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);
1325 page[1].compound_nr = 0;
1327 __ClearPageHead(page);
1330 static void destroy_compound_hugetlb_page_for_demote(struct page *page,
1333 __destroy_compound_gigantic_page(page, order, true);
1336 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1337 static void destroy_compound_gigantic_page(struct page *page,
1340 __destroy_compound_gigantic_page(page, order, false);
1343 static void free_gigantic_page(struct page *page, unsigned int order)
1346 * If the page isn't allocated using the cma allocator,
1347 * cma_release() returns false.
1350 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1354 free_contig_range(page_to_pfn(page), 1 << order);
1357 #ifdef CONFIG_CONTIG_ALLOC
1358 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1359 int nid, nodemask_t *nodemask)
1361 unsigned long nr_pages = pages_per_huge_page(h);
1362 if (nid == NUMA_NO_NODE)
1363 nid = numa_mem_id();
1370 if (hugetlb_cma[nid]) {
1371 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1372 huge_page_order(h), true);
1377 if (!(gfp_mask & __GFP_THISNODE)) {
1378 for_each_node_mask(node, *nodemask) {
1379 if (node == nid || !hugetlb_cma[node])
1382 page = cma_alloc(hugetlb_cma[node], nr_pages,
1383 huge_page_order(h), true);
1391 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1394 #else /* !CONFIG_CONTIG_ALLOC */
1395 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1396 int nid, nodemask_t *nodemask)
1400 #endif /* CONFIG_CONTIG_ALLOC */
1402 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1403 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1404 int nid, nodemask_t *nodemask)
1408 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1409 static inline void destroy_compound_gigantic_page(struct page *page,
1410 unsigned int order) { }
1414 * Remove hugetlb page from lists, and update dtor so that page appears
1415 * as just a compound page.
1417 * A reference is held on the page, except in the case of demote.
1419 * Must be called with hugetlb lock held.
1421 static void __remove_hugetlb_page(struct hstate *h, struct page *page,
1422 bool adjust_surplus,
1425 int nid = page_to_nid(page);
1427 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1428 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1430 lockdep_assert_held(&hugetlb_lock);
1431 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1434 list_del(&page->lru);
1436 if (HPageFreed(page)) {
1437 h->free_huge_pages--;
1438 h->free_huge_pages_node[nid]--;
1440 if (adjust_surplus) {
1441 h->surplus_huge_pages--;
1442 h->surplus_huge_pages_node[nid]--;
1448 * For non-gigantic pages set the destructor to the normal compound
1449 * page dtor. This is needed in case someone takes an additional
1450 * temporary ref to the page, and freeing is delayed until they drop
1453 * For gigantic pages set the destructor to the null dtor. This
1454 * destructor will never be called. Before freeing the gigantic
1455 * page destroy_compound_gigantic_page will turn the compound page
1456 * into a simple group of pages. After this the destructor does not
1459 * This handles the case where more than one ref is held when and
1460 * after update_and_free_page is called.
1462 * In the case of demote we do not ref count the page as it will soon
1463 * be turned into a page of smaller size.
1466 set_page_refcounted(page);
1467 if (hstate_is_gigantic(h))
1468 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1470 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
1473 h->nr_huge_pages_node[nid]--;
1476 static void remove_hugetlb_page(struct hstate *h, struct page *page,
1477 bool adjust_surplus)
1479 __remove_hugetlb_page(h, page, adjust_surplus, false);
1482 static void remove_hugetlb_page_for_demote(struct hstate *h, struct page *page,
1483 bool adjust_surplus)
1485 __remove_hugetlb_page(h, page, adjust_surplus, true);
1488 static void add_hugetlb_page(struct hstate *h, struct page *page,
1489 bool adjust_surplus)
1492 int nid = page_to_nid(page);
1494 VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page);
1496 lockdep_assert_held(&hugetlb_lock);
1498 INIT_LIST_HEAD(&page->lru);
1500 h->nr_huge_pages_node[nid]++;
1502 if (adjust_surplus) {
1503 h->surplus_huge_pages++;
1504 h->surplus_huge_pages_node[nid]++;
1507 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1508 set_page_private(page, 0);
1509 SetHPageVmemmapOptimized(page);
1512 * This page is about to be managed by the hugetlb allocator and
1513 * should have no users. Drop our reference, and check for others
1516 zeroed = put_page_testzero(page);
1519 * It is VERY unlikely soneone else has taken a ref on
1520 * the page. In this case, we simply return as the
1521 * hugetlb destructor (free_huge_page) will be called
1522 * when this other ref is dropped.
1526 arch_clear_hugepage_flags(page);
1527 enqueue_huge_page(h, page);
1530 static void __update_and_free_page(struct hstate *h, struct page *page)
1533 struct page *subpage = page;
1535 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1538 if (alloc_huge_page_vmemmap(h, page)) {
1539 spin_lock_irq(&hugetlb_lock);
1541 * If we cannot allocate vmemmap pages, just refuse to free the
1542 * page and put the page back on the hugetlb free list and treat
1543 * as a surplus page.
1545 add_hugetlb_page(h, page, true);
1546 spin_unlock_irq(&hugetlb_lock);
1550 for (i = 0; i < pages_per_huge_page(h);
1551 i++, subpage = mem_map_next(subpage, page, i)) {
1552 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1553 1 << PG_referenced | 1 << PG_dirty |
1554 1 << PG_active | 1 << PG_private |
1559 * Non-gigantic pages demoted from CMA allocated gigantic pages
1560 * need to be given back to CMA in free_gigantic_page.
1562 if (hstate_is_gigantic(h) ||
1563 hugetlb_cma_page(page, huge_page_order(h))) {
1564 destroy_compound_gigantic_page(page, huge_page_order(h));
1565 free_gigantic_page(page, huge_page_order(h));
1567 __free_pages(page, huge_page_order(h));
1572 * As update_and_free_page() can be called under any context, so we cannot
1573 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1574 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1575 * the vmemmap pages.
1577 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1578 * freed and frees them one-by-one. As the page->mapping pointer is going
1579 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1580 * structure of a lockless linked list of huge pages to be freed.
1582 static LLIST_HEAD(hpage_freelist);
1584 static void free_hpage_workfn(struct work_struct *work)
1586 struct llist_node *node;
1588 node = llist_del_all(&hpage_freelist);
1594 page = container_of((struct address_space **)node,
1595 struct page, mapping);
1597 page->mapping = NULL;
1599 * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate()
1600 * is going to trigger because a previous call to
1601 * remove_hugetlb_page() will set_compound_page_dtor(page,
1602 * NULL_COMPOUND_DTOR), so do not use page_hstate() directly.
1604 h = size_to_hstate(page_size(page));
1606 __update_and_free_page(h, page);
1611 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1613 static inline void flush_free_hpage_work(struct hstate *h)
1615 if (free_vmemmap_pages_per_hpage(h))
1616 flush_work(&free_hpage_work);
1619 static void update_and_free_page(struct hstate *h, struct page *page,
1622 if (!HPageVmemmapOptimized(page) || !atomic) {
1623 __update_and_free_page(h, page);
1628 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1630 * Only call schedule_work() if hpage_freelist is previously
1631 * empty. Otherwise, schedule_work() had been called but the workfn
1632 * hasn't retrieved the list yet.
1634 if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist))
1635 schedule_work(&free_hpage_work);
1638 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1640 struct page *page, *t_page;
1642 list_for_each_entry_safe(page, t_page, list, lru) {
1643 update_and_free_page(h, page, false);
1648 struct hstate *size_to_hstate(unsigned long size)
1652 for_each_hstate(h) {
1653 if (huge_page_size(h) == size)
1659 void free_huge_page(struct page *page)
1662 * Can't pass hstate in here because it is called from the
1663 * compound page destructor.
1665 struct hstate *h = page_hstate(page);
1666 int nid = page_to_nid(page);
1667 struct hugepage_subpool *spool = hugetlb_page_subpool(page);
1668 bool restore_reserve;
1669 unsigned long flags;
1671 VM_BUG_ON_PAGE(page_count(page), page);
1672 VM_BUG_ON_PAGE(page_mapcount(page), page);
1674 hugetlb_set_page_subpool(page, NULL);
1675 page->mapping = NULL;
1676 restore_reserve = HPageRestoreReserve(page);
1677 ClearHPageRestoreReserve(page);
1680 * If HPageRestoreReserve was set on page, page allocation consumed a
1681 * reservation. If the page was associated with a subpool, there
1682 * would have been a page reserved in the subpool before allocation
1683 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1684 * reservation, do not call hugepage_subpool_put_pages() as this will
1685 * remove the reserved page from the subpool.
1687 if (!restore_reserve) {
1689 * A return code of zero implies that the subpool will be
1690 * under its minimum size if the reservation is not restored
1691 * after page is free. Therefore, force restore_reserve
1694 if (hugepage_subpool_put_pages(spool, 1) == 0)
1695 restore_reserve = true;
1698 spin_lock_irqsave(&hugetlb_lock, flags);
1699 ClearHPageMigratable(page);
1700 hugetlb_cgroup_uncharge_page(hstate_index(h),
1701 pages_per_huge_page(h), page);
1702 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1703 pages_per_huge_page(h), page);
1704 if (restore_reserve)
1705 h->resv_huge_pages++;
1707 if (HPageTemporary(page)) {
1708 remove_hugetlb_page(h, page, false);
1709 spin_unlock_irqrestore(&hugetlb_lock, flags);
1710 update_and_free_page(h, page, true);
1711 } else if (h->surplus_huge_pages_node[nid]) {
1712 /* remove the page from active list */
1713 remove_hugetlb_page(h, page, true);
1714 spin_unlock_irqrestore(&hugetlb_lock, flags);
1715 update_and_free_page(h, page, true);
1717 arch_clear_hugepage_flags(page);
1718 enqueue_huge_page(h, page);
1719 spin_unlock_irqrestore(&hugetlb_lock, flags);
1724 * Must be called with the hugetlb lock held
1726 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1728 lockdep_assert_held(&hugetlb_lock);
1730 h->nr_huge_pages_node[nid]++;
1733 static void __prep_new_huge_page(struct hstate *h, struct page *page)
1735 free_huge_page_vmemmap(h, page);
1736 INIT_LIST_HEAD(&page->lru);
1737 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1738 hugetlb_set_page_subpool(page, NULL);
1739 set_hugetlb_cgroup(page, NULL);
1740 set_hugetlb_cgroup_rsvd(page, NULL);
1743 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1745 __prep_new_huge_page(h, page);
1746 spin_lock_irq(&hugetlb_lock);
1747 __prep_account_new_huge_page(h, nid);
1748 spin_unlock_irq(&hugetlb_lock);
1751 static bool __prep_compound_gigantic_page(struct page *page, unsigned int order,
1755 int nr_pages = 1 << order;
1756 struct page *p = page + 1;
1758 /* we rely on prep_new_huge_page to set the destructor */
1759 set_compound_order(page, order);
1760 __ClearPageReserved(page);
1761 __SetPageHead(page);
1762 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1764 * For gigantic hugepages allocated through bootmem at
1765 * boot, it's safer to be consistent with the not-gigantic
1766 * hugepages and clear the PG_reserved bit from all tail pages
1767 * too. Otherwise drivers using get_user_pages() to access tail
1768 * pages may get the reference counting wrong if they see
1769 * PG_reserved set on a tail page (despite the head page not
1770 * having PG_reserved set). Enforcing this consistency between
1771 * head and tail pages allows drivers to optimize away a check
1772 * on the head page when they need know if put_page() is needed
1773 * after get_user_pages().
1775 __ClearPageReserved(p);
1777 * Subtle and very unlikely
1779 * Gigantic 'page allocators' such as memblock or cma will
1780 * return a set of pages with each page ref counted. We need
1781 * to turn this set of pages into a compound page with tail
1782 * page ref counts set to zero. Code such as speculative page
1783 * cache adding could take a ref on a 'to be' tail page.
1784 * We need to respect any increased ref count, and only set
1785 * the ref count to zero if count is currently 1. If count
1786 * is not 1, we return an error. An error return indicates
1787 * the set of pages can not be converted to a gigantic page.
1788 * The caller who allocated the pages should then discard the
1789 * pages using the appropriate free interface.
1791 * In the case of demote, the ref count will be zero.
1794 if (!page_ref_freeze(p, 1)) {
1795 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
1799 VM_BUG_ON_PAGE(page_count(p), p);
1801 set_compound_head(p, page);
1803 atomic_set(compound_mapcount_ptr(page), -1);
1804 atomic_set(compound_pincount_ptr(page), 0);
1808 /* undo tail page modifications made above */
1810 for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) {
1811 clear_compound_head(p);
1812 set_page_refcounted(p);
1814 /* need to clear PG_reserved on remaining tail pages */
1815 for (; j < nr_pages; j++, p = mem_map_next(p, page, j))
1816 __ClearPageReserved(p);
1817 set_compound_order(page, 0);
1819 page[1].compound_nr = 0;
1821 __ClearPageHead(page);
1825 static bool prep_compound_gigantic_page(struct page *page, unsigned int order)
1827 return __prep_compound_gigantic_page(page, order, false);
1830 static bool prep_compound_gigantic_page_for_demote(struct page *page,
1833 return __prep_compound_gigantic_page(page, order, true);
1837 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1838 * transparent huge pages. See the PageTransHuge() documentation for more
1841 int PageHuge(struct page *page)
1843 if (!PageCompound(page))
1846 page = compound_head(page);
1847 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1849 EXPORT_SYMBOL_GPL(PageHuge);
1852 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1853 * normal or transparent huge pages.
1855 int PageHeadHuge(struct page *page_head)
1857 if (!PageHead(page_head))
1860 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1862 EXPORT_SYMBOL_GPL(PageHeadHuge);
1865 * Find and lock address space (mapping) in write mode.
1867 * Upon entry, the page is locked which means that page_mapping() is
1868 * stable. Due to locking order, we can only trylock_write. If we can
1869 * not get the lock, simply return NULL to caller.
1871 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1873 struct address_space *mapping = page_mapping(hpage);
1878 if (i_mmap_trylock_write(mapping))
1884 pgoff_t hugetlb_basepage_index(struct page *page)
1886 struct page *page_head = compound_head(page);
1887 pgoff_t index = page_index(page_head);
1888 unsigned long compound_idx;
1890 if (compound_order(page_head) >= MAX_ORDER)
1891 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1893 compound_idx = page - page_head;
1895 return (index << compound_order(page_head)) + compound_idx;
1898 static struct page *alloc_buddy_huge_page(struct hstate *h,
1899 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1900 nodemask_t *node_alloc_noretry)
1902 int order = huge_page_order(h);
1904 bool alloc_try_hard = true;
1907 * By default we always try hard to allocate the page with
1908 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1909 * a loop (to adjust global huge page counts) and previous allocation
1910 * failed, do not continue to try hard on the same node. Use the
1911 * node_alloc_noretry bitmap to manage this state information.
1913 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1914 alloc_try_hard = false;
1915 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1917 gfp_mask |= __GFP_RETRY_MAYFAIL;
1918 if (nid == NUMA_NO_NODE)
1919 nid = numa_mem_id();
1920 page = __alloc_pages(gfp_mask, order, nid, nmask);
1922 __count_vm_event(HTLB_BUDDY_PGALLOC);
1924 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1927 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1928 * indicates an overall state change. Clear bit so that we resume
1929 * normal 'try hard' allocations.
1931 if (node_alloc_noretry && page && !alloc_try_hard)
1932 node_clear(nid, *node_alloc_noretry);
1935 * If we tried hard to get a page but failed, set bit so that
1936 * subsequent attempts will not try as hard until there is an
1937 * overall state change.
1939 if (node_alloc_noretry && !page && alloc_try_hard)
1940 node_set(nid, *node_alloc_noretry);
1946 * Common helper to allocate a fresh hugetlb page. All specific allocators
1947 * should use this function to get new hugetlb pages
1949 static struct page *alloc_fresh_huge_page(struct hstate *h,
1950 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1951 nodemask_t *node_alloc_noretry)
1957 if (hstate_is_gigantic(h))
1958 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1960 page = alloc_buddy_huge_page(h, gfp_mask,
1961 nid, nmask, node_alloc_noretry);
1965 if (hstate_is_gigantic(h)) {
1966 if (!prep_compound_gigantic_page(page, huge_page_order(h))) {
1968 * Rare failure to convert pages to compound page.
1969 * Free pages and try again - ONCE!
1971 free_gigantic_page(page, huge_page_order(h));
1979 prep_new_huge_page(h, page, page_to_nid(page));
1985 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1988 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1989 nodemask_t *node_alloc_noretry)
1993 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1995 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1996 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1997 node_alloc_noretry);
2005 put_page(page); /* free it into the hugepage allocator */
2011 * Remove huge page from pool from next node to free. Attempt to keep
2012 * persistent huge pages more or less balanced over allowed nodes.
2013 * This routine only 'removes' the hugetlb page. The caller must make
2014 * an additional call to free the page to low level allocators.
2015 * Called with hugetlb_lock locked.
2017 static struct page *remove_pool_huge_page(struct hstate *h,
2018 nodemask_t *nodes_allowed,
2022 struct page *page = NULL;
2024 lockdep_assert_held(&hugetlb_lock);
2025 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2027 * If we're returning unused surplus pages, only examine
2028 * nodes with surplus pages.
2030 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
2031 !list_empty(&h->hugepage_freelists[node])) {
2032 page = list_entry(h->hugepage_freelists[node].next,
2034 remove_hugetlb_page(h, page, acct_surplus);
2043 * Dissolve a given free hugepage into free buddy pages. This function does
2044 * nothing for in-use hugepages and non-hugepages.
2045 * This function returns values like below:
2047 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
2048 * when the system is under memory pressure and the feature of
2049 * freeing unused vmemmap pages associated with each hugetlb page
2051 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
2052 * (allocated or reserved.)
2053 * 0: successfully dissolved free hugepages or the page is not a
2054 * hugepage (considered as already dissolved)
2056 int dissolve_free_huge_page(struct page *page)
2061 /* Not to disrupt normal path by vainly holding hugetlb_lock */
2062 if (!PageHuge(page))
2065 spin_lock_irq(&hugetlb_lock);
2066 if (!PageHuge(page)) {
2071 if (!page_count(page)) {
2072 struct page *head = compound_head(page);
2073 struct hstate *h = page_hstate(head);
2074 if (h->free_huge_pages - h->resv_huge_pages == 0)
2078 * We should make sure that the page is already on the free list
2079 * when it is dissolved.
2081 if (unlikely(!HPageFreed(head))) {
2082 spin_unlock_irq(&hugetlb_lock);
2086 * Theoretically, we should return -EBUSY when we
2087 * encounter this race. In fact, we have a chance
2088 * to successfully dissolve the page if we do a
2089 * retry. Because the race window is quite small.
2090 * If we seize this opportunity, it is an optimization
2091 * for increasing the success rate of dissolving page.
2096 remove_hugetlb_page(h, head, false);
2097 h->max_huge_pages--;
2098 spin_unlock_irq(&hugetlb_lock);
2101 * Normally update_and_free_page will allocate required vmemmmap
2102 * before freeing the page. update_and_free_page will fail to
2103 * free the page if it can not allocate required vmemmap. We
2104 * need to adjust max_huge_pages if the page is not freed.
2105 * Attempt to allocate vmemmmap here so that we can take
2106 * appropriate action on failure.
2108 rc = alloc_huge_page_vmemmap(h, head);
2111 * Move PageHWPoison flag from head page to the raw
2112 * error page, which makes any subpages rather than
2113 * the error page reusable.
2115 if (PageHWPoison(head) && page != head) {
2116 SetPageHWPoison(page);
2117 ClearPageHWPoison(head);
2119 update_and_free_page(h, head, false);
2121 spin_lock_irq(&hugetlb_lock);
2122 add_hugetlb_page(h, head, false);
2123 h->max_huge_pages++;
2124 spin_unlock_irq(&hugetlb_lock);
2130 spin_unlock_irq(&hugetlb_lock);
2135 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2136 * make specified memory blocks removable from the system.
2137 * Note that this will dissolve a free gigantic hugepage completely, if any
2138 * part of it lies within the given range.
2139 * Also note that if dissolve_free_huge_page() returns with an error, all
2140 * free hugepages that were dissolved before that error are lost.
2142 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2148 if (!hugepages_supported())
2151 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
2152 page = pfn_to_page(pfn);
2153 rc = dissolve_free_huge_page(page);
2162 * Allocates a fresh surplus page from the page allocator.
2164 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
2165 int nid, nodemask_t *nmask, bool zero_ref)
2167 struct page *page = NULL;
2170 if (hstate_is_gigantic(h))
2173 spin_lock_irq(&hugetlb_lock);
2174 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2176 spin_unlock_irq(&hugetlb_lock);
2179 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2183 spin_lock_irq(&hugetlb_lock);
2185 * We could have raced with the pool size change.
2186 * Double check that and simply deallocate the new page
2187 * if we would end up overcommiting the surpluses. Abuse
2188 * temporary page to workaround the nasty free_huge_page
2191 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2192 SetHPageTemporary(page);
2193 spin_unlock_irq(&hugetlb_lock);
2200 * Caller requires a page with zero ref count.
2201 * We will drop ref count here. If someone else is holding
2202 * a ref, the page will be freed when they drop it. Abuse
2203 * temporary page flag to accomplish this.
2205 SetHPageTemporary(page);
2206 if (!put_page_testzero(page)) {
2208 * Unexpected inflated ref count on freshly allocated
2211 pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n");
2212 spin_unlock_irq(&hugetlb_lock);
2219 ClearHPageTemporary(page);
2222 h->surplus_huge_pages++;
2223 h->surplus_huge_pages_node[page_to_nid(page)]++;
2226 spin_unlock_irq(&hugetlb_lock);
2231 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
2232 int nid, nodemask_t *nmask)
2236 if (hstate_is_gigantic(h))
2239 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2244 * We do not account these pages as surplus because they are only
2245 * temporary and will be released properly on the last reference
2247 SetHPageTemporary(page);
2253 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2256 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
2257 struct vm_area_struct *vma, unsigned long addr)
2259 struct page *page = NULL;
2260 struct mempolicy *mpol;
2261 gfp_t gfp_mask = htlb_alloc_mask(h);
2263 nodemask_t *nodemask;
2265 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2266 if (mpol_is_preferred_many(mpol)) {
2267 gfp_t gfp = gfp_mask | __GFP_NOWARN;
2269 gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2270 page = alloc_surplus_huge_page(h, gfp, nid, nodemask, false);
2272 /* Fallback to all nodes if page==NULL */
2277 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask, false);
2278 mpol_cond_put(mpol);
2282 /* page migration callback function */
2283 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
2284 nodemask_t *nmask, gfp_t gfp_mask)
2286 spin_lock_irq(&hugetlb_lock);
2287 if (h->free_huge_pages - h->resv_huge_pages > 0) {
2290 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
2292 spin_unlock_irq(&hugetlb_lock);
2296 spin_unlock_irq(&hugetlb_lock);
2298 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2301 /* mempolicy aware migration callback */
2302 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2303 unsigned long address)
2305 struct mempolicy *mpol;
2306 nodemask_t *nodemask;
2311 gfp_mask = htlb_alloc_mask(h);
2312 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2313 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
2314 mpol_cond_put(mpol);
2320 * Increase the hugetlb pool such that it can accommodate a reservation
2323 static int gather_surplus_pages(struct hstate *h, long delta)
2324 __must_hold(&hugetlb_lock)
2326 struct list_head surplus_list;
2327 struct page *page, *tmp;
2330 long needed, allocated;
2331 bool alloc_ok = true;
2333 lockdep_assert_held(&hugetlb_lock);
2334 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2336 h->resv_huge_pages += delta;
2341 INIT_LIST_HEAD(&surplus_list);
2345 spin_unlock_irq(&hugetlb_lock);
2346 for (i = 0; i < needed; i++) {
2347 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2348 NUMA_NO_NODE, NULL, true);
2353 list_add(&page->lru, &surplus_list);
2359 * After retaking hugetlb_lock, we need to recalculate 'needed'
2360 * because either resv_huge_pages or free_huge_pages may have changed.
2362 spin_lock_irq(&hugetlb_lock);
2363 needed = (h->resv_huge_pages + delta) -
2364 (h->free_huge_pages + allocated);
2369 * We were not able to allocate enough pages to
2370 * satisfy the entire reservation so we free what
2371 * we've allocated so far.
2376 * The surplus_list now contains _at_least_ the number of extra pages
2377 * needed to accommodate the reservation. Add the appropriate number
2378 * of pages to the hugetlb pool and free the extras back to the buddy
2379 * allocator. Commit the entire reservation here to prevent another
2380 * process from stealing the pages as they are added to the pool but
2381 * before they are reserved.
2383 needed += allocated;
2384 h->resv_huge_pages += delta;
2387 /* Free the needed pages to the hugetlb pool */
2388 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2391 /* Add the page to the hugetlb allocator */
2392 enqueue_huge_page(h, page);
2395 spin_unlock_irq(&hugetlb_lock);
2398 * Free unnecessary surplus pages to the buddy allocator.
2399 * Pages have no ref count, call free_huge_page directly.
2401 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2402 free_huge_page(page);
2403 spin_lock_irq(&hugetlb_lock);
2409 * This routine has two main purposes:
2410 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2411 * in unused_resv_pages. This corresponds to the prior adjustments made
2412 * to the associated reservation map.
2413 * 2) Free any unused surplus pages that may have been allocated to satisfy
2414 * the reservation. As many as unused_resv_pages may be freed.
2416 static void return_unused_surplus_pages(struct hstate *h,
2417 unsigned long unused_resv_pages)
2419 unsigned long nr_pages;
2421 LIST_HEAD(page_list);
2423 lockdep_assert_held(&hugetlb_lock);
2424 /* Uncommit the reservation */
2425 h->resv_huge_pages -= unused_resv_pages;
2427 /* Cannot return gigantic pages currently */
2428 if (hstate_is_gigantic(h))
2432 * Part (or even all) of the reservation could have been backed
2433 * by pre-allocated pages. Only free surplus pages.
2435 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2438 * We want to release as many surplus pages as possible, spread
2439 * evenly across all nodes with memory. Iterate across these nodes
2440 * until we can no longer free unreserved surplus pages. This occurs
2441 * when the nodes with surplus pages have no free pages.
2442 * remove_pool_huge_page() will balance the freed pages across the
2443 * on-line nodes with memory and will handle the hstate accounting.
2445 while (nr_pages--) {
2446 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2450 list_add(&page->lru, &page_list);
2454 spin_unlock_irq(&hugetlb_lock);
2455 update_and_free_pages_bulk(h, &page_list);
2456 spin_lock_irq(&hugetlb_lock);
2461 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2462 * are used by the huge page allocation routines to manage reservations.
2464 * vma_needs_reservation is called to determine if the huge page at addr
2465 * within the vma has an associated reservation. If a reservation is
2466 * needed, the value 1 is returned. The caller is then responsible for
2467 * managing the global reservation and subpool usage counts. After
2468 * the huge page has been allocated, vma_commit_reservation is called
2469 * to add the page to the reservation map. If the page allocation fails,
2470 * the reservation must be ended instead of committed. vma_end_reservation
2471 * is called in such cases.
2473 * In the normal case, vma_commit_reservation returns the same value
2474 * as the preceding vma_needs_reservation call. The only time this
2475 * is not the case is if a reserve map was changed between calls. It
2476 * is the responsibility of the caller to notice the difference and
2477 * take appropriate action.
2479 * vma_add_reservation is used in error paths where a reservation must
2480 * be restored when a newly allocated huge page must be freed. It is
2481 * to be called after calling vma_needs_reservation to determine if a
2482 * reservation exists.
2484 * vma_del_reservation is used in error paths where an entry in the reserve
2485 * map was created during huge page allocation and must be removed. It is to
2486 * be called after calling vma_needs_reservation to determine if a reservation
2489 enum vma_resv_mode {
2496 static long __vma_reservation_common(struct hstate *h,
2497 struct vm_area_struct *vma, unsigned long addr,
2498 enum vma_resv_mode mode)
2500 struct resv_map *resv;
2503 long dummy_out_regions_needed;
2505 resv = vma_resv_map(vma);
2509 idx = vma_hugecache_offset(h, vma, addr);
2511 case VMA_NEEDS_RESV:
2512 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2513 /* We assume that vma_reservation_* routines always operate on
2514 * 1 page, and that adding to resv map a 1 page entry can only
2515 * ever require 1 region.
2517 VM_BUG_ON(dummy_out_regions_needed != 1);
2519 case VMA_COMMIT_RESV:
2520 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2521 /* region_add calls of range 1 should never fail. */
2525 region_abort(resv, idx, idx + 1, 1);
2529 if (vma->vm_flags & VM_MAYSHARE) {
2530 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2531 /* region_add calls of range 1 should never fail. */
2534 region_abort(resv, idx, idx + 1, 1);
2535 ret = region_del(resv, idx, idx + 1);
2539 if (vma->vm_flags & VM_MAYSHARE) {
2540 region_abort(resv, idx, idx + 1, 1);
2541 ret = region_del(resv, idx, idx + 1);
2543 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2544 /* region_add calls of range 1 should never fail. */
2552 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2555 * We know private mapping must have HPAGE_RESV_OWNER set.
2557 * In most cases, reserves always exist for private mappings.
2558 * However, a file associated with mapping could have been
2559 * hole punched or truncated after reserves were consumed.
2560 * As subsequent fault on such a range will not use reserves.
2561 * Subtle - The reserve map for private mappings has the
2562 * opposite meaning than that of shared mappings. If NO
2563 * entry is in the reserve map, it means a reservation exists.
2564 * If an entry exists in the reserve map, it means the
2565 * reservation has already been consumed. As a result, the
2566 * return value of this routine is the opposite of the
2567 * value returned from reserve map manipulation routines above.
2576 static long vma_needs_reservation(struct hstate *h,
2577 struct vm_area_struct *vma, unsigned long addr)
2579 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2582 static long vma_commit_reservation(struct hstate *h,
2583 struct vm_area_struct *vma, unsigned long addr)
2585 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2588 static void vma_end_reservation(struct hstate *h,
2589 struct vm_area_struct *vma, unsigned long addr)
2591 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2594 static long vma_add_reservation(struct hstate *h,
2595 struct vm_area_struct *vma, unsigned long addr)
2597 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2600 static long vma_del_reservation(struct hstate *h,
2601 struct vm_area_struct *vma, unsigned long addr)
2603 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2607 * This routine is called to restore reservation information on error paths.
2608 * It should ONLY be called for pages allocated via alloc_huge_page(), and
2609 * the hugetlb mutex should remain held when calling this routine.
2611 * It handles two specific cases:
2612 * 1) A reservation was in place and the page consumed the reservation.
2613 * HPageRestoreReserve is set in the page.
2614 * 2) No reservation was in place for the page, so HPageRestoreReserve is
2615 * not set. However, alloc_huge_page always updates the reserve map.
2617 * In case 1, free_huge_page later in the error path will increment the
2618 * global reserve count. But, free_huge_page does not have enough context
2619 * to adjust the reservation map. This case deals primarily with private
2620 * mappings. Adjust the reserve map here to be consistent with global
2621 * reserve count adjustments to be made by free_huge_page. Make sure the
2622 * reserve map indicates there is a reservation present.
2624 * In case 2, simply undo reserve map modifications done by alloc_huge_page.
2626 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2627 unsigned long address, struct page *page)
2629 long rc = vma_needs_reservation(h, vma, address);
2631 if (HPageRestoreReserve(page)) {
2632 if (unlikely(rc < 0))
2634 * Rare out of memory condition in reserve map
2635 * manipulation. Clear HPageRestoreReserve so that
2636 * global reserve count will not be incremented
2637 * by free_huge_page. This will make it appear
2638 * as though the reservation for this page was
2639 * consumed. This may prevent the task from
2640 * faulting in the page at a later time. This
2641 * is better than inconsistent global huge page
2642 * accounting of reserve counts.
2644 ClearHPageRestoreReserve(page);
2646 (void)vma_add_reservation(h, vma, address);
2648 vma_end_reservation(h, vma, address);
2652 * This indicates there is an entry in the reserve map
2653 * not added by alloc_huge_page. We know it was added
2654 * before the alloc_huge_page call, otherwise
2655 * HPageRestoreReserve would be set on the page.
2656 * Remove the entry so that a subsequent allocation
2657 * does not consume a reservation.
2659 rc = vma_del_reservation(h, vma, address);
2662 * VERY rare out of memory condition. Since
2663 * we can not delete the entry, set
2664 * HPageRestoreReserve so that the reserve
2665 * count will be incremented when the page
2666 * is freed. This reserve will be consumed
2667 * on a subsequent allocation.
2669 SetHPageRestoreReserve(page);
2670 } else if (rc < 0) {
2672 * Rare out of memory condition from
2673 * vma_needs_reservation call. Memory allocation is
2674 * only attempted if a new entry is needed. Therefore,
2675 * this implies there is not an entry in the
2678 * For shared mappings, no entry in the map indicates
2679 * no reservation. We are done.
2681 if (!(vma->vm_flags & VM_MAYSHARE))
2683 * For private mappings, no entry indicates
2684 * a reservation is present. Since we can
2685 * not add an entry, set SetHPageRestoreReserve
2686 * on the page so reserve count will be
2687 * incremented when freed. This reserve will
2688 * be consumed on a subsequent allocation.
2690 SetHPageRestoreReserve(page);
2693 * No reservation present, do nothing
2695 vma_end_reservation(h, vma, address);
2700 * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one
2701 * @h: struct hstate old page belongs to
2702 * @old_page: Old page to dissolve
2703 * @list: List to isolate the page in case we need to
2704 * Returns 0 on success, otherwise negated error.
2706 static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page,
2707 struct list_head *list)
2709 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2710 int nid = page_to_nid(old_page);
2711 bool alloc_retry = false;
2712 struct page *new_page;
2716 * Before dissolving the page, we need to allocate a new one for the
2717 * pool to remain stable. Here, we allocate the page and 'prep' it
2718 * by doing everything but actually updating counters and adding to
2719 * the pool. This simplifies and let us do most of the processing
2723 new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL);
2727 * If all goes well, this page will be directly added to the free
2728 * list in the pool. For this the ref count needs to be zero.
2729 * Attempt to drop now, and retry once if needed. It is VERY
2730 * unlikely there is another ref on the page.
2732 * If someone else has a reference to the page, it will be freed
2733 * when they drop their ref. Abuse temporary page flag to accomplish
2734 * this. Retry once if there is an inflated ref count.
2736 SetHPageTemporary(new_page);
2737 if (!put_page_testzero(new_page)) {
2744 ClearHPageTemporary(new_page);
2746 __prep_new_huge_page(h, new_page);
2749 spin_lock_irq(&hugetlb_lock);
2750 if (!PageHuge(old_page)) {
2752 * Freed from under us. Drop new_page too.
2755 } else if (page_count(old_page)) {
2757 * Someone has grabbed the page, try to isolate it here.
2758 * Fail with -EBUSY if not possible.
2760 spin_unlock_irq(&hugetlb_lock);
2761 if (!isolate_huge_page(old_page, list))
2763 spin_lock_irq(&hugetlb_lock);
2765 } else if (!HPageFreed(old_page)) {
2767 * Page's refcount is 0 but it has not been enqueued in the
2768 * freelist yet. Race window is small, so we can succeed here if
2771 spin_unlock_irq(&hugetlb_lock);
2776 * Ok, old_page is still a genuine free hugepage. Remove it from
2777 * the freelist and decrease the counters. These will be
2778 * incremented again when calling __prep_account_new_huge_page()
2779 * and enqueue_huge_page() for new_page. The counters will remain
2780 * stable since this happens under the lock.
2782 remove_hugetlb_page(h, old_page, false);
2785 * Ref count on new page is already zero as it was dropped
2786 * earlier. It can be directly added to the pool free list.
2788 __prep_account_new_huge_page(h, nid);
2789 enqueue_huge_page(h, new_page);
2792 * Pages have been replaced, we can safely free the old one.
2794 spin_unlock_irq(&hugetlb_lock);
2795 update_and_free_page(h, old_page, false);
2801 spin_unlock_irq(&hugetlb_lock);
2802 /* Page has a zero ref count, but needs a ref to be freed */
2803 set_page_refcounted(new_page);
2804 update_and_free_page(h, new_page, false);
2809 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
2816 * The page might have been dissolved from under our feet, so make sure
2817 * to carefully check the state under the lock.
2818 * Return success when racing as if we dissolved the page ourselves.
2820 spin_lock_irq(&hugetlb_lock);
2821 if (PageHuge(page)) {
2822 head = compound_head(page);
2823 h = page_hstate(head);
2825 spin_unlock_irq(&hugetlb_lock);
2828 spin_unlock_irq(&hugetlb_lock);
2831 * Fence off gigantic pages as there is a cyclic dependency between
2832 * alloc_contig_range and them. Return -ENOMEM as this has the effect
2833 * of bailing out right away without further retrying.
2835 if (hstate_is_gigantic(h))
2838 if (page_count(head) && isolate_huge_page(head, list))
2840 else if (!page_count(head))
2841 ret = alloc_and_dissolve_huge_page(h, head, list);
2846 struct page *alloc_huge_page(struct vm_area_struct *vma,
2847 unsigned long addr, int avoid_reserve)
2849 struct hugepage_subpool *spool = subpool_vma(vma);
2850 struct hstate *h = hstate_vma(vma);
2852 long map_chg, map_commit;
2855 struct hugetlb_cgroup *h_cg;
2856 bool deferred_reserve;
2858 idx = hstate_index(h);
2860 * Examine the region/reserve map to determine if the process
2861 * has a reservation for the page to be allocated. A return
2862 * code of zero indicates a reservation exists (no change).
2864 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2866 return ERR_PTR(-ENOMEM);
2869 * Processes that did not create the mapping will have no
2870 * reserves as indicated by the region/reserve map. Check
2871 * that the allocation will not exceed the subpool limit.
2872 * Allocations for MAP_NORESERVE mappings also need to be
2873 * checked against any subpool limit.
2875 if (map_chg || avoid_reserve) {
2876 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2878 vma_end_reservation(h, vma, addr);
2879 return ERR_PTR(-ENOSPC);
2883 * Even though there was no reservation in the region/reserve
2884 * map, there could be reservations associated with the
2885 * subpool that can be used. This would be indicated if the
2886 * return value of hugepage_subpool_get_pages() is zero.
2887 * However, if avoid_reserve is specified we still avoid even
2888 * the subpool reservations.
2894 /* If this allocation is not consuming a reservation, charge it now.
2896 deferred_reserve = map_chg || avoid_reserve;
2897 if (deferred_reserve) {
2898 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2899 idx, pages_per_huge_page(h), &h_cg);
2901 goto out_subpool_put;
2904 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2906 goto out_uncharge_cgroup_reservation;
2908 spin_lock_irq(&hugetlb_lock);
2910 * glb_chg is passed to indicate whether or not a page must be taken
2911 * from the global free pool (global change). gbl_chg == 0 indicates
2912 * a reservation exists for the allocation.
2914 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2916 spin_unlock_irq(&hugetlb_lock);
2917 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2919 goto out_uncharge_cgroup;
2920 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2921 SetHPageRestoreReserve(page);
2922 h->resv_huge_pages--;
2924 spin_lock_irq(&hugetlb_lock);
2925 list_add(&page->lru, &h->hugepage_activelist);
2928 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2929 /* If allocation is not consuming a reservation, also store the
2930 * hugetlb_cgroup pointer on the page.
2932 if (deferred_reserve) {
2933 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2937 spin_unlock_irq(&hugetlb_lock);
2939 hugetlb_set_page_subpool(page, spool);
2941 map_commit = vma_commit_reservation(h, vma, addr);
2942 if (unlikely(map_chg > map_commit)) {
2944 * The page was added to the reservation map between
2945 * vma_needs_reservation and vma_commit_reservation.
2946 * This indicates a race with hugetlb_reserve_pages.
2947 * Adjust for the subpool count incremented above AND
2948 * in hugetlb_reserve_pages for the same page. Also,
2949 * the reservation count added in hugetlb_reserve_pages
2950 * no longer applies.
2954 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2955 hugetlb_acct_memory(h, -rsv_adjust);
2956 if (deferred_reserve)
2957 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2958 pages_per_huge_page(h), page);
2962 out_uncharge_cgroup:
2963 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2964 out_uncharge_cgroup_reservation:
2965 if (deferred_reserve)
2966 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2969 if (map_chg || avoid_reserve)
2970 hugepage_subpool_put_pages(spool, 1);
2971 vma_end_reservation(h, vma, addr);
2972 return ERR_PTR(-ENOSPC);
2975 int alloc_bootmem_huge_page(struct hstate *h, int nid)
2976 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2977 int __alloc_bootmem_huge_page(struct hstate *h, int nid)
2979 struct huge_bootmem_page *m = NULL; /* initialize for clang */
2982 if (nid != NUMA_NO_NODE && nid >= nr_online_nodes)
2984 /* do node specific alloc */
2985 if (nid != NUMA_NO_NODE) {
2986 m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h),
2987 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
2992 /* allocate from next node when distributing huge pages */
2993 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2994 m = memblock_alloc_try_nid_raw(
2995 huge_page_size(h), huge_page_size(h),
2996 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2998 * Use the beginning of the huge page to store the
2999 * huge_bootmem_page struct (until gather_bootmem
3000 * puts them into the mem_map).
3008 /* Put them into a private list first because mem_map is not up yet */
3009 INIT_LIST_HEAD(&m->list);
3010 list_add(&m->list, &huge_boot_pages);
3016 * Put bootmem huge pages into the standard lists after mem_map is up.
3017 * Note: This only applies to gigantic (order > MAX_ORDER) pages.
3019 static void __init gather_bootmem_prealloc(void)
3021 struct huge_bootmem_page *m;
3023 list_for_each_entry(m, &huge_boot_pages, list) {
3024 struct page *page = virt_to_page(m);
3025 struct hstate *h = m->hstate;
3027 VM_BUG_ON(!hstate_is_gigantic(h));
3028 WARN_ON(page_count(page) != 1);
3029 if (prep_compound_gigantic_page(page, huge_page_order(h))) {
3030 WARN_ON(PageReserved(page));
3031 prep_new_huge_page(h, page, page_to_nid(page));
3032 put_page(page); /* add to the hugepage allocator */
3034 /* VERY unlikely inflated ref count on a tail page */
3035 free_gigantic_page(page, huge_page_order(h));
3039 * We need to restore the 'stolen' pages to totalram_pages
3040 * in order to fix confusing memory reports from free(1) and
3041 * other side-effects, like CommitLimit going negative.
3043 adjust_managed_page_count(page, pages_per_huge_page(h));
3047 static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)
3052 for (i = 0; i < h->max_huge_pages_node[nid]; ++i) {
3053 if (hstate_is_gigantic(h)) {
3054 if (!alloc_bootmem_huge_page(h, nid))
3058 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3060 page = alloc_fresh_huge_page(h, gfp_mask, nid,
3061 &node_states[N_MEMORY], NULL);
3064 put_page(page); /* free it into the hugepage allocator */
3068 if (i == h->max_huge_pages_node[nid])
3071 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3072 pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n",
3073 h->max_huge_pages_node[nid], buf, nid, i);
3074 h->max_huge_pages -= (h->max_huge_pages_node[nid] - i);
3075 h->max_huge_pages_node[nid] = i;
3078 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
3081 nodemask_t *node_alloc_noretry;
3082 bool node_specific_alloc = false;
3084 /* skip gigantic hugepages allocation if hugetlb_cma enabled */
3085 if (hstate_is_gigantic(h) && hugetlb_cma_size) {
3086 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
3090 /* do node specific alloc */
3091 for (i = 0; i < nr_online_nodes; i++) {
3092 if (h->max_huge_pages_node[i] > 0) {
3093 hugetlb_hstate_alloc_pages_onenode(h, i);
3094 node_specific_alloc = true;
3098 if (node_specific_alloc)
3101 /* below will do all node balanced alloc */
3102 if (!hstate_is_gigantic(h)) {
3104 * Bit mask controlling how hard we retry per-node allocations.
3105 * Ignore errors as lower level routines can deal with
3106 * node_alloc_noretry == NULL. If this kmalloc fails at boot
3107 * time, we are likely in bigger trouble.
3109 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
3112 /* allocations done at boot time */
3113 node_alloc_noretry = NULL;
3116 /* bit mask controlling how hard we retry per-node allocations */
3117 if (node_alloc_noretry)
3118 nodes_clear(*node_alloc_noretry);
3120 for (i = 0; i < h->max_huge_pages; ++i) {
3121 if (hstate_is_gigantic(h)) {
3122 if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))
3124 } else if (!alloc_pool_huge_page(h,
3125 &node_states[N_MEMORY],
3126 node_alloc_noretry))
3130 if (i < h->max_huge_pages) {
3133 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3134 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
3135 h->max_huge_pages, buf, i);
3136 h->max_huge_pages = i;
3138 kfree(node_alloc_noretry);
3141 static void __init hugetlb_init_hstates(void)
3143 struct hstate *h, *h2;
3145 for_each_hstate(h) {
3146 if (minimum_order > huge_page_order(h))
3147 minimum_order = huge_page_order(h);
3149 /* oversize hugepages were init'ed in early boot */
3150 if (!hstate_is_gigantic(h))
3151 hugetlb_hstate_alloc_pages(h);
3154 * Set demote order for each hstate. Note that
3155 * h->demote_order is initially 0.
3156 * - We can not demote gigantic pages if runtime freeing
3157 * is not supported, so skip this.
3158 * - If CMA allocation is possible, we can not demote
3159 * HUGETLB_PAGE_ORDER or smaller size pages.
3161 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3163 if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER)
3165 for_each_hstate(h2) {
3168 if (h2->order < h->order &&
3169 h2->order > h->demote_order)
3170 h->demote_order = h2->order;
3173 VM_BUG_ON(minimum_order == UINT_MAX);
3176 static void __init report_hugepages(void)
3180 for_each_hstate(h) {
3183 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3184 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
3185 buf, h->free_huge_pages);
3189 #ifdef CONFIG_HIGHMEM
3190 static void try_to_free_low(struct hstate *h, unsigned long count,
3191 nodemask_t *nodes_allowed)
3194 LIST_HEAD(page_list);
3196 lockdep_assert_held(&hugetlb_lock);
3197 if (hstate_is_gigantic(h))
3201 * Collect pages to be freed on a list, and free after dropping lock
3203 for_each_node_mask(i, *nodes_allowed) {
3204 struct page *page, *next;
3205 struct list_head *freel = &h->hugepage_freelists[i];
3206 list_for_each_entry_safe(page, next, freel, lru) {
3207 if (count >= h->nr_huge_pages)
3209 if (PageHighMem(page))
3211 remove_hugetlb_page(h, page, false);
3212 list_add(&page->lru, &page_list);
3217 spin_unlock_irq(&hugetlb_lock);
3218 update_and_free_pages_bulk(h, &page_list);
3219 spin_lock_irq(&hugetlb_lock);
3222 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3223 nodemask_t *nodes_allowed)
3229 * Increment or decrement surplus_huge_pages. Keep node-specific counters
3230 * balanced by operating on them in a round-robin fashion.
3231 * Returns 1 if an adjustment was made.
3233 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3238 lockdep_assert_held(&hugetlb_lock);
3239 VM_BUG_ON(delta != -1 && delta != 1);
3242 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3243 if (h->surplus_huge_pages_node[node])
3247 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3248 if (h->surplus_huge_pages_node[node] <
3249 h->nr_huge_pages_node[node])
3256 h->surplus_huge_pages += delta;
3257 h->surplus_huge_pages_node[node] += delta;
3261 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3262 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3263 nodemask_t *nodes_allowed)
3265 unsigned long min_count, ret;
3267 LIST_HEAD(page_list);
3268 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3271 * Bit mask controlling how hard we retry per-node allocations.
3272 * If we can not allocate the bit mask, do not attempt to allocate
3273 * the requested huge pages.
3275 if (node_alloc_noretry)
3276 nodes_clear(*node_alloc_noretry);
3281 * resize_lock mutex prevents concurrent adjustments to number of
3282 * pages in hstate via the proc/sysfs interfaces.
3284 mutex_lock(&h->resize_lock);
3285 flush_free_hpage_work(h);
3286 spin_lock_irq(&hugetlb_lock);
3289 * Check for a node specific request.
3290 * Changing node specific huge page count may require a corresponding
3291 * change to the global count. In any case, the passed node mask
3292 * (nodes_allowed) will restrict alloc/free to the specified node.
3294 if (nid != NUMA_NO_NODE) {
3295 unsigned long old_count = count;
3297 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3299 * User may have specified a large count value which caused the
3300 * above calculation to overflow. In this case, they wanted
3301 * to allocate as many huge pages as possible. Set count to
3302 * largest possible value to align with their intention.
3304 if (count < old_count)
3309 * Gigantic pages runtime allocation depend on the capability for large
3310 * page range allocation.
3311 * If the system does not provide this feature, return an error when
3312 * the user tries to allocate gigantic pages but let the user free the
3313 * boottime allocated gigantic pages.
3315 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3316 if (count > persistent_huge_pages(h)) {
3317 spin_unlock_irq(&hugetlb_lock);
3318 mutex_unlock(&h->resize_lock);
3319 NODEMASK_FREE(node_alloc_noretry);
3322 /* Fall through to decrease pool */
3326 * Increase the pool size
3327 * First take pages out of surplus state. Then make up the
3328 * remaining difference by allocating fresh huge pages.
3330 * We might race with alloc_surplus_huge_page() here and be unable
3331 * to convert a surplus huge page to a normal huge page. That is
3332 * not critical, though, it just means the overall size of the
3333 * pool might be one hugepage larger than it needs to be, but
3334 * within all the constraints specified by the sysctls.
3336 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3337 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3341 while (count > persistent_huge_pages(h)) {
3343 * If this allocation races such that we no longer need the
3344 * page, free_huge_page will handle it by freeing the page
3345 * and reducing the surplus.
3347 spin_unlock_irq(&hugetlb_lock);
3349 /* yield cpu to avoid soft lockup */
3352 ret = alloc_pool_huge_page(h, nodes_allowed,
3353 node_alloc_noretry);
3354 spin_lock_irq(&hugetlb_lock);
3358 /* Bail for signals. Probably ctrl-c from user */
3359 if (signal_pending(current))
3364 * Decrease the pool size
3365 * First return free pages to the buddy allocator (being careful
3366 * to keep enough around to satisfy reservations). Then place
3367 * pages into surplus state as needed so the pool will shrink
3368 * to the desired size as pages become free.
3370 * By placing pages into the surplus state independent of the
3371 * overcommit value, we are allowing the surplus pool size to
3372 * exceed overcommit. There are few sane options here. Since
3373 * alloc_surplus_huge_page() is checking the global counter,
3374 * though, we'll note that we're not allowed to exceed surplus
3375 * and won't grow the pool anywhere else. Not until one of the
3376 * sysctls are changed, or the surplus pages go out of use.
3378 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3379 min_count = max(count, min_count);
3380 try_to_free_low(h, min_count, nodes_allowed);
3383 * Collect pages to be removed on list without dropping lock
3385 while (min_count < persistent_huge_pages(h)) {
3386 page = remove_pool_huge_page(h, nodes_allowed, 0);
3390 list_add(&page->lru, &page_list);
3392 /* free the pages after dropping lock */
3393 spin_unlock_irq(&hugetlb_lock);
3394 update_and_free_pages_bulk(h, &page_list);
3395 flush_free_hpage_work(h);
3396 spin_lock_irq(&hugetlb_lock);
3398 while (count < persistent_huge_pages(h)) {
3399 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3403 h->max_huge_pages = persistent_huge_pages(h);
3404 spin_unlock_irq(&hugetlb_lock);
3405 mutex_unlock(&h->resize_lock);
3407 NODEMASK_FREE(node_alloc_noretry);
3412 static int demote_free_huge_page(struct hstate *h, struct page *page)
3414 int i, nid = page_to_nid(page);
3415 struct hstate *target_hstate;
3418 target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order);
3420 remove_hugetlb_page_for_demote(h, page, false);
3421 spin_unlock_irq(&hugetlb_lock);
3423 rc = alloc_huge_page_vmemmap(h, page);
3425 /* Allocation of vmemmmap failed, we can not demote page */
3426 spin_lock_irq(&hugetlb_lock);
3427 set_page_refcounted(page);
3428 add_hugetlb_page(h, page, false);
3433 * Use destroy_compound_hugetlb_page_for_demote for all huge page
3434 * sizes as it will not ref count pages.
3436 destroy_compound_hugetlb_page_for_demote(page, huge_page_order(h));
3439 * Taking target hstate mutex synchronizes with set_max_huge_pages.
3440 * Without the mutex, pages added to target hstate could be marked
3443 * Note that we already hold h->resize_lock. To prevent deadlock,
3444 * use the convention of always taking larger size hstate mutex first.
3446 mutex_lock(&target_hstate->resize_lock);
3447 for (i = 0; i < pages_per_huge_page(h);
3448 i += pages_per_huge_page(target_hstate)) {
3449 if (hstate_is_gigantic(target_hstate))
3450 prep_compound_gigantic_page_for_demote(page + i,
3451 target_hstate->order);
3453 prep_compound_page(page + i, target_hstate->order);
3454 set_page_private(page + i, 0);
3455 set_page_refcounted(page + i);
3456 prep_new_huge_page(target_hstate, page + i, nid);
3459 mutex_unlock(&target_hstate->resize_lock);
3461 spin_lock_irq(&hugetlb_lock);
3464 * Not absolutely necessary, but for consistency update max_huge_pages
3465 * based on pool changes for the demoted page.
3467 h->max_huge_pages--;
3468 target_hstate->max_huge_pages += pages_per_huge_page(h);
3473 static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
3474 __must_hold(&hugetlb_lock)
3479 lockdep_assert_held(&hugetlb_lock);
3481 /* We should never get here if no demote order */
3482 if (!h->demote_order) {
3483 pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n");
3484 return -EINVAL; /* internal error */
3487 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3488 list_for_each_entry(page, &h->hugepage_freelists[node], lru) {
3489 if (PageHWPoison(page))
3492 return demote_free_huge_page(h, page);
3497 * Only way to get here is if all pages on free lists are poisoned.
3498 * Return -EBUSY so that caller will not retry.
3503 #define HSTATE_ATTR_RO(_name) \
3504 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3506 #define HSTATE_ATTR_WO(_name) \
3507 static struct kobj_attribute _name##_attr = __ATTR_WO(_name)
3509 #define HSTATE_ATTR(_name) \
3510 static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
3512 static struct kobject *hugepages_kobj;
3513 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3515 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3517 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3521 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3522 if (hstate_kobjs[i] == kobj) {
3524 *nidp = NUMA_NO_NODE;
3528 return kobj_to_node_hstate(kobj, nidp);
3531 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3532 struct kobj_attribute *attr, char *buf)
3535 unsigned long nr_huge_pages;
3538 h = kobj_to_hstate(kobj, &nid);
3539 if (nid == NUMA_NO_NODE)
3540 nr_huge_pages = h->nr_huge_pages;
3542 nr_huge_pages = h->nr_huge_pages_node[nid];
3544 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3547 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3548 struct hstate *h, int nid,
3549 unsigned long count, size_t len)
3552 nodemask_t nodes_allowed, *n_mask;
3554 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3557 if (nid == NUMA_NO_NODE) {
3559 * global hstate attribute
3561 if (!(obey_mempolicy &&
3562 init_nodemask_of_mempolicy(&nodes_allowed)))
3563 n_mask = &node_states[N_MEMORY];
3565 n_mask = &nodes_allowed;
3568 * Node specific request. count adjustment happens in
3569 * set_max_huge_pages() after acquiring hugetlb_lock.
3571 init_nodemask_of_node(&nodes_allowed, nid);
3572 n_mask = &nodes_allowed;
3575 err = set_max_huge_pages(h, count, nid, n_mask);
3577 return err ? err : len;
3580 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3581 struct kobject *kobj, const char *buf,
3585 unsigned long count;
3589 err = kstrtoul(buf, 10, &count);
3593 h = kobj_to_hstate(kobj, &nid);
3594 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3597 static ssize_t nr_hugepages_show(struct kobject *kobj,
3598 struct kobj_attribute *attr, char *buf)
3600 return nr_hugepages_show_common(kobj, attr, buf);
3603 static ssize_t nr_hugepages_store(struct kobject *kobj,
3604 struct kobj_attribute *attr, const char *buf, size_t len)
3606 return nr_hugepages_store_common(false, kobj, buf, len);
3608 HSTATE_ATTR(nr_hugepages);
3613 * hstate attribute for optionally mempolicy-based constraint on persistent
3614 * huge page alloc/free.
3616 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3617 struct kobj_attribute *attr,
3620 return nr_hugepages_show_common(kobj, attr, buf);
3623 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3624 struct kobj_attribute *attr, const char *buf, size_t len)
3626 return nr_hugepages_store_common(true, kobj, buf, len);
3628 HSTATE_ATTR(nr_hugepages_mempolicy);
3632 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3633 struct kobj_attribute *attr, char *buf)
3635 struct hstate *h = kobj_to_hstate(kobj, NULL);
3636 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3639 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3640 struct kobj_attribute *attr, const char *buf, size_t count)
3643 unsigned long input;
3644 struct hstate *h = kobj_to_hstate(kobj, NULL);
3646 if (hstate_is_gigantic(h))
3649 err = kstrtoul(buf, 10, &input);
3653 spin_lock_irq(&hugetlb_lock);
3654 h->nr_overcommit_huge_pages = input;
3655 spin_unlock_irq(&hugetlb_lock);
3659 HSTATE_ATTR(nr_overcommit_hugepages);
3661 static ssize_t free_hugepages_show(struct kobject *kobj,
3662 struct kobj_attribute *attr, char *buf)
3665 unsigned long free_huge_pages;
3668 h = kobj_to_hstate(kobj, &nid);
3669 if (nid == NUMA_NO_NODE)
3670 free_huge_pages = h->free_huge_pages;
3672 free_huge_pages = h->free_huge_pages_node[nid];
3674 return sysfs_emit(buf, "%lu\n", free_huge_pages);
3676 HSTATE_ATTR_RO(free_hugepages);
3678 static ssize_t resv_hugepages_show(struct kobject *kobj,
3679 struct kobj_attribute *attr, char *buf)
3681 struct hstate *h = kobj_to_hstate(kobj, NULL);
3682 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3684 HSTATE_ATTR_RO(resv_hugepages);
3686 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3687 struct kobj_attribute *attr, char *buf)
3690 unsigned long surplus_huge_pages;
3693 h = kobj_to_hstate(kobj, &nid);
3694 if (nid == NUMA_NO_NODE)
3695 surplus_huge_pages = h->surplus_huge_pages;
3697 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3699 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3701 HSTATE_ATTR_RO(surplus_hugepages);
3703 static ssize_t demote_store(struct kobject *kobj,
3704 struct kobj_attribute *attr, const char *buf, size_t len)
3706 unsigned long nr_demote;
3707 unsigned long nr_available;
3708 nodemask_t nodes_allowed, *n_mask;
3713 err = kstrtoul(buf, 10, &nr_demote);
3716 h = kobj_to_hstate(kobj, &nid);
3718 if (nid != NUMA_NO_NODE) {
3719 init_nodemask_of_node(&nodes_allowed, nid);
3720 n_mask = &nodes_allowed;
3722 n_mask = &node_states[N_MEMORY];
3725 /* Synchronize with other sysfs operations modifying huge pages */
3726 mutex_lock(&h->resize_lock);
3727 spin_lock_irq(&hugetlb_lock);
3731 * Check for available pages to demote each time thorough the
3732 * loop as demote_pool_huge_page will drop hugetlb_lock.
3734 if (nid != NUMA_NO_NODE)
3735 nr_available = h->free_huge_pages_node[nid];
3737 nr_available = h->free_huge_pages;
3738 nr_available -= h->resv_huge_pages;
3742 err = demote_pool_huge_page(h, n_mask);
3749 spin_unlock_irq(&hugetlb_lock);
3750 mutex_unlock(&h->resize_lock);
3756 HSTATE_ATTR_WO(demote);
3758 static ssize_t demote_size_show(struct kobject *kobj,
3759 struct kobj_attribute *attr, char *buf)
3762 struct hstate *h = kobj_to_hstate(kobj, &nid);
3763 unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K;
3765 return sysfs_emit(buf, "%lukB\n", demote_size);
3768 static ssize_t demote_size_store(struct kobject *kobj,
3769 struct kobj_attribute *attr,
3770 const char *buf, size_t count)
3772 struct hstate *h, *demote_hstate;
3773 unsigned long demote_size;
3774 unsigned int demote_order;
3777 demote_size = (unsigned long)memparse(buf, NULL);
3779 demote_hstate = size_to_hstate(demote_size);
3782 demote_order = demote_hstate->order;
3783 if (demote_order < HUGETLB_PAGE_ORDER)
3786 /* demote order must be smaller than hstate order */
3787 h = kobj_to_hstate(kobj, &nid);
3788 if (demote_order >= h->order)
3791 /* resize_lock synchronizes access to demote size and writes */
3792 mutex_lock(&h->resize_lock);
3793 h->demote_order = demote_order;
3794 mutex_unlock(&h->resize_lock);
3798 HSTATE_ATTR(demote_size);
3800 static struct attribute *hstate_attrs[] = {
3801 &nr_hugepages_attr.attr,
3802 &nr_overcommit_hugepages_attr.attr,
3803 &free_hugepages_attr.attr,
3804 &resv_hugepages_attr.attr,
3805 &surplus_hugepages_attr.attr,
3807 &nr_hugepages_mempolicy_attr.attr,
3812 static const struct attribute_group hstate_attr_group = {
3813 .attrs = hstate_attrs,
3816 static struct attribute *hstate_demote_attrs[] = {
3817 &demote_size_attr.attr,
3822 static const struct attribute_group hstate_demote_attr_group = {
3823 .attrs = hstate_demote_attrs,
3826 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3827 struct kobject **hstate_kobjs,
3828 const struct attribute_group *hstate_attr_group)
3831 int hi = hstate_index(h);
3833 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3834 if (!hstate_kobjs[hi])
3837 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3839 kobject_put(hstate_kobjs[hi]);
3840 hstate_kobjs[hi] = NULL;
3843 if (h->demote_order) {
3844 if (sysfs_create_group(hstate_kobjs[hi],
3845 &hstate_demote_attr_group))
3846 pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name);
3852 static void __init hugetlb_sysfs_init(void)
3857 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3858 if (!hugepages_kobj)
3861 for_each_hstate(h) {
3862 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3863 hstate_kobjs, &hstate_attr_group);
3865 pr_err("HugeTLB: Unable to add hstate %s", h->name);
3872 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3873 * with node devices in node_devices[] using a parallel array. The array
3874 * index of a node device or _hstate == node id.
3875 * This is here to avoid any static dependency of the node device driver, in
3876 * the base kernel, on the hugetlb module.
3878 struct node_hstate {
3879 struct kobject *hugepages_kobj;
3880 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3882 static struct node_hstate node_hstates[MAX_NUMNODES];
3885 * A subset of global hstate attributes for node devices
3887 static struct attribute *per_node_hstate_attrs[] = {
3888 &nr_hugepages_attr.attr,
3889 &free_hugepages_attr.attr,
3890 &surplus_hugepages_attr.attr,
3894 static const struct attribute_group per_node_hstate_attr_group = {
3895 .attrs = per_node_hstate_attrs,
3899 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3900 * Returns node id via non-NULL nidp.
3902 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3906 for (nid = 0; nid < nr_node_ids; nid++) {
3907 struct node_hstate *nhs = &node_hstates[nid];
3909 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3910 if (nhs->hstate_kobjs[i] == kobj) {
3922 * Unregister hstate attributes from a single node device.
3923 * No-op if no hstate attributes attached.
3925 static void hugetlb_unregister_node(struct node *node)
3928 struct node_hstate *nhs = &node_hstates[node->dev.id];
3930 if (!nhs->hugepages_kobj)
3931 return; /* no hstate attributes */
3933 for_each_hstate(h) {
3934 int idx = hstate_index(h);
3935 if (nhs->hstate_kobjs[idx]) {
3936 kobject_put(nhs->hstate_kobjs[idx]);
3937 nhs->hstate_kobjs[idx] = NULL;
3941 kobject_put(nhs->hugepages_kobj);
3942 nhs->hugepages_kobj = NULL;
3947 * Register hstate attributes for a single node device.
3948 * No-op if attributes already registered.
3950 static void hugetlb_register_node(struct node *node)
3953 struct node_hstate *nhs = &node_hstates[node->dev.id];
3956 if (nhs->hugepages_kobj)
3957 return; /* already allocated */
3959 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3961 if (!nhs->hugepages_kobj)
3964 for_each_hstate(h) {
3965 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3967 &per_node_hstate_attr_group);
3969 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3970 h->name, node->dev.id);
3971 hugetlb_unregister_node(node);
3978 * hugetlb init time: register hstate attributes for all registered node
3979 * devices of nodes that have memory. All on-line nodes should have
3980 * registered their associated device by this time.
3982 static void __init hugetlb_register_all_nodes(void)
3986 for_each_node_state(nid, N_MEMORY) {
3987 struct node *node = node_devices[nid];
3988 if (node->dev.id == nid)
3989 hugetlb_register_node(node);
3993 * Let the node device driver know we're here so it can
3994 * [un]register hstate attributes on node hotplug.
3996 register_hugetlbfs_with_node(hugetlb_register_node,
3997 hugetlb_unregister_node);
3999 #else /* !CONFIG_NUMA */
4001 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4009 static void hugetlb_register_all_nodes(void) { }
4013 static int __init hugetlb_init(void)
4017 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
4020 if (!hugepages_supported()) {
4021 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
4022 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
4027 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
4028 * architectures depend on setup being done here.
4030 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
4031 if (!parsed_default_hugepagesz) {
4033 * If we did not parse a default huge page size, set
4034 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
4035 * number of huge pages for this default size was implicitly
4036 * specified, set that here as well.
4037 * Note that the implicit setting will overwrite an explicit
4038 * setting. A warning will be printed in this case.
4040 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
4041 if (default_hstate_max_huge_pages) {
4042 if (default_hstate.max_huge_pages) {
4045 string_get_size(huge_page_size(&default_hstate),
4046 1, STRING_UNITS_2, buf, 32);
4047 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
4048 default_hstate.max_huge_pages, buf);
4049 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
4050 default_hstate_max_huge_pages);
4052 default_hstate.max_huge_pages =
4053 default_hstate_max_huge_pages;
4055 for (i = 0; i < nr_online_nodes; i++)
4056 default_hstate.max_huge_pages_node[i] =
4057 default_hugepages_in_node[i];
4061 hugetlb_cma_check();
4062 hugetlb_init_hstates();
4063 gather_bootmem_prealloc();
4066 hugetlb_sysfs_init();
4067 hugetlb_register_all_nodes();
4068 hugetlb_cgroup_file_init();
4071 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
4073 num_fault_mutexes = 1;
4075 hugetlb_fault_mutex_table =
4076 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
4078 BUG_ON(!hugetlb_fault_mutex_table);
4080 for (i = 0; i < num_fault_mutexes; i++)
4081 mutex_init(&hugetlb_fault_mutex_table[i]);
4084 subsys_initcall(hugetlb_init);
4086 /* Overwritten by architectures with more huge page sizes */
4087 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
4089 return size == HPAGE_SIZE;
4092 void __init hugetlb_add_hstate(unsigned int order)
4097 if (size_to_hstate(PAGE_SIZE << order)) {
4100 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
4102 h = &hstates[hugetlb_max_hstate++];
4103 mutex_init(&h->resize_lock);
4105 h->mask = ~(huge_page_size(h) - 1);
4106 for (i = 0; i < MAX_NUMNODES; ++i)
4107 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
4108 INIT_LIST_HEAD(&h->hugepage_activelist);
4109 h->next_nid_to_alloc = first_memory_node;
4110 h->next_nid_to_free = first_memory_node;
4111 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
4112 huge_page_size(h)/1024);
4113 hugetlb_vmemmap_init(h);
4118 bool __init __weak hugetlb_node_alloc_supported(void)
4123 * hugepages command line processing
4124 * hugepages normally follows a valid hugepagsz or default_hugepagsz
4125 * specification. If not, ignore the hugepages value. hugepages can also
4126 * be the first huge page command line option in which case it implicitly
4127 * specifies the number of huge pages for the default size.
4129 static int __init hugepages_setup(char *s)
4132 static unsigned long *last_mhp;
4133 int node = NUMA_NO_NODE;
4138 if (!parsed_valid_hugepagesz) {
4139 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
4140 parsed_valid_hugepagesz = true;
4145 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
4146 * yet, so this hugepages= parameter goes to the "default hstate".
4147 * Otherwise, it goes with the previously parsed hugepagesz or
4148 * default_hugepagesz.
4150 else if (!hugetlb_max_hstate)
4151 mhp = &default_hstate_max_huge_pages;
4153 mhp = &parsed_hstate->max_huge_pages;
4155 if (mhp == last_mhp) {
4156 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
4162 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4164 /* Parameter is node format */
4165 if (p[count] == ':') {
4166 if (!hugetlb_node_alloc_supported()) {
4167 pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n");
4170 if (tmp >= nr_online_nodes)
4172 node = array_index_nospec(tmp, nr_online_nodes);
4174 /* Parse hugepages */
4175 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4177 if (!hugetlb_max_hstate)
4178 default_hugepages_in_node[node] = tmp;
4180 parsed_hstate->max_huge_pages_node[node] = tmp;
4182 /* Go to parse next node*/
4183 if (p[count] == ',')
4196 * Global state is always initialized later in hugetlb_init.
4197 * But we need to allocate gigantic hstates here early to still
4198 * use the bootmem allocator.
4200 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
4201 hugetlb_hstate_alloc_pages(parsed_hstate);
4208 pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p);
4211 __setup("hugepages=", hugepages_setup);
4214 * hugepagesz command line processing
4215 * A specific huge page size can only be specified once with hugepagesz.
4216 * hugepagesz is followed by hugepages on the command line. The global
4217 * variable 'parsed_valid_hugepagesz' is used to determine if prior
4218 * hugepagesz argument was valid.
4220 static int __init hugepagesz_setup(char *s)
4225 parsed_valid_hugepagesz = false;
4226 size = (unsigned long)memparse(s, NULL);
4228 if (!arch_hugetlb_valid_size(size)) {
4229 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
4233 h = size_to_hstate(size);
4236 * hstate for this size already exists. This is normally
4237 * an error, but is allowed if the existing hstate is the
4238 * default hstate. More specifically, it is only allowed if
4239 * the number of huge pages for the default hstate was not
4240 * previously specified.
4242 if (!parsed_default_hugepagesz || h != &default_hstate ||
4243 default_hstate.max_huge_pages) {
4244 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
4249 * No need to call hugetlb_add_hstate() as hstate already
4250 * exists. But, do set parsed_hstate so that a following
4251 * hugepages= parameter will be applied to this hstate.
4254 parsed_valid_hugepagesz = true;
4258 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4259 parsed_valid_hugepagesz = true;
4262 __setup("hugepagesz=", hugepagesz_setup);
4265 * default_hugepagesz command line input
4266 * Only one instance of default_hugepagesz allowed on command line.
4268 static int __init default_hugepagesz_setup(char *s)
4273 parsed_valid_hugepagesz = false;
4274 if (parsed_default_hugepagesz) {
4275 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
4279 size = (unsigned long)memparse(s, NULL);
4281 if (!arch_hugetlb_valid_size(size)) {
4282 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
4286 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4287 parsed_valid_hugepagesz = true;
4288 parsed_default_hugepagesz = true;
4289 default_hstate_idx = hstate_index(size_to_hstate(size));
4292 * The number of default huge pages (for this size) could have been
4293 * specified as the first hugetlb parameter: hugepages=X. If so,
4294 * then default_hstate_max_huge_pages is set. If the default huge
4295 * page size is gigantic (>= MAX_ORDER), then the pages must be
4296 * allocated here from bootmem allocator.
4298 if (default_hstate_max_huge_pages) {
4299 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
4300 for (i = 0; i < nr_online_nodes; i++)
4301 default_hstate.max_huge_pages_node[i] =
4302 default_hugepages_in_node[i];
4303 if (hstate_is_gigantic(&default_hstate))
4304 hugetlb_hstate_alloc_pages(&default_hstate);
4305 default_hstate_max_huge_pages = 0;
4310 __setup("default_hugepagesz=", default_hugepagesz_setup);
4312 static unsigned int allowed_mems_nr(struct hstate *h)
4315 unsigned int nr = 0;
4316 nodemask_t *mpol_allowed;
4317 unsigned int *array = h->free_huge_pages_node;
4318 gfp_t gfp_mask = htlb_alloc_mask(h);
4320 mpol_allowed = policy_nodemask_current(gfp_mask);
4322 for_each_node_mask(node, cpuset_current_mems_allowed) {
4323 if (!mpol_allowed || node_isset(node, *mpol_allowed))
4330 #ifdef CONFIG_SYSCTL
4331 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
4332 void *buffer, size_t *length,
4333 loff_t *ppos, unsigned long *out)
4335 struct ctl_table dup_table;
4338 * In order to avoid races with __do_proc_doulongvec_minmax(), we
4339 * can duplicate the @table and alter the duplicate of it.
4342 dup_table.data = out;
4344 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
4347 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
4348 struct ctl_table *table, int write,
4349 void *buffer, size_t *length, loff_t *ppos)
4351 struct hstate *h = &default_hstate;
4352 unsigned long tmp = h->max_huge_pages;
4355 if (!hugepages_supported())
4358 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4364 ret = __nr_hugepages_store_common(obey_mempolicy, h,
4365 NUMA_NO_NODE, tmp, *length);
4370 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
4371 void *buffer, size_t *length, loff_t *ppos)
4374 return hugetlb_sysctl_handler_common(false, table, write,
4375 buffer, length, ppos);
4379 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
4380 void *buffer, size_t *length, loff_t *ppos)
4382 return hugetlb_sysctl_handler_common(true, table, write,
4383 buffer, length, ppos);
4385 #endif /* CONFIG_NUMA */
4387 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
4388 void *buffer, size_t *length, loff_t *ppos)
4390 struct hstate *h = &default_hstate;
4394 if (!hugepages_supported())
4397 tmp = h->nr_overcommit_huge_pages;
4399 if (write && hstate_is_gigantic(h))
4402 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4408 spin_lock_irq(&hugetlb_lock);
4409 h->nr_overcommit_huge_pages = tmp;
4410 spin_unlock_irq(&hugetlb_lock);
4416 #endif /* CONFIG_SYSCTL */
4418 void hugetlb_report_meminfo(struct seq_file *m)
4421 unsigned long total = 0;
4423 if (!hugepages_supported())
4426 for_each_hstate(h) {
4427 unsigned long count = h->nr_huge_pages;
4429 total += huge_page_size(h) * count;
4431 if (h == &default_hstate)
4433 "HugePages_Total: %5lu\n"
4434 "HugePages_Free: %5lu\n"
4435 "HugePages_Rsvd: %5lu\n"
4436 "HugePages_Surp: %5lu\n"
4437 "Hugepagesize: %8lu kB\n",
4441 h->surplus_huge_pages,
4442 huge_page_size(h) / SZ_1K);
4445 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
4448 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
4450 struct hstate *h = &default_hstate;
4452 if (!hugepages_supported())
4455 return sysfs_emit_at(buf, len,
4456 "Node %d HugePages_Total: %5u\n"
4457 "Node %d HugePages_Free: %5u\n"
4458 "Node %d HugePages_Surp: %5u\n",
4459 nid, h->nr_huge_pages_node[nid],
4460 nid, h->free_huge_pages_node[nid],
4461 nid, h->surplus_huge_pages_node[nid]);
4464 void hugetlb_show_meminfo(void)
4469 if (!hugepages_supported())
4472 for_each_node_state(nid, N_MEMORY)
4474 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
4476 h->nr_huge_pages_node[nid],
4477 h->free_huge_pages_node[nid],
4478 h->surplus_huge_pages_node[nid],
4479 huge_page_size(h) / SZ_1K);
4482 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
4484 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
4485 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
4488 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
4489 unsigned long hugetlb_total_pages(void)
4492 unsigned long nr_total_pages = 0;
4495 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
4496 return nr_total_pages;
4499 static int hugetlb_acct_memory(struct hstate *h, long delta)
4506 spin_lock_irq(&hugetlb_lock);
4508 * When cpuset is configured, it breaks the strict hugetlb page
4509 * reservation as the accounting is done on a global variable. Such
4510 * reservation is completely rubbish in the presence of cpuset because
4511 * the reservation is not checked against page availability for the
4512 * current cpuset. Application can still potentially OOM'ed by kernel
4513 * with lack of free htlb page in cpuset that the task is in.
4514 * Attempt to enforce strict accounting with cpuset is almost
4515 * impossible (or too ugly) because cpuset is too fluid that
4516 * task or memory node can be dynamically moved between cpusets.
4518 * The change of semantics for shared hugetlb mapping with cpuset is
4519 * undesirable. However, in order to preserve some of the semantics,
4520 * we fall back to check against current free page availability as
4521 * a best attempt and hopefully to minimize the impact of changing
4522 * semantics that cpuset has.
4524 * Apart from cpuset, we also have memory policy mechanism that
4525 * also determines from which node the kernel will allocate memory
4526 * in a NUMA system. So similar to cpuset, we also should consider
4527 * the memory policy of the current task. Similar to the description
4531 if (gather_surplus_pages(h, delta) < 0)
4534 if (delta > allowed_mems_nr(h)) {
4535 return_unused_surplus_pages(h, delta);
4542 return_unused_surplus_pages(h, (unsigned long) -delta);
4545 spin_unlock_irq(&hugetlb_lock);
4549 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4551 struct resv_map *resv = vma_resv_map(vma);
4554 * This new VMA should share its siblings reservation map if present.
4555 * The VMA will only ever have a valid reservation map pointer where
4556 * it is being copied for another still existing VMA. As that VMA
4557 * has a reference to the reservation map it cannot disappear until
4558 * after this open call completes. It is therefore safe to take a
4559 * new reference here without additional locking.
4561 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
4562 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
4563 kref_get(&resv->refs);
4567 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4569 struct hstate *h = hstate_vma(vma);
4570 struct resv_map *resv = vma_resv_map(vma);
4571 struct hugepage_subpool *spool = subpool_vma(vma);
4572 unsigned long reserve, start, end;
4575 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4578 start = vma_hugecache_offset(h, vma, vma->vm_start);
4579 end = vma_hugecache_offset(h, vma, vma->vm_end);
4581 reserve = (end - start) - region_count(resv, start, end);
4582 hugetlb_cgroup_uncharge_counter(resv, start, end);
4585 * Decrement reserve counts. The global reserve count may be
4586 * adjusted if the subpool has a minimum size.
4588 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4589 hugetlb_acct_memory(h, -gbl_reserve);
4592 kref_put(&resv->refs, resv_map_release);
4595 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4597 if (addr & ~(huge_page_mask(hstate_vma(vma))))
4602 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4604 return huge_page_size(hstate_vma(vma));
4608 * We cannot handle pagefaults against hugetlb pages at all. They cause
4609 * handle_mm_fault() to try to instantiate regular-sized pages in the
4610 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
4613 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4620 * When a new function is introduced to vm_operations_struct and added
4621 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4622 * This is because under System V memory model, mappings created via
4623 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4624 * their original vm_ops are overwritten with shm_vm_ops.
4626 const struct vm_operations_struct hugetlb_vm_ops = {
4627 .fault = hugetlb_vm_op_fault,
4628 .open = hugetlb_vm_op_open,
4629 .close = hugetlb_vm_op_close,
4630 .may_split = hugetlb_vm_op_split,
4631 .pagesize = hugetlb_vm_op_pagesize,
4634 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4638 unsigned int shift = huge_page_shift(hstate_vma(vma));
4641 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4642 vma->vm_page_prot)));
4644 entry = huge_pte_wrprotect(mk_huge_pte(page,
4645 vma->vm_page_prot));
4647 entry = pte_mkyoung(entry);
4648 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4653 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4654 unsigned long address, pte_t *ptep)
4658 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4659 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4660 update_mmu_cache(vma, address, ptep);
4663 bool is_hugetlb_entry_migration(pte_t pte)
4667 if (huge_pte_none(pte) || pte_present(pte))
4669 swp = pte_to_swp_entry(pte);
4670 if (is_migration_entry(swp))
4676 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4680 if (huge_pte_none(pte) || pte_present(pte))
4682 swp = pte_to_swp_entry(pte);
4683 if (is_hwpoison_entry(swp))
4690 hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
4691 struct page *new_page)
4693 __SetPageUptodate(new_page);
4694 hugepage_add_new_anon_rmap(new_page, vma, addr);
4695 set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
4696 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
4697 ClearHPageRestoreReserve(new_page);
4698 SetHPageMigratable(new_page);
4701 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
4702 struct vm_area_struct *vma)
4704 pte_t *src_pte, *dst_pte, entry, dst_entry;
4705 struct page *ptepage;
4707 bool cow = is_cow_mapping(vma->vm_flags);
4708 struct hstate *h = hstate_vma(vma);
4709 unsigned long sz = huge_page_size(h);
4710 unsigned long npages = pages_per_huge_page(h);
4711 struct address_space *mapping = vma->vm_file->f_mapping;
4712 struct mmu_notifier_range range;
4716 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
4719 mmu_notifier_invalidate_range_start(&range);
4722 * For shared mappings i_mmap_rwsem must be held to call
4723 * huge_pte_alloc, otherwise the returned ptep could go
4724 * away if part of a shared pmd and another thread calls
4727 i_mmap_lock_read(mapping);
4730 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
4731 spinlock_t *src_ptl, *dst_ptl;
4732 src_pte = huge_pte_offset(src, addr, sz);
4735 dst_pte = huge_pte_alloc(dst, vma, addr, sz);
4742 * If the pagetables are shared don't copy or take references.
4743 * dst_pte == src_pte is the common case of src/dest sharing.
4745 * However, src could have 'unshared' and dst shares with
4746 * another vma. If dst_pte !none, this implies sharing.
4747 * Check here before taking page table lock, and once again
4748 * after taking the lock below.
4750 dst_entry = huge_ptep_get(dst_pte);
4751 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
4754 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4755 src_ptl = huge_pte_lockptr(h, src, src_pte);
4756 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4757 entry = huge_ptep_get(src_pte);
4758 dst_entry = huge_ptep_get(dst_pte);
4760 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
4762 * Skip if src entry none. Also, skip in the
4763 * unlikely case dst entry !none as this implies
4764 * sharing with another vma.
4767 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
4768 is_hugetlb_entry_hwpoisoned(entry))) {
4769 swp_entry_t swp_entry = pte_to_swp_entry(entry);
4771 if (is_writable_migration_entry(swp_entry) && cow) {
4773 * COW mappings require pages in both
4774 * parent and child to be set to read.
4776 swp_entry = make_readable_migration_entry(
4777 swp_offset(swp_entry));
4778 entry = swp_entry_to_pte(swp_entry);
4779 set_huge_swap_pte_at(src, addr, src_pte,
4782 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
4784 entry = huge_ptep_get(src_pte);
4785 ptepage = pte_page(entry);
4789 * This is a rare case where we see pinned hugetlb
4790 * pages while they're prone to COW. We need to do the
4791 * COW earlier during fork.
4793 * When pre-allocating the page or copying data, we
4794 * need to be without the pgtable locks since we could
4795 * sleep during the process.
4797 if (unlikely(page_needs_cow_for_dma(vma, ptepage))) {
4798 pte_t src_pte_old = entry;
4801 spin_unlock(src_ptl);
4802 spin_unlock(dst_ptl);
4803 /* Do not use reserve as it's private owned */
4804 new = alloc_huge_page(vma, addr, 1);
4810 copy_user_huge_page(new, ptepage, addr, vma,
4814 /* Install the new huge page if src pte stable */
4815 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4816 src_ptl = huge_pte_lockptr(h, src, src_pte);
4817 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4818 entry = huge_ptep_get(src_pte);
4819 if (!pte_same(src_pte_old, entry)) {
4820 restore_reserve_on_error(h, vma, addr,
4823 /* dst_entry won't change as in child */
4826 hugetlb_install_page(vma, dst_pte, addr, new);
4827 spin_unlock(src_ptl);
4828 spin_unlock(dst_ptl);
4834 * No need to notify as we are downgrading page
4835 * table protection not changing it to point
4838 * See Documentation/vm/mmu_notifier.rst
4840 huge_ptep_set_wrprotect(src, addr, src_pte);
4841 entry = huge_pte_wrprotect(entry);
4844 page_dup_rmap(ptepage, true);
4845 set_huge_pte_at(dst, addr, dst_pte, entry);
4846 hugetlb_count_add(npages, dst);
4848 spin_unlock(src_ptl);
4849 spin_unlock(dst_ptl);
4853 mmu_notifier_invalidate_range_end(&range);
4855 i_mmap_unlock_read(mapping);
4860 static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
4861 unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte)
4863 struct hstate *h = hstate_vma(vma);
4864 struct mm_struct *mm = vma->vm_mm;
4865 spinlock_t *src_ptl, *dst_ptl;
4868 dst_ptl = huge_pte_lock(h, mm, dst_pte);
4869 src_ptl = huge_pte_lockptr(h, mm, src_pte);
4872 * We don't have to worry about the ordering of src and dst ptlocks
4873 * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock.
4875 if (src_ptl != dst_ptl)
4876 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4878 pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);
4879 set_huge_pte_at(mm, new_addr, dst_pte, pte);
4881 if (src_ptl != dst_ptl)
4882 spin_unlock(src_ptl);
4883 spin_unlock(dst_ptl);
4886 int move_hugetlb_page_tables(struct vm_area_struct *vma,
4887 struct vm_area_struct *new_vma,
4888 unsigned long old_addr, unsigned long new_addr,
4891 struct hstate *h = hstate_vma(vma);
4892 struct address_space *mapping = vma->vm_file->f_mapping;
4893 unsigned long sz = huge_page_size(h);
4894 struct mm_struct *mm = vma->vm_mm;
4895 unsigned long old_end = old_addr + len;
4896 unsigned long old_addr_copy;
4897 pte_t *src_pte, *dst_pte;
4898 struct mmu_notifier_range range;
4900 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr,
4902 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4903 mmu_notifier_invalidate_range_start(&range);
4904 /* Prevent race with file truncation */
4905 i_mmap_lock_write(mapping);
4906 for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
4907 src_pte = huge_pte_offset(mm, old_addr, sz);
4910 if (huge_pte_none(huge_ptep_get(src_pte)))
4913 /* old_addr arg to huge_pmd_unshare() is a pointer and so the
4914 * arg may be modified. Pass a copy instead to preserve the
4915 * value in old_addr.
4917 old_addr_copy = old_addr;
4919 if (huge_pmd_unshare(mm, vma, &old_addr_copy, src_pte))
4922 dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz);
4926 move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte);
4928 flush_tlb_range(vma, old_end - len, old_end);
4929 mmu_notifier_invalidate_range_end(&range);
4930 i_mmap_unlock_write(mapping);
4932 return len + old_addr - old_end;
4935 static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
4936 unsigned long start, unsigned long end,
4937 struct page *ref_page)
4939 struct mm_struct *mm = vma->vm_mm;
4940 unsigned long address;
4945 struct hstate *h = hstate_vma(vma);
4946 unsigned long sz = huge_page_size(h);
4947 struct mmu_notifier_range range;
4948 bool force_flush = false;
4950 WARN_ON(!is_vm_hugetlb_page(vma));
4951 BUG_ON(start & ~huge_page_mask(h));
4952 BUG_ON(end & ~huge_page_mask(h));
4955 * This is a hugetlb vma, all the pte entries should point
4958 tlb_change_page_size(tlb, sz);
4959 tlb_start_vma(tlb, vma);
4962 * If sharing possible, alert mmu notifiers of worst case.
4964 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
4966 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4967 mmu_notifier_invalidate_range_start(&range);
4969 for (; address < end; address += sz) {
4970 ptep = huge_pte_offset(mm, address, sz);
4974 ptl = huge_pte_lock(h, mm, ptep);
4975 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
4977 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
4982 pte = huge_ptep_get(ptep);
4983 if (huge_pte_none(pte)) {
4989 * Migrating hugepage or HWPoisoned hugepage is already
4990 * unmapped and its refcount is dropped, so just clear pte here.
4992 if (unlikely(!pte_present(pte))) {
4993 huge_pte_clear(mm, address, ptep, sz);
4998 page = pte_page(pte);
5000 * If a reference page is supplied, it is because a specific
5001 * page is being unmapped, not a range. Ensure the page we
5002 * are about to unmap is the actual page of interest.
5005 if (page != ref_page) {
5010 * Mark the VMA as having unmapped its page so that
5011 * future faults in this VMA will fail rather than
5012 * looking like data was lost
5014 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
5017 pte = huge_ptep_get_and_clear(mm, address, ptep);
5018 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
5019 if (huge_pte_dirty(pte))
5020 set_page_dirty(page);
5022 hugetlb_count_sub(pages_per_huge_page(h), mm);
5023 page_remove_rmap(page, vma, true);
5026 tlb_remove_page_size(tlb, page, huge_page_size(h));
5028 * Bail out after unmapping reference page if supplied
5033 mmu_notifier_invalidate_range_end(&range);
5034 tlb_end_vma(tlb, vma);
5037 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
5038 * could defer the flush until now, since by holding i_mmap_rwsem we
5039 * guaranteed that the last refernece would not be dropped. But we must
5040 * do the flushing before we return, as otherwise i_mmap_rwsem will be
5041 * dropped and the last reference to the shared PMDs page might be
5044 * In theory we could defer the freeing of the PMD pages as well, but
5045 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
5046 * detect sharing, so we cannot defer the release of the page either.
5047 * Instead, do flush now.
5050 tlb_flush_mmu_tlbonly(tlb);
5053 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
5054 struct vm_area_struct *vma, unsigned long start,
5055 unsigned long end, struct page *ref_page)
5057 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
5060 * Clear this flag so that x86's huge_pmd_share page_table_shareable
5061 * test will fail on a vma being torn down, and not grab a page table
5062 * on its way out. We're lucky that the flag has such an appropriate
5063 * name, and can in fact be safely cleared here. We could clear it
5064 * before the __unmap_hugepage_range above, but all that's necessary
5065 * is to clear it before releasing the i_mmap_rwsem. This works
5066 * because in the context this is called, the VMA is about to be
5067 * destroyed and the i_mmap_rwsem is held.
5069 vma->vm_flags &= ~VM_MAYSHARE;
5072 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
5073 unsigned long end, struct page *ref_page)
5075 struct mmu_gather tlb;
5077 tlb_gather_mmu(&tlb, vma->vm_mm);
5078 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
5079 tlb_finish_mmu(&tlb);
5083 * This is called when the original mapper is failing to COW a MAP_PRIVATE
5084 * mapping it owns the reserve page for. The intention is to unmap the page
5085 * from other VMAs and let the children be SIGKILLed if they are faulting the
5088 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
5089 struct page *page, unsigned long address)
5091 struct hstate *h = hstate_vma(vma);
5092 struct vm_area_struct *iter_vma;
5093 struct address_space *mapping;
5097 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
5098 * from page cache lookup which is in HPAGE_SIZE units.
5100 address = address & huge_page_mask(h);
5101 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
5103 mapping = vma->vm_file->f_mapping;
5106 * Take the mapping lock for the duration of the table walk. As
5107 * this mapping should be shared between all the VMAs,
5108 * __unmap_hugepage_range() is called as the lock is already held
5110 i_mmap_lock_write(mapping);
5111 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
5112 /* Do not unmap the current VMA */
5113 if (iter_vma == vma)
5117 * Shared VMAs have their own reserves and do not affect
5118 * MAP_PRIVATE accounting but it is possible that a shared
5119 * VMA is using the same page so check and skip such VMAs.
5121 if (iter_vma->vm_flags & VM_MAYSHARE)
5125 * Unmap the page from other VMAs without their own reserves.
5126 * They get marked to be SIGKILLed if they fault in these
5127 * areas. This is because a future no-page fault on this VMA
5128 * could insert a zeroed page instead of the data existing
5129 * from the time of fork. This would look like data corruption
5131 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
5132 unmap_hugepage_range(iter_vma, address,
5133 address + huge_page_size(h), page);
5135 i_mmap_unlock_write(mapping);
5139 * Hugetlb_cow() should be called with page lock of the original hugepage held.
5140 * Called with hugetlb_fault_mutex_table held and pte_page locked so we
5141 * cannot race with other handlers or page migration.
5142 * Keep the pte_same checks anyway to make transition from the mutex easier.
5144 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
5145 unsigned long address, pte_t *ptep,
5146 struct page *pagecache_page, spinlock_t *ptl)
5149 struct hstate *h = hstate_vma(vma);
5150 struct page *old_page, *new_page;
5151 int outside_reserve = 0;
5153 unsigned long haddr = address & huge_page_mask(h);
5154 struct mmu_notifier_range range;
5156 pte = huge_ptep_get(ptep);
5157 old_page = pte_page(pte);
5160 /* If no-one else is actually using this page, avoid the copy
5161 * and just make the page writable */
5162 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
5163 page_move_anon_rmap(old_page, vma);
5164 set_huge_ptep_writable(vma, haddr, ptep);
5169 * If the process that created a MAP_PRIVATE mapping is about to
5170 * perform a COW due to a shared page count, attempt to satisfy
5171 * the allocation without using the existing reserves. The pagecache
5172 * page is used to determine if the reserve at this address was
5173 * consumed or not. If reserves were used, a partial faulted mapping
5174 * at the time of fork() could consume its reserves on COW instead
5175 * of the full address range.
5177 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
5178 old_page != pagecache_page)
5179 outside_reserve = 1;
5184 * Drop page table lock as buddy allocator may be called. It will
5185 * be acquired again before returning to the caller, as expected.
5188 new_page = alloc_huge_page(vma, haddr, outside_reserve);
5190 if (IS_ERR(new_page)) {
5192 * If a process owning a MAP_PRIVATE mapping fails to COW,
5193 * it is due to references held by a child and an insufficient
5194 * huge page pool. To guarantee the original mappers
5195 * reliability, unmap the page from child processes. The child
5196 * may get SIGKILLed if it later faults.
5198 if (outside_reserve) {
5199 struct address_space *mapping = vma->vm_file->f_mapping;
5204 BUG_ON(huge_pte_none(pte));
5206 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
5207 * unmapping. unmapping needs to hold i_mmap_rwsem
5208 * in write mode. Dropping i_mmap_rwsem in read mode
5209 * here is OK as COW mappings do not interact with
5212 * Reacquire both after unmap operation.
5214 idx = vma_hugecache_offset(h, vma, haddr);
5215 hash = hugetlb_fault_mutex_hash(mapping, idx);
5216 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5217 i_mmap_unlock_read(mapping);
5219 unmap_ref_private(mm, vma, old_page, haddr);
5221 i_mmap_lock_read(mapping);
5222 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5224 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5226 pte_same(huge_ptep_get(ptep), pte)))
5227 goto retry_avoidcopy;
5229 * race occurs while re-acquiring page table
5230 * lock, and our job is done.
5235 ret = vmf_error(PTR_ERR(new_page));
5236 goto out_release_old;
5240 * When the original hugepage is shared one, it does not have
5241 * anon_vma prepared.
5243 if (unlikely(anon_vma_prepare(vma))) {
5245 goto out_release_all;
5248 copy_user_huge_page(new_page, old_page, address, vma,
5249 pages_per_huge_page(h));
5250 __SetPageUptodate(new_page);
5252 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
5253 haddr + huge_page_size(h));
5254 mmu_notifier_invalidate_range_start(&range);
5257 * Retake the page table lock to check for racing updates
5258 * before the page tables are altered
5261 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5262 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
5263 ClearHPageRestoreReserve(new_page);
5266 huge_ptep_clear_flush(vma, haddr, ptep);
5267 mmu_notifier_invalidate_range(mm, range.start, range.end);
5268 page_remove_rmap(old_page, vma, true);
5269 hugepage_add_new_anon_rmap(new_page, vma, haddr);
5270 set_huge_pte_at(mm, haddr, ptep,
5271 make_huge_pte(vma, new_page, 1));
5272 SetHPageMigratable(new_page);
5273 /* Make the old page be freed below */
5274 new_page = old_page;
5277 mmu_notifier_invalidate_range_end(&range);
5279 /* No restore in case of successful pagetable update (Break COW) */
5280 if (new_page != old_page)
5281 restore_reserve_on_error(h, vma, haddr, new_page);
5286 spin_lock(ptl); /* Caller expects lock to be held */
5290 /* Return the pagecache page at a given address within a VMA */
5291 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
5292 struct vm_area_struct *vma, unsigned long address)
5294 struct address_space *mapping;
5297 mapping = vma->vm_file->f_mapping;
5298 idx = vma_hugecache_offset(h, vma, address);
5300 return find_lock_page(mapping, idx);
5304 * Return whether there is a pagecache page to back given address within VMA.
5305 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
5307 static bool hugetlbfs_pagecache_present(struct hstate *h,
5308 struct vm_area_struct *vma, unsigned long address)
5310 struct address_space *mapping;
5314 mapping = vma->vm_file->f_mapping;
5315 idx = vma_hugecache_offset(h, vma, address);
5317 page = find_get_page(mapping, idx);
5320 return page != NULL;
5323 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
5326 struct inode *inode = mapping->host;
5327 struct hstate *h = hstate_inode(inode);
5328 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
5332 ClearHPageRestoreReserve(page);
5335 * set page dirty so that it will not be removed from cache/file
5336 * by non-hugetlbfs specific code paths.
5338 set_page_dirty(page);
5340 spin_lock(&inode->i_lock);
5341 inode->i_blocks += blocks_per_huge_page(h);
5342 spin_unlock(&inode->i_lock);
5346 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
5347 struct address_space *mapping,
5350 unsigned long haddr,
5352 unsigned long reason)
5356 struct vm_fault vmf = {
5359 .real_address = addr,
5363 * Hard to debug if it ends up being
5364 * used by a callee that assumes
5365 * something about the other
5366 * uninitialized fields... same as in
5372 * hugetlb_fault_mutex and i_mmap_rwsem must be
5373 * dropped before handling userfault. Reacquire
5374 * after handling fault to make calling code simpler.
5376 hash = hugetlb_fault_mutex_hash(mapping, idx);
5377 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5378 i_mmap_unlock_read(mapping);
5379 ret = handle_userfault(&vmf, reason);
5380 i_mmap_lock_read(mapping);
5381 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5386 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
5387 struct vm_area_struct *vma,
5388 struct address_space *mapping, pgoff_t idx,
5389 unsigned long address, pte_t *ptep, unsigned int flags)
5391 struct hstate *h = hstate_vma(vma);
5392 vm_fault_t ret = VM_FAULT_SIGBUS;
5398 unsigned long haddr = address & huge_page_mask(h);
5399 bool new_page, new_pagecache_page = false;
5402 * Currently, we are forced to kill the process in the event the
5403 * original mapper has unmapped pages from the child due to a failed
5404 * COW. Warn that such a situation has occurred as it may not be obvious
5406 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
5407 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
5413 * We can not race with truncation due to holding i_mmap_rwsem.
5414 * i_size is modified when holding i_mmap_rwsem, so check here
5415 * once for faults beyond end of file.
5417 size = i_size_read(mapping->host) >> huge_page_shift(h);
5423 page = find_lock_page(mapping, idx);
5425 /* Check for page in userfault range */
5426 if (userfaultfd_missing(vma)) {
5427 ret = hugetlb_handle_userfault(vma, mapping, idx,
5428 flags, haddr, address,
5433 page = alloc_huge_page(vma, haddr, 0);
5436 * Returning error will result in faulting task being
5437 * sent SIGBUS. The hugetlb fault mutex prevents two
5438 * tasks from racing to fault in the same page which
5439 * could result in false unable to allocate errors.
5440 * Page migration does not take the fault mutex, but
5441 * does a clear then write of pte's under page table
5442 * lock. Page fault code could race with migration,
5443 * notice the clear pte and try to allocate a page
5444 * here. Before returning error, get ptl and make
5445 * sure there really is no pte entry.
5447 ptl = huge_pte_lock(h, mm, ptep);
5449 if (huge_pte_none(huge_ptep_get(ptep)))
5450 ret = vmf_error(PTR_ERR(page));
5454 clear_huge_page(page, address, pages_per_huge_page(h));
5455 __SetPageUptodate(page);
5458 if (vma->vm_flags & VM_MAYSHARE) {
5459 int err = huge_add_to_page_cache(page, mapping, idx);
5466 new_pagecache_page = true;
5469 if (unlikely(anon_vma_prepare(vma))) {
5471 goto backout_unlocked;
5477 * If memory error occurs between mmap() and fault, some process
5478 * don't have hwpoisoned swap entry for errored virtual address.
5479 * So we need to block hugepage fault by PG_hwpoison bit check.
5481 if (unlikely(PageHWPoison(page))) {
5482 ret = VM_FAULT_HWPOISON_LARGE |
5483 VM_FAULT_SET_HINDEX(hstate_index(h));
5484 goto backout_unlocked;
5487 /* Check for page in userfault range. */
5488 if (userfaultfd_minor(vma)) {
5491 ret = hugetlb_handle_userfault(vma, mapping, idx,
5492 flags, haddr, address,
5499 * If we are going to COW a private mapping later, we examine the
5500 * pending reservations for this page now. This will ensure that
5501 * any allocations necessary to record that reservation occur outside
5504 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5505 if (vma_needs_reservation(h, vma, haddr) < 0) {
5507 goto backout_unlocked;
5509 /* Just decrements count, does not deallocate */
5510 vma_end_reservation(h, vma, haddr);
5513 ptl = huge_pte_lock(h, mm, ptep);
5515 if (!huge_pte_none(huge_ptep_get(ptep)))
5519 ClearHPageRestoreReserve(page);
5520 hugepage_add_new_anon_rmap(page, vma, haddr);
5522 page_dup_rmap(page, true);
5523 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
5524 && (vma->vm_flags & VM_SHARED)));
5525 set_huge_pte_at(mm, haddr, ptep, new_pte);
5527 hugetlb_count_add(pages_per_huge_page(h), mm);
5528 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5529 /* Optimization, do the COW without a second fault */
5530 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
5536 * Only set HPageMigratable in newly allocated pages. Existing pages
5537 * found in the pagecache may not have HPageMigratableset if they have
5538 * been isolated for migration.
5541 SetHPageMigratable(page);
5551 /* restore reserve for newly allocated pages not in page cache */
5552 if (new_page && !new_pagecache_page)
5553 restore_reserve_on_error(h, vma, haddr, page);
5559 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5561 unsigned long key[2];
5564 key[0] = (unsigned long) mapping;
5567 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
5569 return hash & (num_fault_mutexes - 1);
5573 * For uniprocessor systems we always use a single mutex, so just
5574 * return 0 and avoid the hashing overhead.
5576 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5582 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
5583 unsigned long address, unsigned int flags)
5590 struct page *page = NULL;
5591 struct page *pagecache_page = NULL;
5592 struct hstate *h = hstate_vma(vma);
5593 struct address_space *mapping;
5594 int need_wait_lock = 0;
5595 unsigned long haddr = address & huge_page_mask(h);
5597 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5600 * Since we hold no locks, ptep could be stale. That is
5601 * OK as we are only making decisions based on content and
5602 * not actually modifying content here.
5604 entry = huge_ptep_get(ptep);
5605 if (unlikely(is_hugetlb_entry_migration(entry))) {
5606 migration_entry_wait_huge(vma, mm, ptep);
5608 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
5609 return VM_FAULT_HWPOISON_LARGE |
5610 VM_FAULT_SET_HINDEX(hstate_index(h));
5614 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
5615 * until finished with ptep. This serves two purposes:
5616 * 1) It prevents huge_pmd_unshare from being called elsewhere
5617 * and making the ptep no longer valid.
5618 * 2) It synchronizes us with i_size modifications during truncation.
5620 * ptep could have already be assigned via huge_pte_offset. That
5621 * is OK, as huge_pte_alloc will return the same value unless
5622 * something has changed.
5624 mapping = vma->vm_file->f_mapping;
5625 i_mmap_lock_read(mapping);
5626 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
5628 i_mmap_unlock_read(mapping);
5629 return VM_FAULT_OOM;
5633 * Serialize hugepage allocation and instantiation, so that we don't
5634 * get spurious allocation failures if two CPUs race to instantiate
5635 * the same page in the page cache.
5637 idx = vma_hugecache_offset(h, vma, haddr);
5638 hash = hugetlb_fault_mutex_hash(mapping, idx);
5639 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5641 entry = huge_ptep_get(ptep);
5642 if (huge_pte_none(entry)) {
5643 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
5650 * entry could be a migration/hwpoison entry at this point, so this
5651 * check prevents the kernel from going below assuming that we have
5652 * an active hugepage in pagecache. This goto expects the 2nd page
5653 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
5654 * properly handle it.
5656 if (!pte_present(entry))
5660 * If we are going to COW the mapping later, we examine the pending
5661 * reservations for this page now. This will ensure that any
5662 * allocations necessary to record that reservation occur outside the
5663 * spinlock. For private mappings, we also lookup the pagecache
5664 * page now as it is used to determine if a reservation has been
5667 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
5668 if (vma_needs_reservation(h, vma, haddr) < 0) {
5672 /* Just decrements count, does not deallocate */
5673 vma_end_reservation(h, vma, haddr);
5675 if (!(vma->vm_flags & VM_MAYSHARE))
5676 pagecache_page = hugetlbfs_pagecache_page(h,
5680 ptl = huge_pte_lock(h, mm, ptep);
5682 /* Check for a racing update before calling hugetlb_cow */
5683 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
5687 * hugetlb_cow() requires page locks of pte_page(entry) and
5688 * pagecache_page, so here we need take the former one
5689 * when page != pagecache_page or !pagecache_page.
5691 page = pte_page(entry);
5692 if (page != pagecache_page)
5693 if (!trylock_page(page)) {
5700 if (flags & FAULT_FLAG_WRITE) {
5701 if (!huge_pte_write(entry)) {
5702 ret = hugetlb_cow(mm, vma, address, ptep,
5703 pagecache_page, ptl);
5706 entry = huge_pte_mkdirty(entry);
5708 entry = pte_mkyoung(entry);
5709 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
5710 flags & FAULT_FLAG_WRITE))
5711 update_mmu_cache(vma, haddr, ptep);
5713 if (page != pagecache_page)
5719 if (pagecache_page) {
5720 unlock_page(pagecache_page);
5721 put_page(pagecache_page);
5724 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5725 i_mmap_unlock_read(mapping);
5727 * Generally it's safe to hold refcount during waiting page lock. But
5728 * here we just wait to defer the next page fault to avoid busy loop and
5729 * the page is not used after unlocked before returning from the current
5730 * page fault. So we are safe from accessing freed page, even if we wait
5731 * here without taking refcount.
5734 wait_on_page_locked(page);
5738 #ifdef CONFIG_USERFAULTFD
5740 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
5741 * modifications for huge pages.
5743 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
5745 struct vm_area_struct *dst_vma,
5746 unsigned long dst_addr,
5747 unsigned long src_addr,
5748 enum mcopy_atomic_mode mode,
5749 struct page **pagep)
5751 bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE);
5752 struct hstate *h = hstate_vma(dst_vma);
5753 struct address_space *mapping = dst_vma->vm_file->f_mapping;
5754 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
5756 int vm_shared = dst_vma->vm_flags & VM_SHARED;
5762 bool page_in_pagecache = false;
5766 page = find_lock_page(mapping, idx);
5769 page_in_pagecache = true;
5770 } else if (!*pagep) {
5771 /* If a page already exists, then it's UFFDIO_COPY for
5772 * a non-missing case. Return -EEXIST.
5775 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5780 page = alloc_huge_page(dst_vma, dst_addr, 0);
5786 ret = copy_huge_page_from_user(page,
5787 (const void __user *) src_addr,
5788 pages_per_huge_page(h), false);
5790 /* fallback to copy_from_user outside mmap_lock */
5791 if (unlikely(ret)) {
5793 /* Free the allocated page which may have
5794 * consumed a reservation.
5796 restore_reserve_on_error(h, dst_vma, dst_addr, page);
5799 /* Allocate a temporary page to hold the copied
5802 page = alloc_huge_page_vma(h, dst_vma, dst_addr);
5808 /* Set the outparam pagep and return to the caller to
5809 * copy the contents outside the lock. Don't free the
5816 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5823 page = alloc_huge_page(dst_vma, dst_addr, 0);
5829 copy_user_huge_page(page, *pagep, dst_addr, dst_vma,
5830 pages_per_huge_page(h));
5836 * The memory barrier inside __SetPageUptodate makes sure that
5837 * preceding stores to the page contents become visible before
5838 * the set_pte_at() write.
5840 __SetPageUptodate(page);
5842 /* Add shared, newly allocated pages to the page cache. */
5843 if (vm_shared && !is_continue) {
5844 size = i_size_read(mapping->host) >> huge_page_shift(h);
5847 goto out_release_nounlock;
5850 * Serialization between remove_inode_hugepages() and
5851 * huge_add_to_page_cache() below happens through the
5852 * hugetlb_fault_mutex_table that here must be hold by
5855 ret = huge_add_to_page_cache(page, mapping, idx);
5857 goto out_release_nounlock;
5858 page_in_pagecache = true;
5861 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
5865 * Recheck the i_size after holding PT lock to make sure not
5866 * to leave any page mapped (as page_mapped()) beyond the end
5867 * of the i_size (remove_inode_hugepages() is strict about
5868 * enforcing that). If we bail out here, we'll also leave a
5869 * page in the radix tree in the vm_shared case beyond the end
5870 * of the i_size, but remove_inode_hugepages() will take care
5871 * of it as soon as we drop the hugetlb_fault_mutex_table.
5873 size = i_size_read(mapping->host) >> huge_page_shift(h);
5876 goto out_release_unlock;
5879 if (!huge_pte_none(huge_ptep_get(dst_pte)))
5880 goto out_release_unlock;
5883 page_dup_rmap(page, true);
5885 ClearHPageRestoreReserve(page);
5886 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
5889 /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */
5890 if (is_continue && !vm_shared)
5893 writable = dst_vma->vm_flags & VM_WRITE;
5895 _dst_pte = make_huge_pte(dst_vma, page, writable);
5897 _dst_pte = huge_pte_mkdirty(_dst_pte);
5898 _dst_pte = pte_mkyoung(_dst_pte);
5900 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
5902 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
5903 dst_vma->vm_flags & VM_WRITE);
5904 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
5906 /* No need to invalidate - it was non-present before */
5907 update_mmu_cache(dst_vma, dst_addr, dst_pte);
5911 SetHPageMigratable(page);
5912 if (vm_shared || is_continue)
5919 if (vm_shared || is_continue)
5921 out_release_nounlock:
5922 if (!page_in_pagecache)
5923 restore_reserve_on_error(h, dst_vma, dst_addr, page);
5927 #endif /* CONFIG_USERFAULTFD */
5929 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
5930 int refs, struct page **pages,
5931 struct vm_area_struct **vmas)
5935 for (nr = 0; nr < refs; nr++) {
5937 pages[nr] = mem_map_offset(page, nr);
5943 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
5944 struct page **pages, struct vm_area_struct **vmas,
5945 unsigned long *position, unsigned long *nr_pages,
5946 long i, unsigned int flags, int *locked)
5948 unsigned long pfn_offset;
5949 unsigned long vaddr = *position;
5950 unsigned long remainder = *nr_pages;
5951 struct hstate *h = hstate_vma(vma);
5952 int err = -EFAULT, refs;
5954 while (vaddr < vma->vm_end && remainder) {
5956 spinlock_t *ptl = NULL;
5961 * If we have a pending SIGKILL, don't keep faulting pages and
5962 * potentially allocating memory.
5964 if (fatal_signal_pending(current)) {
5970 * Some archs (sparc64, sh*) have multiple pte_ts to
5971 * each hugepage. We have to make sure we get the
5972 * first, for the page indexing below to work.
5974 * Note that page table lock is not held when pte is null.
5976 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
5979 ptl = huge_pte_lock(h, mm, pte);
5980 absent = !pte || huge_pte_none(huge_ptep_get(pte));
5983 * When coredumping, it suits get_dump_page if we just return
5984 * an error where there's an empty slot with no huge pagecache
5985 * to back it. This way, we avoid allocating a hugepage, and
5986 * the sparse dumpfile avoids allocating disk blocks, but its
5987 * huge holes still show up with zeroes where they need to be.
5989 if (absent && (flags & FOLL_DUMP) &&
5990 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
5998 * We need call hugetlb_fault for both hugepages under migration
5999 * (in which case hugetlb_fault waits for the migration,) and
6000 * hwpoisoned hugepages (in which case we need to prevent the
6001 * caller from accessing to them.) In order to do this, we use
6002 * here is_swap_pte instead of is_hugetlb_entry_migration and
6003 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
6004 * both cases, and because we can't follow correct pages
6005 * directly from any kind of swap entries.
6007 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
6008 ((flags & FOLL_WRITE) &&
6009 !huge_pte_write(huge_ptep_get(pte)))) {
6011 unsigned int fault_flags = 0;
6015 if (flags & FOLL_WRITE)
6016 fault_flags |= FAULT_FLAG_WRITE;
6018 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
6019 FAULT_FLAG_KILLABLE;
6020 if (flags & FOLL_NOWAIT)
6021 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
6022 FAULT_FLAG_RETRY_NOWAIT;
6023 if (flags & FOLL_TRIED) {
6025 * Note: FAULT_FLAG_ALLOW_RETRY and
6026 * FAULT_FLAG_TRIED can co-exist
6028 fault_flags |= FAULT_FLAG_TRIED;
6030 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
6031 if (ret & VM_FAULT_ERROR) {
6032 err = vm_fault_to_errno(ret, flags);
6036 if (ret & VM_FAULT_RETRY) {
6038 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
6042 * VM_FAULT_RETRY must not return an
6043 * error, it will return zero
6046 * No need to update "position" as the
6047 * caller will not check it after
6048 * *nr_pages is set to 0.
6055 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
6056 page = pte_page(huge_ptep_get(pte));
6059 * If subpage information not requested, update counters
6060 * and skip the same_page loop below.
6062 if (!pages && !vmas && !pfn_offset &&
6063 (vaddr + huge_page_size(h) < vma->vm_end) &&
6064 (remainder >= pages_per_huge_page(h))) {
6065 vaddr += huge_page_size(h);
6066 remainder -= pages_per_huge_page(h);
6067 i += pages_per_huge_page(h);
6072 /* vaddr may not be aligned to PAGE_SIZE */
6073 refs = min3(pages_per_huge_page(h) - pfn_offset, remainder,
6074 (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT);
6077 record_subpages_vmas(mem_map_offset(page, pfn_offset),
6079 likely(pages) ? pages + i : NULL,
6080 vmas ? vmas + i : NULL);
6084 * try_grab_folio() should always succeed here,
6085 * because: a) we hold the ptl lock, and b) we've just
6086 * checked that the huge page is present in the page
6087 * tables. If the huge page is present, then the tail
6088 * pages must also be present. The ptl prevents the
6089 * head page and tail pages from being rearranged in
6090 * any way. So this page must be available at this
6091 * point, unless the page refcount overflowed:
6093 if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs,
6102 vaddr += (refs << PAGE_SHIFT);
6108 *nr_pages = remainder;
6110 * setting position is actually required only if remainder is
6111 * not zero but it's faster not to add a "if (remainder)"
6119 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
6120 unsigned long address, unsigned long end, pgprot_t newprot)
6122 struct mm_struct *mm = vma->vm_mm;
6123 unsigned long start = address;
6126 struct hstate *h = hstate_vma(vma);
6127 unsigned long pages = 0;
6128 bool shared_pmd = false;
6129 struct mmu_notifier_range range;
6132 * In the case of shared PMDs, the area to flush could be beyond
6133 * start/end. Set range.start/range.end to cover the maximum possible
6134 * range if PMD sharing is possible.
6136 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
6137 0, vma, mm, start, end);
6138 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
6140 BUG_ON(address >= end);
6141 flush_cache_range(vma, range.start, range.end);
6143 mmu_notifier_invalidate_range_start(&range);
6144 i_mmap_lock_write(vma->vm_file->f_mapping);
6145 for (; address < end; address += huge_page_size(h)) {
6147 ptep = huge_pte_offset(mm, address, huge_page_size(h));
6150 ptl = huge_pte_lock(h, mm, ptep);
6151 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
6157 pte = huge_ptep_get(ptep);
6158 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
6162 if (unlikely(is_hugetlb_entry_migration(pte))) {
6163 swp_entry_t entry = pte_to_swp_entry(pte);
6165 if (is_writable_migration_entry(entry)) {
6168 entry = make_readable_migration_entry(
6170 newpte = swp_entry_to_pte(entry);
6171 set_huge_swap_pte_at(mm, address, ptep,
6172 newpte, huge_page_size(h));
6178 if (!huge_pte_none(pte)) {
6180 unsigned int shift = huge_page_shift(hstate_vma(vma));
6182 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
6183 pte = huge_pte_modify(old_pte, newprot);
6184 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
6185 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
6191 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
6192 * may have cleared our pud entry and done put_page on the page table:
6193 * once we release i_mmap_rwsem, another task can do the final put_page
6194 * and that page table be reused and filled with junk. If we actually
6195 * did unshare a page of pmds, flush the range corresponding to the pud.
6198 flush_hugetlb_tlb_range(vma, range.start, range.end);
6200 flush_hugetlb_tlb_range(vma, start, end);
6202 * No need to call mmu_notifier_invalidate_range() we are downgrading
6203 * page table protection not changing it to point to a new page.
6205 * See Documentation/vm/mmu_notifier.rst
6207 i_mmap_unlock_write(vma->vm_file->f_mapping);
6208 mmu_notifier_invalidate_range_end(&range);
6210 return pages << h->order;
6213 /* Return true if reservation was successful, false otherwise. */
6214 bool hugetlb_reserve_pages(struct inode *inode,
6216 struct vm_area_struct *vma,
6217 vm_flags_t vm_flags)
6220 struct hstate *h = hstate_inode(inode);
6221 struct hugepage_subpool *spool = subpool_inode(inode);
6222 struct resv_map *resv_map;
6223 struct hugetlb_cgroup *h_cg = NULL;
6224 long gbl_reserve, regions_needed = 0;
6226 /* This should never happen */
6228 VM_WARN(1, "%s called with a negative range\n", __func__);
6233 * Only apply hugepage reservation if asked. At fault time, an
6234 * attempt will be made for VM_NORESERVE to allocate a page
6235 * without using reserves
6237 if (vm_flags & VM_NORESERVE)
6241 * Shared mappings base their reservation on the number of pages that
6242 * are already allocated on behalf of the file. Private mappings need
6243 * to reserve the full area even if read-only as mprotect() may be
6244 * called to make the mapping read-write. Assume !vma is a shm mapping
6246 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6248 * resv_map can not be NULL as hugetlb_reserve_pages is only
6249 * called for inodes for which resv_maps were created (see
6250 * hugetlbfs_get_inode).
6252 resv_map = inode_resv_map(inode);
6254 chg = region_chg(resv_map, from, to, ®ions_needed);
6257 /* Private mapping. */
6258 resv_map = resv_map_alloc();
6264 set_vma_resv_map(vma, resv_map);
6265 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
6271 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
6272 chg * pages_per_huge_page(h), &h_cg) < 0)
6275 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
6276 /* For private mappings, the hugetlb_cgroup uncharge info hangs
6279 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
6283 * There must be enough pages in the subpool for the mapping. If
6284 * the subpool has a minimum size, there may be some global
6285 * reservations already in place (gbl_reserve).
6287 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
6288 if (gbl_reserve < 0)
6289 goto out_uncharge_cgroup;
6292 * Check enough hugepages are available for the reservation.
6293 * Hand the pages back to the subpool if there are not
6295 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
6299 * Account for the reservations made. Shared mappings record regions
6300 * that have reservations as they are shared by multiple VMAs.
6301 * When the last VMA disappears, the region map says how much
6302 * the reservation was and the page cache tells how much of
6303 * the reservation was consumed. Private mappings are per-VMA and
6304 * only the consumed reservations are tracked. When the VMA
6305 * disappears, the original reservation is the VMA size and the
6306 * consumed reservations are stored in the map. Hence, nothing
6307 * else has to be done for private mappings here
6309 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6310 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
6312 if (unlikely(add < 0)) {
6313 hugetlb_acct_memory(h, -gbl_reserve);
6315 } else if (unlikely(chg > add)) {
6317 * pages in this range were added to the reserve
6318 * map between region_chg and region_add. This
6319 * indicates a race with alloc_huge_page. Adjust
6320 * the subpool and reserve counts modified above
6321 * based on the difference.
6326 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
6327 * reference to h_cg->css. See comment below for detail.
6329 hugetlb_cgroup_uncharge_cgroup_rsvd(
6331 (chg - add) * pages_per_huge_page(h), h_cg);
6333 rsv_adjust = hugepage_subpool_put_pages(spool,
6335 hugetlb_acct_memory(h, -rsv_adjust);
6338 * The file_regions will hold their own reference to
6339 * h_cg->css. So we should release the reference held
6340 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
6343 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
6349 /* put back original number of pages, chg */
6350 (void)hugepage_subpool_put_pages(spool, chg);
6351 out_uncharge_cgroup:
6352 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
6353 chg * pages_per_huge_page(h), h_cg);
6355 if (!vma || vma->vm_flags & VM_MAYSHARE)
6356 /* Only call region_abort if the region_chg succeeded but the
6357 * region_add failed or didn't run.
6359 if (chg >= 0 && add < 0)
6360 region_abort(resv_map, from, to, regions_needed);
6361 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
6362 kref_put(&resv_map->refs, resv_map_release);
6366 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
6369 struct hstate *h = hstate_inode(inode);
6370 struct resv_map *resv_map = inode_resv_map(inode);
6372 struct hugepage_subpool *spool = subpool_inode(inode);
6376 * Since this routine can be called in the evict inode path for all
6377 * hugetlbfs inodes, resv_map could be NULL.
6380 chg = region_del(resv_map, start, end);
6382 * region_del() can fail in the rare case where a region
6383 * must be split and another region descriptor can not be
6384 * allocated. If end == LONG_MAX, it will not fail.
6390 spin_lock(&inode->i_lock);
6391 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
6392 spin_unlock(&inode->i_lock);
6395 * If the subpool has a minimum size, the number of global
6396 * reservations to be released may be adjusted.
6398 * Note that !resv_map implies freed == 0. So (chg - freed)
6399 * won't go negative.
6401 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
6402 hugetlb_acct_memory(h, -gbl_reserve);
6407 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
6408 static unsigned long page_table_shareable(struct vm_area_struct *svma,
6409 struct vm_area_struct *vma,
6410 unsigned long addr, pgoff_t idx)
6412 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
6414 unsigned long sbase = saddr & PUD_MASK;
6415 unsigned long s_end = sbase + PUD_SIZE;
6417 /* Allow segments to share if only one is marked locked */
6418 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
6419 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
6422 * match the virtual addresses, permission and the alignment of the
6425 if (pmd_index(addr) != pmd_index(saddr) ||
6426 vm_flags != svm_flags ||
6427 !range_in_vma(svma, sbase, s_end))
6433 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
6435 unsigned long base = addr & PUD_MASK;
6436 unsigned long end = base + PUD_SIZE;
6439 * check on proper vm_flags and page table alignment
6441 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
6446 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6448 #ifdef CONFIG_USERFAULTFD
6449 if (uffd_disable_huge_pmd_share(vma))
6452 return vma_shareable(vma, addr);
6456 * Determine if start,end range within vma could be mapped by shared pmd.
6457 * If yes, adjust start and end to cover range associated with possible
6458 * shared pmd mappings.
6460 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
6461 unsigned long *start, unsigned long *end)
6463 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
6464 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6467 * vma needs to span at least one aligned PUD size, and the range
6468 * must be at least partially within in.
6470 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
6471 (*end <= v_start) || (*start >= v_end))
6474 /* Extend the range to be PUD aligned for a worst case scenario */
6475 if (*start > v_start)
6476 *start = ALIGN_DOWN(*start, PUD_SIZE);
6479 *end = ALIGN(*end, PUD_SIZE);
6483 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
6484 * and returns the corresponding pte. While this is not necessary for the
6485 * !shared pmd case because we can allocate the pmd later as well, it makes the
6486 * code much cleaner.
6488 * This routine must be called with i_mmap_rwsem held in at least read mode if
6489 * sharing is possible. For hugetlbfs, this prevents removal of any page
6490 * table entries associated with the address space. This is important as we
6491 * are setting up sharing based on existing page table entries (mappings).
6493 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
6494 unsigned long addr, pud_t *pud)
6496 struct address_space *mapping = vma->vm_file->f_mapping;
6497 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
6499 struct vm_area_struct *svma;
6500 unsigned long saddr;
6505 i_mmap_assert_locked(mapping);
6506 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
6510 saddr = page_table_shareable(svma, vma, addr, idx);
6512 spte = huge_pte_offset(svma->vm_mm, saddr,
6513 vma_mmu_pagesize(svma));
6515 get_page(virt_to_page(spte));
6524 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
6525 if (pud_none(*pud)) {
6526 pud_populate(mm, pud,
6527 (pmd_t *)((unsigned long)spte & PAGE_MASK));
6530 put_page(virt_to_page(spte));
6534 pte = (pte_t *)pmd_alloc(mm, pud, addr);
6539 * unmap huge page backed by shared pte.
6541 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
6542 * indicated by page_count > 1, unmap is achieved by clearing pud and
6543 * decrementing the ref count. If count == 1, the pte page is not shared.
6545 * Called with page table lock held and i_mmap_rwsem held in write mode.
6547 * returns: 1 successfully unmapped a shared pte page
6548 * 0 the underlying pte page is not shared, or it is the last user
6550 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
6551 unsigned long *addr, pte_t *ptep)
6553 pgd_t *pgd = pgd_offset(mm, *addr);
6554 p4d_t *p4d = p4d_offset(pgd, *addr);
6555 pud_t *pud = pud_offset(p4d, *addr);
6557 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
6558 BUG_ON(page_count(virt_to_page(ptep)) == 0);
6559 if (page_count(virt_to_page(ptep)) == 1)
6563 put_page(virt_to_page(ptep));
6565 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
6569 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
6570 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
6571 unsigned long addr, pud_t *pud)
6576 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
6577 unsigned long *addr, pte_t *ptep)
6582 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
6583 unsigned long *start, unsigned long *end)
6587 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6591 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
6593 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
6594 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
6595 unsigned long addr, unsigned long sz)
6602 pgd = pgd_offset(mm, addr);
6603 p4d = p4d_alloc(mm, pgd, addr);
6606 pud = pud_alloc(mm, p4d, addr);
6608 if (sz == PUD_SIZE) {
6611 BUG_ON(sz != PMD_SIZE);
6612 if (want_pmd_share(vma, addr) && pud_none(*pud))
6613 pte = huge_pmd_share(mm, vma, addr, pud);
6615 pte = (pte_t *)pmd_alloc(mm, pud, addr);
6618 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
6624 * huge_pte_offset() - Walk the page table to resolve the hugepage
6625 * entry at address @addr
6627 * Return: Pointer to page table entry (PUD or PMD) for
6628 * address @addr, or NULL if a !p*d_present() entry is encountered and the
6629 * size @sz doesn't match the hugepage size at this level of the page
6632 pte_t *huge_pte_offset(struct mm_struct *mm,
6633 unsigned long addr, unsigned long sz)
6640 pgd = pgd_offset(mm, addr);
6641 if (!pgd_present(*pgd))
6643 p4d = p4d_offset(pgd, addr);
6644 if (!p4d_present(*p4d))
6647 pud = pud_offset(p4d, addr);
6649 /* must be pud huge, non-present or none */
6650 return (pte_t *)pud;
6651 if (!pud_present(*pud))
6653 /* must have a valid entry and size to go further */
6655 pmd = pmd_offset(pud, addr);
6656 /* must be pmd huge, non-present or none */
6657 return (pte_t *)pmd;
6660 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
6663 * These functions are overwritable if your architecture needs its own
6666 struct page * __weak
6667 follow_huge_addr(struct mm_struct *mm, unsigned long address,
6670 return ERR_PTR(-EINVAL);
6673 struct page * __weak
6674 follow_huge_pd(struct vm_area_struct *vma,
6675 unsigned long address, hugepd_t hpd, int flags, int pdshift)
6677 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
6681 struct page * __weak
6682 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
6683 pmd_t *pmd, int flags)
6685 struct page *page = NULL;
6689 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
6690 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
6691 (FOLL_PIN | FOLL_GET)))
6695 ptl = pmd_lockptr(mm, pmd);
6698 * make sure that the address range covered by this pmd is not
6699 * unmapped from other threads.
6701 if (!pmd_huge(*pmd))
6703 pte = huge_ptep_get((pte_t *)pmd);
6704 if (pte_present(pte)) {
6705 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
6707 * try_grab_page() should always succeed here, because: a) we
6708 * hold the pmd (ptl) lock, and b) we've just checked that the
6709 * huge pmd (head) page is present in the page tables. The ptl
6710 * prevents the head page and tail pages from being rearranged
6711 * in any way. So this page must be available at this point,
6712 * unless the page refcount overflowed:
6714 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
6719 if (is_hugetlb_entry_migration(pte)) {
6721 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
6725 * hwpoisoned entry is treated as no_page_table in
6726 * follow_page_mask().
6734 struct page * __weak
6735 follow_huge_pud(struct mm_struct *mm, unsigned long address,
6736 pud_t *pud, int flags)
6738 if (flags & (FOLL_GET | FOLL_PIN))
6741 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
6744 struct page * __weak
6745 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
6747 if (flags & (FOLL_GET | FOLL_PIN))
6750 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
6753 bool isolate_huge_page(struct page *page, struct list_head *list)
6757 spin_lock_irq(&hugetlb_lock);
6758 if (!PageHeadHuge(page) ||
6759 !HPageMigratable(page) ||
6760 !get_page_unless_zero(page)) {
6764 ClearHPageMigratable(page);
6765 list_move_tail(&page->lru, list);
6767 spin_unlock_irq(&hugetlb_lock);
6771 int get_hwpoison_huge_page(struct page *page, bool *hugetlb)
6776 spin_lock_irq(&hugetlb_lock);
6777 if (PageHeadHuge(page)) {
6779 if (HPageFreed(page) || HPageMigratable(page))
6780 ret = get_page_unless_zero(page);
6784 spin_unlock_irq(&hugetlb_lock);
6788 int get_huge_page_for_hwpoison(unsigned long pfn, int flags)
6792 spin_lock_irq(&hugetlb_lock);
6793 ret = __get_huge_page_for_hwpoison(pfn, flags);
6794 spin_unlock_irq(&hugetlb_lock);
6798 void putback_active_hugepage(struct page *page)
6800 spin_lock_irq(&hugetlb_lock);
6801 SetHPageMigratable(page);
6802 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
6803 spin_unlock_irq(&hugetlb_lock);
6807 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
6809 struct hstate *h = page_hstate(oldpage);
6811 hugetlb_cgroup_migrate(oldpage, newpage);
6812 set_page_owner_migrate_reason(newpage, reason);
6815 * transfer temporary state of the new huge page. This is
6816 * reverse to other transitions because the newpage is going to
6817 * be final while the old one will be freed so it takes over
6818 * the temporary status.
6820 * Also note that we have to transfer the per-node surplus state
6821 * here as well otherwise the global surplus count will not match
6824 if (HPageTemporary(newpage)) {
6825 int old_nid = page_to_nid(oldpage);
6826 int new_nid = page_to_nid(newpage);
6828 SetHPageTemporary(oldpage);
6829 ClearHPageTemporary(newpage);
6832 * There is no need to transfer the per-node surplus state
6833 * when we do not cross the node.
6835 if (new_nid == old_nid)
6837 spin_lock_irq(&hugetlb_lock);
6838 if (h->surplus_huge_pages_node[old_nid]) {
6839 h->surplus_huge_pages_node[old_nid]--;
6840 h->surplus_huge_pages_node[new_nid]++;
6842 spin_unlock_irq(&hugetlb_lock);
6847 * This function will unconditionally remove all the shared pmd pgtable entries
6848 * within the specific vma for a hugetlbfs memory range.
6850 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
6852 struct hstate *h = hstate_vma(vma);
6853 unsigned long sz = huge_page_size(h);
6854 struct mm_struct *mm = vma->vm_mm;
6855 struct mmu_notifier_range range;
6856 unsigned long address, start, end;
6860 if (!(vma->vm_flags & VM_MAYSHARE))
6863 start = ALIGN(vma->vm_start, PUD_SIZE);
6864 end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6870 * No need to call adjust_range_if_pmd_sharing_possible(), because
6871 * we have already done the PUD_SIZE alignment.
6873 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
6875 mmu_notifier_invalidate_range_start(&range);
6876 i_mmap_lock_write(vma->vm_file->f_mapping);
6877 for (address = start; address < end; address += PUD_SIZE) {
6878 unsigned long tmp = address;
6880 ptep = huge_pte_offset(mm, address, sz);
6883 ptl = huge_pte_lock(h, mm, ptep);
6884 /* We don't want 'address' to be changed */
6885 huge_pmd_unshare(mm, vma, &tmp, ptep);
6888 flush_hugetlb_tlb_range(vma, start, end);
6889 i_mmap_unlock_write(vma->vm_file->f_mapping);
6891 * No need to call mmu_notifier_invalidate_range(), see
6892 * Documentation/vm/mmu_notifier.rst.
6894 mmu_notifier_invalidate_range_end(&range);
6898 static bool cma_reserve_called __initdata;
6900 static int __init cmdline_parse_hugetlb_cma(char *p)
6907 if (sscanf(s, "%lu%n", &tmp, &count) != 1)
6910 if (s[count] == ':') {
6911 if (tmp >= MAX_NUMNODES)
6913 nid = array_index_nospec(tmp, MAX_NUMNODES);
6916 tmp = memparse(s, &s);
6917 hugetlb_cma_size_in_node[nid] = tmp;
6918 hugetlb_cma_size += tmp;
6921 * Skip the separator if have one, otherwise
6922 * break the parsing.
6929 hugetlb_cma_size = memparse(p, &p);
6937 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
6939 void __init hugetlb_cma_reserve(int order)
6941 unsigned long size, reserved, per_node;
6942 bool node_specific_cma_alloc = false;
6945 cma_reserve_called = true;
6947 if (!hugetlb_cma_size)
6950 for (nid = 0; nid < MAX_NUMNODES; nid++) {
6951 if (hugetlb_cma_size_in_node[nid] == 0)
6954 if (!node_state(nid, N_ONLINE)) {
6955 pr_warn("hugetlb_cma: invalid node %d specified\n", nid);
6956 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
6957 hugetlb_cma_size_in_node[nid] = 0;
6961 if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) {
6962 pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n",
6963 nid, (PAGE_SIZE << order) / SZ_1M);
6964 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
6965 hugetlb_cma_size_in_node[nid] = 0;
6967 node_specific_cma_alloc = true;
6971 /* Validate the CMA size again in case some invalid nodes specified. */
6972 if (!hugetlb_cma_size)
6975 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
6976 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
6977 (PAGE_SIZE << order) / SZ_1M);
6978 hugetlb_cma_size = 0;
6982 if (!node_specific_cma_alloc) {
6984 * If 3 GB area is requested on a machine with 4 numa nodes,
6985 * let's allocate 1 GB on first three nodes and ignore the last one.
6987 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
6988 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
6989 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
6993 for_each_node_state(nid, N_ONLINE) {
6995 char name[CMA_MAX_NAME];
6997 if (node_specific_cma_alloc) {
6998 if (hugetlb_cma_size_in_node[nid] == 0)
7001 size = hugetlb_cma_size_in_node[nid];
7003 size = min(per_node, hugetlb_cma_size - reserved);
7006 size = round_up(size, PAGE_SIZE << order);
7008 snprintf(name, sizeof(name), "hugetlb%d", nid);
7010 * Note that 'order per bit' is based on smallest size that
7011 * may be returned to CMA allocator in the case of
7012 * huge page demotion.
7014 res = cma_declare_contiguous_nid(0, size, 0,
7015 PAGE_SIZE << HUGETLB_PAGE_ORDER,
7017 &hugetlb_cma[nid], nid);
7019 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
7025 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
7028 if (reserved >= hugetlb_cma_size)
7034 * hugetlb_cma_size is used to determine if allocations from
7035 * cma are possible. Set to zero if no cma regions are set up.
7037 hugetlb_cma_size = 0;
7040 void __init hugetlb_cma_check(void)
7042 if (!hugetlb_cma_size || cma_reserve_called)
7045 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
7048 #endif /* CONFIG_CMA */