1
2
3
4
5#include <linux/list.h>
6#include <linux/init.h>
7#include <linux/mm.h>
8#include <linux/seq_file.h>
9#include <linux/sysctl.h>
10#include <linux/highmem.h>
11#include <linux/mmu_notifier.h>
12#include <linux/nodemask.h>
13#include <linux/pagemap.h>
14#include <linux/mempolicy.h>
15#include <linux/compiler.h>
16#include <linux/cpuset.h>
17#include <linux/mutex.h>
18#include <linux/bootmem.h>
19#include <linux/sysfs.h>
20#include <linux/slab.h>
21#include <linux/mmdebug.h>
22#include <linux/sched/signal.h>
23#include <linux/rmap.h>
24#include <linux/string_helpers.h>
25#include <linux/swap.h>
26#include <linux/swapops.h>
27#include <linux/jhash.h>
28
29#include <asm/page.h>
30#include <asm/pgtable.h>
31#include <asm/tlb.h>
32
33#include <linux/io.h>
34#include <linux/hugetlb.h>
35#include <linux/hugetlb_cgroup.h>
36#include <linux/node.h>
37#include <linux/userfaultfd_k.h>
38#include <linux/page_owner.h>
39#include "internal.h"
40
41int hugetlb_max_hstate __read_mostly;
42unsigned int default_hstate_idx;
43struct hstate hstates[HUGE_MAX_HSTATE];
44
45
46
47
48static unsigned int minimum_order __read_mostly = UINT_MAX;
49
50__initdata LIST_HEAD(huge_boot_pages);
51
52
53static struct hstate * __initdata parsed_hstate;
54static unsigned long __initdata default_hstate_max_huge_pages;
55static unsigned long __initdata default_hstate_size;
56static bool __initdata parsed_valid_hugepagesz = true;
57
58
59
60
61
62DEFINE_SPINLOCK(hugetlb_lock);
63
64
65
66
67
68static int num_fault_mutexes;
69struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
70
71
72static int hugetlb_acct_memory(struct hstate *h, long delta);
73
74static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
75{
76 bool free = (spool->count == 0) && (spool->used_hpages == 0);
77
78 spin_unlock(&spool->lock);
79
80
81
82
83 if (free) {
84 if (spool->min_hpages != -1)
85 hugetlb_acct_memory(spool->hstate,
86 -spool->min_hpages);
87 kfree(spool);
88 }
89}
90
91struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
92 long min_hpages)
93{
94 struct hugepage_subpool *spool;
95
96 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
97 if (!spool)
98 return NULL;
99
100 spin_lock_init(&spool->lock);
101 spool->count = 1;
102 spool->max_hpages = max_hpages;
103 spool->hstate = h;
104 spool->min_hpages = min_hpages;
105
106 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
107 kfree(spool);
108 return NULL;
109 }
110 spool->rsv_hpages = min_hpages;
111
112 return spool;
113}
114
115void hugepage_put_subpool(struct hugepage_subpool *spool)
116{
117 spin_lock(&spool->lock);
118 BUG_ON(!spool->count);
119 spool->count--;
120 unlock_or_release_subpool(spool);
121}
122
123
124
125
126
127
128
129
130
131static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
132 long delta)
133{
134 long ret = delta;
135
136 if (!spool)
137 return ret;
138
139 spin_lock(&spool->lock);
140
141 if (spool->max_hpages != -1) {
142 if ((spool->used_hpages + delta) <= spool->max_hpages)
143 spool->used_hpages += delta;
144 else {
145 ret = -ENOMEM;
146 goto unlock_ret;
147 }
148 }
149
150
151 if (spool->min_hpages != -1 && spool->rsv_hpages) {
152 if (delta > spool->rsv_hpages) {
153
154
155
156
157 ret = delta - spool->rsv_hpages;
158 spool->rsv_hpages = 0;
159 } else {
160 ret = 0;
161 spool->rsv_hpages -= delta;
162 }
163 }
164
165unlock_ret:
166 spin_unlock(&spool->lock);
167 return ret;
168}
169
170
171
172
173
174
175
176static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
177 long delta)
178{
179 long ret = delta;
180
181 if (!spool)
182 return delta;
183
184 spin_lock(&spool->lock);
185
186 if (spool->max_hpages != -1)
187 spool->used_hpages -= delta;
188
189
190 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
191 if (spool->rsv_hpages + delta <= spool->min_hpages)
192 ret = 0;
193 else
194 ret = spool->rsv_hpages + delta - spool->min_hpages;
195
196 spool->rsv_hpages += delta;
197 if (spool->rsv_hpages > spool->min_hpages)
198 spool->rsv_hpages = spool->min_hpages;
199 }
200
201
202
203
204
205 unlock_or_release_subpool(spool);
206
207 return ret;
208}
209
210static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
211{
212 return HUGETLBFS_SB(inode->i_sb)->spool;
213}
214
215static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
216{
217 return subpool_inode(file_inode(vma->vm_file));
218}
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239struct file_region {
240 struct list_head link;
241 long from;
242 long to;
243};
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259static long region_add(struct resv_map *resv, long f, long t)
260{
261 struct list_head *head = &resv->regions;
262 struct file_region *rg, *nrg, *trg;
263 long add = 0;
264
265 spin_lock(&resv->lock);
266
267 list_for_each_entry(rg, head, link)
268 if (f <= rg->to)
269 break;
270
271
272
273
274
275
276
277 if (&rg->link == head || t < rg->from) {
278 VM_BUG_ON(resv->region_cache_count <= 0);
279
280 resv->region_cache_count--;
281 nrg = list_first_entry(&resv->region_cache, struct file_region,
282 link);
283 list_del(&nrg->link);
284
285 nrg->from = f;
286 nrg->to = t;
287 list_add(&nrg->link, rg->link.prev);
288
289 add += t - f;
290 goto out_locked;
291 }
292
293
294 if (f > rg->from)
295 f = rg->from;
296
297
298 nrg = rg;
299 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
300 if (&rg->link == head)
301 break;
302 if (rg->from > t)
303 break;
304
305
306
307
308 if (rg->to > t)
309 t = rg->to;
310 if (rg != nrg) {
311
312
313
314
315 add -= (rg->to - rg->from);
316 list_del(&rg->link);
317 kfree(rg);
318 }
319 }
320
321 add += (nrg->from - f);
322 nrg->from = f;
323 add += t - nrg->to;
324 nrg->to = t;
325
326out_locked:
327 resv->adds_in_progress--;
328 spin_unlock(&resv->lock);
329 VM_BUG_ON(add < 0);
330 return add;
331}
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355static long region_chg(struct resv_map *resv, long f, long t)
356{
357 struct list_head *head = &resv->regions;
358 struct file_region *rg, *nrg = NULL;
359 long chg = 0;
360
361retry:
362 spin_lock(&resv->lock);
363retry_locked:
364 resv->adds_in_progress++;
365
366
367
368
369
370 if (resv->adds_in_progress > resv->region_cache_count) {
371 struct file_region *trg;
372
373 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
374
375 resv->adds_in_progress--;
376 spin_unlock(&resv->lock);
377
378 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
379 if (!trg) {
380 kfree(nrg);
381 return -ENOMEM;
382 }
383
384 spin_lock(&resv->lock);
385 list_add(&trg->link, &resv->region_cache);
386 resv->region_cache_count++;
387 goto retry_locked;
388 }
389
390
391 list_for_each_entry(rg, head, link)
392 if (f <= rg->to)
393 break;
394
395
396
397
398 if (&rg->link == head || t < rg->from) {
399 if (!nrg) {
400 resv->adds_in_progress--;
401 spin_unlock(&resv->lock);
402 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
403 if (!nrg)
404 return -ENOMEM;
405
406 nrg->from = f;
407 nrg->to = f;
408 INIT_LIST_HEAD(&nrg->link);
409 goto retry;
410 }
411
412 list_add(&nrg->link, rg->link.prev);
413 chg = t - f;
414 goto out_nrg;
415 }
416
417
418 if (f > rg->from)
419 f = rg->from;
420 chg = t - f;
421
422
423 list_for_each_entry(rg, rg->link.prev, link) {
424 if (&rg->link == head)
425 break;
426 if (rg->from > t)
427 goto out;
428
429
430
431
432 if (rg->to > t) {
433 chg += rg->to - t;
434 t = rg->to;
435 }
436 chg -= rg->to - rg->from;
437 }
438
439out:
440 spin_unlock(&resv->lock);
441
442 kfree(nrg);
443 return chg;
444out_nrg:
445 spin_unlock(&resv->lock);
446 return chg;
447}
448
449
450
451
452
453
454
455
456
457
458
459
460static void region_abort(struct resv_map *resv, long f, long t)
461{
462 spin_lock(&resv->lock);
463 VM_BUG_ON(!resv->region_cache_count);
464 resv->adds_in_progress--;
465 spin_unlock(&resv->lock);
466}
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482static long region_del(struct resv_map *resv, long f, long t)
483{
484 struct list_head *head = &resv->regions;
485 struct file_region *rg, *trg;
486 struct file_region *nrg = NULL;
487 long del = 0;
488
489retry:
490 spin_lock(&resv->lock);
491 list_for_each_entry_safe(rg, trg, head, link) {
492
493
494
495
496
497
498
499 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
500 continue;
501
502 if (rg->from >= t)
503 break;
504
505 if (f > rg->from && t < rg->to) {
506
507
508
509
510 if (!nrg &&
511 resv->region_cache_count > resv->adds_in_progress) {
512 nrg = list_first_entry(&resv->region_cache,
513 struct file_region,
514 link);
515 list_del(&nrg->link);
516 resv->region_cache_count--;
517 }
518
519 if (!nrg) {
520 spin_unlock(&resv->lock);
521 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
522 if (!nrg)
523 return -ENOMEM;
524 goto retry;
525 }
526
527 del += t - f;
528
529
530 nrg->from = t;
531 nrg->to = rg->to;
532 INIT_LIST_HEAD(&nrg->link);
533
534
535 rg->to = f;
536
537 list_add(&nrg->link, &rg->link);
538 nrg = NULL;
539 break;
540 }
541
542 if (f <= rg->from && t >= rg->to) {
543 del += rg->to - rg->from;
544 list_del(&rg->link);
545 kfree(rg);
546 continue;
547 }
548
549 if (f <= rg->from) {
550 del += t - rg->from;
551 rg->from = t;
552 } else {
553 del += rg->to - f;
554 rg->to = f;
555 }
556 }
557
558 spin_unlock(&resv->lock);
559 kfree(nrg);
560 return del;
561}
562
563
564
565
566
567
568
569
570
571
572void hugetlb_fix_reserve_counts(struct inode *inode)
573{
574 struct hugepage_subpool *spool = subpool_inode(inode);
575 long rsv_adjust;
576
577 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
578 if (rsv_adjust) {
579 struct hstate *h = hstate_inode(inode);
580
581 hugetlb_acct_memory(h, 1);
582 }
583}
584
585
586
587
588
589static long region_count(struct resv_map *resv, long f, long t)
590{
591 struct list_head *head = &resv->regions;
592 struct file_region *rg;
593 long chg = 0;
594
595 spin_lock(&resv->lock);
596
597 list_for_each_entry(rg, head, link) {
598 long seg_from;
599 long seg_to;
600
601 if (rg->to <= f)
602 continue;
603 if (rg->from >= t)
604 break;
605
606 seg_from = max(rg->from, f);
607 seg_to = min(rg->to, t);
608
609 chg += seg_to - seg_from;
610 }
611 spin_unlock(&resv->lock);
612
613 return chg;
614}
615
616
617
618
619
620static pgoff_t vma_hugecache_offset(struct hstate *h,
621 struct vm_area_struct *vma, unsigned long address)
622{
623 return ((address - vma->vm_start) >> huge_page_shift(h)) +
624 (vma->vm_pgoff >> huge_page_order(h));
625}
626
627pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
628 unsigned long address)
629{
630 return vma_hugecache_offset(hstate_vma(vma), vma, address);
631}
632EXPORT_SYMBOL_GPL(linear_hugepage_index);
633
634
635
636
637
638unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
639{
640 struct hstate *hstate;
641
642 if (!is_vm_hugetlb_page(vma))
643 return PAGE_SIZE;
644
645 hstate = hstate_vma(vma);
646
647 return 1UL << huge_page_shift(hstate);
648}
649EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
650
651
652
653
654
655
656
657#ifndef vma_mmu_pagesize
658unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
659{
660 return vma_kernel_pagesize(vma);
661}
662#endif
663
664
665
666
667
668
669#define HPAGE_RESV_OWNER (1UL << 0)
670#define HPAGE_RESV_UNMAPPED (1UL << 1)
671#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692static unsigned long get_vma_private_data(struct vm_area_struct *vma)
693{
694 return (unsigned long)vma->vm_private_data;
695}
696
697static void set_vma_private_data(struct vm_area_struct *vma,
698 unsigned long value)
699{
700 vma->vm_private_data = (void *)value;
701}
702
703struct resv_map *resv_map_alloc(void)
704{
705 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
706 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
707
708 if (!resv_map || !rg) {
709 kfree(resv_map);
710 kfree(rg);
711 return NULL;
712 }
713
714 kref_init(&resv_map->refs);
715 spin_lock_init(&resv_map->lock);
716 INIT_LIST_HEAD(&resv_map->regions);
717
718 resv_map->adds_in_progress = 0;
719
720 INIT_LIST_HEAD(&resv_map->region_cache);
721 list_add(&rg->link, &resv_map->region_cache);
722 resv_map->region_cache_count = 1;
723
724 return resv_map;
725}
726
727void resv_map_release(struct kref *ref)
728{
729 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
730 struct list_head *head = &resv_map->region_cache;
731 struct file_region *rg, *trg;
732
733
734 region_del(resv_map, 0, LONG_MAX);
735
736
737 list_for_each_entry_safe(rg, trg, head, link) {
738 list_del(&rg->link);
739 kfree(rg);
740 }
741
742 VM_BUG_ON(resv_map->adds_in_progress);
743
744 kfree(resv_map);
745}
746
747static inline struct resv_map *inode_resv_map(struct inode *inode)
748{
749 return inode->i_mapping->private_data;
750}
751
752static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
753{
754 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
755 if (vma->vm_flags & VM_MAYSHARE) {
756 struct address_space *mapping = vma->vm_file->f_mapping;
757 struct inode *inode = mapping->host;
758
759 return inode_resv_map(inode);
760
761 } else {
762 return (struct resv_map *)(get_vma_private_data(vma) &
763 ~HPAGE_RESV_MASK);
764 }
765}
766
767static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
768{
769 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
770 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
771
772 set_vma_private_data(vma, (get_vma_private_data(vma) &
773 HPAGE_RESV_MASK) | (unsigned long)map);
774}
775
776static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
777{
778 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
779 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
780
781 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
782}
783
784static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
785{
786 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
787
788 return (get_vma_private_data(vma) & flag) != 0;
789}
790
791
792void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
793{
794 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
795 if (!(vma->vm_flags & VM_MAYSHARE))
796 vma->vm_private_data = (void *)0;
797}
798
799
800static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
801{
802 if (vma->vm_flags & VM_NORESERVE) {
803
804
805
806
807
808
809
810
811
812 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
813 return true;
814 else
815 return false;
816 }
817
818
819 if (vma->vm_flags & VM_MAYSHARE) {
820
821
822
823
824
825
826
827 if (chg)
828 return false;
829 else
830 return true;
831 }
832
833
834
835
836
837 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853 if (chg)
854 return false;
855 else
856 return true;
857 }
858
859 return false;
860}
861
862static void enqueue_huge_page(struct hstate *h, struct page *page)
863{
864 int nid = page_to_nid(page);
865 list_move(&page->lru, &h->hugepage_freelists[nid]);
866 h->free_huge_pages++;
867 h->free_huge_pages_node[nid]++;
868}
869
870static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
871{
872 struct page *page;
873
874 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
875 if (!PageHWPoison(page))
876 break;
877
878
879
880
881 if (&h->hugepage_freelists[nid] == &page->lru)
882 return NULL;
883 list_move(&page->lru, &h->hugepage_activelist);
884 set_page_refcounted(page);
885 h->free_huge_pages--;
886 h->free_huge_pages_node[nid]--;
887 return page;
888}
889
890static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
891 nodemask_t *nmask)
892{
893 unsigned int cpuset_mems_cookie;
894 struct zonelist *zonelist;
895 struct zone *zone;
896 struct zoneref *z;
897 int node = -1;
898
899 zonelist = node_zonelist(nid, gfp_mask);
900
901retry_cpuset:
902 cpuset_mems_cookie = read_mems_allowed_begin();
903 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
904 struct page *page;
905
906 if (!cpuset_zone_allowed(zone, gfp_mask))
907 continue;
908
909
910
911
912 if (zone_to_nid(zone) == node)
913 continue;
914 node = zone_to_nid(zone);
915
916 page = dequeue_huge_page_node_exact(h, node);
917 if (page)
918 return page;
919 }
920 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
921 goto retry_cpuset;
922
923 return NULL;
924}
925
926
927static inline gfp_t htlb_alloc_mask(struct hstate *h)
928{
929 if (hugepage_migration_supported(h))
930 return GFP_HIGHUSER_MOVABLE;
931 else
932 return GFP_HIGHUSER;
933}
934
935static struct page *dequeue_huge_page_vma(struct hstate *h,
936 struct vm_area_struct *vma,
937 unsigned long address, int avoid_reserve,
938 long chg)
939{
940 struct page *page;
941 struct mempolicy *mpol;
942 gfp_t gfp_mask;
943 nodemask_t *nodemask;
944 int nid;
945
946
947
948
949
950
951 if (!vma_has_reserves(vma, chg) &&
952 h->free_huge_pages - h->resv_huge_pages == 0)
953 goto err;
954
955
956 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
957 goto err;
958
959 gfp_mask = htlb_alloc_mask(h);
960 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
961 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
962 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
963 SetPagePrivate(page);
964 h->resv_huge_pages--;
965 }
966
967 mpol_cond_put(mpol);
968 return page;
969
970err:
971 return NULL;
972}
973
974
975
976
977
978
979
980
981static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
982{
983 nid = next_node_in(nid, *nodes_allowed);
984 VM_BUG_ON(nid >= MAX_NUMNODES);
985
986 return nid;
987}
988
989static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
990{
991 if (!node_isset(nid, *nodes_allowed))
992 nid = next_node_allowed(nid, nodes_allowed);
993 return nid;
994}
995
996
997
998
999
1000
1001
1002static int hstate_next_node_to_alloc(struct hstate *h,
1003 nodemask_t *nodes_allowed)
1004{
1005 int nid;
1006
1007 VM_BUG_ON(!nodes_allowed);
1008
1009 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1010 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1011
1012 return nid;
1013}
1014
1015
1016
1017
1018
1019
1020
1021static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1022{
1023 int nid;
1024
1025 VM_BUG_ON(!nodes_allowed);
1026
1027 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1028 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1029
1030 return nid;
1031}
1032
1033#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1034 for (nr_nodes = nodes_weight(*mask); \
1035 nr_nodes > 0 && \
1036 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1037 nr_nodes--)
1038
1039#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1040 for (nr_nodes = nodes_weight(*mask); \
1041 nr_nodes > 0 && \
1042 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1043 nr_nodes--)
1044
1045#ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1046static void destroy_compound_gigantic_page(struct page *page,
1047 unsigned int order)
1048{
1049 int i;
1050 int nr_pages = 1 << order;
1051 struct page *p = page + 1;
1052
1053 atomic_set(compound_mapcount_ptr(page), 0);
1054 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1055 clear_compound_head(p);
1056 set_page_refcounted(p);
1057 }
1058
1059 set_compound_order(page, 0);
1060 __ClearPageHead(page);
1061}
1062
1063static void free_gigantic_page(struct page *page, unsigned int order)
1064{
1065 free_contig_range(page_to_pfn(page), 1 << order);
1066}
1067
1068static int __alloc_gigantic_page(unsigned long start_pfn,
1069 unsigned long nr_pages, gfp_t gfp_mask)
1070{
1071 unsigned long end_pfn = start_pfn + nr_pages;
1072 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
1073 gfp_mask);
1074}
1075
1076static bool pfn_range_valid_gigantic(struct zone *z,
1077 unsigned long start_pfn, unsigned long nr_pages)
1078{
1079 unsigned long i, end_pfn = start_pfn + nr_pages;
1080 struct page *page;
1081
1082 for (i = start_pfn; i < end_pfn; i++) {
1083 if (!pfn_valid(i))
1084 return false;
1085
1086 page = pfn_to_page(i);
1087
1088 if (page_zone(page) != z)
1089 return false;
1090
1091 if (PageReserved(page))
1092 return false;
1093
1094 if (page_count(page) > 0)
1095 return false;
1096
1097 if (PageHuge(page))
1098 return false;
1099 }
1100
1101 return true;
1102}
1103
1104static bool zone_spans_last_pfn(const struct zone *zone,
1105 unsigned long start_pfn, unsigned long nr_pages)
1106{
1107 unsigned long last_pfn = start_pfn + nr_pages - 1;
1108 return zone_spans_pfn(zone, last_pfn);
1109}
1110
1111static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1112 int nid, nodemask_t *nodemask)
1113{
1114 unsigned int order = huge_page_order(h);
1115 unsigned long nr_pages = 1 << order;
1116 unsigned long ret, pfn, flags;
1117 struct zonelist *zonelist;
1118 struct zone *zone;
1119 struct zoneref *z;
1120
1121 zonelist = node_zonelist(nid, gfp_mask);
1122 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) {
1123 spin_lock_irqsave(&zone->lock, flags);
1124
1125 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
1126 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
1127 if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) {
1128
1129
1130
1131
1132
1133
1134
1135 spin_unlock_irqrestore(&zone->lock, flags);
1136 ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask);
1137 if (!ret)
1138 return pfn_to_page(pfn);
1139 spin_lock_irqsave(&zone->lock, flags);
1140 }
1141 pfn += nr_pages;
1142 }
1143
1144 spin_unlock_irqrestore(&zone->lock, flags);
1145 }
1146
1147 return NULL;
1148}
1149
1150static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1151static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1152
1153#else
1154static inline bool gigantic_page_supported(void) { return false; }
1155static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1156 int nid, nodemask_t *nodemask) { return NULL; }
1157static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1158static inline void destroy_compound_gigantic_page(struct page *page,
1159 unsigned int order) { }
1160#endif
1161
1162static void update_and_free_page(struct hstate *h, struct page *page)
1163{
1164 int i;
1165
1166 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1167 return;
1168
1169 h->nr_huge_pages--;
1170 h->nr_huge_pages_node[page_to_nid(page)]--;
1171 for (i = 0; i < pages_per_huge_page(h); i++) {
1172 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1173 1 << PG_referenced | 1 << PG_dirty |
1174 1 << PG_active | 1 << PG_private |
1175 1 << PG_writeback);
1176 }
1177 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1178 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1179 set_page_refcounted(page);
1180 if (hstate_is_gigantic(h)) {
1181 destroy_compound_gigantic_page(page, huge_page_order(h));
1182 free_gigantic_page(page, huge_page_order(h));
1183 } else {
1184 __free_pages(page, huge_page_order(h));
1185 }
1186}
1187
1188struct hstate *size_to_hstate(unsigned long size)
1189{
1190 struct hstate *h;
1191
1192 for_each_hstate(h) {
1193 if (huge_page_size(h) == size)
1194 return h;
1195 }
1196 return NULL;
1197}
1198
1199
1200
1201
1202
1203
1204
1205bool page_huge_active(struct page *page)
1206{
1207 VM_BUG_ON_PAGE(!PageHuge(page), page);
1208 return PageHead(page) && PagePrivate(&page[1]);
1209}
1210
1211
1212static void set_page_huge_active(struct page *page)
1213{
1214 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1215 SetPagePrivate(&page[1]);
1216}
1217
1218static void clear_page_huge_active(struct page *page)
1219{
1220 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1221 ClearPagePrivate(&page[1]);
1222}
1223
1224
1225
1226
1227
1228static inline bool PageHugeTemporary(struct page *page)
1229{
1230 if (!PageHuge(page))
1231 return false;
1232
1233 return (unsigned long)page[2].mapping == -1U;
1234}
1235
1236static inline void SetPageHugeTemporary(struct page *page)
1237{
1238 page[2].mapping = (void *)-1U;
1239}
1240
1241static inline void ClearPageHugeTemporary(struct page *page)
1242{
1243 page[2].mapping = NULL;
1244}
1245
1246void free_huge_page(struct page *page)
1247{
1248
1249
1250
1251
1252 struct hstate *h = page_hstate(page);
1253 int nid = page_to_nid(page);
1254 struct hugepage_subpool *spool =
1255 (struct hugepage_subpool *)page_private(page);
1256 bool restore_reserve;
1257
1258 set_page_private(page, 0);
1259 page->mapping = NULL;
1260 VM_BUG_ON_PAGE(page_count(page), page);
1261 VM_BUG_ON_PAGE(page_mapcount(page), page);
1262 restore_reserve = PagePrivate(page);
1263 ClearPagePrivate(page);
1264
1265
1266
1267
1268
1269
1270 if (hugepage_subpool_put_pages(spool, 1) == 0)
1271 restore_reserve = true;
1272
1273 spin_lock(&hugetlb_lock);
1274 clear_page_huge_active(page);
1275 hugetlb_cgroup_uncharge_page(hstate_index(h),
1276 pages_per_huge_page(h), page);
1277 if (restore_reserve)
1278 h->resv_huge_pages++;
1279
1280 if (PageHugeTemporary(page)) {
1281 list_del(&page->lru);
1282 ClearPageHugeTemporary(page);
1283 update_and_free_page(h, page);
1284 } else if (h->surplus_huge_pages_node[nid]) {
1285
1286 list_del(&page->lru);
1287 update_and_free_page(h, page);
1288 h->surplus_huge_pages--;
1289 h->surplus_huge_pages_node[nid]--;
1290 } else {
1291 arch_clear_hugepage_flags(page);
1292 enqueue_huge_page(h, page);
1293 }
1294 spin_unlock(&hugetlb_lock);
1295}
1296
1297static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1298{
1299 INIT_LIST_HEAD(&page->lru);
1300 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1301 spin_lock(&hugetlb_lock);
1302 set_hugetlb_cgroup(page, NULL);
1303 h->nr_huge_pages++;
1304 h->nr_huge_pages_node[nid]++;
1305 spin_unlock(&hugetlb_lock);
1306}
1307
1308static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1309{
1310 int i;
1311 int nr_pages = 1 << order;
1312 struct page *p = page + 1;
1313
1314
1315 set_compound_order(page, order);
1316 __ClearPageReserved(page);
1317 __SetPageHead(page);
1318 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331 __ClearPageReserved(p);
1332 set_page_count(p, 0);
1333 set_compound_head(p, page);
1334 }
1335 atomic_set(compound_mapcount_ptr(page), -1);
1336}
1337
1338
1339
1340
1341
1342
1343int PageHuge(struct page *page)
1344{
1345 if (!PageCompound(page))
1346 return 0;
1347
1348 page = compound_head(page);
1349 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1350}
1351EXPORT_SYMBOL_GPL(PageHuge);
1352
1353
1354
1355
1356
1357int PageHeadHuge(struct page *page_head)
1358{
1359 if (!PageHead(page_head))
1360 return 0;
1361
1362 return get_compound_page_dtor(page_head) == free_huge_page;
1363}
1364
1365pgoff_t __basepage_index(struct page *page)
1366{
1367 struct page *page_head = compound_head(page);
1368 pgoff_t index = page_index(page_head);
1369 unsigned long compound_idx;
1370
1371 if (!PageHuge(page_head))
1372 return page_index(page);
1373
1374 if (compound_order(page_head) >= MAX_ORDER)
1375 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1376 else
1377 compound_idx = page - page_head;
1378
1379 return (index << compound_order(page_head)) + compound_idx;
1380}
1381
1382static struct page *alloc_buddy_huge_page(struct hstate *h,
1383 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1384{
1385 int order = huge_page_order(h);
1386 struct page *page;
1387
1388 gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
1389 if (nid == NUMA_NO_NODE)
1390 nid = numa_mem_id();
1391 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1392 if (page)
1393 __count_vm_event(HTLB_BUDDY_PGALLOC);
1394 else
1395 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1396
1397 return page;
1398}
1399
1400
1401
1402
1403
1404static struct page *alloc_fresh_huge_page(struct hstate *h,
1405 gfp_t gfp_mask, int nid, nodemask_t *nmask)
1406{
1407 struct page *page;
1408
1409 if (hstate_is_gigantic(h))
1410 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1411 else
1412 page = alloc_buddy_huge_page(h, gfp_mask,
1413 nid, nmask);
1414 if (!page)
1415 return NULL;
1416
1417 if (hstate_is_gigantic(h))
1418 prep_compound_gigantic_page(page, huge_page_order(h));
1419 prep_new_huge_page(h, page, page_to_nid(page));
1420
1421 return page;
1422}
1423
1424
1425
1426
1427
1428static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1429{
1430 struct page *page;
1431 int nr_nodes, node;
1432 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1433
1434 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1435 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed);
1436 if (page)
1437 break;
1438 }
1439
1440 if (!page)
1441 return 0;
1442
1443 put_page(page);
1444
1445 return 1;
1446}
1447
1448
1449
1450
1451
1452
1453
1454static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1455 bool acct_surplus)
1456{
1457 int nr_nodes, node;
1458 int ret = 0;
1459
1460 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1461
1462
1463
1464
1465 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1466 !list_empty(&h->hugepage_freelists[node])) {
1467 struct page *page =
1468 list_entry(h->hugepage_freelists[node].next,
1469 struct page, lru);
1470 list_del(&page->lru);
1471 h->free_huge_pages--;
1472 h->free_huge_pages_node[node]--;
1473 if (acct_surplus) {
1474 h->surplus_huge_pages--;
1475 h->surplus_huge_pages_node[node]--;
1476 }
1477 update_and_free_page(h, page);
1478 ret = 1;
1479 break;
1480 }
1481 }
1482
1483 return ret;
1484}
1485
1486
1487
1488
1489
1490
1491
1492int dissolve_free_huge_page(struct page *page)
1493{
1494 int rc = 0;
1495
1496 spin_lock(&hugetlb_lock);
1497 if (PageHuge(page) && !page_count(page)) {
1498 struct page *head = compound_head(page);
1499 struct hstate *h = page_hstate(head);
1500 int nid = page_to_nid(head);
1501 if (h->free_huge_pages - h->resv_huge_pages == 0) {
1502 rc = -EBUSY;
1503 goto out;
1504 }
1505
1506
1507
1508
1509 if (PageHWPoison(head) && page != head) {
1510 SetPageHWPoison(page);
1511 ClearPageHWPoison(head);
1512 }
1513 list_del(&head->lru);
1514 h->free_huge_pages--;
1515 h->free_huge_pages_node[nid]--;
1516 h->max_huge_pages--;
1517 update_and_free_page(h, head);
1518 }
1519out:
1520 spin_unlock(&hugetlb_lock);
1521 return rc;
1522}
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1533{
1534 unsigned long pfn;
1535 struct page *page;
1536 int rc = 0;
1537
1538 if (!hugepages_supported())
1539 return rc;
1540
1541 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1542 page = pfn_to_page(pfn);
1543 if (PageHuge(page) && !page_count(page)) {
1544 rc = dissolve_free_huge_page(page);
1545 if (rc)
1546 break;
1547 }
1548 }
1549
1550 return rc;
1551}
1552
1553
1554
1555
1556static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1557 int nid, nodemask_t *nmask)
1558{
1559 struct page *page = NULL;
1560
1561 if (hstate_is_gigantic(h))
1562 return NULL;
1563
1564 spin_lock(&hugetlb_lock);
1565 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1566 goto out_unlock;
1567 spin_unlock(&hugetlb_lock);
1568
1569 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1570 if (!page)
1571 return NULL;
1572
1573 spin_lock(&hugetlb_lock);
1574
1575
1576
1577
1578
1579
1580
1581 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1582 SetPageHugeTemporary(page);
1583 put_page(page);
1584 page = NULL;
1585 } else {
1586 h->surplus_huge_pages++;
1587 h->surplus_huge_pages_node[page_to_nid(page)]++;
1588 }
1589
1590out_unlock:
1591 spin_unlock(&hugetlb_lock);
1592
1593 return page;
1594}
1595
1596static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1597 int nid, nodemask_t *nmask)
1598{
1599 struct page *page;
1600
1601 if (hstate_is_gigantic(h))
1602 return NULL;
1603
1604 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
1605 if (!page)
1606 return NULL;
1607
1608
1609
1610
1611
1612 SetPageHugeTemporary(page);
1613
1614 return page;
1615}
1616
1617
1618
1619
1620static
1621struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1622 struct vm_area_struct *vma, unsigned long addr)
1623{
1624 struct page *page;
1625 struct mempolicy *mpol;
1626 gfp_t gfp_mask = htlb_alloc_mask(h);
1627 int nid;
1628 nodemask_t *nodemask;
1629
1630 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1631 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1632 mpol_cond_put(mpol);
1633
1634 return page;
1635}
1636
1637
1638struct page *alloc_huge_page_node(struct hstate *h, int nid)
1639{
1640 gfp_t gfp_mask = htlb_alloc_mask(h);
1641 struct page *page = NULL;
1642
1643 if (nid != NUMA_NO_NODE)
1644 gfp_mask |= __GFP_THISNODE;
1645
1646 spin_lock(&hugetlb_lock);
1647 if (h->free_huge_pages - h->resv_huge_pages > 0)
1648 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1649 spin_unlock(&hugetlb_lock);
1650
1651 if (!page)
1652 page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1653
1654 return page;
1655}
1656
1657
1658struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1659 nodemask_t *nmask)
1660{
1661 gfp_t gfp_mask = htlb_alloc_mask(h);
1662
1663 spin_lock(&hugetlb_lock);
1664 if (h->free_huge_pages - h->resv_huge_pages > 0) {
1665 struct page *page;
1666
1667 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1668 if (page) {
1669 spin_unlock(&hugetlb_lock);
1670 return page;
1671 }
1672 }
1673 spin_unlock(&hugetlb_lock);
1674
1675 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1676}
1677
1678
1679struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1680 unsigned long address)
1681{
1682 struct mempolicy *mpol;
1683 nodemask_t *nodemask;
1684 struct page *page;
1685 gfp_t gfp_mask;
1686 int node;
1687
1688 gfp_mask = htlb_alloc_mask(h);
1689 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1690 page = alloc_huge_page_nodemask(h, node, nodemask);
1691 mpol_cond_put(mpol);
1692
1693 return page;
1694}
1695
1696
1697
1698
1699
1700static int gather_surplus_pages(struct hstate *h, int delta)
1701{
1702 struct list_head surplus_list;
1703 struct page *page, *tmp;
1704 int ret, i;
1705 int needed, allocated;
1706 bool alloc_ok = true;
1707
1708 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1709 if (needed <= 0) {
1710 h->resv_huge_pages += delta;
1711 return 0;
1712 }
1713
1714 allocated = 0;
1715 INIT_LIST_HEAD(&surplus_list);
1716
1717 ret = -ENOMEM;
1718retry:
1719 spin_unlock(&hugetlb_lock);
1720 for (i = 0; i < needed; i++) {
1721 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1722 NUMA_NO_NODE, NULL);
1723 if (!page) {
1724 alloc_ok = false;
1725 break;
1726 }
1727 list_add(&page->lru, &surplus_list);
1728 cond_resched();
1729 }
1730 allocated += i;
1731
1732
1733
1734
1735
1736 spin_lock(&hugetlb_lock);
1737 needed = (h->resv_huge_pages + delta) -
1738 (h->free_huge_pages + allocated);
1739 if (needed > 0) {
1740 if (alloc_ok)
1741 goto retry;
1742
1743
1744
1745
1746
1747 goto free;
1748 }
1749
1750
1751
1752
1753
1754
1755
1756
1757 needed += allocated;
1758 h->resv_huge_pages += delta;
1759 ret = 0;
1760
1761
1762 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1763 if ((--needed) < 0)
1764 break;
1765
1766
1767
1768
1769 put_page_testzero(page);
1770 VM_BUG_ON_PAGE(page_count(page), page);
1771 enqueue_huge_page(h, page);
1772 }
1773free:
1774 spin_unlock(&hugetlb_lock);
1775
1776
1777 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1778 put_page(page);
1779 spin_lock(&hugetlb_lock);
1780
1781 return ret;
1782}
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798static void return_unused_surplus_pages(struct hstate *h,
1799 unsigned long unused_resv_pages)
1800{
1801 unsigned long nr_pages;
1802
1803
1804 if (hstate_is_gigantic(h))
1805 goto out;
1806
1807
1808
1809
1810
1811 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825 while (nr_pages--) {
1826 h->resv_huge_pages--;
1827 unused_resv_pages--;
1828 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1829 goto out;
1830 cond_resched_lock(&hugetlb_lock);
1831 }
1832
1833out:
1834
1835 h->resv_huge_pages -= unused_resv_pages;
1836}
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863enum vma_resv_mode {
1864 VMA_NEEDS_RESV,
1865 VMA_COMMIT_RESV,
1866 VMA_END_RESV,
1867 VMA_ADD_RESV,
1868};
1869static long __vma_reservation_common(struct hstate *h,
1870 struct vm_area_struct *vma, unsigned long addr,
1871 enum vma_resv_mode mode)
1872{
1873 struct resv_map *resv;
1874 pgoff_t idx;
1875 long ret;
1876
1877 resv = vma_resv_map(vma);
1878 if (!resv)
1879 return 1;
1880
1881 idx = vma_hugecache_offset(h, vma, addr);
1882 switch (mode) {
1883 case VMA_NEEDS_RESV:
1884 ret = region_chg(resv, idx, idx + 1);
1885 break;
1886 case VMA_COMMIT_RESV:
1887 ret = region_add(resv, idx, idx + 1);
1888 break;
1889 case VMA_END_RESV:
1890 region_abort(resv, idx, idx + 1);
1891 ret = 0;
1892 break;
1893 case VMA_ADD_RESV:
1894 if (vma->vm_flags & VM_MAYSHARE)
1895 ret = region_add(resv, idx, idx + 1);
1896 else {
1897 region_abort(resv, idx, idx + 1);
1898 ret = region_del(resv, idx, idx + 1);
1899 }
1900 break;
1901 default:
1902 BUG();
1903 }
1904
1905 if (vma->vm_flags & VM_MAYSHARE)
1906 return ret;
1907 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921 if (ret)
1922 return 0;
1923 else
1924 return 1;
1925 }
1926 else
1927 return ret < 0 ? ret : 0;
1928}
1929
1930static long vma_needs_reservation(struct hstate *h,
1931 struct vm_area_struct *vma, unsigned long addr)
1932{
1933 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1934}
1935
1936static long vma_commit_reservation(struct hstate *h,
1937 struct vm_area_struct *vma, unsigned long addr)
1938{
1939 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1940}
1941
1942static void vma_end_reservation(struct hstate *h,
1943 struct vm_area_struct *vma, unsigned long addr)
1944{
1945 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1946}
1947
1948static long vma_add_reservation(struct hstate *h,
1949 struct vm_area_struct *vma, unsigned long addr)
1950{
1951 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1952}
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965static void restore_reserve_on_error(struct hstate *h,
1966 struct vm_area_struct *vma, unsigned long address,
1967 struct page *page)
1968{
1969 if (unlikely(PagePrivate(page))) {
1970 long rc = vma_needs_reservation(h, vma, address);
1971
1972 if (unlikely(rc < 0)) {
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984 ClearPagePrivate(page);
1985 } else if (rc) {
1986 rc = vma_add_reservation(h, vma, address);
1987 if (unlikely(rc < 0))
1988
1989
1990
1991
1992 ClearPagePrivate(page);
1993 } else
1994 vma_end_reservation(h, vma, address);
1995 }
1996}
1997
1998struct page *alloc_huge_page(struct vm_area_struct *vma,
1999 unsigned long addr, int avoid_reserve)
2000{
2001 struct hugepage_subpool *spool = subpool_vma(vma);
2002 struct hstate *h = hstate_vma(vma);
2003 struct page *page;
2004 long map_chg, map_commit;
2005 long gbl_chg;
2006 int ret, idx;
2007 struct hugetlb_cgroup *h_cg;
2008
2009 idx = hstate_index(h);
2010
2011
2012
2013
2014
2015 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2016 if (map_chg < 0)
2017 return ERR_PTR(-ENOMEM);
2018
2019
2020
2021
2022
2023
2024
2025
2026 if (map_chg || avoid_reserve) {
2027 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2028 if (gbl_chg < 0) {
2029 vma_end_reservation(h, vma, addr);
2030 return ERR_PTR(-ENOSPC);
2031 }
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041 if (avoid_reserve)
2042 gbl_chg = 1;
2043 }
2044
2045 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2046 if (ret)
2047 goto out_subpool_put;
2048
2049 spin_lock(&hugetlb_lock);
2050
2051
2052
2053
2054
2055 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2056 if (!page) {
2057 spin_unlock(&hugetlb_lock);
2058 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2059 if (!page)
2060 goto out_uncharge_cgroup;
2061 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2062 SetPagePrivate(page);
2063 h->resv_huge_pages--;
2064 }
2065 spin_lock(&hugetlb_lock);
2066 list_move(&page->lru, &h->hugepage_activelist);
2067
2068 }
2069 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2070 spin_unlock(&hugetlb_lock);
2071
2072 set_page_private(page, (unsigned long)spool);
2073
2074 map_commit = vma_commit_reservation(h, vma, addr);
2075 if (unlikely(map_chg > map_commit)) {
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085 long rsv_adjust;
2086
2087 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2088 hugetlb_acct_memory(h, -rsv_adjust);
2089 }
2090 return page;
2091
2092out_uncharge_cgroup:
2093 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2094out_subpool_put:
2095 if (map_chg || avoid_reserve)
2096 hugepage_subpool_put_pages(spool, 1);
2097 vma_end_reservation(h, vma, addr);
2098 return ERR_PTR(-ENOSPC);
2099}
2100
2101int alloc_bootmem_huge_page(struct hstate *h)
2102 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2103int __alloc_bootmem_huge_page(struct hstate *h)
2104{
2105 struct huge_bootmem_page *m;
2106 int nr_nodes, node;
2107
2108 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2109 void *addr;
2110
2111 addr = memblock_virt_alloc_try_nid_nopanic(
2112 huge_page_size(h), huge_page_size(h),
2113 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
2114 if (addr) {
2115
2116
2117
2118
2119
2120 m = addr;
2121 goto found;
2122 }
2123 }
2124 return 0;
2125
2126found:
2127 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2128
2129 list_add(&m->list, &huge_boot_pages);
2130 m->hstate = h;
2131 return 1;
2132}
2133
2134static void __init prep_compound_huge_page(struct page *page,
2135 unsigned int order)
2136{
2137 if (unlikely(order > (MAX_ORDER - 1)))
2138 prep_compound_gigantic_page(page, order);
2139 else
2140 prep_compound_page(page, order);
2141}
2142
2143
2144static void __init gather_bootmem_prealloc(void)
2145{
2146 struct huge_bootmem_page *m;
2147
2148 list_for_each_entry(m, &huge_boot_pages, list) {
2149 struct hstate *h = m->hstate;
2150 struct page *page;
2151
2152#ifdef CONFIG_HIGHMEM
2153 page = pfn_to_page(m->phys >> PAGE_SHIFT);
2154 memblock_free_late(__pa(m),
2155 sizeof(struct huge_bootmem_page));
2156#else
2157 page = virt_to_page(m);
2158#endif
2159 WARN_ON(page_count(page) != 1);
2160 prep_compound_huge_page(page, h->order);
2161 WARN_ON(PageReserved(page));
2162 prep_new_huge_page(h, page, page_to_nid(page));
2163 put_page(page);
2164
2165
2166
2167
2168
2169
2170
2171 if (hstate_is_gigantic(h))
2172 adjust_managed_page_count(page, 1 << h->order);
2173 }
2174}
2175
2176static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2177{
2178 unsigned long i;
2179
2180 for (i = 0; i < h->max_huge_pages; ++i) {
2181 if (hstate_is_gigantic(h)) {
2182 if (!alloc_bootmem_huge_page(h))
2183 break;
2184 } else if (!alloc_pool_huge_page(h,
2185 &node_states[N_MEMORY]))
2186 break;
2187 cond_resched();
2188 }
2189 if (i < h->max_huge_pages) {
2190 char buf[32];
2191
2192 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2193 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2194 h->max_huge_pages, buf, i);
2195 h->max_huge_pages = i;
2196 }
2197}
2198
2199static void __init hugetlb_init_hstates(void)
2200{
2201 struct hstate *h;
2202
2203 for_each_hstate(h) {
2204 if (minimum_order > huge_page_order(h))
2205 minimum_order = huge_page_order(h);
2206
2207
2208 if (!hstate_is_gigantic(h))
2209 hugetlb_hstate_alloc_pages(h);
2210 }
2211 VM_BUG_ON(minimum_order == UINT_MAX);
2212}
2213
2214static void __init report_hugepages(void)
2215{
2216 struct hstate *h;
2217
2218 for_each_hstate(h) {
2219 char buf[32];
2220
2221 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2222 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2223 buf, h->free_huge_pages);
2224 }
2225}
2226
2227#ifdef CONFIG_HIGHMEM
2228static void try_to_free_low(struct hstate *h, unsigned long count,
2229 nodemask_t *nodes_allowed)
2230{
2231 int i;
2232
2233 if (hstate_is_gigantic(h))
2234 return;
2235
2236 for_each_node_mask(i, *nodes_allowed) {
2237 struct page *page, *next;
2238 struct list_head *freel = &h->hugepage_freelists[i];
2239 list_for_each_entry_safe(page, next, freel, lru) {
2240 if (count >= h->nr_huge_pages)
2241 return;
2242 if (PageHighMem(page))
2243 continue;
2244 list_del(&page->lru);
2245 update_and_free_page(h, page);
2246 h->free_huge_pages--;
2247 h->free_huge_pages_node[page_to_nid(page)]--;
2248 }
2249 }
2250}
2251#else
2252static inline void try_to_free_low(struct hstate *h, unsigned long count,
2253 nodemask_t *nodes_allowed)
2254{
2255}
2256#endif
2257
2258
2259
2260
2261
2262
2263static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2264 int delta)
2265{
2266 int nr_nodes, node;
2267
2268 VM_BUG_ON(delta != -1 && delta != 1);
2269
2270 if (delta < 0) {
2271 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2272 if (h->surplus_huge_pages_node[node])
2273 goto found;
2274 }
2275 } else {
2276 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2277 if (h->surplus_huge_pages_node[node] <
2278 h->nr_huge_pages_node[node])
2279 goto found;
2280 }
2281 }
2282 return 0;
2283
2284found:
2285 h->surplus_huge_pages += delta;
2286 h->surplus_huge_pages_node[node] += delta;
2287 return 1;
2288}
2289
2290#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2291static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2292 nodemask_t *nodes_allowed)
2293{
2294 unsigned long min_count, ret;
2295
2296 if (hstate_is_gigantic(h) && !gigantic_page_supported())
2297 return h->max_huge_pages;
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310 spin_lock(&hugetlb_lock);
2311 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2312 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2313 break;
2314 }
2315
2316 while (count > persistent_huge_pages(h)) {
2317
2318
2319
2320
2321
2322 spin_unlock(&hugetlb_lock);
2323
2324
2325 cond_resched();
2326
2327 ret = alloc_pool_huge_page(h, nodes_allowed);
2328 spin_lock(&hugetlb_lock);
2329 if (!ret)
2330 goto out;
2331
2332
2333 if (signal_pending(current))
2334 goto out;
2335 }
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2353 min_count = max(count, min_count);
2354 try_to_free_low(h, min_count, nodes_allowed);
2355 while (min_count < persistent_huge_pages(h)) {
2356 if (!free_pool_huge_page(h, nodes_allowed, 0))
2357 break;
2358 cond_resched_lock(&hugetlb_lock);
2359 }
2360 while (count < persistent_huge_pages(h)) {
2361 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2362 break;
2363 }
2364out:
2365 ret = persistent_huge_pages(h);
2366 spin_unlock(&hugetlb_lock);
2367 return ret;
2368}
2369
2370#define HSTATE_ATTR_RO(_name) \
2371 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2372
2373#define HSTATE_ATTR(_name) \
2374 static struct kobj_attribute _name##_attr = \
2375 __ATTR(_name, 0644, _name##_show, _name##_store)
2376
2377static struct kobject *hugepages_kobj;
2378static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2379
2380static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2381
2382static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2383{
2384 int i;
2385
2386 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2387 if (hstate_kobjs[i] == kobj) {
2388 if (nidp)
2389 *nidp = NUMA_NO_NODE;
2390 return &hstates[i];
2391 }
2392
2393 return kobj_to_node_hstate(kobj, nidp);
2394}
2395
2396static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2397 struct kobj_attribute *attr, char *buf)
2398{
2399 struct hstate *h;
2400 unsigned long nr_huge_pages;
2401 int nid;
2402
2403 h = kobj_to_hstate(kobj, &nid);
2404 if (nid == NUMA_NO_NODE)
2405 nr_huge_pages = h->nr_huge_pages;
2406 else
2407 nr_huge_pages = h->nr_huge_pages_node[nid];
2408
2409 return sprintf(buf, "%lu\n", nr_huge_pages);
2410}
2411
2412static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2413 struct hstate *h, int nid,
2414 unsigned long count, size_t len)
2415{
2416 int err;
2417 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2418
2419 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2420 err = -EINVAL;
2421 goto out;
2422 }
2423
2424 if (nid == NUMA_NO_NODE) {
2425
2426
2427
2428 if (!(obey_mempolicy &&
2429 init_nodemask_of_mempolicy(nodes_allowed))) {
2430 NODEMASK_FREE(nodes_allowed);
2431 nodes_allowed = &node_states[N_MEMORY];
2432 }
2433 } else if (nodes_allowed) {
2434
2435
2436
2437
2438 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2439 init_nodemask_of_node(nodes_allowed, nid);
2440 } else
2441 nodes_allowed = &node_states[N_MEMORY];
2442
2443 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2444
2445 if (nodes_allowed != &node_states[N_MEMORY])
2446 NODEMASK_FREE(nodes_allowed);
2447
2448 return len;
2449out:
2450 NODEMASK_FREE(nodes_allowed);
2451 return err;
2452}
2453
2454static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2455 struct kobject *kobj, const char *buf,
2456 size_t len)
2457{
2458 struct hstate *h;
2459 unsigned long count;
2460 int nid;
2461 int err;
2462
2463 err = kstrtoul(buf, 10, &count);
2464 if (err)
2465 return err;
2466
2467 h = kobj_to_hstate(kobj, &nid);
2468 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2469}
2470
2471static ssize_t nr_hugepages_show(struct kobject *kobj,
2472 struct kobj_attribute *attr, char *buf)
2473{
2474 return nr_hugepages_show_common(kobj, attr, buf);
2475}
2476
2477static ssize_t nr_hugepages_store(struct kobject *kobj,
2478 struct kobj_attribute *attr, const char *buf, size_t len)
2479{
2480 return nr_hugepages_store_common(false, kobj, buf, len);
2481}
2482HSTATE_ATTR(nr_hugepages);
2483
2484#ifdef CONFIG_NUMA
2485
2486
2487
2488
2489
2490static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2491 struct kobj_attribute *attr, char *buf)
2492{
2493 return nr_hugepages_show_common(kobj, attr, buf);
2494}
2495
2496static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2497 struct kobj_attribute *attr, const char *buf, size_t len)
2498{
2499 return nr_hugepages_store_common(true, kobj, buf, len);
2500}
2501HSTATE_ATTR(nr_hugepages_mempolicy);
2502#endif
2503
2504
2505static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2506 struct kobj_attribute *attr, char *buf)
2507{
2508 struct hstate *h = kobj_to_hstate(kobj, NULL);
2509 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2510}
2511
2512static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2513 struct kobj_attribute *attr, const char *buf, size_t count)
2514{
2515 int err;
2516 unsigned long input;
2517 struct hstate *h = kobj_to_hstate(kobj, NULL);
2518
2519 if (hstate_is_gigantic(h))
2520 return -EINVAL;
2521
2522 err = kstrtoul(buf, 10, &input);
2523 if (err)
2524 return err;
2525
2526 spin_lock(&hugetlb_lock);
2527 h->nr_overcommit_huge_pages = input;
2528 spin_unlock(&hugetlb_lock);
2529
2530 return count;
2531}
2532HSTATE_ATTR(nr_overcommit_hugepages);
2533
2534static ssize_t free_hugepages_show(struct kobject *kobj,
2535 struct kobj_attribute *attr, char *buf)
2536{
2537 struct hstate *h;
2538 unsigned long free_huge_pages;
2539 int nid;
2540
2541 h = kobj_to_hstate(kobj, &nid);
2542 if (nid == NUMA_NO_NODE)
2543 free_huge_pages = h->free_huge_pages;
2544 else
2545 free_huge_pages = h->free_huge_pages_node[nid];
2546
2547 return sprintf(buf, "%lu\n", free_huge_pages);
2548}
2549HSTATE_ATTR_RO(free_hugepages);
2550
2551static ssize_t resv_hugepages_show(struct kobject *kobj,
2552 struct kobj_attribute *attr, char *buf)
2553{
2554 struct hstate *h = kobj_to_hstate(kobj, NULL);
2555 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2556}
2557HSTATE_ATTR_RO(resv_hugepages);
2558
2559static ssize_t surplus_hugepages_show(struct kobject *kobj,
2560 struct kobj_attribute *attr, char *buf)
2561{
2562 struct hstate *h;
2563 unsigned long surplus_huge_pages;
2564 int nid;
2565
2566 h = kobj_to_hstate(kobj, &nid);
2567 if (nid == NUMA_NO_NODE)
2568 surplus_huge_pages = h->surplus_huge_pages;
2569 else
2570 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2571
2572 return sprintf(buf, "%lu\n", surplus_huge_pages);
2573}
2574HSTATE_ATTR_RO(surplus_hugepages);
2575
2576static struct attribute *hstate_attrs[] = {
2577 &nr_hugepages_attr.attr,
2578 &nr_overcommit_hugepages_attr.attr,
2579 &free_hugepages_attr.attr,
2580 &resv_hugepages_attr.attr,
2581 &surplus_hugepages_attr.attr,
2582#ifdef CONFIG_NUMA
2583 &nr_hugepages_mempolicy_attr.attr,
2584#endif
2585 NULL,
2586};
2587
2588static const struct attribute_group hstate_attr_group = {
2589 .attrs = hstate_attrs,
2590};
2591
2592static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2593 struct kobject **hstate_kobjs,
2594 const struct attribute_group *hstate_attr_group)
2595{
2596 int retval;
2597 int hi = hstate_index(h);
2598
2599 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2600 if (!hstate_kobjs[hi])
2601 return -ENOMEM;
2602
2603 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2604 if (retval)
2605 kobject_put(hstate_kobjs[hi]);
2606
2607 return retval;
2608}
2609
2610static void __init hugetlb_sysfs_init(void)
2611{
2612 struct hstate *h;
2613 int err;
2614
2615 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2616 if (!hugepages_kobj)
2617 return;
2618
2619 for_each_hstate(h) {
2620 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2621 hstate_kobjs, &hstate_attr_group);
2622 if (err)
2623 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2624 }
2625}
2626
2627#ifdef CONFIG_NUMA
2628
2629
2630
2631
2632
2633
2634
2635
2636struct node_hstate {
2637 struct kobject *hugepages_kobj;
2638 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2639};
2640static struct node_hstate node_hstates[MAX_NUMNODES];
2641
2642
2643
2644
2645static struct attribute *per_node_hstate_attrs[] = {
2646 &nr_hugepages_attr.attr,
2647 &free_hugepages_attr.attr,
2648 &surplus_hugepages_attr.attr,
2649 NULL,
2650};
2651
2652static const struct attribute_group per_node_hstate_attr_group = {
2653 .attrs = per_node_hstate_attrs,
2654};
2655
2656
2657
2658
2659
2660static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2661{
2662 int nid;
2663
2664 for (nid = 0; nid < nr_node_ids; nid++) {
2665 struct node_hstate *nhs = &node_hstates[nid];
2666 int i;
2667 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2668 if (nhs->hstate_kobjs[i] == kobj) {
2669 if (nidp)
2670 *nidp = nid;
2671 return &hstates[i];
2672 }
2673 }
2674
2675 BUG();
2676 return NULL;
2677}
2678
2679
2680
2681
2682
2683static void hugetlb_unregister_node(struct node *node)
2684{
2685 struct hstate *h;
2686 struct node_hstate *nhs = &node_hstates[node->dev.id];
2687
2688 if (!nhs->hugepages_kobj)
2689 return;
2690
2691 for_each_hstate(h) {
2692 int idx = hstate_index(h);
2693 if (nhs->hstate_kobjs[idx]) {
2694 kobject_put(nhs->hstate_kobjs[idx]);
2695 nhs->hstate_kobjs[idx] = NULL;
2696 }
2697 }
2698
2699 kobject_put(nhs->hugepages_kobj);
2700 nhs->hugepages_kobj = NULL;
2701}
2702
2703
2704
2705
2706
2707
2708static void hugetlb_register_node(struct node *node)
2709{
2710 struct hstate *h;
2711 struct node_hstate *nhs = &node_hstates[node->dev.id];
2712 int err;
2713
2714 if (nhs->hugepages_kobj)
2715 return;
2716
2717 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2718 &node->dev.kobj);
2719 if (!nhs->hugepages_kobj)
2720 return;
2721
2722 for_each_hstate(h) {
2723 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2724 nhs->hstate_kobjs,
2725 &per_node_hstate_attr_group);
2726 if (err) {
2727 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2728 h->name, node->dev.id);
2729 hugetlb_unregister_node(node);
2730 break;
2731 }
2732 }
2733}
2734
2735
2736
2737
2738
2739
2740static void __init hugetlb_register_all_nodes(void)
2741{
2742 int nid;
2743
2744 for_each_node_state(nid, N_MEMORY) {
2745 struct node *node = node_devices[nid];
2746 if (node->dev.id == nid)
2747 hugetlb_register_node(node);
2748 }
2749
2750
2751
2752
2753
2754 register_hugetlbfs_with_node(hugetlb_register_node,
2755 hugetlb_unregister_node);
2756}
2757#else
2758
2759static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2760{
2761 BUG();
2762 if (nidp)
2763 *nidp = -1;
2764 return NULL;
2765}
2766
2767static void hugetlb_register_all_nodes(void) { }
2768
2769#endif
2770
2771static int __init hugetlb_init(void)
2772{
2773 int i;
2774
2775 if (!hugepages_supported())
2776 return 0;
2777
2778 if (!size_to_hstate(default_hstate_size)) {
2779 if (default_hstate_size != 0) {
2780 pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2781 default_hstate_size, HPAGE_SIZE);
2782 }
2783
2784 default_hstate_size = HPAGE_SIZE;
2785 if (!size_to_hstate(default_hstate_size))
2786 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2787 }
2788 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2789 if (default_hstate_max_huge_pages) {
2790 if (!default_hstate.max_huge_pages)
2791 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2792 }
2793
2794 hugetlb_init_hstates();
2795 gather_bootmem_prealloc();
2796 report_hugepages();
2797
2798 hugetlb_sysfs_init();
2799 hugetlb_register_all_nodes();
2800 hugetlb_cgroup_file_init();
2801
2802#ifdef CONFIG_SMP
2803 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2804#else
2805 num_fault_mutexes = 1;
2806#endif
2807 hugetlb_fault_mutex_table =
2808 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2809 BUG_ON(!hugetlb_fault_mutex_table);
2810
2811 for (i = 0; i < num_fault_mutexes; i++)
2812 mutex_init(&hugetlb_fault_mutex_table[i]);
2813 return 0;
2814}
2815subsys_initcall(hugetlb_init);
2816
2817
2818void __init hugetlb_bad_size(void)
2819{
2820 parsed_valid_hugepagesz = false;
2821}
2822
2823void __init hugetlb_add_hstate(unsigned int order)
2824{
2825 struct hstate *h;
2826 unsigned long i;
2827
2828 if (size_to_hstate(PAGE_SIZE << order)) {
2829 pr_warn("hugepagesz= specified twice, ignoring\n");
2830 return;
2831 }
2832 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2833 BUG_ON(order == 0);
2834 h = &hstates[hugetlb_max_hstate++];
2835 h->order = order;
2836 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2837 h->nr_huge_pages = 0;
2838 h->free_huge_pages = 0;
2839 for (i = 0; i < MAX_NUMNODES; ++i)
2840 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2841 INIT_LIST_HEAD(&h->hugepage_activelist);
2842 h->next_nid_to_alloc = first_memory_node;
2843 h->next_nid_to_free = first_memory_node;
2844 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2845 huge_page_size(h)/1024);
2846
2847 parsed_hstate = h;
2848}
2849
2850static int __init hugetlb_nrpages_setup(char *s)
2851{
2852 unsigned long *mhp;
2853 static unsigned long *last_mhp;
2854
2855 if (!parsed_valid_hugepagesz) {
2856 pr_warn("hugepages = %s preceded by "
2857 "an unsupported hugepagesz, ignoring\n", s);
2858 parsed_valid_hugepagesz = true;
2859 return 1;
2860 }
2861
2862
2863
2864
2865 else if (!hugetlb_max_hstate)
2866 mhp = &default_hstate_max_huge_pages;
2867 else
2868 mhp = &parsed_hstate->max_huge_pages;
2869
2870 if (mhp == last_mhp) {
2871 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2872 return 1;
2873 }
2874
2875 if (sscanf(s, "%lu", mhp) <= 0)
2876 *mhp = 0;
2877
2878
2879
2880
2881
2882
2883 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2884 hugetlb_hstate_alloc_pages(parsed_hstate);
2885
2886 last_mhp = mhp;
2887
2888 return 1;
2889}
2890__setup("hugepages=", hugetlb_nrpages_setup);
2891
2892static int __init hugetlb_default_setup(char *s)
2893{
2894 default_hstate_size = memparse(s, &s);
2895 return 1;
2896}
2897__setup("default_hugepagesz=", hugetlb_default_setup);
2898
2899static unsigned int cpuset_mems_nr(unsigned int *array)
2900{
2901 int node;
2902 unsigned int nr = 0;
2903
2904 for_each_node_mask(node, cpuset_current_mems_allowed)
2905 nr += array[node];
2906
2907 return nr;
2908}
2909
2910#ifdef CONFIG_SYSCTL
2911static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2912 struct ctl_table *table, int write,
2913 void __user *buffer, size_t *length, loff_t *ppos)
2914{
2915 struct hstate *h = &default_hstate;
2916 unsigned long tmp = h->max_huge_pages;
2917 int ret;
2918
2919 if (!hugepages_supported())
2920 return -EOPNOTSUPP;
2921
2922 table->data = &tmp;
2923 table->maxlen = sizeof(unsigned long);
2924 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2925 if (ret)
2926 goto out;
2927
2928 if (write)
2929 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2930 NUMA_NO_NODE, tmp, *length);
2931out:
2932 return ret;
2933}
2934
2935int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2936 void __user *buffer, size_t *length, loff_t *ppos)
2937{
2938
2939 return hugetlb_sysctl_handler_common(false, table, write,
2940 buffer, length, ppos);
2941}
2942
2943#ifdef CONFIG_NUMA
2944int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2945 void __user *buffer, size_t *length, loff_t *ppos)
2946{
2947 return hugetlb_sysctl_handler_common(true, table, write,
2948 buffer, length, ppos);
2949}
2950#endif
2951
2952int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2953 void __user *buffer,
2954 size_t *length, loff_t *ppos)
2955{
2956 struct hstate *h = &default_hstate;
2957 unsigned long tmp;
2958 int ret;
2959
2960 if (!hugepages_supported())
2961 return -EOPNOTSUPP;
2962
2963 tmp = h->nr_overcommit_huge_pages;
2964
2965 if (write && hstate_is_gigantic(h))
2966 return -EINVAL;
2967
2968 table->data = &tmp;
2969 table->maxlen = sizeof(unsigned long);
2970 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2971 if (ret)
2972 goto out;
2973
2974 if (write) {
2975 spin_lock(&hugetlb_lock);
2976 h->nr_overcommit_huge_pages = tmp;
2977 spin_unlock(&hugetlb_lock);
2978 }
2979out:
2980 return ret;
2981}
2982
2983#endif
2984
2985void hugetlb_report_meminfo(struct seq_file *m)
2986{
2987 struct hstate *h;
2988 unsigned long total = 0;
2989
2990 if (!hugepages_supported())
2991 return;
2992
2993 for_each_hstate(h) {
2994 unsigned long count = h->nr_huge_pages;
2995
2996 total += (PAGE_SIZE << huge_page_order(h)) * count;
2997
2998 if (h == &default_hstate)
2999 seq_printf(m,
3000 "HugePages_Total: %5lu\n"
3001 "HugePages_Free: %5lu\n"
3002 "HugePages_Rsvd: %5lu\n"
3003 "HugePages_Surp: %5lu\n"
3004 "Hugepagesize: %8lu kB\n",
3005 count,
3006 h->free_huge_pages,
3007 h->resv_huge_pages,
3008 h->surplus_huge_pages,
3009 (PAGE_SIZE << huge_page_order(h)) / 1024);
3010 }
3011
3012 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3013}
3014
3015int hugetlb_report_node_meminfo(int nid, char *buf)
3016{
3017 struct hstate *h = &default_hstate;
3018 if (!hugepages_supported())
3019 return 0;
3020 return sprintf(buf,
3021 "Node %d HugePages_Total: %5u\n"
3022 "Node %d HugePages_Free: %5u\n"
3023 "Node %d HugePages_Surp: %5u\n",
3024 nid, h->nr_huge_pages_node[nid],
3025 nid, h->free_huge_pages_node[nid],
3026 nid, h->surplus_huge_pages_node[nid]);
3027}
3028
3029void hugetlb_show_meminfo(void)
3030{
3031 struct hstate *h;
3032 int nid;
3033
3034 if (!hugepages_supported())
3035 return;
3036
3037 for_each_node_state(nid, N_MEMORY)
3038 for_each_hstate(h)
3039 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3040 nid,
3041 h->nr_huge_pages_node[nid],
3042 h->free_huge_pages_node[nid],
3043 h->surplus_huge_pages_node[nid],
3044 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3045}
3046
3047void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3048{
3049 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3050 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3051}
3052
3053
3054unsigned long hugetlb_total_pages(void)
3055{
3056 struct hstate *h;
3057 unsigned long nr_total_pages = 0;
3058
3059 for_each_hstate(h)
3060 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3061 return nr_total_pages;
3062}
3063
3064static int hugetlb_acct_memory(struct hstate *h, long delta)
3065{
3066 int ret = -ENOMEM;
3067
3068 spin_lock(&hugetlb_lock);
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086 if (delta > 0) {
3087 if (gather_surplus_pages(h, delta) < 0)
3088 goto out;
3089
3090 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3091 return_unused_surplus_pages(h, delta);
3092 goto out;
3093 }
3094 }
3095
3096 ret = 0;
3097 if (delta < 0)
3098 return_unused_surplus_pages(h, (unsigned long) -delta);
3099
3100out:
3101 spin_unlock(&hugetlb_lock);
3102 return ret;
3103}
3104
3105static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3106{
3107 struct resv_map *resv = vma_resv_map(vma);
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3118 kref_get(&resv->refs);
3119}
3120
3121static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3122{
3123 struct hstate *h = hstate_vma(vma);
3124 struct resv_map *resv = vma_resv_map(vma);
3125 struct hugepage_subpool *spool = subpool_vma(vma);
3126 unsigned long reserve, start, end;
3127 long gbl_reserve;
3128
3129 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3130 return;
3131
3132 start = vma_hugecache_offset(h, vma, vma->vm_start);
3133 end = vma_hugecache_offset(h, vma, vma->vm_end);
3134
3135 reserve = (end - start) - region_count(resv, start, end);
3136
3137 kref_put(&resv->refs, resv_map_release);
3138
3139 if (reserve) {
3140
3141
3142
3143
3144 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3145 hugetlb_acct_memory(h, -gbl_reserve);
3146 }
3147}
3148
3149static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3150{
3151 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3152 return -EINVAL;
3153 return 0;
3154}
3155
3156
3157
3158
3159
3160
3161
3162static int hugetlb_vm_op_fault(struct vm_fault *vmf)
3163{
3164 BUG();
3165 return 0;
3166}
3167
3168const struct vm_operations_struct hugetlb_vm_ops = {
3169 .fault = hugetlb_vm_op_fault,
3170 .open = hugetlb_vm_op_open,
3171 .close = hugetlb_vm_op_close,
3172 .split = hugetlb_vm_op_split,
3173};
3174
3175static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3176 int writable)
3177{
3178 pte_t entry;
3179
3180 if (writable) {
3181 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3182 vma->vm_page_prot)));
3183 } else {
3184 entry = huge_pte_wrprotect(mk_huge_pte(page,
3185 vma->vm_page_prot));
3186 }
3187 entry = pte_mkyoung(entry);
3188 entry = pte_mkhuge(entry);
3189 entry = arch_make_huge_pte(entry, vma, page, writable);
3190
3191 return entry;
3192}
3193
3194static void set_huge_ptep_writable(struct vm_area_struct *vma,
3195 unsigned long address, pte_t *ptep)
3196{
3197 pte_t entry;
3198
3199 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3200 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3201 update_mmu_cache(vma, address, ptep);
3202}
3203
3204bool is_hugetlb_entry_migration(pte_t pte)
3205{
3206 swp_entry_t swp;
3207
3208 if (huge_pte_none(pte) || pte_present(pte))
3209 return false;
3210 swp = pte_to_swp_entry(pte);
3211 if (non_swap_entry(swp) && is_migration_entry(swp))
3212 return true;
3213 else
3214 return false;
3215}
3216
3217static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3218{
3219 swp_entry_t swp;
3220
3221 if (huge_pte_none(pte) || pte_present(pte))
3222 return 0;
3223 swp = pte_to_swp_entry(pte);
3224 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3225 return 1;
3226 else
3227 return 0;
3228}
3229
3230int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3231 struct vm_area_struct *vma)
3232{
3233 pte_t *src_pte, *dst_pte, entry;
3234 struct page *ptepage;
3235 unsigned long addr;
3236 int cow;
3237 struct hstate *h = hstate_vma(vma);
3238 unsigned long sz = huge_page_size(h);
3239 unsigned long mmun_start;
3240 unsigned long mmun_end;
3241 int ret = 0;
3242
3243 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3244
3245 mmun_start = vma->vm_start;
3246 mmun_end = vma->vm_end;
3247 if (cow)
3248 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3249
3250 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3251 spinlock_t *src_ptl, *dst_ptl;
3252 src_pte = huge_pte_offset(src, addr, sz);
3253 if (!src_pte)
3254 continue;
3255 dst_pte = huge_pte_alloc(dst, addr, sz);
3256 if (!dst_pte) {
3257 ret = -ENOMEM;
3258 break;
3259 }
3260
3261
3262 if (dst_pte == src_pte)
3263 continue;
3264
3265 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3266 src_ptl = huge_pte_lockptr(h, src, src_pte);
3267 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3268 entry = huge_ptep_get(src_pte);
3269 if (huge_pte_none(entry)) {
3270 ;
3271 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3272 is_hugetlb_entry_hwpoisoned(entry))) {
3273 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3274
3275 if (is_write_migration_entry(swp_entry) && cow) {
3276
3277
3278
3279
3280 make_migration_entry_read(&swp_entry);
3281 entry = swp_entry_to_pte(swp_entry);
3282 set_huge_swap_pte_at(src, addr, src_pte,
3283 entry, sz);
3284 }
3285 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3286 } else {
3287 if (cow) {
3288
3289
3290
3291
3292
3293
3294
3295 huge_ptep_set_wrprotect(src, addr, src_pte);
3296 }
3297 entry = huge_ptep_get(src_pte);
3298 ptepage = pte_page(entry);
3299 get_page(ptepage);
3300 page_dup_rmap(ptepage, true);
3301 set_huge_pte_at(dst, addr, dst_pte, entry);
3302 hugetlb_count_add(pages_per_huge_page(h), dst);
3303 }
3304 spin_unlock(src_ptl);
3305 spin_unlock(dst_ptl);
3306 }
3307
3308 if (cow)
3309 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3310
3311 return ret;
3312}
3313
3314void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3315 unsigned long start, unsigned long end,
3316 struct page *ref_page)
3317{
3318 struct mm_struct *mm = vma->vm_mm;
3319 unsigned long address;
3320 pte_t *ptep;
3321 pte_t pte;
3322 spinlock_t *ptl;
3323 struct page *page;
3324 struct hstate *h = hstate_vma(vma);
3325 unsigned long sz = huge_page_size(h);
3326 const unsigned long mmun_start = start;
3327 const unsigned long mmun_end = end;
3328
3329 WARN_ON(!is_vm_hugetlb_page(vma));
3330 BUG_ON(start & ~huge_page_mask(h));
3331 BUG_ON(end & ~huge_page_mask(h));
3332
3333
3334
3335
3336
3337 tlb_remove_check_page_size_change(tlb, sz);
3338 tlb_start_vma(tlb, vma);
3339 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3340 address = start;
3341 for (; address < end; address += sz) {
3342 ptep = huge_pte_offset(mm, address, sz);
3343 if (!ptep)
3344 continue;
3345
3346 ptl = huge_pte_lock(h, mm, ptep);
3347 if (huge_pmd_unshare(mm, &address, ptep)) {
3348 spin_unlock(ptl);
3349 continue;
3350 }
3351
3352 pte = huge_ptep_get(ptep);
3353 if (huge_pte_none(pte)) {
3354 spin_unlock(ptl);
3355 continue;
3356 }
3357
3358
3359
3360
3361
3362 if (unlikely(!pte_present(pte))) {
3363 huge_pte_clear(mm, address, ptep, sz);
3364 spin_unlock(ptl);
3365 continue;
3366 }
3367
3368 page = pte_page(pte);
3369
3370
3371
3372
3373
3374 if (ref_page) {
3375 if (page != ref_page) {
3376 spin_unlock(ptl);
3377 continue;
3378 }
3379
3380
3381
3382
3383
3384 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3385 }
3386
3387 pte = huge_ptep_get_and_clear(mm, address, ptep);
3388 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3389 if (huge_pte_dirty(pte))
3390 set_page_dirty(page);
3391
3392 hugetlb_count_sub(pages_per_huge_page(h), mm);
3393 page_remove_rmap(page, true);
3394
3395 spin_unlock(ptl);
3396 tlb_remove_page_size(tlb, page, huge_page_size(h));
3397
3398
3399
3400 if (ref_page)
3401 break;
3402 }
3403 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3404 tlb_end_vma(tlb, vma);
3405}
3406
3407void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3408 struct vm_area_struct *vma, unsigned long start,
3409 unsigned long end, struct page *ref_page)
3410{
3411 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423 vma->vm_flags &= ~VM_MAYSHARE;
3424}
3425
3426void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3427 unsigned long end, struct page *ref_page)
3428{
3429 struct mm_struct *mm;
3430 struct mmu_gather tlb;
3431
3432 mm = vma->vm_mm;
3433
3434 tlb_gather_mmu(&tlb, mm, start, end);
3435 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3436 tlb_finish_mmu(&tlb, start, end);
3437}
3438
3439
3440
3441
3442
3443
3444
3445static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3446 struct page *page, unsigned long address)
3447{
3448 struct hstate *h = hstate_vma(vma);
3449 struct vm_area_struct *iter_vma;
3450 struct address_space *mapping;
3451 pgoff_t pgoff;
3452
3453
3454
3455
3456
3457 address = address & huge_page_mask(h);
3458 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3459 vma->vm_pgoff;
3460 mapping = vma->vm_file->f_mapping;
3461
3462
3463
3464
3465
3466
3467 i_mmap_lock_write(mapping);
3468 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3469
3470 if (iter_vma == vma)
3471 continue;
3472
3473
3474
3475
3476
3477
3478 if (iter_vma->vm_flags & VM_MAYSHARE)
3479 continue;
3480
3481
3482
3483
3484
3485
3486
3487
3488 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3489 unmap_hugepage_range(iter_vma, address,
3490 address + huge_page_size(h), page);
3491 }
3492 i_mmap_unlock_write(mapping);
3493}
3494
3495
3496
3497
3498
3499
3500
3501static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3502 unsigned long address, pte_t *ptep,
3503 struct page *pagecache_page, spinlock_t *ptl)
3504{
3505 pte_t pte;
3506 struct hstate *h = hstate_vma(vma);
3507 struct page *old_page, *new_page;
3508 int ret = 0, outside_reserve = 0;
3509 unsigned long mmun_start;
3510 unsigned long mmun_end;
3511
3512 pte = huge_ptep_get(ptep);
3513 old_page = pte_page(pte);
3514
3515retry_avoidcopy:
3516
3517
3518 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3519 page_move_anon_rmap(old_page, vma);
3520 set_huge_ptep_writable(vma, address, ptep);
3521 return 0;
3522 }
3523
3524
3525
3526
3527
3528
3529
3530
3531
3532
3533 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3534 old_page != pagecache_page)
3535 outside_reserve = 1;
3536
3537 get_page(old_page);
3538
3539
3540
3541
3542
3543 spin_unlock(ptl);
3544 new_page = alloc_huge_page(vma, address, outside_reserve);
3545
3546 if (IS_ERR(new_page)) {
3547
3548
3549
3550
3551
3552
3553
3554 if (outside_reserve) {
3555 put_page(old_page);
3556 BUG_ON(huge_pte_none(pte));
3557 unmap_ref_private(mm, vma, old_page, address);
3558 BUG_ON(huge_pte_none(pte));
3559 spin_lock(ptl);
3560 ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3561 huge_page_size(h));
3562 if (likely(ptep &&
3563 pte_same(huge_ptep_get(ptep), pte)))
3564 goto retry_avoidcopy;
3565
3566
3567
3568
3569 return 0;
3570 }
3571
3572 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3573 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3574 goto out_release_old;
3575 }
3576
3577
3578
3579
3580
3581 if (unlikely(anon_vma_prepare(vma))) {
3582 ret = VM_FAULT_OOM;
3583 goto out_release_all;
3584 }
3585
3586 copy_user_huge_page(new_page, old_page, address, vma,
3587 pages_per_huge_page(h));
3588 __SetPageUptodate(new_page);
3589 set_page_huge_active(new_page);
3590
3591 mmun_start = address & huge_page_mask(h);
3592 mmun_end = mmun_start + huge_page_size(h);
3593 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3594
3595
3596
3597
3598
3599 spin_lock(ptl);
3600 ptep = huge_pte_offset(mm, address & huge_page_mask(h),
3601 huge_page_size(h));
3602 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3603 ClearPagePrivate(new_page);
3604
3605
3606 huge_ptep_clear_flush(vma, address, ptep);
3607 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3608 set_huge_pte_at(mm, address, ptep,
3609 make_huge_pte(vma, new_page, 1));
3610 page_remove_rmap(old_page, true);
3611 hugepage_add_new_anon_rmap(new_page, vma, address);
3612
3613 new_page = old_page;
3614 }
3615 spin_unlock(ptl);
3616 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3617out_release_all:
3618 restore_reserve_on_error(h, vma, address, new_page);
3619 put_page(new_page);
3620out_release_old:
3621 put_page(old_page);
3622
3623 spin_lock(ptl);
3624 return ret;
3625}
3626
3627
3628static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3629 struct vm_area_struct *vma, unsigned long address)
3630{
3631 struct address_space *mapping;
3632 pgoff_t idx;
3633
3634 mapping = vma->vm_file->f_mapping;
3635 idx = vma_hugecache_offset(h, vma, address);
3636
3637 return find_lock_page(mapping, idx);
3638}
3639
3640
3641
3642
3643
3644static bool hugetlbfs_pagecache_present(struct hstate *h,
3645 struct vm_area_struct *vma, unsigned long address)
3646{
3647 struct address_space *mapping;
3648 pgoff_t idx;
3649 struct page *page;
3650
3651 mapping = vma->vm_file->f_mapping;
3652 idx = vma_hugecache_offset(h, vma, address);
3653
3654 page = find_get_page(mapping, idx);
3655 if (page)
3656 put_page(page);
3657 return page != NULL;
3658}
3659
3660int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3661 pgoff_t idx)
3662{
3663 struct inode *inode = mapping->host;
3664 struct hstate *h = hstate_inode(inode);
3665 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3666
3667 if (err)
3668 return err;
3669 ClearPagePrivate(page);
3670
3671 spin_lock(&inode->i_lock);
3672 inode->i_blocks += blocks_per_huge_page(h);
3673 spin_unlock(&inode->i_lock);
3674 return 0;
3675}
3676
3677static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3678 struct address_space *mapping, pgoff_t idx,
3679 unsigned long address, pte_t *ptep, unsigned int flags)
3680{
3681 struct hstate *h = hstate_vma(vma);
3682 int ret = VM_FAULT_SIGBUS;
3683 int anon_rmap = 0;
3684 unsigned long size;
3685 struct page *page;
3686 pte_t new_pte;
3687 spinlock_t *ptl;
3688
3689
3690
3691
3692
3693
3694 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3695 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3696 current->pid);
3697 return ret;
3698 }
3699
3700
3701
3702
3703
3704retry:
3705 page = find_lock_page(mapping, idx);
3706 if (!page) {
3707 size = i_size_read(mapping->host) >> huge_page_shift(h);
3708 if (idx >= size)
3709 goto out;
3710
3711
3712
3713
3714 if (userfaultfd_missing(vma)) {
3715 u32 hash;
3716 struct vm_fault vmf = {
3717 .vma = vma,
3718 .address = address,
3719 .flags = flags,
3720
3721
3722
3723
3724
3725
3726
3727 };
3728
3729
3730
3731
3732
3733
3734 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping,
3735 idx, address);
3736 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3737 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3738 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3739 goto out;
3740 }
3741
3742 page = alloc_huge_page(vma, address, 0);
3743 if (IS_ERR(page)) {
3744 ret = PTR_ERR(page);
3745 if (ret == -ENOMEM)
3746 ret = VM_FAULT_OOM;
3747 else
3748 ret = VM_FAULT_SIGBUS;
3749 goto out;
3750 }
3751 clear_huge_page(page, address, pages_per_huge_page(h));
3752 __SetPageUptodate(page);
3753 set_page_huge_active(page);
3754
3755 if (vma->vm_flags & VM_MAYSHARE) {
3756 int err = huge_add_to_page_cache(page, mapping, idx);
3757 if (err) {
3758 put_page(page);
3759 if (err == -EEXIST)
3760 goto retry;
3761 goto out;
3762 }
3763 } else {
3764 lock_page(page);
3765 if (unlikely(anon_vma_prepare(vma))) {
3766 ret = VM_FAULT_OOM;
3767 goto backout_unlocked;
3768 }
3769 anon_rmap = 1;
3770 }
3771 } else {
3772
3773
3774
3775
3776
3777 if (unlikely(PageHWPoison(page))) {
3778 ret = VM_FAULT_HWPOISON |
3779 VM_FAULT_SET_HINDEX(hstate_index(h));
3780 goto backout_unlocked;
3781 }
3782 }
3783
3784
3785
3786
3787
3788
3789
3790 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3791 if (vma_needs_reservation(h, vma, address) < 0) {
3792 ret = VM_FAULT_OOM;
3793 goto backout_unlocked;
3794 }
3795
3796 vma_end_reservation(h, vma, address);
3797 }
3798
3799 ptl = huge_pte_lock(h, mm, ptep);
3800 size = i_size_read(mapping->host) >> huge_page_shift(h);
3801 if (idx >= size)
3802 goto backout;
3803
3804 ret = 0;
3805 if (!huge_pte_none(huge_ptep_get(ptep)))
3806 goto backout;
3807
3808 if (anon_rmap) {
3809 ClearPagePrivate(page);
3810 hugepage_add_new_anon_rmap(page, vma, address);
3811 } else
3812 page_dup_rmap(page, true);
3813 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3814 && (vma->vm_flags & VM_SHARED)));
3815 set_huge_pte_at(mm, address, ptep, new_pte);
3816
3817 hugetlb_count_add(pages_per_huge_page(h), mm);
3818 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3819
3820 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
3821 }
3822
3823 spin_unlock(ptl);
3824 unlock_page(page);
3825out:
3826 return ret;
3827
3828backout:
3829 spin_unlock(ptl);
3830backout_unlocked:
3831 unlock_page(page);
3832 restore_reserve_on_error(h, vma, address, page);
3833 put_page(page);
3834 goto out;
3835}
3836
3837#ifdef CONFIG_SMP
3838u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3839 struct vm_area_struct *vma,
3840 struct address_space *mapping,
3841 pgoff_t idx, unsigned long address)
3842{
3843 unsigned long key[2];
3844 u32 hash;
3845
3846 if (vma->vm_flags & VM_SHARED) {
3847 key[0] = (unsigned long) mapping;
3848 key[1] = idx;
3849 } else {
3850 key[0] = (unsigned long) mm;
3851 key[1] = address >> huge_page_shift(h);
3852 }
3853
3854 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3855
3856 return hash & (num_fault_mutexes - 1);
3857}
3858#else
3859
3860
3861
3862
3863u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3864 struct vm_area_struct *vma,
3865 struct address_space *mapping,
3866 pgoff_t idx, unsigned long address)
3867{
3868 return 0;
3869}
3870#endif
3871
3872int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3873 unsigned long address, unsigned int flags)
3874{
3875 pte_t *ptep, entry;
3876 spinlock_t *ptl;
3877 int ret;
3878 u32 hash;
3879 pgoff_t idx;
3880 struct page *page = NULL;
3881 struct page *pagecache_page = NULL;
3882 struct hstate *h = hstate_vma(vma);
3883 struct address_space *mapping;
3884 int need_wait_lock = 0;
3885
3886 address &= huge_page_mask(h);
3887
3888 ptep = huge_pte_offset(mm, address, huge_page_size(h));
3889 if (ptep) {
3890 entry = huge_ptep_get(ptep);
3891 if (unlikely(is_hugetlb_entry_migration(entry))) {
3892 migration_entry_wait_huge(vma, mm, ptep);
3893 return 0;
3894 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3895 return VM_FAULT_HWPOISON_LARGE |
3896 VM_FAULT_SET_HINDEX(hstate_index(h));
3897 } else {
3898 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3899 if (!ptep)
3900 return VM_FAULT_OOM;
3901 }
3902
3903 mapping = vma->vm_file->f_mapping;
3904 idx = vma_hugecache_offset(h, vma, address);
3905
3906
3907
3908
3909
3910
3911 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3912 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3913
3914 entry = huge_ptep_get(ptep);
3915 if (huge_pte_none(entry)) {
3916 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3917 goto out_mutex;
3918 }
3919
3920 ret = 0;
3921
3922
3923
3924
3925
3926
3927
3928
3929 if (!pte_present(entry))
3930 goto out_mutex;
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3941 if (vma_needs_reservation(h, vma, address) < 0) {
3942 ret = VM_FAULT_OOM;
3943 goto out_mutex;
3944 }
3945
3946 vma_end_reservation(h, vma, address);
3947
3948 if (!(vma->vm_flags & VM_MAYSHARE))
3949 pagecache_page = hugetlbfs_pagecache_page(h,
3950 vma, address);
3951 }
3952
3953 ptl = huge_pte_lock(h, mm, ptep);
3954
3955
3956 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3957 goto out_ptl;
3958
3959
3960
3961
3962
3963
3964 page = pte_page(entry);
3965 if (page != pagecache_page)
3966 if (!trylock_page(page)) {
3967 need_wait_lock = 1;
3968 goto out_ptl;
3969 }
3970
3971 get_page(page);
3972
3973 if (flags & FAULT_FLAG_WRITE) {
3974 if (!huge_pte_write(entry)) {
3975 ret = hugetlb_cow(mm, vma, address, ptep,
3976 pagecache_page, ptl);
3977 goto out_put_page;
3978 }
3979 entry = huge_pte_mkdirty(entry);
3980 }
3981 entry = pte_mkyoung(entry);
3982 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3983 flags & FAULT_FLAG_WRITE))
3984 update_mmu_cache(vma, address, ptep);
3985out_put_page:
3986 if (page != pagecache_page)
3987 unlock_page(page);
3988 put_page(page);
3989out_ptl:
3990 spin_unlock(ptl);
3991
3992 if (pagecache_page) {
3993 unlock_page(pagecache_page);
3994 put_page(pagecache_page);
3995 }
3996out_mutex:
3997 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3998
3999
4000
4001
4002
4003
4004
4005 if (need_wait_lock)
4006 wait_on_page_locked(page);
4007 return ret;
4008}
4009
4010
4011
4012
4013
4014int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4015 pte_t *dst_pte,
4016 struct vm_area_struct *dst_vma,
4017 unsigned long dst_addr,
4018 unsigned long src_addr,
4019 struct page **pagep)
4020{
4021 struct address_space *mapping;
4022 pgoff_t idx;
4023 unsigned long size;
4024 int vm_shared = dst_vma->vm_flags & VM_SHARED;
4025 struct hstate *h = hstate_vma(dst_vma);
4026 pte_t _dst_pte;
4027 spinlock_t *ptl;
4028 int ret;
4029 struct page *page;
4030
4031 if (!*pagep) {
4032 ret = -ENOMEM;
4033 page = alloc_huge_page(dst_vma, dst_addr, 0);
4034 if (IS_ERR(page))
4035 goto out;
4036
4037 ret = copy_huge_page_from_user(page,
4038 (const void __user *) src_addr,
4039 pages_per_huge_page(h), false);
4040
4041
4042 if (unlikely(ret)) {
4043 ret = -EFAULT;
4044 *pagep = page;
4045
4046 goto out;
4047 }
4048 } else {
4049 page = *pagep;
4050 *pagep = NULL;
4051 }
4052
4053
4054
4055
4056
4057
4058 __SetPageUptodate(page);
4059 set_page_huge_active(page);
4060
4061 mapping = dst_vma->vm_file->f_mapping;
4062 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4063
4064
4065
4066
4067 if (vm_shared) {
4068 size = i_size_read(mapping->host) >> huge_page_shift(h);
4069 ret = -EFAULT;
4070 if (idx >= size)
4071 goto out_release_nounlock;
4072
4073
4074
4075
4076
4077
4078
4079 ret = huge_add_to_page_cache(page, mapping, idx);
4080 if (ret)
4081 goto out_release_nounlock;
4082 }
4083
4084 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4085 spin_lock(ptl);
4086
4087
4088
4089
4090
4091
4092
4093
4094
4095
4096 size = i_size_read(mapping->host) >> huge_page_shift(h);
4097 ret = -EFAULT;
4098 if (idx >= size)
4099 goto out_release_unlock;
4100
4101 ret = -EEXIST;
4102 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4103 goto out_release_unlock;
4104
4105 if (vm_shared) {
4106 page_dup_rmap(page, true);
4107 } else {
4108 ClearPagePrivate(page);
4109 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4110 }
4111
4112 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4113 if (dst_vma->vm_flags & VM_WRITE)
4114 _dst_pte = huge_pte_mkdirty(_dst_pte);
4115 _dst_pte = pte_mkyoung(_dst_pte);
4116
4117 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4118
4119 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4120 dst_vma->vm_flags & VM_WRITE);
4121 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4122
4123
4124 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4125
4126 spin_unlock(ptl);
4127 if (vm_shared)
4128 unlock_page(page);
4129 ret = 0;
4130out:
4131 return ret;
4132out_release_unlock:
4133 spin_unlock(ptl);
4134 if (vm_shared)
4135 unlock_page(page);
4136out_release_nounlock:
4137 put_page(page);
4138 goto out;
4139}
4140
4141long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4142 struct page **pages, struct vm_area_struct **vmas,
4143 unsigned long *position, unsigned long *nr_pages,
4144 long i, unsigned int flags, int *nonblocking)
4145{
4146 unsigned long pfn_offset;
4147 unsigned long vaddr = *position;
4148 unsigned long remainder = *nr_pages;
4149 struct hstate *h = hstate_vma(vma);
4150 int err = -EFAULT;
4151
4152 while (vaddr < vma->vm_end && remainder) {
4153 pte_t *pte;
4154 spinlock_t *ptl = NULL;
4155 int absent;
4156 struct page *page;
4157
4158
4159
4160
4161
4162 if (unlikely(fatal_signal_pending(current))) {
4163 remainder = 0;
4164 break;
4165 }
4166
4167
4168
4169
4170
4171
4172
4173
4174 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4175 huge_page_size(h));
4176 if (pte)
4177 ptl = huge_pte_lock(h, mm, pte);
4178 absent = !pte || huge_pte_none(huge_ptep_get(pte));
4179
4180
4181
4182
4183
4184
4185
4186
4187 if (absent && (flags & FOLL_DUMP) &&
4188 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4189 if (pte)
4190 spin_unlock(ptl);
4191 remainder = 0;
4192 break;
4193 }
4194
4195
4196
4197
4198
4199
4200
4201
4202
4203
4204
4205 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4206 ((flags & FOLL_WRITE) &&
4207 !huge_pte_write(huge_ptep_get(pte)))) {
4208 int ret;
4209 unsigned int fault_flags = 0;
4210
4211 if (pte)
4212 spin_unlock(ptl);
4213 if (flags & FOLL_WRITE)
4214 fault_flags |= FAULT_FLAG_WRITE;
4215 if (nonblocking)
4216 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4217 if (flags & FOLL_NOWAIT)
4218 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4219 FAULT_FLAG_RETRY_NOWAIT;
4220 if (flags & FOLL_TRIED) {
4221 VM_WARN_ON_ONCE(fault_flags &
4222 FAULT_FLAG_ALLOW_RETRY);
4223 fault_flags |= FAULT_FLAG_TRIED;
4224 }
4225 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4226 if (ret & VM_FAULT_ERROR) {
4227 err = vm_fault_to_errno(ret, flags);
4228 remainder = 0;
4229 break;
4230 }
4231 if (ret & VM_FAULT_RETRY) {
4232 if (nonblocking)
4233 *nonblocking = 0;
4234 *nr_pages = 0;
4235
4236
4237
4238
4239
4240
4241
4242
4243
4244 return i;
4245 }
4246 continue;
4247 }
4248
4249 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4250 page = pte_page(huge_ptep_get(pte));
4251same_page:
4252 if (pages) {
4253 pages[i] = mem_map_offset(page, pfn_offset);
4254 get_page(pages[i]);
4255 }
4256
4257 if (vmas)
4258 vmas[i] = vma;
4259
4260 vaddr += PAGE_SIZE;
4261 ++pfn_offset;
4262 --remainder;
4263 ++i;
4264 if (vaddr < vma->vm_end && remainder &&
4265 pfn_offset < pages_per_huge_page(h)) {
4266
4267
4268
4269
4270 goto same_page;
4271 }
4272 spin_unlock(ptl);
4273 }
4274 *nr_pages = remainder;
4275
4276
4277
4278
4279
4280 *position = vaddr;
4281
4282 return i ? i : err;
4283}
4284
4285#ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4286
4287
4288
4289
4290#define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4291#endif
4292
4293unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4294 unsigned long address, unsigned long end, pgprot_t newprot)
4295{
4296 struct mm_struct *mm = vma->vm_mm;
4297 unsigned long start = address;
4298 pte_t *ptep;
4299 pte_t pte;
4300 struct hstate *h = hstate_vma(vma);
4301 unsigned long pages = 0;
4302
4303 BUG_ON(address >= end);
4304 flush_cache_range(vma, address, end);
4305
4306 mmu_notifier_invalidate_range_start(mm, start, end);
4307 i_mmap_lock_write(vma->vm_file->f_mapping);
4308 for (; address < end; address += huge_page_size(h)) {
4309 spinlock_t *ptl;
4310 ptep = huge_pte_offset(mm, address, huge_page_size(h));
4311 if (!ptep)
4312 continue;
4313 ptl = huge_pte_lock(h, mm, ptep);
4314 if (huge_pmd_unshare(mm, &address, ptep)) {
4315 pages++;
4316 spin_unlock(ptl);
4317 continue;
4318 }
4319 pte = huge_ptep_get(ptep);
4320 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4321 spin_unlock(ptl);
4322 continue;
4323 }
4324 if (unlikely(is_hugetlb_entry_migration(pte))) {
4325 swp_entry_t entry = pte_to_swp_entry(pte);
4326
4327 if (is_write_migration_entry(entry)) {
4328 pte_t newpte;
4329
4330 make_migration_entry_read(&entry);
4331 newpte = swp_entry_to_pte(entry);
4332 set_huge_swap_pte_at(mm, address, ptep,
4333 newpte, huge_page_size(h));
4334 pages++;
4335 }
4336 spin_unlock(ptl);
4337 continue;
4338 }
4339 if (!huge_pte_none(pte)) {
4340 pte = huge_ptep_get_and_clear(mm, address, ptep);
4341 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
4342 pte = arch_make_huge_pte(pte, vma, NULL, 0);
4343 set_huge_pte_at(mm, address, ptep, pte);
4344 pages++;
4345 }
4346 spin_unlock(ptl);
4347 }
4348
4349
4350
4351
4352
4353
4354 flush_hugetlb_tlb_range(vma, start, end);
4355
4356
4357
4358
4359
4360
4361 i_mmap_unlock_write(vma->vm_file->f_mapping);
4362 mmu_notifier_invalidate_range_end(mm, start, end);
4363
4364 return pages << h->order;
4365}
4366
4367int hugetlb_reserve_pages(struct inode *inode,
4368 long from, long to,
4369 struct vm_area_struct *vma,
4370 vm_flags_t vm_flags)
4371{
4372 long ret, chg;
4373 struct hstate *h = hstate_inode(inode);
4374 struct hugepage_subpool *spool = subpool_inode(inode);
4375 struct resv_map *resv_map;
4376 long gbl_reserve;
4377
4378
4379 if (from > to) {
4380 VM_WARN(1, "%s called with a negative range\n", __func__);
4381 return -EINVAL;
4382 }
4383
4384
4385
4386
4387
4388
4389 if (vm_flags & VM_NORESERVE)
4390 return 0;
4391
4392
4393
4394
4395
4396
4397
4398 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4399 resv_map = inode_resv_map(inode);
4400
4401 chg = region_chg(resv_map, from, to);
4402
4403 } else {
4404 resv_map = resv_map_alloc();
4405 if (!resv_map)
4406 return -ENOMEM;
4407
4408 chg = to - from;
4409
4410 set_vma_resv_map(vma, resv_map);
4411 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4412 }
4413
4414 if (chg < 0) {
4415 ret = chg;
4416 goto out_err;
4417 }
4418
4419
4420
4421
4422
4423
4424 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4425 if (gbl_reserve < 0) {
4426 ret = -ENOSPC;
4427 goto out_err;
4428 }
4429
4430
4431
4432
4433
4434 ret = hugetlb_acct_memory(h, gbl_reserve);
4435 if (ret < 0) {
4436
4437 (void)hugepage_subpool_put_pages(spool, chg);
4438 goto out_err;
4439 }
4440
4441
4442
4443
4444
4445
4446
4447
4448
4449
4450
4451
4452 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4453 long add = region_add(resv_map, from, to);
4454
4455 if (unlikely(chg > add)) {
4456
4457
4458
4459
4460
4461
4462
4463 long rsv_adjust;
4464
4465 rsv_adjust = hugepage_subpool_put_pages(spool,
4466 chg - add);
4467 hugetlb_acct_memory(h, -rsv_adjust);
4468 }
4469 }
4470 return 0;
4471out_err:
4472 if (!vma || vma->vm_flags & VM_MAYSHARE)
4473
4474 if (chg >= 0)
4475 region_abort(resv_map, from, to);
4476 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4477 kref_put(&resv_map->refs, resv_map_release);
4478 return ret;
4479}
4480
4481long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4482 long freed)
4483{
4484 struct hstate *h = hstate_inode(inode);
4485 struct resv_map *resv_map = inode_resv_map(inode);
4486 long chg = 0;
4487 struct hugepage_subpool *spool = subpool_inode(inode);
4488 long gbl_reserve;
4489
4490 if (resv_map) {
4491 chg = region_del(resv_map, start, end);
4492
4493
4494
4495
4496
4497 if (chg < 0)
4498 return chg;
4499 }
4500
4501 spin_lock(&inode->i_lock);
4502 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4503 spin_unlock(&inode->i_lock);
4504
4505
4506
4507
4508
4509 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4510 hugetlb_acct_memory(h, -gbl_reserve);
4511
4512 return 0;
4513}
4514
4515#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4516static unsigned long page_table_shareable(struct vm_area_struct *svma,
4517 struct vm_area_struct *vma,
4518 unsigned long addr, pgoff_t idx)
4519{
4520 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4521 svma->vm_start;
4522 unsigned long sbase = saddr & PUD_MASK;
4523 unsigned long s_end = sbase + PUD_SIZE;
4524
4525
4526 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4527 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4528
4529
4530
4531
4532
4533 if (pmd_index(addr) != pmd_index(saddr) ||
4534 vm_flags != svm_flags ||
4535 sbase < svma->vm_start || svma->vm_end < s_end)
4536 return 0;
4537
4538 return saddr;
4539}
4540
4541static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4542{
4543 unsigned long base = addr & PUD_MASK;
4544 unsigned long end = base + PUD_SIZE;
4545
4546
4547
4548
4549 if (vma->vm_flags & VM_MAYSHARE &&
4550 vma->vm_start <= base && end <= vma->vm_end)
4551 return true;
4552 return false;
4553}
4554
4555
4556
4557
4558
4559
4560
4561
4562
4563
4564pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4565{
4566 struct vm_area_struct *vma = find_vma(mm, addr);
4567 struct address_space *mapping = vma->vm_file->f_mapping;
4568 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4569 vma->vm_pgoff;
4570 struct vm_area_struct *svma;
4571 unsigned long saddr;
4572 pte_t *spte = NULL;
4573 pte_t *pte;
4574 spinlock_t *ptl;
4575
4576 if (!vma_shareable(vma, addr))
4577 return (pte_t *)pmd_alloc(mm, pud, addr);
4578
4579 i_mmap_lock_write(mapping);
4580 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4581 if (svma == vma)
4582 continue;
4583
4584 saddr = page_table_shareable(svma, vma, addr, idx);
4585 if (saddr) {
4586 spte = huge_pte_offset(svma->vm_mm, saddr,
4587 vma_mmu_pagesize(svma));
4588 if (spte) {
4589 get_page(virt_to_page(spte));
4590 break;
4591 }
4592 }
4593 }
4594
4595 if (!spte)
4596 goto out;
4597
4598 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4599 if (pud_none(*pud)) {
4600 pud_populate(mm, pud,
4601 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4602 mm_inc_nr_pmds(mm);
4603 } else {
4604 put_page(virt_to_page(spte));
4605 }
4606 spin_unlock(ptl);
4607out:
4608 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4609 i_mmap_unlock_write(mapping);
4610 return pte;
4611}
4612
4613
4614
4615
4616
4617
4618
4619
4620
4621
4622
4623
4624
4625int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4626{
4627 pgd_t *pgd = pgd_offset(mm, *addr);
4628 p4d_t *p4d = p4d_offset(pgd, *addr);
4629 pud_t *pud = pud_offset(p4d, *addr);
4630
4631 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4632 if (page_count(virt_to_page(ptep)) == 1)
4633 return 0;
4634
4635 pud_clear(pud);
4636 put_page(virt_to_page(ptep));
4637 mm_dec_nr_pmds(mm);
4638 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4639 return 1;
4640}
4641#define want_pmd_share() (1)
4642#else
4643pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4644{
4645 return NULL;
4646}
4647
4648int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4649{
4650 return 0;
4651}
4652#define want_pmd_share() (0)
4653#endif
4654
4655#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4656pte_t *huge_pte_alloc(struct mm_struct *mm,
4657 unsigned long addr, unsigned long sz)
4658{
4659 pgd_t *pgd;
4660 p4d_t *p4d;
4661 pud_t *pud;
4662 pte_t *pte = NULL;
4663
4664 pgd = pgd_offset(mm, addr);
4665 p4d = p4d_alloc(mm, pgd, addr);
4666 if (!p4d)
4667 return NULL;
4668 pud = pud_alloc(mm, p4d, addr);
4669 if (pud) {
4670 if (sz == PUD_SIZE) {
4671 pte = (pte_t *)pud;
4672 } else {
4673 BUG_ON(sz != PMD_SIZE);
4674 if (want_pmd_share() && pud_none(*pud))
4675 pte = huge_pmd_share(mm, addr, pud);
4676 else
4677 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4678 }
4679 }
4680 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
4681
4682 return pte;
4683}
4684
4685
4686
4687
4688
4689
4690
4691
4692
4693
4694pte_t *huge_pte_offset(struct mm_struct *mm,
4695 unsigned long addr, unsigned long sz)
4696{
4697 pgd_t *pgd;
4698 p4d_t *p4d;
4699 pud_t *pud;
4700 pmd_t *pmd;
4701
4702 pgd = pgd_offset(mm, addr);
4703 if (!pgd_present(*pgd))
4704 return NULL;
4705 p4d = p4d_offset(pgd, addr);
4706 if (!p4d_present(*p4d))
4707 return NULL;
4708
4709 pud = pud_offset(p4d, addr);
4710 if (sz != PUD_SIZE && pud_none(*pud))
4711 return NULL;
4712
4713 if (pud_huge(*pud) || !pud_present(*pud))
4714 return (pte_t *)pud;
4715
4716 pmd = pmd_offset(pud, addr);
4717 if (sz != PMD_SIZE && pmd_none(*pmd))
4718 return NULL;
4719
4720 if (pmd_huge(*pmd) || !pmd_present(*pmd))
4721 return (pte_t *)pmd;
4722
4723 return NULL;
4724}
4725
4726#endif
4727
4728
4729
4730
4731
4732struct page * __weak
4733follow_huge_addr(struct mm_struct *mm, unsigned long address,
4734 int write)
4735{
4736 return ERR_PTR(-EINVAL);
4737}
4738
4739struct page * __weak
4740follow_huge_pd(struct vm_area_struct *vma,
4741 unsigned long address, hugepd_t hpd, int flags, int pdshift)
4742{
4743 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4744 return NULL;
4745}
4746
4747struct page * __weak
4748follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4749 pmd_t *pmd, int flags)
4750{
4751 struct page *page = NULL;
4752 spinlock_t *ptl;
4753 pte_t pte;
4754retry:
4755 ptl = pmd_lockptr(mm, pmd);
4756 spin_lock(ptl);
4757
4758
4759
4760
4761 if (!pmd_huge(*pmd))
4762 goto out;
4763 pte = huge_ptep_get((pte_t *)pmd);
4764 if (pte_present(pte)) {
4765 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4766 if (flags & FOLL_GET)
4767 get_page(page);
4768 } else {
4769 if (is_hugetlb_entry_migration(pte)) {
4770 spin_unlock(ptl);
4771 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4772 goto retry;
4773 }
4774
4775
4776
4777
4778 }
4779out:
4780 spin_unlock(ptl);
4781 return page;
4782}
4783
4784struct page * __weak
4785follow_huge_pud(struct mm_struct *mm, unsigned long address,
4786 pud_t *pud, int flags)
4787{
4788 if (flags & FOLL_GET)
4789 return NULL;
4790
4791 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4792}
4793
4794struct page * __weak
4795follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
4796{
4797 if (flags & FOLL_GET)
4798 return NULL;
4799
4800 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
4801}
4802
4803bool isolate_huge_page(struct page *page, struct list_head *list)
4804{
4805 bool ret = true;
4806
4807 VM_BUG_ON_PAGE(!PageHead(page), page);
4808 spin_lock(&hugetlb_lock);
4809 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4810 ret = false;
4811 goto unlock;
4812 }
4813 clear_page_huge_active(page);
4814 list_move_tail(&page->lru, list);
4815unlock:
4816 spin_unlock(&hugetlb_lock);
4817 return ret;
4818}
4819
4820void putback_active_hugepage(struct page *page)
4821{
4822 VM_BUG_ON_PAGE(!PageHead(page), page);
4823 spin_lock(&hugetlb_lock);
4824 set_page_huge_active(page);
4825 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4826 spin_unlock(&hugetlb_lock);
4827 put_page(page);
4828}
4829
4830void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
4831{
4832 struct hstate *h = page_hstate(oldpage);
4833
4834 hugetlb_cgroup_migrate(oldpage, newpage);
4835 set_page_owner_migrate_reason(newpage, reason);
4836
4837
4838
4839
4840
4841
4842
4843
4844
4845
4846
4847 if (PageHugeTemporary(newpage)) {
4848 int old_nid = page_to_nid(oldpage);
4849 int new_nid = page_to_nid(newpage);
4850
4851 SetPageHugeTemporary(oldpage);
4852 ClearPageHugeTemporary(newpage);
4853
4854 spin_lock(&hugetlb_lock);
4855 if (h->surplus_huge_pages_node[old_nid]) {
4856 h->surplus_huge_pages_node[old_nid]--;
4857 h->surplus_huge_pages_node[new_nid]++;
4858 }
4859 spin_unlock(&hugetlb_lock);
4860 }
4861}
4862