum vma_resv_mode {
2492 static long __vma_reservation_common(struct hstate *h,
2493 struct vm_area_struct *vma, unsigned long addr,
2494 enum vma_resv_mode mode)
2496 struct resv_map *resv;
2499 long dummy_out_regions_needed;
2501 resv = vma_resv_map(vma);
2505 idx = vma_hugecache_offset(h, vma, addr);
2507 case VMA_NEEDS_RESV:
2508 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2509 /* We assume that vma_reservation_* routines always operate on
2510 * 1 page, and that adding to resv map a 1 page entry can only
2511 * ever require 1 region.
2513 VM_BUG_ON(dummy_out_regions_needed != 1);
2515 case VMA_COMMIT_RESV:
2516 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2517 /* region_add calls of range 1 should never fail. */
2521 region_abort(resv, idx, idx + 1, 1);
2525 if (vma->vm_flags & VM_MAYSHARE) {
2526 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2527 /* region_add calls of range 1 should never fail. */
2530 region_abort(resv, idx, idx + 1, 1);
2531 ret = region_del(resv, idx, idx + 1);
2535 if (vma->vm_flags & VM_MAYSHARE) {
2536 region_abort(resv, idx, idx + 1, 1);
2537 ret = region_del(resv, idx, idx + 1);
2539 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2540 /* region_add calls of range 1 should never fail. */
2548 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2551 * We know private mapping must have HPAGE_RESV_OWNER set.
2553 * In most cases, reserves always exist for private mappings.
2554 * However, a file associated with mapping could have been
2555 * hole punched or truncated after reserves were consumed.
2556 * As subsequent fault on such a range will not use reserves.
2557 * Subtle - The reserve map for private mappings has the
2558 * opposite meaning than that of shared mappings. If NO
2559 * entry is in the reserve map, it means a reservation exists.
2560 * If an entry exists in the reserve map, it means the
2561 * reservation has already been consumed. As a result, the
2562 * return value of this routine is the opposite of the
2563 * value returned from reserve map manipulation routines above.
2572 static long vma_needs_reservation(struct hstate *h,
2573 struct vm_area_struct *vma, unsigned long addr)
2575 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2578 static long vma_commit_reservation(struct hstate *h,
2579 struct vm_area_struct *vma, unsigned long addr)
2581 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2584 static void vma_end_reservation(struct hstate *h,
2585 struct vm_area_struct *vma, unsigned long addr)
2587 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2590 static long vma_add_reservation(struct hstate *h,
2591 struct vm_area_struct *vma, unsigned long addr)
2593 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2596 static long vma_del_reservation(struct hstate *h,
2597 struct vm_area_struct *vma, unsigned long addr)
2599 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2603 * This routine is called to restore reservation information on error paths.
2604 * It should ONLY be called for pages allocated via alloc_huge_page(), and
2605 * the hugetlb mutex should remain held when calling this routine.
2607 * It handles two specific cases:
2608 * 1) A reservation was in place and the page consumed the reservation.
2609 * HPageRestoreReserve is set in the page.
2610 * 2) No reservation was in place for the page, so HPageRestoreReserve is
2611 * not set. However, alloc_huge_page always updates the reserve map.
2613 * In case 1, free_huge_page later in the error path will increment the
2614 * global reserve count. But, free_huge_page does not have enough context
2615 * to adjust the reservation map. This case deals primarily with private
2616 * mappings. Adjust the reserve map here to be consistent with global
2617 * reserve count adjustments to be made by free_huge_page. Make sure the
2618 * reserve map indicates there is a reservation present.
2620 * In case 2, simply undo reserve map modifications done by alloc_huge_page.
2622 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2623 unsigned long address, struct page *page)
2625 long rc = vma_needs_reservation(h, vma, address);
2627 if (HPageRestoreReserve(page)) {
2628 if (unlikely(rc < 0))
2630 * Rare out of memory condition in reserve map
2631 * manipulation. Clear HPageRestoreReserve so that
2632 * global reserve count will not be incremented
2633 * by free_huge_page. This will make it appear
2634 * as though the reservation for this page was
2635 * consumed. This may prevent the task from
2636 * faulting in the page at a later time. This
2637 * is better than inconsistent global huge page
2638 * accounting of reserve counts.
2640 ClearHPageRestoreReserve(page);
2642 (void)vma_add_reservation(h, vma, address);
2644 vma_end_reservation(h, vma, address);
2648 * This indicates there is an entry in the reserve map
2649 * not added by alloc_huge_page. We know it was added
2650 * before the alloc_huge_page call, otherwise
2651 * HPageRestoreReserve would be set on the page.
2652 * Remove the entry so that a subsequent allocation
2653 * does not consume a reservation.
2655 rc = vma_del_reservation(h, vma, address);
2658 * VERY rare out of memory condition. Since
2659 * we can not delete the entry, set
2660 * HPageRestoreReserve so that the reserve
2661 * count will be incremented when the page
2662 * is freed. This reserve will be consumed
2663 * on a subsequent allocation.
2665 SetHPageRestoreReserve(page);
2666 } else if (rc < 0) {
2668 * Rare out of memory condition from
2669 * vma_needs_reservation call. Memory allocation is
2670 * only attempted if a new entry is needed. Therefore,
2671 * this implies there is not an entry in the
2674 * For shared mappings, no entry in the map indicates
2675 * no reservation. We are done.
2677 if (!(vma->vm_flags & VM_MAYSHARE))
2679 * For private mappings, no entry indicates
2680 * a reservation is present. Since we can
2681 * not add an entry, set SetHPageRestoreReserve
2682 * on the page so reserve count will be
2683 * incremented when freed. This reserve will
2684 * be consumed on a subsequent allocation.
2686 SetHPageRestoreReserve(page);
2689 * No reservation present, do nothing
2691 vma_end_reservation(h, vma, address);
2696 * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one
2697 * @h: struct hstate old page belongs to
2698 * @old_page: Old page to dissolve
2699 * @list: List to isolate the page in case we need to
2700 * Returns 0 on success, otherwise negated error.
2702 static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page,
2703 struct list_head *list)
2705 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2706 int nid = page_to_nid(old_page);
2707 bool alloc_retry = false;
2708 struct page *new_page;
2712 * Before dissolving the page, we need to allocate a new one for the
2713 * pool to remain stable. Here, we allocate the page and 'prep' it
2714 * by doing everything but actually updating counters and adding to
2715 * the pool. This simplifies and let us do most of the processing
2719 new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL);
2723 * If all goes well, this page will be directly added to the free
2724 * list in the pool. For this the ref count needs to be zero.
2725 * Attempt to drop now, and retry once if needed. It is VERY
2726 * unlikely there is another ref on the page.
2728 * If someone else has a reference to the page, it will be freed
2729 * when they drop their ref. Abuse temporary page flag to accomplish
2730 * this. Retry once if there is an inflated ref count.
2732 SetHPageTemporary(new_page);
2733 if (!put_page_testzero(new_page)) {
2740 ClearHPageTemporary(new_page);
2742 __prep_new_huge_page(h, new_page);
2745 spin_lock_irq(&hugetlb_lock);
2746 if (!PageHuge(old_page)) {
2748 * Freed from under us. Drop new_page too.
2751 } else if (page_count(old_page)) {
2753 * Someone has grabbed the page, try to isolate it here.
2754 * Fail with -EBUSY if not possible.
2756 spin_unlock_irq(&hugetlb_lock);
2757 if (!isolate_huge_page(old_page, list))
2759 spin_lock_irq(&hugetlb_lock);
2761 } else if (!HPageFreed(old_page)) {
2763 * Page's refcount is 0 but it has not been enqueued in the
2764 * freelist yet. Race window is small, so we can succeed here if
2767 spin_unlock_irq(&hugetlb_lock);
2772 * Ok, old_page is still a genuine free hugepage. Remove it from
2773 * the freelist and decrease the counters. These will be
2774 * incremented again when calling __prep_account_new_huge_page()
2775 * and enqueue_huge_page() for new_page. The counters will remain
2776 * stable since this happens under the lock.
2778 remove_hugetlb_page(h, old_page, false);
2781 * Ref count on new page is already zero as it was dropped
2782 * earlier. It can be directly added to the pool free list.
2784 __prep_account_new_huge_page(h, nid);
2785 enqueue_huge_page(h, new_page);
2788 * Pages have been replaced, we can safely free the old one.
2790 spin_unlock_irq(&hugetlb_lock);
2791 update_and_free_page(h, old_page, false);
2797 spin_unlock_irq(&hugetlb_lock);
2798 /* Page has a zero ref count, but needs a ref to be freed */
2799 set_page_refcounted(new_page);
2800 update_and_free_page(h, new_page, false);
2805 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
2812 * The page might have been dissolved from under our feet, so make sure
2813 * to carefully check the state under the lock.
2814 * Return success when racing as if we dissolved the page ourselves.
2816 spin_lock_irq(&hugetlb_lock);
2817 if (PageHuge(page)) {
2818 head = compound_head(page);
2819 h = page_hstate(head);
2821 spin_unlock_irq(&hugetlb_lock);
2824 spin_unlock_irq(&hugetlb_lock);
2827 * Fence off gigantic pages as there is a cyclic dependency between
2828 * alloc_contig_range and them. Return -ENOMEM as this has the effect
2829 * of bailing out right away without further retrying.
2831 if (hstate_is_gigantic(h))
2834 if (page_count(head) && isolate_huge_page(head, list))
2836 else if (!page_count(head))
2837 ret = alloc_and_dissolve_huge_page(h, head, list);
2842 struct page *alloc_huge_page(struct vm_area_struct *vma,
2843 unsigned long addr, int avoid_reserve)
2845 struct hugepage_subpool *spool = subpool_vma(vma);
2846 struct hstate *h = hstate_vma(vma);
2848 long map_chg, map_commit;
2851 struct hugetlb_cgroup *h_cg;
2852 bool deferred_reserve;
2854 idx = hstate_index(h);
2856 * Examine the region/reserve map to determine if the process
2857 * has a reservation for the page to be allocated. A return
2858 * code of zero indicates a reservation exists (no change).
2860 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2862 return ERR_PTR(-ENOMEM);
2865 * Processes that did not create the mapping will have no
2866 * reserves as indicated by the region/reserve map. Check
2867 * that the allocation will not exceed the subpool limit.
2868 * Allocations for MAP_NORESERVE mappings also need to be
2869 * checked against any subpool limit.
2871 if (map_chg || avoid_reserve) {
2872 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2874 vma_end_reservation(h, vma, addr);
2875 return ERR_PTR(-ENOSPC);
2879 * Even though there was no reservation in the region/reserve
2880 * map, there could be reservations associated with the
2881 * subpool that can be used. This would be indicated if the
2882 * return value of hugepage_subpool_get_pages() is zero.
2883 * However, if avoid_reserve is specified we still avoid even
2884 * the subpool reservations.
2890 /* If this allocation is not consuming a reservation, charge it now.
2892 deferred_reserve = map_chg || avoid_reserve;
2893 if (deferred_reserve) {
2894 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2895 idx, pages_per_huge_page(h), &h_cg);
2897 goto out_subpool_put;
2900 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2902 goto out_uncharge_cgroup_reservation;
2904 spin_lock_irq(&hugetlb_lock);
2906 * glb_chg is passed to indicate whether or not a page must be taken
2907 * from the global free pool (global change). gbl_chg == 0 indicates
2908 * a reservation exists for the allocation.
2910 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2912 spin_unlock_irq(&hugetlb_lock);
2913 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2915 goto out_uncharge_cgroup;
2916 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2917 SetHPageRestoreReserve(page);
2918 h->resv_huge_pages--;
2920 spin_lock_irq(&hugetlb_lock);
2921 list_add(&page->lru, &h->hugepage_activelist);
2924 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2925 /* If allocation is not consuming a reservation, also store the
2926 * hugetlb_cgroup pointer on the page.
2928 if (deferred_reserve) {
2929 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2933 spin_unlock_irq(&hugetlb_lock);
2935 hugetlb_set_page_subpool(page, spool);
2937 map_commit = vma_commit_reservation(h, vma, addr);
2938 if (unlikely(map_chg > map_commit)) {
2940 * The page was added to the reservation map between
2941 * vma_needs_reservation and vma_commit_reservation.
2942 * This indicates a race with hugetlb_reserve_pages.
2943 * Adjust for the subpool count incremented above AND
2944 * in hugetlb_reserve_pages for the same page. Also,
2945 * the reservation count added in hugetlb_reserve_pages
2946 * no longer applies.
2950 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2951 hugetlb_acct_memory(h, -rsv_adjust);
2952 if (deferred_reserve)
2953 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2954 pages_per_huge_page(h), page);
2958 out_uncharge_cgroup:
2959 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2960 out_uncharge_cgroup_reservation:
2961 if (deferred_reserve)
2962 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2965 if (map_chg || avoid_reserve)
2966 hugepage_subpool_put_pages(spool, 1);
2967 vma_end_reservation(h, vma, addr);
2968 return ERR_PTR(-ENOSPC);
2971 int alloc_bootmem_huge_page(struct hstate *h, int nid)
2972 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2973 int __alloc_bootmem_huge_page(struct hstate *h, int nid)
2975 struct huge_bootmem_page *m = NULL; /* initialize for clang */
2978 if (nid != NUMA_NO_NODE && nid >= nr_online_nodes)
2980 /* do node specific alloc */
2981 if (nid != NUMA_NO_NODE) {
2982 m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h),
2983 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
2988 /* allocate from next node when distributing huge pages */
2989 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2990 m = memblock_alloc_try_nid_raw(
2991 huge_page_size(h), huge_page_size(h),
2992 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2994 * Use the beginning of the huge page to store the
2995 * huge_bootmem_page struct (until gather_bootmem
2996 * puts them into the mem_map).
3004 /* Put them into a private list first because mem_map is not up yet */
3005 INIT_LIST_HEAD(&m->list);
3006 list_add(&m->list, &huge_boot_pages);
3012 * Put bootmem huge pages into the standard lists after mem_map is up.
3013 * Note: This only applies to gigantic (order > MAX_ORDER) pages.
3015 static void __init gather_bootmem_prealloc(void)
3017 struct huge_bootmem_page *m;
3019 list_for_each_entry(m, &huge_boot_pages, list) {
3020 struct page *page = virt_to_page(m);
3021 struct hstate *h = m->hstate;
3023 VM_BUG_ON(!hstate_is_gigantic(h));
3024 WARN_ON(page_count(page) != 1);
3025 if (prep_compound_gigantic_page(page, huge_page_order(h))) {
3026 WARN_ON(PageReserved(page));
3027 prep_new_huge_page(h, page, page_to_nid(page));
3028 put_page(page); /* add to the hugepage allocator */
3030 /* VERY unlikely inflated ref count on a tail page */
3031 free_gigantic_page(page, huge_page_order(h));
3035 * We need to restore the 'stolen' pages to totalram_pages
3036 * in order to fix confusing memory reports from free(1) and
3037 * other side-effects, like CommitLimit going negative.
3039 adjust_managed_page_count(page, pages_per_huge_page(h));
3043 static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)
3048 for (i = 0; i < h->max_huge_pages_node[nid]; ++i) {
3049 if (hstate_is_gigantic(h)) {
3050 if (!alloc_bootmem_huge_page(h, nid))
3054 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3056 page = alloc_fresh_huge_page(h, gfp_mask, nid,
3057 &node_states[N_MEMORY], NULL);
3060 put_page(page); /* free it into the hugepage allocator */
3064 if (i == h->max_huge_pages_node[nid])
3067 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3068 pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n",
3069 h->max_huge_pages_node[nid], buf, nid, i);
3070 h->max_huge_pages -= (h->max_huge_pages_node[nid] - i);
3071 h->max_huge_pages_node[nid] = i;
3074 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
3077 nodemask_t *node_alloc_noretry;
3078 bool node_specific_alloc = false;
3080 /* skip gigantic hugepages allocation if hugetlb_cma enabled */
3081 if (hstate_is_gigantic(h) && hugetlb_cma_size) {
3082 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
3086 /* do node specific alloc */
3087 for (i = 0; i < nr_online_nodes; i++) {
3088 if (h->max_huge_pages_node[i] > 0) {
3089 hugetlb_hstate_alloc_pages_onenode(h, i);
3090 node_specific_alloc = true;
3094 if (node_specific_alloc)
3097 /* below will do all node balanced alloc */
3098 if (!hstate_is_gigantic(h)) {
3100 * Bit mask controlling how hard we retry per-node allocations.
3101 * Ignore errors as lower level routines can deal with
3102 * node_alloc_noretry == NULL. If this kmalloc fails at boot
3103 * time, we are likely in bigger trouble.
3105 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
3108 /* allocations done at boot time */
3109 node_alloc_noretry = NULL;
3112 /* bit mask controlling how hard we retry per-node allocations */
3113 if (node_alloc_noretry)
3114 nodes_clear(*node_alloc_noretry);
3116 for (i = 0; i < h->max_huge_pages; ++i) {
3117 if (hstate_is_gigantic(h)) {
3118 if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))
3120 } else if (!alloc_pool_huge_page(h,
3121 &node_states[N_MEMORY],
3122 node_alloc_noretry))
3126 if (i < h->max_huge_pages) {
3129 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3130 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
3131 h->max_huge_pages, buf, i);
3132 h->max_huge_pages = i;
3134 kfree(node_alloc_noretry);
3137 static void __init hugetlb_init_hstates(void)
3139 struct hstate *h, *h2;
3141 for_each_hstate(h) {
3142 if (minimum_order > huge_page_order(h))
3143 minimum_order = huge_page_order(h);
3145 /* oversize hugepages were init'ed in early boot */
3146 if (!hstate_is_gigantic(h))
3147 hugetlb_hstate_alloc_pages(h);
3150 * Set demote order for each hstate. Note that
3151 * h->demote_order is initially 0.
3152 * - We can not demote gigantic pages if runtime freeing
3153 * is not supported, so skip this.
3154 * - If CMA allocation is possible, we can not demote
3155 * HUGETLB_PAGE_ORDER or smaller size pages.
3157 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3159 if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER)
3161 for_each_hstate(h2) {
3164 if (h2->order < h->order &&
3165 h2->order > h->demote_order)
3166 h->demote_order = h2->order;
3169 VM_BUG_ON(minimum_order == UINT_MAX);
3172 static void __init report_hugepages(void)
3176 for_each_hstate(h) {
3179 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3180 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
3181 buf, h->free_huge_pages);
3185 #ifdef CONFIG_HIGHMEM
3186 static void try_to_free_low(struct hstate *h, unsigned long count,
3187 nodemask_t *nodes_allowed)
3190 LIST_HEAD(page_list);
3192 lockdep_assert_held(&hugetlb_lock);
3193 if (hstate_is_gigantic(h))
3197 * Collect pages to be freed on a list, and free after dropping lock
3199 for_each_node_mask(i, *nodes_allowed) {
3200 struct page *page, *next;
3201 struct list_head *freel = &h->hugepage_freelists[i];
3202 list_for_each_entry_safe(page, next, freel, lru) {
3203 if (count >= h->nr_huge_pages)
3205 if (PageHighMem(page))
3207 remove_hugetlb_page(h, page, false);
3208 list_add(&page->lru, &page_list);
3213 spin_unlock_irq(&hugetlb_lock);
3214 update_and_free_pages_bulk(h, &page_list);
3215 spin_lock_irq(&hugetlb_lock);
3218 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3219 nodemask_t *nodes_allowed)
3225 * Increment or decrement surplus_huge_pages. Keep node-specific counters
3226 * balanced by operating on them in a round-robin fashion.
3227 * Returns 1 if an adjustment was made.
3229 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3234 lockdep_assert_held(&hugetlb_lock);
3235 VM_BUG_ON(delta != -1 && delta != 1);
3238 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3239 if (h->surplus_huge_pages_node[node])
3243 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3244 if (h->surplus_huge_pages_node[node] <
3245 h->nr_huge_pages_node[node])
3252 h->surplus_huge_pages += delta;
3253 h->surplus_huge_pages_node[node] += delta;
3257 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3258 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3259 nodemask_t *nodes_allowed)
3261 unsigned long min_count, ret;
3263 LIST_HEAD(page_list);
3264 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3267 * Bit mask controlling how hard we retry per-node allocations.
3268 * If we can not allocate the bit mask, do not attempt to allocate
3269 * the requested huge pages.
3271 if (node_alloc_noretry)
3272 nodes_clear(*node_alloc_noretry);
3277 * resize_lock mutex prevents concurrent adjustments to number of
3278 * pages in hstate via the proc/sysfs interfaces.
3280 mutex_lock(&h->resize_lock);
3281 flush_free_hpage_work(h);
3282 spin_lock_irq(&hugetlb_lock);
3285 * Check for a node specific request.
3286 * Changing node specific huge page count may require a corresponding
3287 * change to the global count. In any case, the passed node mask
3288 * (nodes_allowed) will restrict alloc/free to the specified node.
3290 if (nid != NUMA_NO_NODE) {
3291 unsigned long old_count = count;
3293 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3295 * User may have specified a large count value which caused the
3296 * above calculation to overflow. In this case, they wanted
3297 * to allocate as many huge pages as possible. Set count to
3298 * largest possible value to align with their intention.
3300 if (count < old_count)
3305 * Gigantic pages runtime allocation depend on the capability for large
3306 * page range allocation.
3307 * If the system does not provide this feature, return an error when
3308 * the user tries to allocate gigantic pages but let the user free the
3309 * boottime allocated gigantic pages.
3311 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3312 if (count > persistent_huge_pages(h)) {
3313 spin_unlock_irq(&hugetlb_lock);
3314 mutex_unlock(&h->resize_lock);
3315 NODEMASK_FREE(node_alloc_noretry);
3318 /* Fall through to decrease pool */
3322 * Increase the pool size
3323 * First take pages out of surplus state. Then make up the
3324 * remaining difference by allocating fresh huge pages.
3326 * We might race with alloc_surplus_huge_page() here and be unable
3327 * to convert a surplus huge page to a normal huge page. That is
3328 * not critical, though, it just means the overall size of the
3329 * pool might be one hugepage larger than it needs to be, but
3330 * within all the constraints specified by the sysctls.
3332 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3333 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3337 while (count > persistent_huge_pages(h)) {
3339 * If this allocation races such that we no longer need the
3340 * page, free_huge_page will handle it by freeing the page
3341 * and reducing the surplus.
3343 spin_unlock_irq(&hugetlb_lock);
3345 /* yield cpu to avoid soft lockup */
3348 ret = alloc_pool_huge_page(h, nodes_allowed,
3349 node_alloc_noretry);
3350 spin_lock_irq(&hugetlb_lock);
3354 /* Bail for signals. Probably ctrl-c from user */
3355 if (signal_pending(current))
3360 * Decrease the pool size
3361 * First return free pages to the buddy allocator (being careful
3362 * to keep enough around to satisfy reservations). Then place
3363 * pages into surplus state as needed so the pool will shrink
3364 * to the desired size as pages become free.
3366 * By placing pages into the surplus state independent of the
3367 * overcommit value, we are allowing the surplus pool size to
3368 * exceed overcommit. There are few sane options here. Since
3369 * alloc_surplus_huge_page() is checking the global counter,
3370 * though, we'll note that we're not allowed to exceed surplus
3371 * and won't grow the pool anywhere else. Not until one of the
3372 * sysctls are changed, or the surplus pages go out of use.
3374 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3375 min_count = max(count, min_count);
3376 try_to_free_low(h, min_count, nodes_allowed);
3379 * Collect pages to be removed on list without dropping lock
3381 while (min_count < persistent_huge_pages(h)) {
3382 page = remove_pool_huge_page(h, nodes_allowed, 0);
3386 list_add(&page->lru, &page_list);
3388 /* free the pages after dropping lock */
3389 spin_unlock_irq(&hugetlb_lock);
3390 update_and_free_pages_bulk(h, &page_list);
3391 flush_free_hpage_work(h);
3392 spin_lock_irq(&hugetlb_lock);
3394 while (count < persistent_huge_pages(h)) {
3395 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3399 h->max_huge_pages = persistent_huge_pages(h);
3400 spin_unlock_irq(&hugetlb_lock);
3401 mutex_unlock(&h->resize_lock);
3403 NODEMASK_FREE(node_alloc_noretry);
3408 static int demote_free_huge_page(struct hstate *h, struct page *page)
3410 int i, nid = page_to_nid(page);
3411 struct hstate *target_hstate;
3414 target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order);
3416 remove_hugetlb_page_for_demote(h, page, false);
3417 spin_unlock_irq(&hugetlb_lock);
3419 rc = alloc_huge_page_vmemmap(h, page);
3421 /* Allocation of vmemmmap failed, we can not demote page */
3422 spin_lock_irq(&hugetlb_lock);
3423 set_page_refcounted(page);
3424 add_hugetlb_page(h, page, false);
3429 * Use destroy_compound_hugetlb_page_for_demote for all huge page
3430 * sizes as it will not ref count pages.
3432 destroy_compound_hugetlb_page_for_demote(page, huge_page_order(h));
3435 * Taking target hstate mutex synchronizes with set_max_huge_pages.
3436 * Without the mutex, pages added to target hstate could be marked
3439 * Note that we already hold h->resize_lock. To prevent deadlock,
3440 * use the convention of always taking larger size hstate mutex first.
3442 mutex_lock(&target_hstate->resize_lock);
3443 for (i = 0; i < pages_per_huge_page(h);
3444 i += pages_per_huge_page(target_hstate)) {
3445 if (hstate_is_gigantic(target_hstate))
3446 prep_compound_gigantic_page_for_demote(page + i,
3447 target_hstate->order);
3449 prep_compound_page(page + i, target_hstate->order);
3450 set_page_private(page + i, 0);
3451 set_page_refcounted(page + i);
3452 prep_new_huge_page(target_hstate, page + i, nid);
3455 mutex_unlock(&target_hstate->resize_lock);
3457 spin_lock_irq(&hugetlb_lock);
3460 * Not absolutely necessary, but for consistency update max_huge_pages
3461 * based on pool changes for the demoted page.
3463 h->max_huge_pages--;
3464 target_hstate->max_huge_pages += pages_per_huge_page(h);
3469 static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
3470 __must_hold(&hugetlb_lock)
3476 lockdep_assert_held(&hugetlb_lock);
3478 /* We should never get here if no demote order */
3479 if (!h->demote_order) {
3480 pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n");
3481 return -EINVAL; /* internal error */
3484 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3485 if (!list_empty(&h->hugepage_freelists[node])) {
3486 page = list_entry(h->hugepage_freelists[node].next,
3488 rc = demote_free_huge_page(h, page);
3496 #define HSTATE_ATTR_RO(_name) \
3497 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3499 #define HSTATE_ATTR_WO(_name) \
3500 static struct kobj_attribute _name##_attr = __ATTR_WO(_name)
3502 #define HSTATE_ATTR(_name) \
3503 static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
3505 static struct kobject *hugepages_kobj;
3506 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3508 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3510 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3514 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3515 if (hstate_kobjs[i] == kobj) {
3517 *nidp = NUMA_NO_NODE;
3521 return kobj_to_node_hstate(kobj, nidp);
3524 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3525 struct kobj_attribute *attr, char *buf)
3528 unsigned long nr_huge_pages;
3531 h = kobj_to_hstate(kobj, &nid);
3532 if (nid == NUMA_NO_NODE)
3533 nr_huge_pages = h->nr_huge_pages;
3535 nr_huge_pages = h->nr_huge_pages_node[nid];
3537 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3540 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3541 struct hstate *h, int nid,
3542 unsigned long count, size_t len)
3545 nodemask_t nodes_allowed, *n_mask;
3547 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3550 if (nid == NUMA_NO_NODE) {
3552 * global hstate attribute
3554 if (!(obey_mempolicy &&
3555 init_nodemask_of_mempolicy(&nodes_allowed)))
3556 n_mask = &node_states[N_MEMORY];
3558 n_mask = &nodes_allowed;
3561 * Node specific request. count adjustment happens in
3562 * set_max_huge_pages() after acquiring hugetlb_lock.
3564 init_nodemask_of_node(&nodes_allowed, nid);
3565 n_mask = &nodes_allowed;
3568 err = set_max_huge_pages(h, count, nid, n_mask);
3570 return err ? err : len;
3573 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3574 struct kobject *kobj, const char *buf,
3578 unsigned long count;
3582 err = kstrtoul(buf, 10, &count);
3586 h = kobj_to_hstate(kobj, &nid);
3587 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3590 static ssize_t nr_hugepages_show(struct kobject *kobj,
3591 struct kobj_attribute *attr, char *buf)
3593 return nr_hugepages_show_common(kobj, attr, buf);
3596 static ssize_t nr_hugepages_store(struct kobject *kobj,
3597 struct kobj_attribute *attr, const char *buf, size_t len)
3599 return nr_hugepages_store_common(false, kobj, buf, len);
3601 HSTATE_ATTR(nr_hugepages);
3606 * hstate attribute for optionally mempolicy-based constraint on persistent
3607 * huge page alloc/free.
3609 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3610 struct kobj_attribute *attr,
3613 return nr_hugepages_show_common(kobj, attr, buf);
3616 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3617 struct kobj_attribute *attr, const char *buf, size_t len)
3619 return nr_hugepages_store_common(true, kobj, buf, len);
3621 HSTATE_ATTR(nr_hugepages_mempolicy);
3625 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3626 struct kobj_attribute *attr, char *buf)
3628 struct hstate *h = kobj_to_hstate(kobj, NULL);
3629 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3632 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3633 struct kobj_attribute *attr, const char *buf, size_t count)
3636 unsigned long input;
3637 struct hstate *h = kobj_to_hstate(kobj, NULL);
3639 if (hstate_is_gigantic(h))
3642 err = kstrtoul(buf, 10, &input);
3646 spin_lock_irq(&hugetlb_lock);
3647 h->nr_overcommit_huge_pages = input;
3648 spin_unlock_irq(&hugetlb_lock);
3652 HSTATE_ATTR(nr_overcommit_hugepages);
3654 static ssize_t free_hugepages_show(struct kobject *kobj,
3655 struct kobj_attribute *attr, char *buf)
3658 unsigned long free_huge_pages;
3661 h = kobj_to_hstate(kobj, &nid);
3662 if (nid == NUMA_NO_NODE)
3663 free_huge_pages = h->free_huge_pages;
3665 free_huge_pages = h->free_huge_pages_node[nid];
3667 return sysfs_emit(buf, "%lu\n", free_huge_pages);
3669 HSTATE_ATTR_RO(free_hugepages);
3671 static ssize_t resv_hugepages_show(struct kobject *kobj,
3672 struct kobj_attribute *attr, char *buf)
3674 struct hstate *h = kobj_to_hstate(kobj, NULL);
3675 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3677 HSTATE_ATTR_RO(resv_hugepages);
3679 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3680 struct kobj_attribute *attr, char *buf)
3683 unsigned long surplus_huge_pages;
3686 h = kobj_to_hstate(kobj, &nid);
3687 if (nid == NUMA_NO_NODE)
3688 surplus_huge_pages = h->surplus_huge_pages;
3690 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3692 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3694 HSTATE_ATTR_RO(surplus_hugepages);
3696 static ssize_t demote_store(struct kobject *kobj,
3697 struct kobj_attribute *attr, const char *buf, size_t len)
3699 unsigned long nr_demote;
3700 unsigned long nr_available;
3701 nodemask_t nodes_allowed, *n_mask;
3706 err = kstrtoul(buf, 10, &nr_demote);
3709 h = kobj_to_hstate(kobj, &nid);
3711 if (nid != NUMA_NO_NODE) {
3712 init_nodemask_of_node(&nodes_allowed, nid);
3713 n_mask = &nodes_allowed;
3715 n_mask = &node_states[N_MEMORY];
3718 /* Synchronize with other sysfs operations modifying huge pages */
3719 mutex_lock(&h->resize_lock);
3720 spin_lock_irq(&hugetlb_lock);
3724 * Check for available pages to demote each time thorough the
3725 * loop as demote_pool_huge_page will drop hugetlb_lock.
3727 if (nid != NUMA_NO_NODE)
3728 nr_available = h->free_huge_pages_node[nid];
3730 nr_available = h->free_huge_pages;
3731 nr_available -= h->resv_huge_pages;
3735 err = demote_pool_huge_page(h, n_mask);
3742 spin_unlock_irq(&hugetlb_lock);
3743 mutex_unlock(&h->resize_lock);
3749 HSTATE_ATTR_WO(demote);
3751 static ssize_t demote_size_show(struct kobject *kobj,
3752 struct kobj_attribute *attr, char *buf)
3755 struct hstate *h = kobj_to_hstate(kobj, &nid);
3756 unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K;
3758 return sysfs_emit(buf, "%lukB\n", demote_size);
3761 static ssize_t demote_size_store(struct kobject *kobj,
3762 struct kobj_attribute *attr,
3763 const char *buf, size_t count)
3765 struct hstate *h, *demote_hstate;
3766 unsigned long demote_size;
3767 unsigned int demote_order;
3770 demote_size = (unsigned long)memparse(buf, NULL);
3772 demote_hstate = size_to_hstate(demote_size);
3775 demote_order = demote_hstate->order;
3776 if (demote_order < HUGETLB_PAGE_ORDER)
3779 /* demote order must be smaller than hstate order */
3780 h = kobj_to_hstate(kobj, &nid);
3781 if (demote_order >= h->order)
3784 /* resize_lock synchronizes access to demote size and writes */
3785 mutex_lock(&h->resize_lock);
3786 h->demote_order = demote_order;
3787 mutex_unlock(&h->resize_lock);
3791 HSTATE_ATTR(demote_size);
3793 static struct attribute *hstate_attrs[] = {
3794 &nr_hugepages_attr.attr,
3795 &nr_overcommit_hugepages_attr.attr,
3796 &free_hugepages_attr.attr,
3797 &resv_hugepages_attr.attr,
3798 &surplus_hugepages_attr.attr,
3800 &nr_hugepages_mempolicy_attr.attr,
3805 static const struct attribute_group hstate_attr_group = {
3806 .attrs = hstate_attrs,
3809 static struct attribute *hstate_demote_attrs[] = {
3810 &demote_size_attr.attr,
3815 static const struct attribute_group hstate_demote_attr_group = {
3816 .attrs = hstate_demote_attrs,
3819 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3820 struct kobject **hstate_kobjs,
3821 const struct attribute_group *hstate_attr_group)
3824 int hi = hstate_index(h);
3826 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3827 if (!hstate_kobjs[hi])
3830 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3832 kobject_put(hstate_kobjs[hi]);
3833 hstate_kobjs[hi] = NULL;
3836 if (h->demote_order) {
3837 if (sysfs_create_group(hstate_kobjs[hi],
3838 &hstate_demote_attr_group))
3839 pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name);
3845 static void __init hugetlb_sysfs_init(void)
3850 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3851 if (!hugepages_kobj)
3854 for_each_hstate(h) {
3855 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3856 hstate_kobjs, &hstate_attr_group);
3858 pr_err("HugeTLB: Unable to add hstate %s", h->name);
3865 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3866 * with node devices in node_devices[] using a parallel array. The array
3867 * index of a node device or _hstate == node id.
3868 * This is here to avoid any static dependency of the node device driver, in
3869 * the base kernel, on the hugetlb module.
3871 struct node_hstate {
3872 struct kobject *hugepages_kobj;
3873 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3875 static struct node_hstate node_hstates[MAX_NUMNODES];
3878 * A subset of global hstate attributes for node devices
3880 static struct attribute *per_node_hstate_attrs[] = {
3881 &nr_hugepages_attr.attr,
3882 &free_hugepages_attr.attr,
3883 &surplus_hugepages_attr.attr,
3887 static const struct attribute_group per_node_hstate_attr_group = {
3888 .attrs = per_node_hstate_attrs,
3892 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3893 * Returns node id via non-NULL nidp.
3895 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3899 for (nid = 0; nid < nr_node_ids; nid++) {
3900 struct node_hstate *nhs = &node_hstates[nid];
3902 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3903 if (nhs->hstate_kobjs[i] == kobj) {
3915 * Unregister hstate attributes from a single node device.
3916 * No-op if no hstate attributes attached.
3918 static void hugetlb_unregister_node(struct node *node)
3921 struct node_hstate *nhs = &node_hstates[node->dev.id];
3923 if (!nhs->hugepages_kobj)
3924 return; /* no hstate attributes */
3926 for_each_hstate(h) {
3927 int idx = hstate_index(h);
3928 if (nhs->hstate_kobjs[idx]) {
3929 kobject_put(nhs->hstate_kobjs[idx]);
3930 nhs->hstate_kobjs[idx] = NULL;
3934 kobject_put(nhs->hugepages_kobj);
3935 nhs->hugepages_kobj = NULL;
3940 * Register hstate attributes for a single node device.
3941 * No-op if attributes already registered.
3943 static void hugetlb_register_node(struct node *node)
3946 struct node_hstate *nhs = &node_hstates[node->dev.id];
3949 if (nhs->hugepages_kobj)
3950 return; /* already allocated */
3952 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3954 if (!nhs->hugepages_kobj)
3957 for_each_hstate(h) {
3958 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3960 &per_node_hstate_attr_group);
3962 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3963 h->name, node->dev.id);
3964 hugetlb_unregister_node(node);
3971 * hugetlb init time: register hstate attributes for all registered node
3972 * devices of nodes that have memory. All on-line nodes should have
3973 * registered their associated device by this time.
3975 static void __init hugetlb_register_all_nodes(void)
3979 for_each_node_state(nid, N_MEMORY) {
3980 struct node *node = node_devices[nid];
3981 if (node->dev.id == nid)
3982 hugetlb_register_node(node);
3986 * Let the node device driver know we're here so it can
3987 * [un]register hstate attributes on node hotplug.
3989 register_hugetlbfs_with_node(hugetlb_register_node,
3990 hugetlb_unregister_node);
3992 #else /* !CONFIG_NUMA */
3994 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4002 static void hugetlb_register_all_nodes(void) { }
4006 static int __init hugetlb_init(void)
4010 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
4013 if (!hugepages_supported()) {
4014 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
4015 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
4020 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
4021 * architectures depend on setup being done here.
4023 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
4024 if (!parsed_default_hugepagesz) {
4026 * If we did not parse a default huge page size, set
4027 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
4028 * number of huge pages for this default size was implicitly
4029 * specified, set that here as well.
4030 * Note that the implicit setting will overwrite an explicit
4031 * setting. A warning will be printed in this case.
4033 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
4034 if (default_hstate_max_huge_pages) {
4035 if (default_hstate.max_huge_pages) {
4038 string_get_size(huge_page_size(&default_hstate),
4039 1, STRING_UNITS_2, buf, 32);
4040 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
4041 default_hstate.max_huge_pages, buf);
4042 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
4043 default_hstate_max_huge_pages);
4045 default_hstate.max_huge_pages =
4046 default_hstate_max_huge_pages;
4048 for (i = 0; i < nr_online_nodes; i++)
4049 default_hstate.max_huge_pages_node[i] =
4050 default_hugepages_in_node[i];
4054 hugetlb_cma_check();
4055 hugetlb_init_hstates();
4056 gather_bootmem_prealloc();
4059 hugetlb_sysfs_init();
4060 hugetlb_register_all_nodes();
4061 hugetlb_cgroup_file_init();
4064 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
4066 num_fault_mutexes = 1;
4068 hugetlb_fault_mutex_table =
4069 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
4071 BUG_ON(!hugetlb_fault_mutex_table);
4073 for (i = 0; i < num_fault_mutexes; i++)
4074 mutex_init(&hugetlb_fault_mutex_table[i]);
4077 subsys_initcall(hugetlb_init);
4079 /* Overwritten by architectures with more huge page sizes */
4080 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
4082 return size == HPAGE_SIZE;
4085 void __init hugetlb_add_hstate(unsigned int order)
4090 if (size_to_hstate(PAGE_SIZE << order)) {
4093 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
4095 h = &hstates[hugetlb_max_hstate++];
4096 mutex_init(&h->resize_lock);
4098 h->mask = ~(huge_page_size(h) - 1);
4099 for (i = 0; i < MAX_NUMNODES; ++i)
4100 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
4101 INIT_LIST_HEAD(&h->hugepage_activelist);
4102 h->next_nid_to_alloc = first_memory_node;
4103 h->next_nid_to_free = first_memory_node;
4104 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
4105 huge_page_size(h)/1024);
4106 hugetlb_vmemmap_init(h);
4111 bool __init __weak hugetlb_node_alloc_supported(void)
4116 * hugepages command line processing
4117 * hugepages normally follows a valid hugepagsz or default_hugepagsz
4118 * specification. If not, ignore the hugepages value. hugepages can also
4119 * be the first huge page command line option in which case it implicitly
4120 * specifies the number of huge pages for the default size.
4122 static int __init hugepages_setup(char *s)
4125 static unsigned long *last_mhp;
4126 int node = NUMA_NO_NODE;
4131 if (!parsed_valid_hugepagesz) {
4132 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
4133 parsed_valid_hugepagesz = true;
4138 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
4139 * yet, so this hugepages= parameter goes to the "default hstate".
4140 * Otherwise, it goes with the previously parsed hugepagesz or
4141 * default_hugepagesz.
4143 else if (!hugetlb_max_hstate)
4144 mhp = &default_hstate_max_huge_pages;
4146 mhp = &parsed_hstate->max_huge_pages;
4148 if (mhp == last_mhp) {
4149 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
4155 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4157 /* Parameter is node format */
4158 if (p[count] == ':') {
4159 if (!hugetlb_node_alloc_supported()) {
4160 pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n");
4163 if (tmp >= nr_online_nodes)
4165 node = array_index_nospec(tmp, nr_online_nodes);
4167 /* Parse hugepages */
4168 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4170 if (!hugetlb_max_hstate)
4171 default_hugepages_in_node[node] = tmp;
4173 parsed_hstate->max_huge_pages_node[node] = tmp;
4175 /* Go to parse next node*/
4176 if (p[count] == ',')
4189 * Global state is always initialized later in hugetlb_init.
4190 * But we need to allocate gigantic hstates here early to still
4191 * use the bootmem allocator.
4193 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
4194 hugetlb_hstate_alloc_pages(parsed_hstate);
4201 pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p);
4204 __setup("hugepages=", hugepages_setup);
4207 * hugepagesz command line processing
4208 * A specific huge page size can only be specified once with hugepagesz.
4209 * hugepagesz is followed by hugepages on the command line. The global
4210 * variable 'parsed_valid_hugepagesz' is used to determine if prior
4211 * hugepagesz argument was valid.
4213 static int __init hugepagesz_setup(char *s)
4218 parsed_valid_hugepagesz = false;
4219 size = (unsigned long)memparse(s, NULL);
4221 if (!arch_hugetlb_valid_size(size)) {
4222 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
4226 h = size_to_hstate(size);
4229 * hstate for this size already exists. This is normally
4230 * an error, but is allowed if the existing hstate is the
4231 * default hstate. More specifically, it is only allowed if
4232 * the number of huge pages for the default hstate was not
4233 * previously specified.
4235 if (!parsed_default_hugepagesz || h != &default_hstate ||
4236 default_hstate.max_huge_pages) {
4237 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
4242 * No need to call hugetlb_add_hstate() as hstate already
4243 * exists. But, do set parsed_hstate so that a following
4244 * hugepages= parameter will be applied to this hstate.
4247 parsed_valid_hugepagesz = true;
4251 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4252 parsed_valid_hugepagesz = true;
4255 __setup("hugepagesz=", hugepagesz_setup);
4258 * default_hugepagesz command line input
4259 * Only one instance of default_hugepagesz allowed on command line.
4261 static int __init default_hugepagesz_setup(char *s)
4266 parsed_valid_hugepagesz = false;
4267 if (parsed_default_hugepagesz) {
4268 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
4272 size = (unsigned long)memparse(s, NULL);
4274 if (!arch_hugetlb_valid_size(size)) {
4275 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
4279 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4280 parsed_valid_hugepagesz = true;
4281 parsed_default_hugepagesz = true;
4282 default_hstate_idx = hstate_index(size_to_hstate(size));
4285 * The number of default huge pages (for this size) could have been
4286 * specified as the first hugetlb parameter: hugepages=X. If so,
4287 * then default_hstate_max_huge_pages is set. If the default huge
4288 * page size is gigantic (>= MAX_ORDER), then the pages must be
4289 * allocated here from bootmem allocator.
4291 if (default_hstate_max_huge_pages) {
4292 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
4293 for (i = 0; i < nr_online_nodes; i++)
4294 default_hstate.max_huge_pages_node[i] =
4295 default_hugepages_in_node[i];
4296 if (hstate_is_gigantic(&default_hstate))
4297 hugetlb_hstate_alloc_pages(&default_hstate);
4298 default_hstate_max_huge_pages = 0;
4303 __setup("default_hugepagesz=", default_hugepagesz_setup);
4305 static unsigned int allowed_mems_nr(struct hstate *h)
4308 unsigned int nr = 0;
4309 nodemask_t *mpol_allowed;
4310 unsigned int *array = h->free_huge_pages_node;
4311 gfp_t gfp_mask = htlb_alloc_mask(h);
4313 mpol_allowed = policy_nodemask_current(gfp_mask);
4315 for_each_node_mask(node, cpuset_current_mems_allowed) {
4316 if (!mpol_allowed || node_isset(node, *mpol_allowed))
4323 #ifdef CONFIG_SYSCTL
4324 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
4325 void *buffer, size_t *length,
4326 loff_t *ppos, unsigned long *out)
4328 struct ctl_table dup_table;
4331 * In order to avoid races with __do_proc_doulongvec_minmax(), we
4332 * can duplicate the @table and alter the duplicate of it.
4335 dup_table.data = out;
4337 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
4340 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
4341 struct ctl_table *table, int write,
4342 void *buffer, size_t *length, loff_t *ppos)
4344 struct hstate *h = &default_hstate;
4345 unsigned long tmp = h->max_huge_pages;
4348 if (!hugepages_supported())
4351 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4357 ret = __nr_hugepages_store_common(obey_mempolicy, h,
4358 NUMA_NO_NODE, tmp, *length);
4363 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
4364 void *buffer, size_t *length, loff_t *ppos)
4367 return hugetlb_sysctl_handler_common(false, table, write,
4368 buffer, length, ppos);
4372 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
4373 void *buffer, size_t *length, loff_t *ppos)
4375 return hugetlb_sysctl_handler_common(true, table, write,
4376 buffer, length, ppos);
4378 #endif /* CONFIG_NUMA */
4380 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
4381 void *buffer, size_t *length, loff_t *ppos)
4383 struct hstate *h = &default_hstate;
4387 if (!hugepages_supported())
4390 tmp = h->nr_overcommit_huge_pages;
4392 if (write && hstate_is_gigantic(h))
4395 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4401 spin_lock_irq(&hugetlb_lock);
4402 h->nr_overcommit_huge_pages = tmp;
4403 spin_unlock_irq(&hugetlb_lock);
4409 #endif /* CONFIG_SYSCTL */
4411 void hugetlb_report_meminfo(struct seq_file *m)
4414 unsigned long total = 0;
4416 if (!hugepages_supported())
4419 for_each_hstate(h) {
4420 unsigned long count = h->nr_huge_pages;
4422 total += huge_page_size(h) * count;
4424 if (h == &default_hstate)
4426 "HugePages_Total: %5lu\n"
4427 "HugePages_Free: %5lu\n"
4428 "HugePages_Rsvd: %5lu\n"
4429 "HugePages_Surp: %5lu\n"
4430 "Hugepagesize: %8lu kB\n",
4434 h->surplus_huge_pages,
4435 huge_page_size(h) / SZ_1K);
4438 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
4441 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
4443 struct hstate *h = &default_hstate;
4445 if (!hugepages_supported())
4448 return sysfs_emit_at(buf, len,
4449 "Node %d HugePages_Total: %5u\n"
4450 "Node %d HugePages_Free: %5u\n"
4451 "Node %d HugePages_Surp: %5u\n",
4452 nid, h->nr_huge_pages_node[nid],
4453 nid, h->free_huge_pages_node[nid],
4454 nid, h->surplus_huge_pages_node[nid]);
4457 void hugetlb_show_meminfo(void)
4462 if (!hugepages_supported())
4465 for_each_node_state(nid, N_MEMORY)
4467 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
4469 h->nr_huge_pages_node[nid],
4470 h->free_huge_pages_node[nid],
4471 h->surplus_huge_pages_node[nid],
4472 huge_page_size(h) / SZ_1K);
4475 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
4477 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
4478 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
4481 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
4482 unsigned long hugetlb_total_pages(void)
4485 unsigned long nr_total_pages = 0;
4488 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
4489 return nr_total_pages;
4492 static int hugetlb_acct_memory(struct hstate *h, long delta)
4499 spin_lock_irq(&hugetlb_lock);
4501 * When cpuset is configured, it breaks the strict hugetlb page
4502 * reservation as the accounting is done on a global variable. Such
4503 * reservation is completely rubbish in the presence of cpuset because
4504 * the reservation is not checked against page availability for the
4505 * current cpuset. Application can still potentially OOM'ed by kernel
4506 * with lack of free htlb page in cpuset that the task is in.
4507 * Attempt to enforce strict accounting with cpuset is almost
4508 * impossible (or too ugly) because cpuset is too fluid that
4509 * task or memory node can be dynamically moved between cpusets.
4511 * The change of semantics for shared hugetlb mapping with cpuset is
4512 * undesirable. However, in order to preserve some of the semantics,
4513 * we fall back to check against current free page availability as
4514 * a best attempt and hopefully to minimize the impact of changing
4515 * semantics that cpuset has.
4517 * Apart from cpuset, we also have memory policy mechanism that
4518 * also determines from which node the kernel will allocate memory
4519 * in a NUMA system. So similar to cpuset, we also should consider
4520 * the memory policy of the current task. Similar to the description
4524 if (gather_surplus_pages(h, delta) < 0)
4527 if (delta > allowed_mems_nr(h)) {
4528 return_unused_surplus_pages(h, delta);
4535 return_unused_surplus_pages(h, (unsigned long) -delta);
4538 spin_unlock_irq(&hugetlb_lock);
4542 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4544 struct resv_map *resv = vma_resv_map(vma);
4547 * This new VMA should share its siblings reservation map if present.
4548 * The VMA will only ever have a valid reservation map pointer where
4549 * it is being copied for another still existing VMA. As that VMA
4550 * has a reference to the reservation map it cannot disappear until
4551 * after this open call completes. It is therefore safe to take a
4552 * new reference here without additional locking.
4554 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
4555 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
4556 kref_get(&resv->refs);
4560 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4562 struct hstate *h = hstate_vma(vma);
4563 struct resv_map *resv = vma_resv_map(vma);
4564 struct hugepage_subpool *spool = subpool_vma(vma);
4565 unsigned long reserve, start, end;
4568 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4571 start = vma_hugecache_offset(h, vma, vma->vm_start);
4572 end = vma_hugecache_offset(h, vma, vma->vm_end);
4574 reserve = (end - start) - region_count(resv, start, end);
4575 hugetlb_cgroup_uncharge_counter(resv, start, end);
4578 * Decrement reserve counts. The global reserve count may be
4579 * adjusted if the subpool has a minimum size.
4581 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4582 hugetlb_acct_memory(h, -gbl_reserve);
4585 kref_put(&resv->refs, resv_map_release);
4588 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4590 if (addr & ~(huge_page_mask(hstate_vma(vma))))
4595 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4597 return huge_page_size(hstate_vma(vma));
4601 * We cannot handle pagefaults against hugetlb pages at all. They cause
4602 * handle_mm_fault() to try to instantiate regular-sized pages in the
4603 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
4606 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4613 * When a new function is introduced to vm_operations_struct and added
4614 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4615 * This is because under System V memory model, mappings created via
4616 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4617 * their original vm_ops are overwritten with shm_vm_ops.
4619 const struct vm_operations_struct hugetlb_vm_ops = {
4620 .fault = hugetlb_vm_op_fault,
4621 .open = hugetlb_vm_op_open,
4622 .close = hugetlb_vm_op_close,
4623 .may_split = hugetlb_vm_op_split,
4624 .pagesize = hugetlb_vm_op_pagesize,
4627 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4631 unsigned int shift = huge_page_shift(hstate_vma(vma));
4634 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4635 vma->vm_page_prot)));
4637 entry = huge_pte_wrprotect(mk_huge_pte(page,
4638 vma->vm_page_prot));
4640 entry = pte_mkyoung(entry);
4641 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4646 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4647 unsigned long address, pte_t *ptep)
4651 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4652 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4653 update_mmu_cache(vma, address, ptep);
4656 bool is_hugetlb_entry_migration(pte_t pte)
4660 if (huge_pte_none(pte) || pte_present(pte))
4662 swp = pte_to_swp_entry(pte);
4663 if (is_migration_entry(swp))
4669 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4673 if (huge_pte_none(pte) || pte_present(pte))
4675 swp = pte_to_swp_entry(pte);
4676 if (is_hwpoison_entry(swp))
4683 hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
4684 struct page *new_page)
4686 __SetPageUptodate(new_page);
4687 hugepage_add_new_anon_rmap(new_page, vma, addr);
4688 set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
4689 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
4690 ClearHPageRestoreReserve(new_page);
4691 SetHPageMigratable(new_page);
4694 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
4695 struct vm_area_struct *vma)
4697 pte_t *src_pte, *dst_pte, entry, dst_entry;
4698 struct page *ptepage;
4700 bool cow = is_cow_mapping(vma->vm_flags);
4701 struct hstate *h = hstate_vma(vma);
4702 unsigned long sz = huge_page_size(h);
4703 unsigned long npages = pages_per_huge_page(h);
4704 struct address_space *mapping = vma->vm_file->f_mapping;
4705 struct mmu_notifier_range range;
4709 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
4712 mmu_notifier_invalidate_range_start(&range);
4715 * For shared mappings i_mmap_rwsem must be held to call
4716 * huge_pte_alloc, otherwise the returned ptep could go
4717 * away if part of a shared pmd and another thread calls
4720 i_mmap_lock_read(mapping);
4723 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
4724 spinlock_t *src_ptl, *dst_ptl;
4725 src_pte = huge_pte_offset(src, addr, sz);
4728 dst_pte = huge_pte_alloc(dst, vma, addr, sz);
4735 * If the pagetables are shared don't copy or take references.
4736 * dst_pte == src_pte is the common case of src/dest sharing.
4738 * However, src could have 'unshared' and dst shares with
4739 * another vma. If dst_pte !none, this implies sharing.
4740 * Check here before taking page table lock, and once again
4741 * after taking the lock below.
4743 dst_entry = huge_ptep_get(dst_pte);
4744 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
4747 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4748 src_ptl = huge_pte_lockptr(h, src, src_pte);
4749 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4750 entry = huge_ptep_get(src_pte);
4751 dst_entry = huge_ptep_get(dst_pte);
4753 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
4755 * Skip if src entry none. Also, skip in the
4756 * unlikely case dst entry !none as this implies
4757 * sharing with another vma.
4760 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
4761 is_hugetlb_entry_hwpoisoned(entry))) {
4762 swp_entry_t swp_entry = pte_to_swp_entry(entry);
4764 if (is_writable_migration_entry(swp_entry) && cow) {
4766 * COW mappings require pages in both
4767 * parent and child to be set to read.
4769 swp_entry = make_readable_migration_entry(
4770 swp_offset(swp_entry));
4771 entry = swp_entry_to_pte(swp_entry);
4772 set_huge_swap_pte_at(src, addr, src_pte,
4775 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
4777 entry = huge_ptep_get(src_pte);
4778 ptepage = pte_page(entry);
4782 * This is a rare case where we see pinned hugetlb
4783 * pages while they're prone to COW. We need to do the
4784 * COW earlier during fork.
4786 * When pre-allocating the page or copying data, we
4787 * need to be without the pgtable locks since we could
4788 * sleep during the process.
4790 if (unlikely(page_needs_cow_for_dma(vma, ptepage))) {
4791 pte_t src_pte_old = entry;
4794 spin_unlock(src_ptl);
4795 spin_unlock(dst_ptl);
4796 /* Do not use reserve as it's private owned */
4797 new = alloc_huge_page(vma, addr, 1);
4803 copy_user_huge_page(new, ptepage, addr, vma,
4807 /* Install the new huge page if src pte stable */
4808 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4809 src_ptl = huge_pte_lockptr(h, src, src_pte);
4810 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4811 entry = huge_ptep_get(src_pte);
4812 if (!pte_same(src_pte_old, entry)) {
4813 restore_reserve_on_error(h, vma, addr,
4816 /* dst_entry won't change as in child */
4819 hugetlb_install_page(vma, dst_pte, addr, new);
4820 spin_unlock(src_ptl);
4821 spin_unlock(dst_ptl);
4827 * No need to notify as we are downgrading page
4828 * table protection not changing it to point
4831 * See Documentation/vm/mmu_notifier.rst
4833 huge_ptep_set_wrprotect(src, addr, src_pte);
4834 entry = huge_pte_wrprotect(entry);
4837 page_dup_rmap(ptepage, true);
4838 set_huge_pte_at(dst, addr, dst_pte, entry);
4839 hugetlb_count_add(npages, dst);
4841 spin_unlock(src_ptl);
4842 spin_unlock(dst_ptl);
4846 mmu_notifier_invalidate_range_end(&range);
4848 i_mmap_unlock_read(mapping);
4853 static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
4854 unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte)
4856 struct hstate *h = hstate_vma(vma);
4857 struct mm_struct *mm = vma->vm_mm;
4858 spinlock_t *src_ptl, *dst_ptl;
4861 dst_ptl = huge_pte_lock(h, mm, dst_pte);
4862 src_ptl = huge_pte_lockptr(h, mm, src_pte);
4865 * We don't have to worry about the ordering of src and dst ptlocks
4866 * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock.
4868 if (src_ptl != dst_ptl)
4869 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4871 pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);
4872 set_huge_pte_at(mm, new_addr, dst_pte, pte);
4874 if (src_ptl != dst_ptl)
4875 spin_unlock(src_ptl);
4876 spin_unlock(dst_ptl);
4879 int move_hugetlb_page_tables(struct vm_area_struct *vma,
4880 struct vm_area_struct *new_vma,
4881 unsigned long old_addr, unsigned long new_addr,
4884 struct hstate *h = hstate_vma(vma);
4885 struct address_space *mapping = vma->vm_file->f_mapping;
4886 unsigned long sz = huge_page_size(h);
4887 struct mm_struct *mm = vma->vm_mm;
4888 unsigned long old_end = old_addr + len;
4889 unsigned long old_addr_copy;
4890 pte_t *src_pte, *dst_pte;
4891 struct mmu_notifier_range range;
4893 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr,
4895 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4896 mmu_notifier_invalidate_range_start(&range);
4897 /* Prevent race with file truncation */
4898 i_mmap_lock_write(mapping);
4899 for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
4900 src_pte = huge_pte_offset(mm, old_addr, sz);
4903 if (huge_pte_none(huge_ptep_get(src_pte)))
4906 /* old_addr arg to huge_pmd_unshare() is a pointer and so the
4907 * arg may be modified. Pass a copy instead to preserve the
4908 * value in old_addr.
4910 old_addr_copy = old_addr;
4912 if (huge_pmd_unshare(mm, vma, &old_addr_copy, src_pte))
4915 dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz);
4919 move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte);
4921 flush_tlb_range(vma, old_end - len, old_end);
4922 mmu_notifier_invalidate_range_end(&range);
4923 i_mmap_unlock_write(mapping);
4925 return len + old_addr - old_end;
4928 static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
4929 unsigned long start, unsigned long end,
4930 struct page *ref_page)
4932 struct mm_struct *mm = vma->vm_mm;
4933 unsigned long address;
4938 struct hstate *h = hstate_vma(vma);
4939 unsigned long sz = huge_page_size(h);
4940 struct mmu_notifier_range range;
4941 bool force_flush = false;
4943 WARN_ON(!is_vm_hugetlb_page(vma));
4944 BUG_ON(start & ~huge_page_mask(h));
4945 BUG_ON(end & ~huge_page_mask(h));
4948 * This is a hugetlb vma, all the pte entries should point
4951 tlb_change_page_size(tlb, sz);
4952 tlb_start_vma(tlb, vma);
4955 * If sharing possible, alert mmu notifiers of worst case.
4957 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
4959 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4960 mmu_notifier_invalidate_range_start(&range);
4962 for (; address < end; address += sz) {
4963 ptep = huge_pte_offset(mm, address, sz);
4967 ptl = huge_pte_lock(h, mm, ptep);
4968 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
4970 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
4975 pte = huge_ptep_get(ptep);
4976 if (huge_pte_none(pte)) {
4982 * Migrating hugepage or HWPoisoned hugepage is already
4983 * unmapped and its refcount is dropped, so just clear pte here.
4985 if (unlikely(!pte_present(pte))) {
4986 huge_pte_clear(mm, address, ptep, sz);
4991 page = pte_page(pte);
4993 * If a reference page is supplied, it is because a specific
4994 * page is being unmapped, not a range. Ensure the page we
4995 * are about to unmap is the actual page of interest.
4998 if (page != ref_page) {
5003 * Mark the VMA as having unmapped its page so that
5004 * future faults in this VMA will fail rather than
5005 * looking like data was lost
5007 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
5010 pte = huge_ptep_get_and_clear(mm, address, ptep);
5011 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
5012 if (huge_pte_dirty(pte))
5013 set_page_dirty(page);
5015 hugetlb_count_sub(pages_per_huge_page(h), mm);
5016 page_remove_rmap(page, true);
5019 tlb_remove_page_size(tlb, page, huge_page_size(h));
5021 * Bail out after unmapping reference page if supplied
5026 mmu_notifier_invalidate_range_end(&range);
5027 tlb_end_vma(tlb, vma);
5030 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
5031 * could defer the flush until now, since by holding i_mmap_rwsem we
5032 * guaranteed that the last refernece would not be dropped. But we must
5033 * do the flushing before we return, as otherwise i_mmap_rwsem will be
5034 * dropped and the last reference to the shared PMDs page might be
5037 * In theory we could defer the freeing of the PMD pages as well, but
5038 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
5039 * detect sharing, so we cannot defer the release of the page either.
5040 * Instead, do flush now.
5043 tlb_flush_mmu_tlbonly(tlb);
5046 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
5047 struct vm_area_struct *vma, unsigned long start,
5048 unsigned long end, struct page *ref_page)
5050 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
5053 * Clear this flag so that x86's huge_pmd_share page_table_shareable
5054 * test will fail on a vma being torn down, and not grab a page table
5055 * on its way out. We're lucky that the flag has such an appropriate
5056 * name, and can in fact be safely cleared here. We could clear it
5057 * before the __unmap_hugepage_range above, but all that's necessary
5058 * is to clear it before releasing the i_mmap_rwsem. This works
5059 * because in the context this is called, the VMA is about to be
5060 * destroyed and the i_mmap_rwsem is held.
5062 vma->vm_flags &= ~VM_MAYSHARE;
5065 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
5066 unsigned long end, struct page *ref_page)
5068 struct mmu_gather tlb;
5070 tlb_gather_mmu(&tlb, vma->vm_mm);
5071 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
5072 tlb_finish_mmu(&tlb);
5076 * This is called when the original mapper is failing to COW a MAP_PRIVATE
5077 * mapping it owns the reserve page for. The intention is to unmap the page
5078 * from other VMAs and let the children be SIGKILLed if they are faulting the
5081 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
5082 struct page *page, unsigned long address)
5084 struct hstate *h = hstate_vma(vma);
5085 struct vm_area_struct *iter_vma;
5086 struct address_space *mapping;
5090 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
5091 * from page cache lookup which is in HPAGE_SIZE units.
5093 address = address & huge_page_mask(h);
5094 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
5096 mapping = vma->vm_file->f_mapping;
5099 * Take the mapping lock for the duration of the table walk. As
5100 * this mapping should be shared between all the VMAs,
5101 * __unmap_hugepage_range() is called as the lock is already held
5103 i_mmap_lock_write(mapping);
5104 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
5105 /* Do not unmap the current VMA */
5106 if (iter_vma == vma)
5110 * Shared VMAs have their own reserves and do not affect
5111 * MAP_PRIVATE accounting but it is possible that a shared
5112 * VMA is using the same page so check and skip such VMAs.
5114 if (iter_vma->vm_flags & VM_MAYSHARE)
5118 * Unmap the page from other VMAs without their own reserves.
5119 * They get marked to be SIGKILLed if they fault in these
5120 * areas. This is because a future no-page fault on this VMA
5121 * could insert a zeroed page instead of the data existing
5122 * from the time of fork. This would look like data corruption
5124 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
5125 unmap_hugepage_range(iter_vma, address,
5126 address + huge_page_size(h), page);
5128 i_mmap_unlock_write(mapping);
5132 * Hugetlb_cow() should be called with page lock of the original hugepage held.
5133 * Called with hugetlb_fault_mutex_table held and pte_page locked so we
5134 * cannot race with other handlers or page migration.
5135 * Keep the pte_same checks anyway to make transition from the mutex easier.
5137 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
5138 unsigned long address, pte_t *ptep,
5139 struct page *pagecache_page, spinlock_t *ptl)
5142 struct hstate *h = hstate_vma(vma);
5143 struct page *old_page, *new_page;
5144 int outside_reserve = 0;
5146 unsigned long haddr = address & huge_page_mask(h);
5147 struct mmu_notifier_range range;
5149 pte = huge_ptep_get(ptep);
5150 old_page = pte_page(pte);
5153 /* If no-one else is actually using this page, avoid the copy
5154 * and just make the page writable */
5155 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
5156 page_move_anon_rmap(old_page, vma);
5157 set_huge_ptep_writable(vma, haddr, ptep);
5162 * If the process that created a MAP_PRIVATE mapping is about to
5163 * perform a COW due to a shared page count, attempt to satisfy
5164 * the allocation without using the existing reserves. The pagecache
5165 * page is used to determine if the reserve at this address was
5166 * consumed or not. If reserves were used, a partial faulted mapping
5167 * at the time of fork() could consume its reserves on COW instead
5168 * of the full address range.
5170 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
5171 old_page != pagecache_page)
5172 outside_reserve = 1;
5177 * Drop page table lock as buddy allocator may be called. It will
5178 * be acquired again before returning to the caller, as expected.
5181 new_page = alloc_huge_page(vma, haddr, outside_reserve);
5183 if (IS_ERR(new_page)) {
5185 * If a process owning a MAP_PRIVATE mapping fails to COW,
5186 * it is due to references held by a child and an insufficient
5187 * huge page pool. To guarantee the original mappers
5188 * reliability, unmap the page from child processes. The child
5189 * may get SIGKILLed if it later faults.
5191 if (outside_reserve) {
5192 struct address_space *mapping = vma->vm_file->f_mapping;
5197 BUG_ON(huge_pte_none(pte));
5199 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
5200 * unmapping. unmapping needs to hold i_mmap_rwsem
5201 * in write mode. Dropping i_mmap_rwsem in read mode
5202 * here is OK as COW mappings do not interact with
5205 * Reacquire both after unmap operation.
5207 idx = vma_hugecache_offset(h, vma, haddr);
5208 hash = hugetlb_fault_mutex_hash(mapping, idx);
5209 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5210 i_mmap_unlock_read(mapping);
5212 unmap_ref_private(mm, vma, old_page, haddr);
5214 i_mmap_lock_read(mapping);
5215 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5217 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5219 pte_same(huge_ptep_get(ptep), pte)))
5220 goto retry_avoidcopy;
5222 * race occurs while re-acquiring page table
5223 * lock, and our job is done.
5228 ret = vmf_error(PTR_ERR(new_page));
5229 goto out_release_old;
5233 * When the original hugepage is shared one, it does not have
5234 * anon_vma prepared.
5236 if (unlikely(anon_vma_prepare(vma))) {
5238 goto out_release_all;
5241 copy_user_huge_page(new_page, old_page, address, vma,
5242 pages_per_huge_page(h));
5243 __SetPageUptodate(new_page);
5245 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
5246 haddr + huge_page_size(h));
5247 mmu_notifier_invalidate_range_start(&range);
5250 * Retake the page table lock to check for racing updates
5251 * before the page tables are altered
5254 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5255 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
5256 ClearHPageRestoreReserve(new_page);
5259 huge_ptep_clear_flush(vma, haddr, ptep);
5260 mmu_notifier_invalidate_range(mm, range.start, range.end);
5261 page_remove_rmap(old_page, true);
5262 hugepage_add_new_anon_rmap(new_page, vma, haddr);
5263 set_huge_pte_at(mm, haddr, ptep,
5264 make_huge_pte(vma, new_page, 1));
5265 SetHPageMigratable(new_page);
5266 /* Make the old page be freed below */
5267 new_page = old_page;
5270 mmu_notifier_invalidate_range_end(&range);
5272 /* No restore in case of successful pagetable update (Break COW) */
5273 if (new_page != old_page)
5274 restore_reserve_on_error(h, vma, haddr, new_page);
5279 spin_lock(ptl); /* Caller expects lock to be held */
5283 /* Return the pagecache page at a given address within a VMA */
5284 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
5285 struct vm_area_struct *vma, unsigned long address)
5287 struct address_space *mapping;
5290 mapping = vma->vm_file->f_mapping;
5291 idx = vma_hugecache_offset(h, vma, address);
5293 return find_lock_page(mapping, idx);
5297 * Return whether there is a pagecache page to back given address within VMA.
5298 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
5300 static bool hugetlbfs_pagecache_present(struct hstate *h,
5301 struct vm_area_struct *vma, unsigned long address)
5303 struct address_space *mapping;
5307 mapping = vma->vm_file->f_mapping;
5308 idx = vma_hugecache_offset(h, vma, address);
5310 page = find_get_page(mapping, idx);
5313 return page != NULL;
5316 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
5319 struct inode *inode = mapping->host;
5320 struct hstate *h = hstate_inode(inode);
5321 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
5325 ClearHPageRestoreReserve(page);
5328 * set page dirty so that it will not be removed from cache/file
5329 * by non-hugetlbfs specific code paths.
5331 set_page_dirty(page);
5333 spin_lock(&inode->i_lock);
5334 inode->i_blocks += blocks_per_huge_page(h);
5335 spin_unlock(&inode->i_lock);
5339 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
5340 struct address_space *mapping,
5343 unsigned long haddr,
5345 unsigned long reason)
5349 struct vm_fault vmf = {
5352 .real_address = addr,
5356 * Hard to debug if it ends up being
5357 * used by a callee that assumes
5358 * something about the other
5359 * uninitialized fields... same as in
5365 * hugetlb_fault_mutex and i_mmap_rwsem must be
5366 * dropped before handling userfault. Reacquire
5367 * after handling fault to make calling code simpler.
5369 hash = hugetlb_fault_mutex_hash(mapping, idx);
5370 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5371 i_mmap_unlock_read(mapping);
5372 ret = handle_userfault(&vmf, reason);
5373 i_mmap_lock_read(mapping);
5374 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5379 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
5380 struct vm_area_struct *vma,
5381 struct address_space *mapping, pgoff_t idx,
5382 unsigned long address, pte_t *ptep, unsigned int flags)
5384 struct hstate *h = hstate_vma(vma);
5385 vm_fault_t ret = VM_FAULT_SIGBUS;
5391 unsigned long haddr = address & huge_page_mask(h);
5392 bool new_page, new_pagecache_page = false;
5395 * Currently, we are forced to kill the process in the event the
5396 * original mapper has unmapped pages from the child due to a failed
5397 * COW. Warn that such a situation has occurred as it may not be obvious
5399 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
5400 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
5406 * We can not race with truncation due to holding i_mmap_rwsem.
5407 * i_size is modified when holding i_mmap_rwsem, so check here
5408 * once for faults beyond end of file.
5410 size = i_size_read(mapping->host) >> huge_page_shift(h);
5416 page = find_lock_page(mapping, idx);
5418 /* Check for page in userfault range */
5419 if (userfaultfd_missing(vma)) {
5420 ret = hugetlb_handle_userfault(vma, mapping, idx,
5421 flags, haddr, address,
5426 page = alloc_huge_page(vma, haddr, 0);
5429 * Returning error will result in faulting task being
5430 * sent SIGBUS. The hugetlb fault mutex prevents two
5431 * tasks from racing to fault in the same page which
5432 * could result in false unable to allocate errors.
5433 * Page migration does not take the fault mutex, but
5434 * does a clear then write of pte's under page table
5435 * lock. Page fault code could race with migration,
5436 * notice the clear pte and try to allocate a page
5437 * here. Before returning error, get ptl and make
5438 * sure there really is no pte entry.
5440 ptl = huge_pte_lock(h, mm, ptep);
5442 if (huge_pte_none(huge_ptep_get(ptep)))
5443 ret = vmf_error(PTR_ERR(page));
5447 clear_huge_page(page, address, pages_per_huge_page(h));
5448 __SetPageUptodate(page);
5451 if (vma->vm_flags & VM_MAYSHARE) {
5452 int err = huge_add_to_page_cache(page, mapping, idx);
5459 new_pagecache_page = true;
5462 if (unlikely(anon_vma_prepare(vma))) {
5464 goto backout_unlocked;
5470 * If memory error occurs between mmap() and fault, some process
5471 * don't have hwpoisoned swap entry for errored virtual address.
5472 * So we need to block hugepage fault by PG_hwpoison bit check.
5474 if (unlikely(PageHWPoison(page))) {
5475 ret = VM_FAULT_HWPOISON_LARGE |
5476 VM_FAULT_SET_HINDEX(hstate_index(h));
5477 goto backout_unlocked;
5480 /* Check for page in userfault range. */
5481 if (userfaultfd_minor(vma)) {
5484 ret = hugetlb_handle_userfault(vma, mapping, idx,
5485 flags, haddr, address,
5492 * If we are going to COW a private mapping later, we examine the
5493 * pending reservations for this page now. This will ensure that
5494 * any allocations necessary to record that reservation occur outside
5497 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5498 if (vma_needs_reservation(h, vma, haddr) < 0) {
5500 goto backout_unlocked;
5502 /* Just decrements count, does not deallocate */
5503 vma_end_reservation(h, vma, haddr);
5506 ptl = huge_pte_lock(h, mm, ptep);
5508 if (!huge_pte_none(huge_ptep_get(ptep)))
5512 ClearHPageRestoreReserve(page);
5513 hugepage_add_new_anon_rmap(page, vma, haddr);
5515 page_dup_rmap(page, true);
5516 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
5517 && (vma->vm_flags & VM_SHARED)));
5518 set_huge_pte_at(mm, haddr, ptep, new_pte);
5520 hugetlb_count_add(pages_per_huge_page(h), mm);
5521 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5522 /* Optimization, do the COW without a second fault */
5523 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
5529 * Only set HPageMigratable in newly allocated pages. Existing pages
5530 * found in the pagecache may not have HPageMigratableset if they have
5531 * been isolated for migration.
5534 SetHPageMigratable(page);
5544 /* restore reserve for newly allocated pages not in page cache */
5545 if (new_page && !new_pagecache_page)
5546 restore_reserve_on_error(h, vma, haddr, page);
5552 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5554 unsigned long key[2];
5557 key[0] = (unsigned long) mapping;
5560 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
5562 return hash & (num_fault_mutexes - 1);
5566 * For uniprocessor systems we always use a single mutex, so just
5567 * return 0 and avoid the hashing overhead.
5569 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5575 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
5576 unsigned long address, unsigned int flags)
5583 struct page *page = NULL;
5584 struct page *pagecache_page = NULL;
5585 struct hstate *h = hstate_vma(vma);
5586 struct address_space *mapping;
5587 int need_wait_lock = 0;
5588 unsigned long haddr = address & huge_page_mask(h);
5590 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5593 * Since we hold no locks, ptep could be stale. That is
5594 * OK as we are only making decisions based on content and
5595 * not actually modifying content here.
5597 entry = huge_ptep_get(ptep);
5598 if (unlikely(is_hugetlb_entry_migration(entry))) {
5599 migration_entry_wait_huge(vma, mm, ptep);
5601 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
5602 return VM_FAULT_HWPOISON_LARGE |
5603 VM_FAULT_SET_HINDEX(hstate_index(h));
5607 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
5608 * until finished with ptep. This serves two purposes:
5609 * 1) It prevents huge_pmd_unshare from being called elsewhere
5610 * and making the ptep no longer valid.
5611 * 2) It synchronizes us with i_size modifications during truncation.
5613 * ptep could have already be assigned via huge_pte_offset. That
5614 * is OK, as huge_pte_alloc will return the same value unless
5615 * something has changed.
5617 mapping = vma->vm_file->f_mapping;
5618 i_mmap_lock_read(mapping);
5619 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
5621 i_mmap_unlock_read(mapping);
5622 return VM_FAULT_OOM;
5626 * Serialize hugepage allocation and instantiation, so that we don't
5627 * get spurious allocation failures if two CPUs race to instantiate
5628 * the same page in the page cache.
5630 idx = vma_hugecache_offset(h, vma, haddr);
5631 hash = hugetlb_fault_mutex_hash(mapping, idx);
5632 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5634 entry = huge_ptep_get(ptep);
5635 if (huge_pte_none(entry)) {
5636 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
5643 * entry could be a migration/hwpoison entry at this point, so this
5644 * check prevents the kernel from going below assuming that we have
5645 * an active hugepage in pagecache. This goto expects the 2nd page
5646 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
5647 * properly handle it.
5649 if (!pte_present(entry))
5653 * If we are going to COW the mapping later, we examine the pending
5654 * reservations for this page now. This will ensure that any
5655 * allocations necessary to record that reservation occur outside the
5656 * spinlock. For private mappings, we also lookup the pagecache
5657 * page now as it is used to determine if a reservation has been
5660 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
5661 if (vma_needs_reservation(h, vma, haddr) < 0) {
5665 /* Just decrements count, does not deallocate */
5666 vma_end_reservation(h, vma, haddr);
5668 if (!(vma->vm_flags & VM_MAYSHARE))
5669 pagecache_page = hugetlbfs_pagecache_page(h,
5673 ptl = huge_pte_lock(h, mm, ptep);
5675 /* Check for a racing update before calling hugetlb_cow */
5676 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
5680 * hugetlb_cow() requires page locks of pte_page(entry) and
5681 * pagecache_page, so here we need take the former one
5682 * when page != pagecache_page or !pagecache_page.
5684 page = pte_page(entry);
5685 if (page != pagecache_page)
5686 if (!trylock_page(page)) {
5693 if (flags & FAULT_FLAG_WRITE) {
5694 if (!huge_pte_write(entry)) {
5695 ret = hugetlb_cow(mm, vma, address, ptep,
5696 pagecache_page, ptl);
5699 entry = huge_pte_mkdirty(entry);
5701 entry = pte_mkyoung(entry);
5702 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
5703 flags & FAULT_FLAG_WRITE))
5704 update_mmu_cache(vma, haddr, ptep);
5706 if (page != pagecache_page)
5712 if (pagecache_page) {
5713 unlock_page(pagecache_page);
5714 put_page(pagecache_page);
5717 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5718 i_mmap_unlock_read(mapping);
5720 * Generally it's safe to hold refcount during waiting page lock. But
5721 * here we just wait to defer the next page fault to avoid busy loop and
5722 * the page is not used after unlocked before returning from the current
5723 * page fault. So we are safe from accessing freed page, even if we wait
5724 * here without taking refcount.
5727 wait_on_page_locked(page);
5731 #ifdef CONFIG_USERFAULTFD
5733 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
5734 * modifications for huge pages.
5736 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
5738 struct vm_area_struct *dst_vma,
5739 unsigned long dst_addr,
5740 unsigned long src_addr,
5741 enum mcopy_atomic_mode mode,
5742 struct page **pagep)
5744 bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE);
5745 struct hstate *h = hstate_vma(dst_vma);
5746 struct address_space *mapping = dst_vma->vm_file->f_mapping;
5747 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
5749 int vm_shared = dst_vma->vm_flags & VM_SHARED;
5755 bool page_in_pagecache = false;
5759 page = find_lock_page(mapping, idx);
5762 page_in_pagecache = true;
5763 } else if (!*pagep) {
5764 /* If a page already exists, then it's UFFDIO_COPY for
5765 * a non-missing case. Return -EEXIST.
5768 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5773 page = alloc_huge_page(dst_vma, dst_addr, 0);
5779 ret = copy_huge_page_from_user(page,
5780 (const void __user *) src_addr,
5781 pages_per_huge_page(h), false);
5783 /* fallback to copy_from_user outside mmap_lock */
5784 if (unlikely(ret)) {
5786 /* Free the allocated page which may have
5787 * consumed a reservation.
5789 restore_reserve_on_error(h, dst_vma, dst_addr, page);
5792 /* Allocate a temporary page to hold the copied
5795 page = alloc_huge_page_vma(h, dst_vma, dst_addr);
5801 /* Set the outparam pagep and return to the caller to
5802 * copy the contents outside the lock. Don't free the
5809 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5816 page = alloc_huge_page(dst_vma, dst_addr, 0);
5822 copy_user_huge_page(page, *pagep, dst_addr, dst_vma,
5823 pages_per_huge_page(h));
5829 * The memory barrier inside __SetPageUptodate makes sure that
5830 * preceding stores to the page contents become visible before
5831 * the set_pte_at() write.
5833 __SetPageUptodate(page);
5835 /* Add shared, newly allocated pages to the page cache. */
5836 if (vm_shared && !is_continue) {
5837 size = i_size_read(mapping->host) >> huge_page_shift(h);
5840 goto out_release_nounlock;
5843 * Serialization between remove_inode_hugepages() and
5844 * huge_add_to_page_cache() below happens through the
5845 * hugetlb_fault_mutex_table that here must be hold by
5848 ret = huge_add_to_page_cache(page, mapping, idx);
5850 goto out_release_nounlock;
5851 page_in_pagecache = true;
5854 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
5858 * Recheck the i_size after holding PT lock to make sure not
5859 * to leave any page mapped (as page_mapped()) beyond the end
5860 * of the i_size (remove_inode_hugepages() is strict about
5861 * enforcing that). If we bail out here, we'll also leave a
5862 * page in the radix tree in the vm_shared case beyond the end
5863 * of the i_size, but remove_inode_hugepages() will take care
5864 * of it as soon as we drop the hugetlb_fault_mutex_table.
5866 size = i_size_read(mapping->host) >> huge_page_shift(h);
5869 goto out_release_unlock;
5872 if (!huge_pte_none(huge_ptep_get(dst_pte)))
5873 goto out_release_unlock;
5876 page_dup_rmap(page, true);
5878 ClearHPageRestoreReserve(page);
5879 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
5882 /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */
5883 if (is_continue && !vm_shared)
5886 writable = dst_vma->vm_flags & VM_WRITE;
5888 _dst_pte = make_huge_pte(dst_vma, page, writable);
5890 _dst_pte = huge_pte_mkdirty(_dst_pte);
5891 _dst_pte = pte_mkyoung(_dst_pte);
5893 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
5895 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
5896 dst_vma->vm_flags & VM_WRITE);
5897 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
5899 /* No need to invalidate - it was non-present before */
5900 update_mmu_cache(dst_vma, dst_addr, dst_pte);
5904 SetHPageMigratable(page);
5905 if (vm_shared || is_continue)
5912 if (vm_shared || is_continue)
5914 out_release_nounlock:
5915 if (!page_in_pagecache)
5916 restore_reserve_on_error(h, dst_vma, dst_addr, page);
5920 #endif /* CONFIG_USERFAULTFD */
5922 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
5923 int refs, struct page **pages,
5924 struct vm_area_struct **vmas)
5928 for (nr = 0; nr < refs; nr++) {
5930 pages[nr] = mem_map_offset(page, nr);
5936 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
5937 struct page **pages, struct vm_area_struct **vmas,
5938 unsigned long *position, unsigned long *nr_pages,
5939 long i, unsigned int flags, int *locked)
5941 unsigned long pfn_offset;
5942 unsigned long vaddr = *position;
5943 unsigned long remainder = *nr_pages;
5944 struct hstate *h = hstate_vma(vma);
5945 int err = -EFAULT, refs;
5947 while (vaddr < vma->vm_end && remainder) {
5949 spinlock_t *ptl = NULL;
5954 * If we have a pending SIGKILL, don't keep faulting pages and
5955 * potentially allocating memory.
5957 if (fatal_signal_pending(current)) {
5963 * Some archs (sparc64, sh*) have multiple pte_ts to
5964 * each hugepage. We have to make sure we get the
5965 * first, for the page indexing below to work.
5967 * Note that page table lock is not held when pte is null.
5969 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
5972 ptl = huge_pte_lock(h, mm, pte);
5973 absent = !pte || huge_pte_none(huge_ptep_get(pte));
5976 * When coredumping, it suits get_dump_page if we just return
5977 * an error where there's an empty slot with no huge pagecache
5978 * to back it. This way, we avoid allocating a hugepage, and
5979 * the sparse dumpfile avoids allocating disk blocks, but its
5980 * huge holes still show up with zeroes where they need to be.
5982 if (absent && (flags & FOLL_DUMP) &&
5983 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
5991 * We need call hugetlb_fault for both hugepages under migration
5992 * (in which case hugetlb_fault waits for the migration,) and
5993 * hwpoisoned hugepages (in which case we need to prevent the
5994 * caller from accessing to them.) In order to do this, we use
5995 * here is_swap_pte instead of is_hugetlb_entry_migration and
5996 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
5997 * both cases, and because we can't follow correct pages
5998 * directly from any kind of swap entries.
6000 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
6001 ((flags & FOLL_WRITE) &&
6002 !huge_pte_write(huge_ptep_get(pte)))) {
6004 unsigned int fault_flags = 0;
6008 if (flags & FOLL_WRITE)
6009 fault_flags |= FAULT_FLAG_WRITE;
6011 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
6012 FAULT_FLAG_KILLABLE;
6013 if (flags & FOLL_NOWAIT)
6014 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
6015 FAULT_FLAG_RETRY_NOWAIT;
6016 if (flags & FOLL_TRIED) {
6018 * Note: FAULT_FLAG_ALLOW_RETRY and
6019 * FAULT_FLAG_TRIED can co-exist
6021 fault_flags |= FAULT_FLAG_TRIED;
6023 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
6024 if (ret & VM_FAULT_ERROR) {
6025 err = vm_fault_to_errno(ret, flags);
6029 if (ret & VM_FAULT_RETRY) {
6031 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
6035 * VM_FAULT_RETRY must not return an
6036 * error, it will return zero
6039 * No need to update "position" as the
6040 * caller will not check it after
6041 * *nr_pages is set to 0.
6048 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
6049 page = pte_page(huge_ptep_get(pte));
6052 * If subpage information not requested, update counters
6053 * and skip the same_page loop below.
6055 if (!pages && !vmas && !pfn_offset &&
6056 (vaddr + huge_page_size(h) < vma->vm_end) &&
6057 (remainder >= pages_per_huge_page(h))) {
6058 vaddr += huge_page_size(h);
6059 remainder -= pages_per_huge_page(h);
6060 i += pages_per_huge_page(h);
6065 /* vaddr may not be aligned to PAGE_SIZE */
6066 refs = min3(pages_per_huge_page(h) - pfn_offset, remainder,
6067 (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT);
6070 record_subpages_vmas(mem_map_offset(page, pfn_offset),
6072 likely(pages) ? pages + i : NULL,
6073 vmas ? vmas + i : NULL);
6077 * try_grab_compound_head() should always succeed here,
6078 * because: a) we hold the ptl lock, and b) we've just
6079 * checked that the huge page is present in the page
6080 * tables. If the huge page is present, then the tail
6081 * pages must also be present. The ptl prevents the
6082 * head page and tail pages from being rearranged in
6083 * any way. So this page must be available at this
6084 * point, unless the page refcount overflowed:
6086 if (WARN_ON_ONCE(!try_grab_compound_head(pages[i],
6096 vaddr += (refs << PAGE_SHIFT);
6102 *nr_pages = remainder;
6104 * setting position is actually required only if remainder is
6105 * not zero but it's faster not to add a "if (remainder)"
6113 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
6114 unsigned long address, unsigned long end, pgprot_t newprot)
6116 struct mm_struct *mm = vma->vm_mm;
6117 unsigned long start = address;
6120 struct hstate *h = hstate_vma(vma);
6121 unsigned long pages = 0;
6122 bool shared_pmd = false;
6123 struct mmu_notifier_range range;
6126 * In the case of shared PMDs, the area to flush could be beyond
6127 * start/end. Set range.start/range.end to cover the maximum possible
6128 * range if PMD sharing is possible.
6130 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
6131 0, vma, mm, start, end);
6132 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
6134 BUG_ON(address >= end);
6135 flush_cache_range(vma, range.start, range.end);
6137 mmu_notifier_invalidate_range_start(&range);
6138 i_mmap_lock_write(vma->vm_file->f_mapping);
6139 for (; address < end; address += huge_page_size(h)) {
6141 ptep = huge_pte_offset(mm, address, huge_page_size(h));
6144 ptl = huge_pte_lock(h, mm, ptep);
6145 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
6151 pte = huge_ptep_get(ptep);
6152 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
6156 if (unlikely(is_hugetlb_entry_migration(pte))) {
6157 swp_entry_t entry = pte_to_swp_entry(pte);
6159 if (is_writable_migration_entry(entry)) {
6162 entry = make_readable_migration_entry(
6164 newpte = swp_entry_to_pte(entry);
6165 set_huge_swap_pte_at(mm, address, ptep,
6166 newpte, huge_page_size(h));
6172 if (!huge_pte_none(pte)) {
6174 unsigned int shift = huge_page_shift(hstate_vma(vma));
6176 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
6177 pte = huge_pte_modify(old_pte, newprot);
6178 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
6179 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
6185 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
6186 * may have cleared our pud entry and done put_page on the page table:
6187 * once we release i_mmap_rwsem, another task can do the final put_page
6188 * and that page table be reused and filled with junk. If we actually
6189 * did unshare a page of pmds, flush the range corresponding to the pud.
6192 flush_hugetlb_tlb_range(vma, range.start, range.end);
6194 flush_hugetlb_tlb_range(vma, start, end);
6196 * No need to call mmu_notifier_invalidate_range() we are downgrading
6197 * page table protection not changing it to point to a new page.
6199 * See Documentation/vm/mmu_notifier.rst
6201 i_mmap_unlock_write(vma->vm_file->f_mapping);
6202 mmu_notifier_invalidate_range_end(&range);
6204 return pages << h->order;
6207 /* Return true if reservation was successful, false otherwise. */
6208 bool hugetlb_reserve_pages(struct inode *inode,
6210 struct vm_area_struct *vma,
6211 vm_flags_t vm_flags)
6214 struct hstate *h = hstate_inode(inode);
6215 struct hugepage_subpool *spool = subpool_inode(inode);
6216 struct resv_map *resv_map;
6217 struct hugetlb_cgroup *h_cg = NULL;
6218 long gbl_reserve, regions_needed = 0;
6220 /* This should never happen */
6222 VM_WARN(1, "%s called with a negative range\n", __func__);
6227 * Only apply hugepage reservation if asked. At fault time, an
6228 * attempt will be made for VM_NORESERVE to allocate a page
6229 * without using reserves
6231 if (vm_flags & VM_NORESERVE)
6235 * Shared mappings base their reservation on the number of pages that
6236 * are already allocated on behalf of the file. Private mappings need
6237 * to reserve the full area even if read-only as mprotect() may be
6238 * called to make the mapping read-write. Assume !vma is a shm mapping
6240 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6242 * resv_map can not be NULL as hugetlb_reserve_pages is only
6243 * called for inodes for which resv_maps were created (see
6244 * hugetlbfs_get_inode).
6246 resv_map = inode_resv_map(inode);
6248 chg = region_chg(resv_map, from, to, ®ions_needed);
6251 /* Private mapping. */
6252 resv_map = resv_map_alloc();
6258 set_vma_resv_map(vma, resv_map);
6259 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
6265 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
6266 chg * pages_per_huge_page(h), &h_cg) < 0)
6269 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
6270 /* For private mappings, the hugetlb_cgroup uncharge info hangs
6273 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
6277 * There must be enough pages in the subpool for the mapping. If
6278 * the subpool has a minimum size, there may be some global
6279 * reservations already in place (gbl_reserve).
6281 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
6282 if (gbl_reserve < 0)
6283 goto out_uncharge_cgroup;
6286 * Check enough hugepages are available for the reservation.
6287 * Hand the pages back to the subpool if there are not
6289 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
6293 * Account for the reservations made. Shared mappings record regions
6294 * that have reservations as they are shared by multiple VMAs.
6295 * When the last VMA disappears, the region map says how much
6296 * the reservation was and the page cache tells how much of
6297 * the reservation was consumed. Private mappings are per-VMA and
6298 * only the consumed reservations are tracked. When the VMA
6299 * disappears, the original reservation is the VMA size and the
6300 * consumed reservations are stored in the map. Hence, nothing
6301 * else has to be done for private mappings here
6303 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6304 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
6306 if (unlikely(add < 0)) {
6307 hugetlb_acct_memory(h, -gbl_reserve);
6309 } else if (unlikely(chg > add)) {
6311 * pages in this range were added to the reserve
6312 * map between region_chg and region_add. This
6313 * indicates a race with alloc_huge_page. Adjust
6314 * the subpool and reserve counts modified above
6315 * based on the difference.
6320 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
6321 * reference to h_cg->css. See comment below for detail.
6323 hugetlb_cgroup_uncharge_cgroup_rsvd(
6325 (chg - add) * pages_per_huge_page(h), h_cg);
6327 rsv_adjust = hugepage_subpool_put_pages(spool,
6329 hugetlb_acct_memory(h, -rsv_adjust);
6332 * The file_regions will hold their own reference to
6333 * h_cg->css. So we should release the reference held
6334 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
6337 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
6343 /* put back original number of pages, chg */
6344 (void)hugepage_subpool_put_pages(spool, chg);
6345 out_uncharge_cgroup:
6346 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
6347 chg * pages_per_huge_page(h), h_cg);
6349 if (!vma || vma->vm_flags & VM_MAYSHARE)
6350 /* Only call region_abort if the region_chg succeeded but the
6351 * region_add failed or didn't run.
6353 if (chg >= 0 && add < 0)
6354 region_abort(resv_map, from, to, regions_needed);
6355 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
6356 kref_put(&resv_map->refs, resv_map_release);
6360 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
6363 struct hstate *h = hstate_inode(inode);
6364 struct resv_map *resv_map = inode_resv_map(inode);
6366 struct hugepage_subpool *spool = subpool_inode(inode);
6370 * Since this routine can be called in the evict inode path for all
6371 * hugetlbfs inodes, resv_map could be NULL.
6374 chg = region_del(resv_map, start, end);
6376 * region_del() can fail in the rare case where a region
6377 * must be split and another region descriptor can not be
6378 * allocated. If end == LONG_MAX, it will not fail.
6384 spin_lock(&inode->i_lock);
6385 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
6386 spin_unlock(&inode->i_lock);
6389 * If the subpool has a minimum size, the number of global
6390 * reservations to be released may be adjusted.
6392 * Note that !resv_map implies freed == 0. So (chg - freed)
6393 * won't go negative.
6395 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
6396 hugetlb_acct_memory(h, -gbl_reserve);
6401 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
6402 static unsigned long page_table_shareable(struct vm_area_struct *svma,
6403 struct vm_area_struct *vma,
6404 unsigned long addr, pgoff_t idx)
6406 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
6408 unsigned long sbase = saddr & PUD_MASK;
6409 unsigned long s_end = sbase + PUD_SIZE;
6411 /* Allow segments to share if only one is marked locked */
6412 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
6413 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
6416 * match the virtual addresses, permission and the alignment of the
6419 if (pmd_index(addr) != pmd_index(saddr) ||
6420 vm_flags != svm_flags ||
6421 !range_in_vma(svma, sbase, s_end))
6427 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
6429 unsigned long base = addr & PUD_MASK;
6430 unsigned long end = base + PUD_SIZE;
6433 * check on proper vm_flags and page table alignment
6435 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
6440 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6442 #ifdef CONFIG_USERFAULTFD
6443 if (uffd_disable_huge_pmd_share(vma))
6446 return vma_shareable(vma, addr);
6450 * Determine if start,end range within vma could be mapped by shared pmd.
6451 * If yes, adjust start and end to cover range associated with possible
6452 * shared pmd mappings.
6454 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
6455 unsigned long *start, unsigned long *end)
6457 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
6458 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6461 * vma needs to span at least one aligned PUD size, and the range
6462 * must be at least partially within in.
6464 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
6465 (*end <= v_start) || (*start >= v_end))
6468 /* Extend the range to be PUD aligned for a worst case scenario */
6469 if (*start > v_start)
6470 *start = ALIGN_DOWN(*start, PUD_SIZE);
6473 *end = ALIGN(*end, PUD_SIZE);
6477 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
6478 * and returns the corresponding pte. While this is not necessary for the
6479 * !shared pmd case because we can allocate the pmd later as well, it makes the
6480 * code much cleaner.
6482 * This routine must be called with i_mmap_rwsem held in at least read mode if
6483 * sharing is possible. For hugetlbfs, this prevents removal of any page
6484 * table entries associated with the address space. This is important as we
6485 * are setting up sharing based on existing page table entries (mappings).
6487 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
6488 unsigned long addr, pud_t *pud)
6490 struct address_space *mapping = vma->vm_file->f_mapping;
6491 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
6493 struct vm_area_struct *svma;
6494 unsigned long saddr;
6499 i_mmap_assert_locked(mapping);
6500 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
6504 saddr = page_table_shareable(svma, vma, addr, idx);
6506 spte = huge_pte_offset(svma->vm_mm, saddr,
6507 vma_mmu_pagesize(svma));
6509 get_page(virt_to_page(spte));
6518 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
6519 if (pud_none(*pud)) {
6520 pud_populate(mm, pud,
6521 (pmd_t *)((unsigned long)spte & PAGE_MASK));
6524 put_page(virt_to_page(spte));
6528 pte = (pte_t *)pmd_alloc(mm, pud, addr);
6533 * unmap huge page backed by shared pte.
6535 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
6536 * indicated by page_count > 1, unmap is achieved by clearing pud and
6537 * decrementing the ref count. If count == 1, the pte page is not shared.
6539 * Called with page table lock held and i_mmap_rwsem held in write mode.
6541 * returns: 1 successfully unmapped a shared pte page
6542 * 0 the underlying pte page is not shared, or it is the last user
6544 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
6545 unsigned long *addr, pte_t *ptep)
6547 pgd_t *pgd = pgd_offset(mm, *addr);
6548 p4d_t *p4d = p4d_offset(pgd, *addr);
6549 pud_t *pud = pud_offset(p4d, *addr);
6551 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
6552 BUG_ON(page_count(virt_to_page(ptep)) == 0);
6553 if (page_count(virt_to_page(ptep)) == 1)
6557 put_page(virt_to_page(ptep));
6559 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
6563 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
6564 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
6565 unsigned long addr, pud_t *pud)
6570 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
6571 unsigned long *addr, pte_t *ptep)
6576 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
6577 unsigned long *start, unsigned long *end)
6581 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6585 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
6587 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
6588 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
6589 unsigned long addr, unsigned long sz)
6596 pgd = pgd_offset(mm, addr);
6597 p4d = p4d_alloc(mm, pgd, addr);
6600 pud = pud_alloc(mm, p4d, addr);
6602 if (sz == PUD_SIZE) {
6605 BUG_ON(sz != PMD_SIZE);
6606 if (want_pmd_share(vma, addr) && pud_none(*pud))
6607 pte = huge_pmd_share(mm, vma, addr, pud);
6609 pte = (pte_t *)pmd_alloc(mm, pud, addr);
6612 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
6618 * huge_pte_offset() - Walk the page table to resolve the hugepage
6619 * entry at address @addr
6621 * Return: Pointer to page table entry (PUD or PMD) for
6622 * address @addr, or NULL if a !p*d_present() entry is encountered and the
6623 * size @sz doesn't match the hugepage size at this level of the page
6626 pte_t *huge_pte_offset(struct mm_struct *mm,
6627 unsigned long addr, unsigned long sz)
6634 pgd = pgd_offset(mm, addr);
6635 if (!pgd_present(*pgd))
6637 p4d = p4d_offset(pgd, addr);
6638 if (!p4d_present(*p4d))
6641 pud = pud_offset(p4d, addr);
6643 /* must be pud huge, non-present or none */
6644 return (pte_t *)pud;
6645 if (!pud_present(*pud))
6647 /* must have a valid entry and size to go further */
6649 pmd = pmd_offset(pud, addr);
6650 /* must be pmd huge, non-present or none */
6651 return (pte_t *)pmd;
6654 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
6657 * These functions are overwritable if your architecture needs its own
6660 struct page * __weak
6661 follow_huge_addr(struct mm_struct *mm, unsigned long address,
6664 return ERR_PTR(-EINVAL);
6667 struct page * __weak
6668 follow_huge_pd(struct vm_area_struct *vma,
6669 unsigned long address, hugepd_t hpd, int flags, int pdshift)
6671 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
6675 struct page * __weak
6676 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
6677 pmd_t *pmd, int flags)
6679 struct page *page = NULL;
6683 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
6684 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
6685 (FOLL_PIN | FOLL_GET)))
6689 ptl = pmd_lockptr(mm, pmd);
6692 * make sure that the address range covered by this pmd is not
6693 * unmapped from other threads.
6695 if (!pmd_huge(*pmd))
6697 pte = huge_ptep_get((pte_t *)pmd);
6698 if (pte_present(pte)) {
6699 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
6701 * try_grab_page() should always succeed here, because: a) we
6702 * hold the pmd (ptl) lock, and b) we've just checked that the
6703 * huge pmd (head) page is present in the page tables. The ptl
6704 * prevents the head page and tail pages from being rearranged
6705 * in any way. So this page must be available at this point,
6706 * unless the page refcount overflowed:
6708 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
6713 if (is_hugetlb_entry_migration(pte)) {
6715 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
6719 * hwpoisoned entry is treated as no_page_table in
6720 * follow_page_mask().
6728 struct page * __weak
6729 follow_huge_pud(struct mm_struct *mm, unsigned long address,
6730 pud_t *pud, int flags)
6732 if (flags & (FOLL_GET | FOLL_PIN))
6735 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
6738 struct page * __weak
6739 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
6741 if (flags & (FOLL_GET | FOLL_PIN))
6744 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
6747 bool isolate_huge_page(struct page *page, struct list_head *list)
6751 spin_lock_irq(&hugetlb_lock);
6752 if (!PageHeadHuge(page) ||
6753 !HPageMigratable(page) ||
6754 !get_page_unless_zero(page)) {
6758 ClearHPageMigratable(page);
6759 list_move_tail(&page->lru, list);
6761 spin_unlock_irq(&hugetlb_lock);
6765 int get_hwpoison_huge_page(struct page *page, bool *hugetlb)
6770 spin_lock_irq(&hugetlb_lock);
6771 if (PageHeadHuge(page)) {
6773 if (HPageFreed(page) || HPageMigratable(page))
6774 ret = get_page_unless_zero(page);
6778 spin_unlock_irq(&hugetlb_lock);
6782 void putback_active_hugepage(struct page *page)
6784 spin_lock_irq(&hugetlb_lock);
6785 SetHPageMigratable(page);
6786 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
6787 spin_unlock_irq(&hugetlb_lock);
6791 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
6793 struct hstate *h = page_hstate(oldpage);
6795 hugetlb_cgroup_migrate(oldpage, newpage);
6796 set_page_owner_migrate_reason(newpage, reason);
6799 * transfer temporary state of the new huge page. This is
6800 * reverse to other transitions because the newpage is going to
6801 * be final while the old one will be freed so it takes over
6802 * the temporary status.
6804 * Also note that we have to transfer the per-node surplus state
6805 * here as well otherwise the global surplus count will not match
6808 if (HPageTemporary(newpage)) {
6809 int old_nid = page_to_nid(oldpage);
6810 int new_nid = page_to_nid(newpage);
6812 SetHPageTemporary(oldpage);
6813 ClearHPageTemporary(newpage);
6816 * There is no need to transfer the per-node surplus state
6817 * when we do not cross the node.
6819 if (new_nid == old_nid)
6821 spin_lock_irq(&hugetlb_lock);
6822 if (h->surplus_huge_pages_node[old_nid]) {
6823 h->surplus_huge_pages_node[old_nid]--;
6824 h->surplus_huge_pages_node[new_nid]++;
6826 spin_unlock_irq(&hugetlb_lock);
6831 * This function will unconditionally remove all the shared pmd pgtable entries
6832 * within the specific vma for a hugetlbfs memory range.
6834 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
6836 struct hstate *h = hstate_vma(vma);
6837 unsigned long sz = huge_page_size(h);
6838 struct mm_struct *mm = vma->vm_mm;
6839 struct mmu_notifier_range range;
6840 unsigned long address, start, end;
6844 if (!(vma->vm_flags & VM_MAYSHARE))
6847 start = ALIGN(vma->vm_start, PUD_SIZE);
6848 end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6854 * No need to call adjust_range_if_pmd_sharing_possible(), because
6855 * we have already done the PUD_SIZE alignment.
6857 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
6859 mmu_notifier_invalidate_range_start(&range);
6860 i_mmap_lock_write(vma->vm_file->f_mapping);
6861 for (address = start; address < end; address += PUD_SIZE) {
6862 unsigned long tmp = address;
6864 ptep = huge_pte_offset(mm, address, sz);
6867 ptl = huge_pte_lock(h, mm, ptep);
6868 /* We don't want 'address' to be changed */
6869 huge_pmd_unshare(mm, vma, &tmp, ptep);
6872 flush_hugetlb_tlb_range(vma, start, end);
6873 i_mmap_unlock_write(vma->vm_file->f_mapping);
6875 * No need to call mmu_notifier_invalidate_range(), see
6876 * Documentation/vm/mmu_notifier.rst.
6878 mmu_notifier_invalidate_range_end(&range);
6882 static bool cma_reserve_called __initdata;
6884 static int __init cmdline_parse_hugetlb_cma(char *p)
6891 if (sscanf(s, "%lu%n", &tmp, &count) != 1)
6894 if (s[count] == ':') {
6895 if (tmp >= MAX_NUMNODES)
6897 nid = array_index_nospec(tmp, MAX_NUMNODES);
6900 tmp = memparse(s, &s);
6901 hugetlb_cma_size_in_node[nid] = tmp;
6902 hugetlb_cma_size += tmp;
6905 * Skip the separator if have one, otherwise
6906 * break the parsing.
6913 hugetlb_cma_size = memparse(p, &p);
6921 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
6923 void __init hugetlb_cma_reserve(int order)
6925 unsigned long size, reserved, per_node;
6926 bool node_specific_cma_alloc = false;
6929 cma_reserve_called = true;
6931 if (!hugetlb_cma_size)
6934 for (nid = 0; nid < MAX_NUMNODES; nid++) {
6935 if (hugetlb_cma_size_in_node[nid] == 0)
6938 if (!node_state(nid, N_ONLINE)) {
6939 pr_warn("hugetlb_cma: invalid node %d specified\n", nid);
6940 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
6941 hugetlb_cma_size_in_node[nid] = 0;
6945 if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) {
6946 pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n",
6947 nid, (PAGE_SIZE << order) / SZ_1M);
6948 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
6949 hugetlb_cma_size_in_node[nid] = 0;
6951 node_specific_cma_alloc = true;
6955 /* Validate the CMA size again in case some invalid nodes specified. */
6956 if (!hugetlb_cma_size)
6959 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
6960 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
6961 (PAGE_SIZE << order) / SZ_1M);
6962 hugetlb_cma_size = 0;
6966 if (!node_specific_cma_alloc) {
6968 * If 3 GB area is requested on a machine with 4 numa nodes,
6969 * let's allocate 1 GB on first three nodes and ignore the last one.
6971 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
6972 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
6973 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
6977 for_each_node_state(nid, N_ONLINE) {
6979 char name[CMA_MAX_NAME];
6981 if (node_specific_cma_alloc) {
6982 if (hugetlb_cma_size_in_node[nid] == 0)
6985 size = hugetlb_cma_size_in_node[nid];
6987 size = min(per_node, hugetlb_cma_size - reserved);
6990 size = round_up(size, PAGE_SIZE << order);
6992 snprintf(name, sizeof(name), "hugetlb%d", nid);
6994 * Note that 'order per bit' is based on smallest size that
6995 * may be returned to CMA allocator in the case of
6996 * huge page demotion.
6998 res = cma_declare_contiguous_nid(0, size, 0,
6999 PAGE_SIZE << HUGETLB_PAGE_ORDER,
7001 &hugetlb_cma[nid], nid);
7003 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
7009 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
7012 if (reserved >= hugetlb_cma_size)
7018 * hugetlb_cma_size is used to determine if allocations from
7019 * cma are possible. Set to zero if no cma regions are set up.
7021 hugetlb_cma_size = 0;
7024 void __init hugetlb_cma_check(void)
7026 if (!hugetlb_cma_size || cma_reserve_called)
7029 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
7032 #endif /* CONFIG_CMA */