1
2
3
4
5#include <linux/list.h>
6#include <linux/init.h>
7#include <linux/module.h>
8#include <linux/mm.h>
9#include <linux/seq_file.h>
10#include <linux/sysctl.h>
11#include <linux/highmem.h>
12#include <linux/mmu_notifier.h>
13#include <linux/nodemask.h>
14#include <linux/pagemap.h>
15#include <linux/mempolicy.h>
16#include <linux/compiler.h>
17#include <linux/cpuset.h>
18#include <linux/mutex.h>
19#include <linux/bootmem.h>
20#include <linux/sysfs.h>
21#include <linux/slab.h>
22#include <linux/rmap.h>
23#include <linux/swap.h>
24#include <linux/swapops.h>
25#include <linux/page-isolation.h>
26#include <linux/jhash.h>
27
28#include <asm/page.h>
29#include <asm/pgtable.h>
30#include <asm/tlb.h>
31
32#include <linux/io.h>
33#include <linux/hugetlb.h>
34#include <linux/hugetlb_cgroup.h>
35#include <linux/node.h>
36#include "internal.h"
37
38int hugepages_treat_as_movable;
39
40int hugetlb_max_hstate __read_mostly;
41unsigned int default_hstate_idx;
42struct hstate hstates[HUGE_MAX_HSTATE];
43
44__initdata LIST_HEAD(huge_boot_pages);
45
46
47static struct hstate * __initdata parsed_hstate;
48static unsigned long __initdata default_hstate_max_huge_pages;
49static unsigned long __initdata default_hstate_size;
50
51
52
53
54
55DEFINE_SPINLOCK(hugetlb_lock);
56
57
58
59
60
61static int num_fault_mutexes;
62static struct mutex *htlb_fault_mutex_table ____cacheline_aligned_in_smp;
63
64static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
65{
66 bool free = (spool->count == 0) && (spool->used_hpages == 0);
67
68 spin_unlock(&spool->lock);
69
70
71
72 if (free)
73 kfree(spool);
74}
75
76struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
77{
78 struct hugepage_subpool *spool;
79
80 spool = kmalloc(sizeof(*spool), GFP_KERNEL);
81 if (!spool)
82 return NULL;
83
84 spin_lock_init(&spool->lock);
85 spool->count = 1;
86 spool->max_hpages = nr_blocks;
87 spool->used_hpages = 0;
88
89 return spool;
90}
91
92void hugepage_put_subpool(struct hugepage_subpool *spool)
93{
94 spin_lock(&spool->lock);
95 BUG_ON(!spool->count);
96 spool->count--;
97 unlock_or_release_subpool(spool);
98}
99
100static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
101 long delta)
102{
103 int ret = 0;
104
105 if (!spool)
106 return 0;
107
108 spin_lock(&spool->lock);
109 if ((spool->used_hpages + delta) <= spool->max_hpages) {
110 spool->used_hpages += delta;
111 } else {
112 ret = -ENOMEM;
113 }
114 spin_unlock(&spool->lock);
115
116 return ret;
117}
118
119static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
120 long delta)
121{
122 if (!spool)
123 return;
124
125 spin_lock(&spool->lock);
126 spool->used_hpages -= delta;
127
128
129 unlock_or_release_subpool(spool);
130}
131
132static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
133{
134 return HUGETLBFS_SB(inode->i_sb)->spool;
135}
136
137static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
138{
139 return subpool_inode(file_inode(vma->vm_file));
140}
141
142
143
144
145
146
147
148
149struct file_region {
150 struct list_head link;
151 long from;
152 long to;
153};
154
155static long region_add(struct resv_map *resv, long f, long t)
156{
157 struct list_head *head = &resv->regions;
158 struct file_region *rg, *nrg, *trg;
159
160 spin_lock(&resv->lock);
161
162 list_for_each_entry(rg, head, link)
163 if (f <= rg->to)
164 break;
165
166
167 if (f > rg->from)
168 f = rg->from;
169
170
171 nrg = rg;
172 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
173 if (&rg->link == head)
174 break;
175 if (rg->from > t)
176 break;
177
178
179
180
181 if (rg->to > t)
182 t = rg->to;
183 if (rg != nrg) {
184 list_del(&rg->link);
185 kfree(rg);
186 }
187 }
188 nrg->from = f;
189 nrg->to = t;
190 spin_unlock(&resv->lock);
191 return 0;
192}
193
194static long region_chg(struct resv_map *resv, long f, long t)
195{
196 struct list_head *head = &resv->regions;
197 struct file_region *rg, *nrg = NULL;
198 long chg = 0;
199
200retry:
201 spin_lock(&resv->lock);
202
203 list_for_each_entry(rg, head, link)
204 if (f <= rg->to)
205 break;
206
207
208
209
210 if (&rg->link == head || t < rg->from) {
211 if (!nrg) {
212 spin_unlock(&resv->lock);
213 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
214 if (!nrg)
215 return -ENOMEM;
216
217 nrg->from = f;
218 nrg->to = f;
219 INIT_LIST_HEAD(&nrg->link);
220 goto retry;
221 }
222
223 list_add(&nrg->link, rg->link.prev);
224 chg = t - f;
225 goto out_nrg;
226 }
227
228
229 if (f > rg->from)
230 f = rg->from;
231 chg = t - f;
232
233
234 list_for_each_entry(rg, rg->link.prev, link) {
235 if (&rg->link == head)
236 break;
237 if (rg->from > t)
238 goto out;
239
240
241
242
243 if (rg->to > t) {
244 chg += rg->to - t;
245 t = rg->to;
246 }
247 chg -= rg->to - rg->from;
248 }
249
250out:
251 spin_unlock(&resv->lock);
252
253 kfree(nrg);
254 return chg;
255out_nrg:
256 spin_unlock(&resv->lock);
257 return chg;
258}
259
260static long region_truncate(struct resv_map *resv, long end)
261{
262 struct list_head *head = &resv->regions;
263 struct file_region *rg, *trg;
264 long chg = 0;
265
266 spin_lock(&resv->lock);
267
268 list_for_each_entry(rg, head, link)
269 if (end <= rg->to)
270 break;
271 if (&rg->link == head)
272 goto out;
273
274
275 if (end > rg->from) {
276 chg = rg->to - end;
277 rg->to = end;
278 rg = list_entry(rg->link.next, typeof(*rg), link);
279 }
280
281
282 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
283 if (&rg->link == head)
284 break;
285 chg += rg->to - rg->from;
286 list_del(&rg->link);
287 kfree(rg);
288 }
289
290out:
291 spin_unlock(&resv->lock);
292 return chg;
293}
294
295static long region_count(struct resv_map *resv, long f, long t)
296{
297 struct list_head *head = &resv->regions;
298 struct file_region *rg;
299 long chg = 0;
300
301 spin_lock(&resv->lock);
302
303 list_for_each_entry(rg, head, link) {
304 long seg_from;
305 long seg_to;
306
307 if (rg->to <= f)
308 continue;
309 if (rg->from >= t)
310 break;
311
312 seg_from = max(rg->from, f);
313 seg_to = min(rg->to, t);
314
315 chg += seg_to - seg_from;
316 }
317 spin_unlock(&resv->lock);
318
319 return chg;
320}
321
322
323
324
325
326static pgoff_t vma_hugecache_offset(struct hstate *h,
327 struct vm_area_struct *vma, unsigned long address)
328{
329 return ((address - vma->vm_start) >> huge_page_shift(h)) +
330 (vma->vm_pgoff >> huge_page_order(h));
331}
332
333pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
334 unsigned long address)
335{
336 return vma_hugecache_offset(hstate_vma(vma), vma, address);
337}
338
339
340
341
342
343unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
344{
345 struct hstate *hstate;
346
347 if (!is_vm_hugetlb_page(vma))
348 return PAGE_SIZE;
349
350 hstate = hstate_vma(vma);
351
352 return 1UL << huge_page_shift(hstate);
353}
354EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
355
356
357
358
359
360
361
362#ifndef vma_mmu_pagesize
363unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
364{
365 return vma_kernel_pagesize(vma);
366}
367#endif
368
369
370
371
372
373
374#define HPAGE_RESV_OWNER (1UL << 0)
375#define HPAGE_RESV_UNMAPPED (1UL << 1)
376#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397static unsigned long get_vma_private_data(struct vm_area_struct *vma)
398{
399 return (unsigned long)vma->vm_private_data;
400}
401
402static void set_vma_private_data(struct vm_area_struct *vma,
403 unsigned long value)
404{
405 vma->vm_private_data = (void *)value;
406}
407
408struct resv_map *resv_map_alloc(void)
409{
410 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
411 if (!resv_map)
412 return NULL;
413
414 kref_init(&resv_map->refs);
415 spin_lock_init(&resv_map->lock);
416 INIT_LIST_HEAD(&resv_map->regions);
417
418 return resv_map;
419}
420
421void resv_map_release(struct kref *ref)
422{
423 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
424
425
426 region_truncate(resv_map, 0);
427 kfree(resv_map);
428}
429
430static inline struct resv_map *inode_resv_map(struct inode *inode)
431{
432 return inode->i_mapping->private_data;
433}
434
435static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
436{
437 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
438 if (vma->vm_flags & VM_MAYSHARE) {
439 struct address_space *mapping = vma->vm_file->f_mapping;
440 struct inode *inode = mapping->host;
441
442 return inode_resv_map(inode);
443
444 } else {
445 return (struct resv_map *)(get_vma_private_data(vma) &
446 ~HPAGE_RESV_MASK);
447 }
448}
449
450static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
451{
452 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
453 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
454
455 set_vma_private_data(vma, (get_vma_private_data(vma) &
456 HPAGE_RESV_MASK) | (unsigned long)map);
457}
458
459static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
460{
461 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
462 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
463
464 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
465}
466
467static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
468{
469 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
470
471 return (get_vma_private_data(vma) & flag) != 0;
472}
473
474
475void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
476{
477 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
478 if (!(vma->vm_flags & VM_MAYSHARE))
479 vma->vm_private_data = (void *)0;
480}
481
482
483static int vma_has_reserves(struct vm_area_struct *vma, long chg)
484{
485 if (vma->vm_flags & VM_NORESERVE) {
486
487
488
489
490
491
492
493
494
495 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
496 return 1;
497 else
498 return 0;
499 }
500
501
502 if (vma->vm_flags & VM_MAYSHARE)
503 return 1;
504
505
506
507
508
509 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
510 return 1;
511
512 return 0;
513}
514
515static void enqueue_huge_page(struct hstate *h, struct page *page)
516{
517 int nid = page_to_nid(page);
518 list_move(&page->lru, &h->hugepage_freelists[nid]);
519 h->free_huge_pages++;
520 h->free_huge_pages_node[nid]++;
521}
522
523static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
524{
525 struct page *page;
526
527 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
528 if (!is_migrate_isolate_page(page))
529 break;
530
531
532
533
534 if (&h->hugepage_freelists[nid] == &page->lru)
535 return NULL;
536 list_move(&page->lru, &h->hugepage_activelist);
537 set_page_refcounted(page);
538 h->free_huge_pages--;
539 h->free_huge_pages_node[nid]--;
540 return page;
541}
542
543
544static inline gfp_t htlb_alloc_mask(struct hstate *h)
545{
546 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
547 return GFP_HIGHUSER_MOVABLE;
548 else
549 return GFP_HIGHUSER;
550}
551
552static struct page *dequeue_huge_page_vma(struct hstate *h,
553 struct vm_area_struct *vma,
554 unsigned long address, int avoid_reserve,
555 long chg)
556{
557 struct page *page = NULL;
558 struct mempolicy *mpol;
559 nodemask_t *nodemask;
560 struct zonelist *zonelist;
561 struct zone *zone;
562 struct zoneref *z;
563 unsigned int cpuset_mems_cookie;
564
565
566
567
568
569
570 if (!vma_has_reserves(vma, chg) &&
571 h->free_huge_pages - h->resv_huge_pages == 0)
572 goto err;
573
574
575 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
576 goto err;
577
578retry_cpuset:
579 cpuset_mems_cookie = read_mems_allowed_begin();
580 zonelist = huge_zonelist(vma, address,
581 htlb_alloc_mask(h), &mpol, &nodemask);
582
583 for_each_zone_zonelist_nodemask(zone, z, zonelist,
584 MAX_NR_ZONES - 1, nodemask) {
585 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
586 page = dequeue_huge_page_node(h, zone_to_nid(zone));
587 if (page) {
588 if (avoid_reserve)
589 break;
590 if (!vma_has_reserves(vma, chg))
591 break;
592
593 SetPagePrivate(page);
594 h->resv_huge_pages--;
595 break;
596 }
597 }
598 }
599
600 mpol_cond_put(mpol);
601 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
602 goto retry_cpuset;
603 return page;
604
605err:
606 return NULL;
607}
608
609
610
611
612
613
614
615
616static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
617{
618 nid = next_node(nid, *nodes_allowed);
619 if (nid == MAX_NUMNODES)
620 nid = first_node(*nodes_allowed);
621 VM_BUG_ON(nid >= MAX_NUMNODES);
622
623 return nid;
624}
625
626static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
627{
628 if (!node_isset(nid, *nodes_allowed))
629 nid = next_node_allowed(nid, nodes_allowed);
630 return nid;
631}
632
633
634
635
636
637
638
639static int hstate_next_node_to_alloc(struct hstate *h,
640 nodemask_t *nodes_allowed)
641{
642 int nid;
643
644 VM_BUG_ON(!nodes_allowed);
645
646 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
647 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
648
649 return nid;
650}
651
652
653
654
655
656
657
658static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
659{
660 int nid;
661
662 VM_BUG_ON(!nodes_allowed);
663
664 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
665 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
666
667 return nid;
668}
669
670#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
671 for (nr_nodes = nodes_weight(*mask); \
672 nr_nodes > 0 && \
673 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
674 nr_nodes--)
675
676#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
677 for (nr_nodes = nodes_weight(*mask); \
678 nr_nodes > 0 && \
679 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
680 nr_nodes--)
681
682#if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
683static void destroy_compound_gigantic_page(struct page *page,
684 unsigned long order)
685{
686 int i;
687 int nr_pages = 1 << order;
688 struct page *p = page + 1;
689
690 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
691 __ClearPageTail(p);
692 set_page_refcounted(p);
693 p->first_page = NULL;
694 }
695
696 set_compound_order(page, 0);
697 __ClearPageHead(page);
698}
699
700static void free_gigantic_page(struct page *page, unsigned order)
701{
702 free_contig_range(page_to_pfn(page), 1 << order);
703}
704
705static int __alloc_gigantic_page(unsigned long start_pfn,
706 unsigned long nr_pages)
707{
708 unsigned long end_pfn = start_pfn + nr_pages;
709 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
710}
711
712static bool pfn_range_valid_gigantic(unsigned long start_pfn,
713 unsigned long nr_pages)
714{
715 unsigned long i, end_pfn = start_pfn + nr_pages;
716 struct page *page;
717
718 for (i = start_pfn; i < end_pfn; i++) {
719 if (!pfn_valid(i))
720 return false;
721
722 page = pfn_to_page(i);
723
724 if (PageReserved(page))
725 return false;
726
727 if (page_count(page) > 0)
728 return false;
729
730 if (PageHuge(page))
731 return false;
732 }
733
734 return true;
735}
736
737static bool zone_spans_last_pfn(const struct zone *zone,
738 unsigned long start_pfn, unsigned long nr_pages)
739{
740 unsigned long last_pfn = start_pfn + nr_pages - 1;
741 return zone_spans_pfn(zone, last_pfn);
742}
743
744static struct page *alloc_gigantic_page(int nid, unsigned order)
745{
746 unsigned long nr_pages = 1 << order;
747 unsigned long ret, pfn, flags;
748 struct zone *z;
749
750 z = NODE_DATA(nid)->node_zones;
751 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
752 spin_lock_irqsave(&z->lock, flags);
753
754 pfn = ALIGN(z->zone_start_pfn, nr_pages);
755 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
756 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
757
758
759
760
761
762
763
764 spin_unlock_irqrestore(&z->lock, flags);
765 ret = __alloc_gigantic_page(pfn, nr_pages);
766 if (!ret)
767 return pfn_to_page(pfn);
768 spin_lock_irqsave(&z->lock, flags);
769 }
770 pfn += nr_pages;
771 }
772
773 spin_unlock_irqrestore(&z->lock, flags);
774 }
775
776 return NULL;
777}
778
779static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
780static void prep_compound_gigantic_page(struct page *page, unsigned long order);
781
782static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
783{
784 struct page *page;
785
786 page = alloc_gigantic_page(nid, huge_page_order(h));
787 if (page) {
788 prep_compound_gigantic_page(page, huge_page_order(h));
789 prep_new_huge_page(h, page, nid);
790 }
791
792 return page;
793}
794
795static int alloc_fresh_gigantic_page(struct hstate *h,
796 nodemask_t *nodes_allowed)
797{
798 struct page *page = NULL;
799 int nr_nodes, node;
800
801 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
802 page = alloc_fresh_gigantic_page_node(h, node);
803 if (page)
804 return 1;
805 }
806
807 return 0;
808}
809
810static inline bool gigantic_page_supported(void) { return true; }
811#else
812static inline bool gigantic_page_supported(void) { return false; }
813static inline void free_gigantic_page(struct page *page, unsigned order) { }
814static inline void destroy_compound_gigantic_page(struct page *page,
815 unsigned long order) { }
816static inline int alloc_fresh_gigantic_page(struct hstate *h,
817 nodemask_t *nodes_allowed) { return 0; }
818#endif
819
820static void update_and_free_page(struct hstate *h, struct page *page)
821{
822 int i;
823
824 if (hstate_is_gigantic(h) && !gigantic_page_supported())
825 return;
826
827 h->nr_huge_pages--;
828 h->nr_huge_pages_node[page_to_nid(page)]--;
829 for (i = 0; i < pages_per_huge_page(h); i++) {
830 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
831 1 << PG_referenced | 1 << PG_dirty |
832 1 << PG_active | 1 << PG_private |
833 1 << PG_writeback);
834 }
835 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
836 set_compound_page_dtor(page, NULL);
837 set_page_refcounted(page);
838 if (hstate_is_gigantic(h)) {
839 destroy_compound_gigantic_page(page, huge_page_order(h));
840 free_gigantic_page(page, huge_page_order(h));
841 } else {
842 arch_release_hugepage(page);
843 __free_pages(page, huge_page_order(h));
844 }
845}
846
847struct hstate *size_to_hstate(unsigned long size)
848{
849 struct hstate *h;
850
851 for_each_hstate(h) {
852 if (huge_page_size(h) == size)
853 return h;
854 }
855 return NULL;
856}
857
858void free_huge_page(struct page *page)
859{
860
861
862
863
864 struct hstate *h = page_hstate(page);
865 int nid = page_to_nid(page);
866 struct hugepage_subpool *spool =
867 (struct hugepage_subpool *)page_private(page);
868 bool restore_reserve;
869
870 set_page_private(page, 0);
871 page->mapping = NULL;
872 BUG_ON(page_count(page));
873 BUG_ON(page_mapcount(page));
874 restore_reserve = PagePrivate(page);
875 ClearPagePrivate(page);
876
877 spin_lock(&hugetlb_lock);
878 hugetlb_cgroup_uncharge_page(hstate_index(h),
879 pages_per_huge_page(h), page);
880 if (restore_reserve)
881 h->resv_huge_pages++;
882
883 if (h->surplus_huge_pages_node[nid]) {
884
885 list_del(&page->lru);
886 update_and_free_page(h, page);
887 h->surplus_huge_pages--;
888 h->surplus_huge_pages_node[nid]--;
889 } else {
890 arch_clear_hugepage_flags(page);
891 enqueue_huge_page(h, page);
892 }
893 spin_unlock(&hugetlb_lock);
894 hugepage_subpool_put_pages(spool, 1);
895}
896
897static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
898{
899 INIT_LIST_HEAD(&page->lru);
900 set_compound_page_dtor(page, free_huge_page);
901 spin_lock(&hugetlb_lock);
902 set_hugetlb_cgroup(page, NULL);
903 h->nr_huge_pages++;
904 h->nr_huge_pages_node[nid]++;
905 spin_unlock(&hugetlb_lock);
906 put_page(page);
907}
908
909static void prep_compound_gigantic_page(struct page *page, unsigned long order)
910{
911 int i;
912 int nr_pages = 1 << order;
913 struct page *p = page + 1;
914
915
916 set_compound_order(page, order);
917 __SetPageHead(page);
918 __ClearPageReserved(page);
919 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
920
921
922
923
924
925
926
927
928
929
930
931
932 __ClearPageReserved(p);
933 set_page_count(p, 0);
934 p->first_page = page;
935
936 smp_wmb();
937 __SetPageTail(p);
938 }
939}
940
941
942
943
944
945
946int PageHuge(struct page *page)
947{
948 if (!PageCompound(page))
949 return 0;
950
951 page = compound_head(page);
952 return get_compound_page_dtor(page) == free_huge_page;
953}
954EXPORT_SYMBOL_GPL(PageHuge);
955
956
957
958
959
960int PageHeadHuge(struct page *page_head)
961{
962 if (!PageHead(page_head))
963 return 0;
964
965 return get_compound_page_dtor(page_head) == free_huge_page;
966}
967
968pgoff_t __basepage_index(struct page *page)
969{
970 struct page *page_head = compound_head(page);
971 pgoff_t index = page_index(page_head);
972 unsigned long compound_idx;
973
974 if (!PageHuge(page_head))
975 return page_index(page);
976
977 if (compound_order(page_head) >= MAX_ORDER)
978 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
979 else
980 compound_idx = page - page_head;
981
982 return (index << compound_order(page_head)) + compound_idx;
983}
984
985static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
986{
987 struct page *page;
988
989 page = alloc_pages_exact_node(nid,
990 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
991 __GFP_REPEAT|__GFP_NOWARN,
992 huge_page_order(h));
993 if (page) {
994 if (arch_prepare_hugepage(page)) {
995 __free_pages(page, huge_page_order(h));
996 return NULL;
997 }
998 prep_new_huge_page(h, page, nid);
999 }
1000
1001 return page;
1002}
1003
1004static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1005{
1006 struct page *page;
1007 int nr_nodes, node;
1008 int ret = 0;
1009
1010 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1011 page = alloc_fresh_huge_page_node(h, node);
1012 if (page) {
1013 ret = 1;
1014 break;
1015 }
1016 }
1017
1018 if (ret)
1019 count_vm_event(HTLB_BUDDY_PGALLOC);
1020 else
1021 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1022
1023 return ret;
1024}
1025
1026
1027
1028
1029
1030
1031
1032static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1033 bool acct_surplus)
1034{
1035 int nr_nodes, node;
1036 int ret = 0;
1037
1038 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1039
1040
1041
1042
1043 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1044 !list_empty(&h->hugepage_freelists[node])) {
1045 struct page *page =
1046 list_entry(h->hugepage_freelists[node].next,
1047 struct page, lru);
1048 list_del(&page->lru);
1049 h->free_huge_pages--;
1050 h->free_huge_pages_node[node]--;
1051 if (acct_surplus) {
1052 h->surplus_huge_pages--;
1053 h->surplus_huge_pages_node[node]--;
1054 }
1055 update_and_free_page(h, page);
1056 ret = 1;
1057 break;
1058 }
1059 }
1060
1061 return ret;
1062}
1063
1064
1065
1066
1067
1068static void dissolve_free_huge_page(struct page *page)
1069{
1070 spin_lock(&hugetlb_lock);
1071 if (PageHuge(page) && !page_count(page)) {
1072 struct hstate *h = page_hstate(page);
1073 int nid = page_to_nid(page);
1074 list_del(&page->lru);
1075 h->free_huge_pages--;
1076 h->free_huge_pages_node[nid]--;
1077 update_and_free_page(h, page);
1078 }
1079 spin_unlock(&hugetlb_lock);
1080}
1081
1082
1083
1084
1085
1086
1087void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1088{
1089 unsigned int order = 8 * sizeof(void *);
1090 unsigned long pfn;
1091 struct hstate *h;
1092
1093 if (!hugepages_supported())
1094 return;
1095
1096
1097 for_each_hstate(h)
1098 if (order > huge_page_order(h))
1099 order = huge_page_order(h);
1100 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << order));
1101 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order)
1102 dissolve_free_huge_page(pfn_to_page(pfn));
1103}
1104
1105static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
1106{
1107 struct page *page;
1108 unsigned int r_nid;
1109
1110 if (hstate_is_gigantic(h))
1111 return NULL;
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136 spin_lock(&hugetlb_lock);
1137 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1138 spin_unlock(&hugetlb_lock);
1139 return NULL;
1140 } else {
1141 h->nr_huge_pages++;
1142 h->surplus_huge_pages++;
1143 }
1144 spin_unlock(&hugetlb_lock);
1145
1146 if (nid == NUMA_NO_NODE)
1147 page = alloc_pages(htlb_alloc_mask(h)|__GFP_COMP|
1148 __GFP_REPEAT|__GFP_NOWARN,
1149 huge_page_order(h));
1150 else
1151 page = alloc_pages_exact_node(nid,
1152 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1153 __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
1154
1155 if (page && arch_prepare_hugepage(page)) {
1156 __free_pages(page, huge_page_order(h));
1157 page = NULL;
1158 }
1159
1160 spin_lock(&hugetlb_lock);
1161 if (page) {
1162 INIT_LIST_HEAD(&page->lru);
1163 r_nid = page_to_nid(page);
1164 set_compound_page_dtor(page, free_huge_page);
1165 set_hugetlb_cgroup(page, NULL);
1166
1167
1168
1169 h->nr_huge_pages_node[r_nid]++;
1170 h->surplus_huge_pages_node[r_nid]++;
1171 __count_vm_event(HTLB_BUDDY_PGALLOC);
1172 } else {
1173 h->nr_huge_pages--;
1174 h->surplus_huge_pages--;
1175 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1176 }
1177 spin_unlock(&hugetlb_lock);
1178
1179 return page;
1180}
1181
1182
1183
1184
1185
1186
1187struct page *alloc_huge_page_node(struct hstate *h, int nid)
1188{
1189 struct page *page = NULL;
1190
1191 spin_lock(&hugetlb_lock);
1192 if (h->free_huge_pages - h->resv_huge_pages > 0)
1193 page = dequeue_huge_page_node(h, nid);
1194 spin_unlock(&hugetlb_lock);
1195
1196 if (!page)
1197 page = alloc_buddy_huge_page(h, nid);
1198
1199 return page;
1200}
1201
1202
1203
1204
1205
1206static int gather_surplus_pages(struct hstate *h, int delta)
1207{
1208 struct list_head surplus_list;
1209 struct page *page, *tmp;
1210 int ret, i;
1211 int needed, allocated;
1212 bool alloc_ok = true;
1213
1214 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1215 if (needed <= 0) {
1216 h->resv_huge_pages += delta;
1217 return 0;
1218 }
1219
1220 allocated = 0;
1221 INIT_LIST_HEAD(&surplus_list);
1222
1223 ret = -ENOMEM;
1224retry:
1225 spin_unlock(&hugetlb_lock);
1226 for (i = 0; i < needed; i++) {
1227 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1228 if (!page) {
1229 alloc_ok = false;
1230 break;
1231 }
1232 list_add(&page->lru, &surplus_list);
1233 }
1234 allocated += i;
1235
1236
1237
1238
1239
1240 spin_lock(&hugetlb_lock);
1241 needed = (h->resv_huge_pages + delta) -
1242 (h->free_huge_pages + allocated);
1243 if (needed > 0) {
1244 if (alloc_ok)
1245 goto retry;
1246
1247
1248
1249
1250
1251 goto free;
1252 }
1253
1254
1255
1256
1257
1258
1259
1260
1261 needed += allocated;
1262 h->resv_huge_pages += delta;
1263 ret = 0;
1264
1265
1266 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1267 if ((--needed) < 0)
1268 break;
1269
1270
1271
1272
1273 put_page_testzero(page);
1274 VM_BUG_ON_PAGE(page_count(page), page);
1275 enqueue_huge_page(h, page);
1276 }
1277free:
1278 spin_unlock(&hugetlb_lock);
1279
1280
1281 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1282 put_page(page);
1283 spin_lock(&hugetlb_lock);
1284
1285 return ret;
1286}
1287
1288
1289
1290
1291
1292
1293
1294static void return_unused_surplus_pages(struct hstate *h,
1295 unsigned long unused_resv_pages)
1296{
1297 unsigned long nr_pages;
1298
1299
1300 h->resv_huge_pages -= unused_resv_pages;
1301
1302
1303 if (hstate_is_gigantic(h))
1304 return;
1305
1306 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316 while (nr_pages--) {
1317 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1318 break;
1319 cond_resched_lock(&hugetlb_lock);
1320 }
1321}
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333static long vma_needs_reservation(struct hstate *h,
1334 struct vm_area_struct *vma, unsigned long addr)
1335{
1336 struct resv_map *resv;
1337 pgoff_t idx;
1338 long chg;
1339
1340 resv = vma_resv_map(vma);
1341 if (!resv)
1342 return 1;
1343
1344 idx = vma_hugecache_offset(h, vma, addr);
1345 chg = region_chg(resv, idx, idx + 1);
1346
1347 if (vma->vm_flags & VM_MAYSHARE)
1348 return chg;
1349 else
1350 return chg < 0 ? chg : 0;
1351}
1352static void vma_commit_reservation(struct hstate *h,
1353 struct vm_area_struct *vma, unsigned long addr)
1354{
1355 struct resv_map *resv;
1356 pgoff_t idx;
1357
1358 resv = vma_resv_map(vma);
1359 if (!resv)
1360 return;
1361
1362 idx = vma_hugecache_offset(h, vma, addr);
1363 region_add(resv, idx, idx + 1);
1364}
1365
1366static struct page *alloc_huge_page(struct vm_area_struct *vma,
1367 unsigned long addr, int avoid_reserve)
1368{
1369 struct hugepage_subpool *spool = subpool_vma(vma);
1370 struct hstate *h = hstate_vma(vma);
1371 struct page *page;
1372 long chg;
1373 int ret, idx;
1374 struct hugetlb_cgroup *h_cg;
1375
1376 idx = hstate_index(h);
1377
1378
1379
1380
1381
1382
1383
1384
1385 chg = vma_needs_reservation(h, vma, addr);
1386 if (chg < 0)
1387 return ERR_PTR(-ENOMEM);
1388 if (chg || avoid_reserve)
1389 if (hugepage_subpool_get_pages(spool, 1))
1390 return ERR_PTR(-ENOSPC);
1391
1392 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1393 if (ret)
1394 goto out_subpool_put;
1395
1396 spin_lock(&hugetlb_lock);
1397 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, chg);
1398 if (!page) {
1399 spin_unlock(&hugetlb_lock);
1400 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1401 if (!page)
1402 goto out_uncharge_cgroup;
1403
1404 spin_lock(&hugetlb_lock);
1405 list_move(&page->lru, &h->hugepage_activelist);
1406
1407 }
1408 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1409 spin_unlock(&hugetlb_lock);
1410
1411 set_page_private(page, (unsigned long)spool);
1412
1413 vma_commit_reservation(h, vma, addr);
1414 return page;
1415
1416out_uncharge_cgroup:
1417 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1418out_subpool_put:
1419 if (chg || avoid_reserve)
1420 hugepage_subpool_put_pages(spool, 1);
1421 return ERR_PTR(-ENOSPC);
1422}
1423
1424
1425
1426
1427
1428
1429struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1430 unsigned long addr, int avoid_reserve)
1431{
1432 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1433 if (IS_ERR(page))
1434 page = NULL;
1435 return page;
1436}
1437
1438int __weak alloc_bootmem_huge_page(struct hstate *h)
1439{
1440 struct huge_bootmem_page *m;
1441 int nr_nodes, node;
1442
1443 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1444 void *addr;
1445
1446 addr = memblock_virt_alloc_try_nid_nopanic(
1447 huge_page_size(h), huge_page_size(h),
1448 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1449 if (addr) {
1450
1451
1452
1453
1454
1455 m = addr;
1456 goto found;
1457 }
1458 }
1459 return 0;
1460
1461found:
1462 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
1463
1464 list_add(&m->list, &huge_boot_pages);
1465 m->hstate = h;
1466 return 1;
1467}
1468
1469static void __init prep_compound_huge_page(struct page *page, int order)
1470{
1471 if (unlikely(order > (MAX_ORDER - 1)))
1472 prep_compound_gigantic_page(page, order);
1473 else
1474 prep_compound_page(page, order);
1475}
1476
1477
1478static void __init gather_bootmem_prealloc(void)
1479{
1480 struct huge_bootmem_page *m;
1481
1482 list_for_each_entry(m, &huge_boot_pages, list) {
1483 struct hstate *h = m->hstate;
1484 struct page *page;
1485
1486#ifdef CONFIG_HIGHMEM
1487 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1488 memblock_free_late(__pa(m),
1489 sizeof(struct huge_bootmem_page));
1490#else
1491 page = virt_to_page(m);
1492#endif
1493 WARN_ON(page_count(page) != 1);
1494 prep_compound_huge_page(page, h->order);
1495 WARN_ON(PageReserved(page));
1496 prep_new_huge_page(h, page, page_to_nid(page));
1497
1498
1499
1500
1501
1502
1503 if (hstate_is_gigantic(h))
1504 adjust_managed_page_count(page, 1 << h->order);
1505 }
1506}
1507
1508static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1509{
1510 unsigned long i;
1511
1512 for (i = 0; i < h->max_huge_pages; ++i) {
1513 if (hstate_is_gigantic(h)) {
1514 if (!alloc_bootmem_huge_page(h))
1515 break;
1516 } else if (!alloc_fresh_huge_page(h,
1517 &node_states[N_MEMORY]))
1518 break;
1519 }
1520 h->max_huge_pages = i;
1521}
1522
1523static void __init hugetlb_init_hstates(void)
1524{
1525 struct hstate *h;
1526
1527 for_each_hstate(h) {
1528
1529 if (!hstate_is_gigantic(h))
1530 hugetlb_hstate_alloc_pages(h);
1531 }
1532}
1533
1534static char * __init memfmt(char *buf, unsigned long n)
1535{
1536 if (n >= (1UL << 30))
1537 sprintf(buf, "%lu GB", n >> 30);
1538 else if (n >= (1UL << 20))
1539 sprintf(buf, "%lu MB", n >> 20);
1540 else
1541 sprintf(buf, "%lu KB", n >> 10);
1542 return buf;
1543}
1544
1545static void __init report_hugepages(void)
1546{
1547 struct hstate *h;
1548
1549 for_each_hstate(h) {
1550 char buf[32];
1551 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1552 memfmt(buf, huge_page_size(h)),
1553 h->free_huge_pages);
1554 }
1555}
1556
1557#ifdef CONFIG_HIGHMEM
1558static void try_to_free_low(struct hstate *h, unsigned long count,
1559 nodemask_t *nodes_allowed)
1560{
1561 int i;
1562
1563 if (hstate_is_gigantic(h))
1564 return;
1565
1566 for_each_node_mask(i, *nodes_allowed) {
1567 struct page *page, *next;
1568 struct list_head *freel = &h->hugepage_freelists[i];
1569 list_for_each_entry_safe(page, next, freel, lru) {
1570 if (count >= h->nr_huge_pages)
1571 return;
1572 if (PageHighMem(page))
1573 continue;
1574 list_del(&page->lru);
1575 update_and_free_page(h, page);
1576 h->free_huge_pages--;
1577 h->free_huge_pages_node[page_to_nid(page)]--;
1578 }
1579 }
1580}
1581#else
1582static inline void try_to_free_low(struct hstate *h, unsigned long count,
1583 nodemask_t *nodes_allowed)
1584{
1585}
1586#endif
1587
1588
1589
1590
1591
1592
1593static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1594 int delta)
1595{
1596 int nr_nodes, node;
1597
1598 VM_BUG_ON(delta != -1 && delta != 1);
1599
1600 if (delta < 0) {
1601 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1602 if (h->surplus_huge_pages_node[node])
1603 goto found;
1604 }
1605 } else {
1606 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1607 if (h->surplus_huge_pages_node[node] <
1608 h->nr_huge_pages_node[node])
1609 goto found;
1610 }
1611 }
1612 return 0;
1613
1614found:
1615 h->surplus_huge_pages += delta;
1616 h->surplus_huge_pages_node[node] += delta;
1617 return 1;
1618}
1619
1620#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1621static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1622 nodemask_t *nodes_allowed)
1623{
1624 unsigned long min_count, ret;
1625
1626 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1627 return h->max_huge_pages;
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640 spin_lock(&hugetlb_lock);
1641 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1642 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1643 break;
1644 }
1645
1646 while (count > persistent_huge_pages(h)) {
1647
1648
1649
1650
1651
1652 spin_unlock(&hugetlb_lock);
1653 if (hstate_is_gigantic(h))
1654 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
1655 else
1656 ret = alloc_fresh_huge_page(h, nodes_allowed);
1657 spin_lock(&hugetlb_lock);
1658 if (!ret)
1659 goto out;
1660
1661
1662 if (signal_pending(current))
1663 goto out;
1664 }
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1682 min_count = max(count, min_count);
1683 try_to_free_low(h, min_count, nodes_allowed);
1684 while (min_count < persistent_huge_pages(h)) {
1685 if (!free_pool_huge_page(h, nodes_allowed, 0))
1686 break;
1687 cond_resched_lock(&hugetlb_lock);
1688 }
1689 while (count < persistent_huge_pages(h)) {
1690 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1691 break;
1692 }
1693out:
1694 ret = persistent_huge_pages(h);
1695 spin_unlock(&hugetlb_lock);
1696 return ret;
1697}
1698
1699#define HSTATE_ATTR_RO(_name) \
1700 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1701
1702#define HSTATE_ATTR(_name) \
1703 static struct kobj_attribute _name##_attr = \
1704 __ATTR(_name, 0644, _name##_show, _name##_store)
1705
1706static struct kobject *hugepages_kobj;
1707static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1708
1709static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1710
1711static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1712{
1713 int i;
1714
1715 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1716 if (hstate_kobjs[i] == kobj) {
1717 if (nidp)
1718 *nidp = NUMA_NO_NODE;
1719 return &hstates[i];
1720 }
1721
1722 return kobj_to_node_hstate(kobj, nidp);
1723}
1724
1725static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1726 struct kobj_attribute *attr, char *buf)
1727{
1728 struct hstate *h;
1729 unsigned long nr_huge_pages;
1730 int nid;
1731
1732 h = kobj_to_hstate(kobj, &nid);
1733 if (nid == NUMA_NO_NODE)
1734 nr_huge_pages = h->nr_huge_pages;
1735 else
1736 nr_huge_pages = h->nr_huge_pages_node[nid];
1737
1738 return sprintf(buf, "%lu\n", nr_huge_pages);
1739}
1740
1741static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
1742 struct hstate *h, int nid,
1743 unsigned long count, size_t len)
1744{
1745 int err;
1746 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1747
1748 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
1749 err = -EINVAL;
1750 goto out;
1751 }
1752
1753 if (nid == NUMA_NO_NODE) {
1754
1755
1756
1757 if (!(obey_mempolicy &&
1758 init_nodemask_of_mempolicy(nodes_allowed))) {
1759 NODEMASK_FREE(nodes_allowed);
1760 nodes_allowed = &node_states[N_MEMORY];
1761 }
1762 } else if (nodes_allowed) {
1763
1764
1765
1766
1767 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1768 init_nodemask_of_node(nodes_allowed, nid);
1769 } else
1770 nodes_allowed = &node_states[N_MEMORY];
1771
1772 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1773
1774 if (nodes_allowed != &node_states[N_MEMORY])
1775 NODEMASK_FREE(nodes_allowed);
1776
1777 return len;
1778out:
1779 NODEMASK_FREE(nodes_allowed);
1780 return err;
1781}
1782
1783static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1784 struct kobject *kobj, const char *buf,
1785 size_t len)
1786{
1787 struct hstate *h;
1788 unsigned long count;
1789 int nid;
1790 int err;
1791
1792 err = kstrtoul(buf, 10, &count);
1793 if (err)
1794 return err;
1795
1796 h = kobj_to_hstate(kobj, &nid);
1797 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
1798}
1799
1800static ssize_t nr_hugepages_show(struct kobject *kobj,
1801 struct kobj_attribute *attr, char *buf)
1802{
1803 return nr_hugepages_show_common(kobj, attr, buf);
1804}
1805
1806static ssize_t nr_hugepages_store(struct kobject *kobj,
1807 struct kobj_attribute *attr, const char *buf, size_t len)
1808{
1809 return nr_hugepages_store_common(false, kobj, buf, len);
1810}
1811HSTATE_ATTR(nr_hugepages);
1812
1813#ifdef CONFIG_NUMA
1814
1815
1816
1817
1818
1819static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1820 struct kobj_attribute *attr, char *buf)
1821{
1822 return nr_hugepages_show_common(kobj, attr, buf);
1823}
1824
1825static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1826 struct kobj_attribute *attr, const char *buf, size_t len)
1827{
1828 return nr_hugepages_store_common(true, kobj, buf, len);
1829}
1830HSTATE_ATTR(nr_hugepages_mempolicy);
1831#endif
1832
1833
1834static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1835 struct kobj_attribute *attr, char *buf)
1836{
1837 struct hstate *h = kobj_to_hstate(kobj, NULL);
1838 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1839}
1840
1841static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1842 struct kobj_attribute *attr, const char *buf, size_t count)
1843{
1844 int err;
1845 unsigned long input;
1846 struct hstate *h = kobj_to_hstate(kobj, NULL);
1847
1848 if (hstate_is_gigantic(h))
1849 return -EINVAL;
1850
1851 err = kstrtoul(buf, 10, &input);
1852 if (err)
1853 return err;
1854
1855 spin_lock(&hugetlb_lock);
1856 h->nr_overcommit_huge_pages = input;
1857 spin_unlock(&hugetlb_lock);
1858
1859 return count;
1860}
1861HSTATE_ATTR(nr_overcommit_hugepages);
1862
1863static ssize_t free_hugepages_show(struct kobject *kobj,
1864 struct kobj_attribute *attr, char *buf)
1865{
1866 struct hstate *h;
1867 unsigned long free_huge_pages;
1868 int nid;
1869
1870 h = kobj_to_hstate(kobj, &nid);
1871 if (nid == NUMA_NO_NODE)
1872 free_huge_pages = h->free_huge_pages;
1873 else
1874 free_huge_pages = h->free_huge_pages_node[nid];
1875
1876 return sprintf(buf, "%lu\n", free_huge_pages);
1877}
1878HSTATE_ATTR_RO(free_hugepages);
1879
1880static ssize_t resv_hugepages_show(struct kobject *kobj,
1881 struct kobj_attribute *attr, char *buf)
1882{
1883 struct hstate *h = kobj_to_hstate(kobj, NULL);
1884 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1885}
1886HSTATE_ATTR_RO(resv_hugepages);
1887
1888static ssize_t surplus_hugepages_show(struct kobject *kobj,
1889 struct kobj_attribute *attr, char *buf)
1890{
1891 struct hstate *h;
1892 unsigned long surplus_huge_pages;
1893 int nid;
1894
1895 h = kobj_to_hstate(kobj, &nid);
1896 if (nid == NUMA_NO_NODE)
1897 surplus_huge_pages = h->surplus_huge_pages;
1898 else
1899 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1900
1901 return sprintf(buf, "%lu\n", surplus_huge_pages);
1902}
1903HSTATE_ATTR_RO(surplus_hugepages);
1904
1905static struct attribute *hstate_attrs[] = {
1906 &nr_hugepages_attr.attr,
1907 &nr_overcommit_hugepages_attr.attr,
1908 &free_hugepages_attr.attr,
1909 &resv_hugepages_attr.attr,
1910 &surplus_hugepages_attr.attr,
1911#ifdef CONFIG_NUMA
1912 &nr_hugepages_mempolicy_attr.attr,
1913#endif
1914 NULL,
1915};
1916
1917static struct attribute_group hstate_attr_group = {
1918 .attrs = hstate_attrs,
1919};
1920
1921static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1922 struct kobject **hstate_kobjs,
1923 struct attribute_group *hstate_attr_group)
1924{
1925 int retval;
1926 int hi = hstate_index(h);
1927
1928 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1929 if (!hstate_kobjs[hi])
1930 return -ENOMEM;
1931
1932 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1933 if (retval)
1934 kobject_put(hstate_kobjs[hi]);
1935
1936 return retval;
1937}
1938
1939static void __init hugetlb_sysfs_init(void)
1940{
1941 struct hstate *h;
1942 int err;
1943
1944 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1945 if (!hugepages_kobj)
1946 return;
1947
1948 for_each_hstate(h) {
1949 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1950 hstate_kobjs, &hstate_attr_group);
1951 if (err)
1952 pr_err("Hugetlb: Unable to add hstate %s", h->name);
1953 }
1954}
1955
1956#ifdef CONFIG_NUMA
1957
1958
1959
1960
1961
1962
1963
1964
1965struct node_hstate {
1966 struct kobject *hugepages_kobj;
1967 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1968};
1969struct node_hstate node_hstates[MAX_NUMNODES];
1970
1971
1972
1973
1974static struct attribute *per_node_hstate_attrs[] = {
1975 &nr_hugepages_attr.attr,
1976 &free_hugepages_attr.attr,
1977 &surplus_hugepages_attr.attr,
1978 NULL,
1979};
1980
1981static struct attribute_group per_node_hstate_attr_group = {
1982 .attrs = per_node_hstate_attrs,
1983};
1984
1985
1986
1987
1988
1989static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1990{
1991 int nid;
1992
1993 for (nid = 0; nid < nr_node_ids; nid++) {
1994 struct node_hstate *nhs = &node_hstates[nid];
1995 int i;
1996 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1997 if (nhs->hstate_kobjs[i] == kobj) {
1998 if (nidp)
1999 *nidp = nid;
2000 return &hstates[i];
2001 }
2002 }
2003
2004 BUG();
2005 return NULL;
2006}
2007
2008
2009
2010
2011
2012static void hugetlb_unregister_node(struct node *node)
2013{
2014 struct hstate *h;
2015 struct node_hstate *nhs = &node_hstates[node->dev.id];
2016
2017 if (!nhs->hugepages_kobj)
2018 return;
2019
2020 for_each_hstate(h) {
2021 int idx = hstate_index(h);
2022 if (nhs->hstate_kobjs[idx]) {
2023 kobject_put(nhs->hstate_kobjs[idx]);
2024 nhs->hstate_kobjs[idx] = NULL;
2025 }
2026 }
2027
2028 kobject_put(nhs->hugepages_kobj);
2029 nhs->hugepages_kobj = NULL;
2030}
2031
2032
2033
2034
2035
2036static void hugetlb_unregister_all_nodes(void)
2037{
2038 int nid;
2039
2040
2041
2042
2043 register_hugetlbfs_with_node(NULL, NULL);
2044
2045
2046
2047
2048 for (nid = 0; nid < nr_node_ids; nid++)
2049 hugetlb_unregister_node(node_devices[nid]);
2050}
2051
2052
2053
2054
2055
2056static void hugetlb_register_node(struct node *node)
2057{
2058 struct hstate *h;
2059 struct node_hstate *nhs = &node_hstates[node->dev.id];
2060 int err;
2061
2062 if (nhs->hugepages_kobj)
2063 return;
2064
2065 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2066 &node->dev.kobj);
2067 if (!nhs->hugepages_kobj)
2068 return;
2069
2070 for_each_hstate(h) {
2071 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2072 nhs->hstate_kobjs,
2073 &per_node_hstate_attr_group);
2074 if (err) {
2075 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2076 h->name, node->dev.id);
2077 hugetlb_unregister_node(node);
2078 break;
2079 }
2080 }
2081}
2082
2083
2084
2085
2086
2087
2088static void __init hugetlb_register_all_nodes(void)
2089{
2090 int nid;
2091
2092 for_each_node_state(nid, N_MEMORY) {
2093 struct node *node = node_devices[nid];
2094 if (node->dev.id == nid)
2095 hugetlb_register_node(node);
2096 }
2097
2098
2099
2100
2101
2102 register_hugetlbfs_with_node(hugetlb_register_node,
2103 hugetlb_unregister_node);
2104}
2105#else
2106
2107static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2108{
2109 BUG();
2110 if (nidp)
2111 *nidp = -1;
2112 return NULL;
2113}
2114
2115static void hugetlb_unregister_all_nodes(void) { }
2116
2117static void hugetlb_register_all_nodes(void) { }
2118
2119#endif
2120
2121static void __exit hugetlb_exit(void)
2122{
2123 struct hstate *h;
2124
2125 hugetlb_unregister_all_nodes();
2126
2127 for_each_hstate(h) {
2128 kobject_put(hstate_kobjs[hstate_index(h)]);
2129 }
2130
2131 kobject_put(hugepages_kobj);
2132 kfree(htlb_fault_mutex_table);
2133}
2134module_exit(hugetlb_exit);
2135
2136static int __init hugetlb_init(void)
2137{
2138 int i;
2139
2140 if (!hugepages_supported())
2141 return 0;
2142
2143 if (!size_to_hstate(default_hstate_size)) {
2144 default_hstate_size = HPAGE_SIZE;
2145 if (!size_to_hstate(default_hstate_size))
2146 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2147 }
2148 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2149 if (default_hstate_max_huge_pages)
2150 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2151
2152 hugetlb_init_hstates();
2153 gather_bootmem_prealloc();
2154 report_hugepages();
2155
2156 hugetlb_sysfs_init();
2157 hugetlb_register_all_nodes();
2158 hugetlb_cgroup_file_init();
2159
2160#ifdef CONFIG_SMP
2161 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2162#else
2163 num_fault_mutexes = 1;
2164#endif
2165 htlb_fault_mutex_table =
2166 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2167 BUG_ON(!htlb_fault_mutex_table);
2168
2169 for (i = 0; i < num_fault_mutexes; i++)
2170 mutex_init(&htlb_fault_mutex_table[i]);
2171 return 0;
2172}
2173module_init(hugetlb_init);
2174
2175
2176void __init hugetlb_add_hstate(unsigned order)
2177{
2178 struct hstate *h;
2179 unsigned long i;
2180
2181 if (size_to_hstate(PAGE_SIZE << order)) {
2182 pr_warning("hugepagesz= specified twice, ignoring\n");
2183 return;
2184 }
2185 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2186 BUG_ON(order == 0);
2187 h = &hstates[hugetlb_max_hstate++];
2188 h->order = order;
2189 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2190 h->nr_huge_pages = 0;
2191 h->free_huge_pages = 0;
2192 for (i = 0; i < MAX_NUMNODES; ++i)
2193 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2194 INIT_LIST_HEAD(&h->hugepage_activelist);
2195 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2196 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2197 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2198 huge_page_size(h)/1024);
2199
2200 parsed_hstate = h;
2201}
2202
2203static int __init hugetlb_nrpages_setup(char *s)
2204{
2205 unsigned long *mhp;
2206 static unsigned long *last_mhp;
2207
2208
2209
2210
2211
2212 if (!hugetlb_max_hstate)
2213 mhp = &default_hstate_max_huge_pages;
2214 else
2215 mhp = &parsed_hstate->max_huge_pages;
2216
2217 if (mhp == last_mhp) {
2218 pr_warning("hugepages= specified twice without "
2219 "interleaving hugepagesz=, ignoring\n");
2220 return 1;
2221 }
2222
2223 if (sscanf(s, "%lu", mhp) <= 0)
2224 *mhp = 0;
2225
2226
2227
2228
2229
2230
2231 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2232 hugetlb_hstate_alloc_pages(parsed_hstate);
2233
2234 last_mhp = mhp;
2235
2236 return 1;
2237}
2238__setup("hugepages=", hugetlb_nrpages_setup);
2239
2240static int __init hugetlb_default_setup(char *s)
2241{
2242 default_hstate_size = memparse(s, &s);
2243 return 1;
2244}
2245__setup("default_hugepagesz=", hugetlb_default_setup);
2246
2247static unsigned int cpuset_mems_nr(unsigned int *array)
2248{
2249 int node;
2250 unsigned int nr = 0;
2251
2252 for_each_node_mask(node, cpuset_current_mems_allowed)
2253 nr += array[node];
2254
2255 return nr;
2256}
2257
2258#ifdef CONFIG_SYSCTL
2259static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2260 struct ctl_table *table, int write,
2261 void __user *buffer, size_t *length, loff_t *ppos)
2262{
2263 struct hstate *h = &default_hstate;
2264 unsigned long tmp = h->max_huge_pages;
2265 int ret;
2266
2267 if (!hugepages_supported())
2268 return -ENOTSUPP;
2269
2270 table->data = &tmp;
2271 table->maxlen = sizeof(unsigned long);
2272 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2273 if (ret)
2274 goto out;
2275
2276 if (write)
2277 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2278 NUMA_NO_NODE, tmp, *length);
2279out:
2280 return ret;
2281}
2282
2283int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2284 void __user *buffer, size_t *length, loff_t *ppos)
2285{
2286
2287 return hugetlb_sysctl_handler_common(false, table, write,
2288 buffer, length, ppos);
2289}
2290
2291#ifdef CONFIG_NUMA
2292int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2293 void __user *buffer, size_t *length, loff_t *ppos)
2294{
2295 return hugetlb_sysctl_handler_common(true, table, write,
2296 buffer, length, ppos);
2297}
2298#endif
2299
2300int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2301 void __user *buffer,
2302 size_t *length, loff_t *ppos)
2303{
2304 struct hstate *h = &default_hstate;
2305 unsigned long tmp;
2306 int ret;
2307
2308 if (!hugepages_supported())
2309 return -ENOTSUPP;
2310
2311 tmp = h->nr_overcommit_huge_pages;
2312
2313 if (write && hstate_is_gigantic(h))
2314 return -EINVAL;
2315
2316 table->data = &tmp;
2317 table->maxlen = sizeof(unsigned long);
2318 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2319 if (ret)
2320 goto out;
2321
2322 if (write) {
2323 spin_lock(&hugetlb_lock);
2324 h->nr_overcommit_huge_pages = tmp;
2325 spin_unlock(&hugetlb_lock);
2326 }
2327out:
2328 return ret;
2329}
2330
2331#endif
2332
2333void hugetlb_report_meminfo(struct seq_file *m)
2334{
2335 struct hstate *h = &default_hstate;
2336 if (!hugepages_supported())
2337 return;
2338 seq_printf(m,
2339 "HugePages_Total: %5lu\n"
2340 "HugePages_Free: %5lu\n"
2341 "HugePages_Rsvd: %5lu\n"
2342 "HugePages_Surp: %5lu\n"
2343 "Hugepagesize: %8lu kB\n",
2344 h->nr_huge_pages,
2345 h->free_huge_pages,
2346 h->resv_huge_pages,
2347 h->surplus_huge_pages,
2348 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2349}
2350
2351int hugetlb_report_node_meminfo(int nid, char *buf)
2352{
2353 struct hstate *h = &default_hstate;
2354 if (!hugepages_supported())
2355 return 0;
2356 return sprintf(buf,
2357 "Node %d HugePages_Total: %5u\n"
2358 "Node %d HugePages_Free: %5u\n"
2359 "Node %d HugePages_Surp: %5u\n",
2360 nid, h->nr_huge_pages_node[nid],
2361 nid, h->free_huge_pages_node[nid],
2362 nid, h->surplus_huge_pages_node[nid]);
2363}
2364
2365void hugetlb_show_meminfo(void)
2366{
2367 struct hstate *h;
2368 int nid;
2369
2370 if (!hugepages_supported())
2371 return;
2372
2373 for_each_node_state(nid, N_MEMORY)
2374 for_each_hstate(h)
2375 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2376 nid,
2377 h->nr_huge_pages_node[nid],
2378 h->free_huge_pages_node[nid],
2379 h->surplus_huge_pages_node[nid],
2380 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2381}
2382
2383
2384unsigned long hugetlb_total_pages(void)
2385{
2386 struct hstate *h;
2387 unsigned long nr_total_pages = 0;
2388
2389 for_each_hstate(h)
2390 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2391 return nr_total_pages;
2392}
2393
2394static int hugetlb_acct_memory(struct hstate *h, long delta)
2395{
2396 int ret = -ENOMEM;
2397
2398 spin_lock(&hugetlb_lock);
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416 if (delta > 0) {
2417 if (gather_surplus_pages(h, delta) < 0)
2418 goto out;
2419
2420 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2421 return_unused_surplus_pages(h, delta);
2422 goto out;
2423 }
2424 }
2425
2426 ret = 0;
2427 if (delta < 0)
2428 return_unused_surplus_pages(h, (unsigned long) -delta);
2429
2430out:
2431 spin_unlock(&hugetlb_lock);
2432 return ret;
2433}
2434
2435static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2436{
2437 struct resv_map *resv = vma_resv_map(vma);
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2448 kref_get(&resv->refs);
2449}
2450
2451static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2452{
2453 struct hstate *h = hstate_vma(vma);
2454 struct resv_map *resv = vma_resv_map(vma);
2455 struct hugepage_subpool *spool = subpool_vma(vma);
2456 unsigned long reserve, start, end;
2457
2458 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2459 return;
2460
2461 start = vma_hugecache_offset(h, vma, vma->vm_start);
2462 end = vma_hugecache_offset(h, vma, vma->vm_end);
2463
2464 reserve = (end - start) - region_count(resv, start, end);
2465
2466 kref_put(&resv->refs, resv_map_release);
2467
2468 if (reserve) {
2469 hugetlb_acct_memory(h, -reserve);
2470 hugepage_subpool_put_pages(spool, reserve);
2471 }
2472}
2473
2474
2475
2476
2477
2478
2479
2480static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2481{
2482 BUG();
2483 return 0;
2484}
2485
2486const struct vm_operations_struct hugetlb_vm_ops = {
2487 .fault = hugetlb_vm_op_fault,
2488 .open = hugetlb_vm_op_open,
2489 .close = hugetlb_vm_op_close,
2490};
2491
2492static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2493 int writable)
2494{
2495 pte_t entry;
2496
2497 if (writable) {
2498 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2499 vma->vm_page_prot)));
2500 } else {
2501 entry = huge_pte_wrprotect(mk_huge_pte(page,
2502 vma->vm_page_prot));
2503 }
2504 entry = pte_mkyoung(entry);
2505 entry = pte_mkhuge(entry);
2506 entry = arch_make_huge_pte(entry, vma, page, writable);
2507
2508 return entry;
2509}
2510
2511static void set_huge_ptep_writable(struct vm_area_struct *vma,
2512 unsigned long address, pte_t *ptep)
2513{
2514 pte_t entry;
2515
2516 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2517 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2518 update_mmu_cache(vma, address, ptep);
2519}
2520
2521static int is_hugetlb_entry_migration(pte_t pte)
2522{
2523 swp_entry_t swp;
2524
2525 if (huge_pte_none(pte) || pte_present(pte))
2526 return 0;
2527 swp = pte_to_swp_entry(pte);
2528 if (non_swap_entry(swp) && is_migration_entry(swp))
2529 return 1;
2530 else
2531 return 0;
2532}
2533
2534static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2535{
2536 swp_entry_t swp;
2537
2538 if (huge_pte_none(pte) || pte_present(pte))
2539 return 0;
2540 swp = pte_to_swp_entry(pte);
2541 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2542 return 1;
2543 else
2544 return 0;
2545}
2546
2547int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2548 struct vm_area_struct *vma)
2549{
2550 pte_t *src_pte, *dst_pte, entry;
2551 struct page *ptepage;
2552 unsigned long addr;
2553 int cow;
2554 struct hstate *h = hstate_vma(vma);
2555 unsigned long sz = huge_page_size(h);
2556 unsigned long mmun_start;
2557 unsigned long mmun_end;
2558 int ret = 0;
2559
2560 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2561
2562 mmun_start = vma->vm_start;
2563 mmun_end = vma->vm_end;
2564 if (cow)
2565 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
2566
2567 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2568 spinlock_t *src_ptl, *dst_ptl;
2569 src_pte = huge_pte_offset(src, addr);
2570 if (!src_pte)
2571 continue;
2572 dst_pte = huge_pte_alloc(dst, addr, sz);
2573 if (!dst_pte) {
2574 ret = -ENOMEM;
2575 break;
2576 }
2577
2578
2579 if (dst_pte == src_pte)
2580 continue;
2581
2582 dst_ptl = huge_pte_lock(h, dst, dst_pte);
2583 src_ptl = huge_pte_lockptr(h, src, src_pte);
2584 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
2585 entry = huge_ptep_get(src_pte);
2586 if (huge_pte_none(entry)) {
2587 ;
2588 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2589 is_hugetlb_entry_hwpoisoned(entry))) {
2590 swp_entry_t swp_entry = pte_to_swp_entry(entry);
2591
2592 if (is_write_migration_entry(swp_entry) && cow) {
2593
2594
2595
2596
2597 make_migration_entry_read(&swp_entry);
2598 entry = swp_entry_to_pte(swp_entry);
2599 set_huge_pte_at(src, addr, src_pte, entry);
2600 }
2601 set_huge_pte_at(dst, addr, dst_pte, entry);
2602 } else {
2603 if (cow) {
2604 huge_ptep_set_wrprotect(src, addr, src_pte);
2605 mmu_notifier_invalidate_range(src, mmun_start,
2606 mmun_end);
2607 }
2608 entry = huge_ptep_get(src_pte);
2609 ptepage = pte_page(entry);
2610 get_page(ptepage);
2611 page_dup_rmap(ptepage);
2612 set_huge_pte_at(dst, addr, dst_pte, entry);
2613 }
2614 spin_unlock(src_ptl);
2615 spin_unlock(dst_ptl);
2616 }
2617
2618 if (cow)
2619 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
2620
2621 return ret;
2622}
2623
2624void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2625 unsigned long start, unsigned long end,
2626 struct page *ref_page)
2627{
2628 int force_flush = 0;
2629 struct mm_struct *mm = vma->vm_mm;
2630 unsigned long address;
2631 pte_t *ptep;
2632 pte_t pte;
2633 spinlock_t *ptl;
2634 struct page *page;
2635 struct hstate *h = hstate_vma(vma);
2636 unsigned long sz = huge_page_size(h);
2637 const unsigned long mmun_start = start;
2638 const unsigned long mmun_end = end;
2639
2640 WARN_ON(!is_vm_hugetlb_page(vma));
2641 BUG_ON(start & ~huge_page_mask(h));
2642 BUG_ON(end & ~huge_page_mask(h));
2643
2644 tlb_start_vma(tlb, vma);
2645 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2646 address = start;
2647again:
2648 for (; address < end; address += sz) {
2649 ptep = huge_pte_offset(mm, address);
2650 if (!ptep)
2651 continue;
2652
2653 ptl = huge_pte_lock(h, mm, ptep);
2654 if (huge_pmd_unshare(mm, &address, ptep))
2655 goto unlock;
2656
2657 pte = huge_ptep_get(ptep);
2658 if (huge_pte_none(pte))
2659 goto unlock;
2660
2661
2662
2663
2664
2665 if (unlikely(!pte_present(pte))) {
2666 huge_pte_clear(mm, address, ptep);
2667 goto unlock;
2668 }
2669
2670 page = pte_page(pte);
2671
2672
2673
2674
2675
2676 if (ref_page) {
2677 if (page != ref_page)
2678 goto unlock;
2679
2680
2681
2682
2683
2684
2685 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2686 }
2687
2688 pte = huge_ptep_get_and_clear(mm, address, ptep);
2689 tlb_remove_tlb_entry(tlb, ptep, address);
2690 if (huge_pte_dirty(pte))
2691 set_page_dirty(page);
2692
2693 page_remove_rmap(page);
2694 force_flush = !__tlb_remove_page(tlb, page);
2695 if (force_flush) {
2696 address += sz;
2697 spin_unlock(ptl);
2698 break;
2699 }
2700
2701 if (ref_page) {
2702 spin_unlock(ptl);
2703 break;
2704 }
2705unlock:
2706 spin_unlock(ptl);
2707 }
2708
2709
2710
2711
2712
2713 if (force_flush) {
2714 force_flush = 0;
2715 tlb_flush_mmu(tlb);
2716 if (address < end && !ref_page)
2717 goto again;
2718 }
2719 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2720 tlb_end_vma(tlb, vma);
2721}
2722
2723void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2724 struct vm_area_struct *vma, unsigned long start,
2725 unsigned long end, struct page *ref_page)
2726{
2727 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739 vma->vm_flags &= ~VM_MAYSHARE;
2740}
2741
2742void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2743 unsigned long end, struct page *ref_page)
2744{
2745 struct mm_struct *mm;
2746 struct mmu_gather tlb;
2747
2748 mm = vma->vm_mm;
2749
2750 tlb_gather_mmu(&tlb, mm, start, end);
2751 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2752 tlb_finish_mmu(&tlb, start, end);
2753}
2754
2755
2756
2757
2758
2759
2760
2761static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2762 struct page *page, unsigned long address)
2763{
2764 struct hstate *h = hstate_vma(vma);
2765 struct vm_area_struct *iter_vma;
2766 struct address_space *mapping;
2767 pgoff_t pgoff;
2768
2769
2770
2771
2772
2773 address = address & huge_page_mask(h);
2774 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2775 vma->vm_pgoff;
2776 mapping = file_inode(vma->vm_file)->i_mapping;
2777
2778
2779
2780
2781
2782
2783 i_mmap_lock_write(mapping);
2784 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2785
2786 if (iter_vma == vma)
2787 continue;
2788
2789
2790
2791
2792
2793
2794
2795
2796 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2797 unmap_hugepage_range(iter_vma, address,
2798 address + huge_page_size(h), page);
2799 }
2800 i_mmap_unlock_write(mapping);
2801}
2802
2803
2804
2805
2806
2807
2808
2809static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2810 unsigned long address, pte_t *ptep, pte_t pte,
2811 struct page *pagecache_page, spinlock_t *ptl)
2812{
2813 struct hstate *h = hstate_vma(vma);
2814 struct page *old_page, *new_page;
2815 int ret = 0, outside_reserve = 0;
2816 unsigned long mmun_start;
2817 unsigned long mmun_end;
2818
2819 old_page = pte_page(pte);
2820
2821retry_avoidcopy:
2822
2823
2824 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
2825 page_move_anon_rmap(old_page, vma, address);
2826 set_huge_ptep_writable(vma, address, ptep);
2827 return 0;
2828 }
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2840 old_page != pagecache_page)
2841 outside_reserve = 1;
2842
2843 page_cache_get(old_page);
2844
2845
2846
2847
2848
2849 spin_unlock(ptl);
2850 new_page = alloc_huge_page(vma, address, outside_reserve);
2851
2852 if (IS_ERR(new_page)) {
2853
2854
2855
2856
2857
2858
2859
2860 if (outside_reserve) {
2861 page_cache_release(old_page);
2862 BUG_ON(huge_pte_none(pte));
2863 unmap_ref_private(mm, vma, old_page, address);
2864 BUG_ON(huge_pte_none(pte));
2865 spin_lock(ptl);
2866 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2867 if (likely(ptep &&
2868 pte_same(huge_ptep_get(ptep), pte)))
2869 goto retry_avoidcopy;
2870
2871
2872
2873
2874 return 0;
2875 }
2876
2877 ret = (PTR_ERR(new_page) == -ENOMEM) ?
2878 VM_FAULT_OOM : VM_FAULT_SIGBUS;
2879 goto out_release_old;
2880 }
2881
2882
2883
2884
2885
2886 if (unlikely(anon_vma_prepare(vma))) {
2887 ret = VM_FAULT_OOM;
2888 goto out_release_all;
2889 }
2890
2891 copy_user_huge_page(new_page, old_page, address, vma,
2892 pages_per_huge_page(h));
2893 __SetPageUptodate(new_page);
2894
2895 mmun_start = address & huge_page_mask(h);
2896 mmun_end = mmun_start + huge_page_size(h);
2897 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2898
2899
2900
2901
2902
2903 spin_lock(ptl);
2904 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2905 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
2906 ClearPagePrivate(new_page);
2907
2908
2909 huge_ptep_clear_flush(vma, address, ptep);
2910 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
2911 set_huge_pte_at(mm, address, ptep,
2912 make_huge_pte(vma, new_page, 1));
2913 page_remove_rmap(old_page);
2914 hugepage_add_new_anon_rmap(new_page, vma, address);
2915
2916 new_page = old_page;
2917 }
2918 spin_unlock(ptl);
2919 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2920out_release_all:
2921 page_cache_release(new_page);
2922out_release_old:
2923 page_cache_release(old_page);
2924
2925 spin_lock(ptl);
2926 return ret;
2927}
2928
2929
2930static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2931 struct vm_area_struct *vma, unsigned long address)
2932{
2933 struct address_space *mapping;
2934 pgoff_t idx;
2935
2936 mapping = vma->vm_file->f_mapping;
2937 idx = vma_hugecache_offset(h, vma, address);
2938
2939 return find_lock_page(mapping, idx);
2940}
2941
2942
2943
2944
2945
2946static bool hugetlbfs_pagecache_present(struct hstate *h,
2947 struct vm_area_struct *vma, unsigned long address)
2948{
2949 struct address_space *mapping;
2950 pgoff_t idx;
2951 struct page *page;
2952
2953 mapping = vma->vm_file->f_mapping;
2954 idx = vma_hugecache_offset(h, vma, address);
2955
2956 page = find_get_page(mapping, idx);
2957 if (page)
2958 put_page(page);
2959 return page != NULL;
2960}
2961
2962static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2963 struct address_space *mapping, pgoff_t idx,
2964 unsigned long address, pte_t *ptep, unsigned int flags)
2965{
2966 struct hstate *h = hstate_vma(vma);
2967 int ret = VM_FAULT_SIGBUS;
2968 int anon_rmap = 0;
2969 unsigned long size;
2970 struct page *page;
2971 pte_t new_pte;
2972 spinlock_t *ptl;
2973
2974
2975
2976
2977
2978
2979 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2980 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2981 current->pid);
2982 return ret;
2983 }
2984
2985
2986
2987
2988
2989retry:
2990 page = find_lock_page(mapping, idx);
2991 if (!page) {
2992 size = i_size_read(mapping->host) >> huge_page_shift(h);
2993 if (idx >= size)
2994 goto out;
2995 page = alloc_huge_page(vma, address, 0);
2996 if (IS_ERR(page)) {
2997 ret = PTR_ERR(page);
2998 if (ret == -ENOMEM)
2999 ret = VM_FAULT_OOM;
3000 else
3001 ret = VM_FAULT_SIGBUS;
3002 goto out;
3003 }
3004 clear_huge_page(page, address, pages_per_huge_page(h));
3005 __SetPageUptodate(page);
3006
3007 if (vma->vm_flags & VM_MAYSHARE) {
3008 int err;
3009 struct inode *inode = mapping->host;
3010
3011 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3012 if (err) {
3013 put_page(page);
3014 if (err == -EEXIST)
3015 goto retry;
3016 goto out;
3017 }
3018 ClearPagePrivate(page);
3019
3020 spin_lock(&inode->i_lock);
3021 inode->i_blocks += blocks_per_huge_page(h);
3022 spin_unlock(&inode->i_lock);
3023 } else {
3024 lock_page(page);
3025 if (unlikely(anon_vma_prepare(vma))) {
3026 ret = VM_FAULT_OOM;
3027 goto backout_unlocked;
3028 }
3029 anon_rmap = 1;
3030 }
3031 } else {
3032
3033
3034
3035
3036
3037 if (unlikely(PageHWPoison(page))) {
3038 ret = VM_FAULT_HWPOISON |
3039 VM_FAULT_SET_HINDEX(hstate_index(h));
3040 goto backout_unlocked;
3041 }
3042 }
3043
3044
3045
3046
3047
3048
3049
3050 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
3051 if (vma_needs_reservation(h, vma, address) < 0) {
3052 ret = VM_FAULT_OOM;
3053 goto backout_unlocked;
3054 }
3055
3056 ptl = huge_pte_lockptr(h, mm, ptep);
3057 spin_lock(ptl);
3058 size = i_size_read(mapping->host) >> huge_page_shift(h);
3059 if (idx >= size)
3060 goto backout;
3061
3062 ret = 0;
3063 if (!huge_pte_none(huge_ptep_get(ptep)))
3064 goto backout;
3065
3066 if (anon_rmap) {
3067 ClearPagePrivate(page);
3068 hugepage_add_new_anon_rmap(page, vma, address);
3069 } else
3070 page_dup_rmap(page);
3071 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3072 && (vma->vm_flags & VM_SHARED)));
3073 set_huge_pte_at(mm, address, ptep, new_pte);
3074
3075 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3076
3077 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
3078 }
3079
3080 spin_unlock(ptl);
3081 unlock_page(page);
3082out:
3083 return ret;
3084
3085backout:
3086 spin_unlock(ptl);
3087backout_unlocked:
3088 unlock_page(page);
3089 put_page(page);
3090 goto out;
3091}
3092
3093#ifdef CONFIG_SMP
3094static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3095 struct vm_area_struct *vma,
3096 struct address_space *mapping,
3097 pgoff_t idx, unsigned long address)
3098{
3099 unsigned long key[2];
3100 u32 hash;
3101
3102 if (vma->vm_flags & VM_SHARED) {
3103 key[0] = (unsigned long) mapping;
3104 key[1] = idx;
3105 } else {
3106 key[0] = (unsigned long) mm;
3107 key[1] = address >> huge_page_shift(h);
3108 }
3109
3110 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3111
3112 return hash & (num_fault_mutexes - 1);
3113}
3114#else
3115
3116
3117
3118
3119static u32 fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3120 struct vm_area_struct *vma,
3121 struct address_space *mapping,
3122 pgoff_t idx, unsigned long address)
3123{
3124 return 0;
3125}
3126#endif
3127
3128int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3129 unsigned long address, unsigned int flags)
3130{
3131 pte_t *ptep, entry;
3132 spinlock_t *ptl;
3133 int ret;
3134 u32 hash;
3135 pgoff_t idx;
3136 struct page *page = NULL;
3137 struct page *pagecache_page = NULL;
3138 struct hstate *h = hstate_vma(vma);
3139 struct address_space *mapping;
3140 int need_wait_lock = 0;
3141
3142 address &= huge_page_mask(h);
3143
3144 ptep = huge_pte_offset(mm, address);
3145 if (ptep) {
3146 entry = huge_ptep_get(ptep);
3147 if (unlikely(is_hugetlb_entry_migration(entry))) {
3148 migration_entry_wait_huge(vma, mm, ptep);
3149 return 0;
3150 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3151 return VM_FAULT_HWPOISON_LARGE |
3152 VM_FAULT_SET_HINDEX(hstate_index(h));
3153 }
3154
3155 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3156 if (!ptep)
3157 return VM_FAULT_OOM;
3158
3159 mapping = vma->vm_file->f_mapping;
3160 idx = vma_hugecache_offset(h, vma, address);
3161
3162
3163
3164
3165
3166
3167 hash = fault_mutex_hash(h, mm, vma, mapping, idx, address);
3168 mutex_lock(&htlb_fault_mutex_table[hash]);
3169
3170 entry = huge_ptep_get(ptep);
3171 if (huge_pte_none(entry)) {
3172 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3173 goto out_mutex;
3174 }
3175
3176 ret = 0;
3177
3178
3179
3180
3181
3182
3183
3184
3185 if (!pte_present(entry))
3186 goto out_mutex;
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3197 if (vma_needs_reservation(h, vma, address) < 0) {
3198 ret = VM_FAULT_OOM;
3199 goto out_mutex;
3200 }
3201
3202 if (!(vma->vm_flags & VM_MAYSHARE))
3203 pagecache_page = hugetlbfs_pagecache_page(h,
3204 vma, address);
3205 }
3206
3207 ptl = huge_pte_lock(h, mm, ptep);
3208
3209
3210 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3211 goto out_ptl;
3212
3213
3214
3215
3216
3217
3218 page = pte_page(entry);
3219 if (page != pagecache_page)
3220 if (!trylock_page(page)) {
3221 need_wait_lock = 1;
3222 goto out_ptl;
3223 }
3224
3225 get_page(page);
3226
3227 if (flags & FAULT_FLAG_WRITE) {
3228 if (!huge_pte_write(entry)) {
3229 ret = hugetlb_cow(mm, vma, address, ptep, entry,
3230 pagecache_page, ptl);
3231 goto out_put_page;
3232 }
3233 entry = huge_pte_mkdirty(entry);
3234 }
3235 entry = pte_mkyoung(entry);
3236 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3237 flags & FAULT_FLAG_WRITE))
3238 update_mmu_cache(vma, address, ptep);
3239out_put_page:
3240 if (page != pagecache_page)
3241 unlock_page(page);
3242 put_page(page);
3243out_ptl:
3244 spin_unlock(ptl);
3245
3246 if (pagecache_page) {
3247 unlock_page(pagecache_page);
3248 put_page(pagecache_page);
3249 }
3250out_mutex:
3251 mutex_unlock(&htlb_fault_mutex_table[hash]);
3252
3253
3254
3255
3256
3257
3258
3259 if (need_wait_lock)
3260 wait_on_page_locked(page);
3261 return ret;
3262}
3263
3264long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3265 struct page **pages, struct vm_area_struct **vmas,
3266 unsigned long *position, unsigned long *nr_pages,
3267 long i, unsigned int flags)
3268{
3269 unsigned long pfn_offset;
3270 unsigned long vaddr = *position;
3271 unsigned long remainder = *nr_pages;
3272 struct hstate *h = hstate_vma(vma);
3273
3274 while (vaddr < vma->vm_end && remainder) {
3275 pte_t *pte;
3276 spinlock_t *ptl = NULL;
3277 int absent;
3278 struct page *page;
3279
3280
3281
3282
3283
3284
3285
3286
3287 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3288 if (pte)
3289 ptl = huge_pte_lock(h, mm, pte);
3290 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3291
3292
3293
3294
3295
3296
3297
3298
3299 if (absent && (flags & FOLL_DUMP) &&
3300 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3301 if (pte)
3302 spin_unlock(ptl);
3303 remainder = 0;
3304 break;
3305 }
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3318 ((flags & FOLL_WRITE) &&
3319 !huge_pte_write(huge_ptep_get(pte)))) {
3320 int ret;
3321
3322 if (pte)
3323 spin_unlock(ptl);
3324 ret = hugetlb_fault(mm, vma, vaddr,
3325 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3326 if (!(ret & VM_FAULT_ERROR))
3327 continue;
3328
3329 remainder = 0;
3330 break;
3331 }
3332
3333 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3334 page = pte_page(huge_ptep_get(pte));
3335same_page:
3336 if (pages) {
3337 pages[i] = mem_map_offset(page, pfn_offset);
3338 get_page_foll(pages[i]);
3339 }
3340
3341 if (vmas)
3342 vmas[i] = vma;
3343
3344 vaddr += PAGE_SIZE;
3345 ++pfn_offset;
3346 --remainder;
3347 ++i;
3348 if (vaddr < vma->vm_end && remainder &&
3349 pfn_offset < pages_per_huge_page(h)) {
3350
3351
3352
3353
3354 goto same_page;
3355 }
3356 spin_unlock(ptl);
3357 }
3358 *nr_pages = remainder;
3359 *position = vaddr;
3360
3361 return i ? i : -EFAULT;
3362}
3363
3364unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3365 unsigned long address, unsigned long end, pgprot_t newprot)
3366{
3367 struct mm_struct *mm = vma->vm_mm;
3368 unsigned long start = address;
3369 pte_t *ptep;
3370 pte_t pte;
3371 struct hstate *h = hstate_vma(vma);
3372 unsigned long pages = 0;
3373
3374 BUG_ON(address >= end);
3375 flush_cache_range(vma, address, end);
3376
3377 mmu_notifier_invalidate_range_start(mm, start, end);
3378 i_mmap_lock_write(vma->vm_file->f_mapping);
3379 for (; address < end; address += huge_page_size(h)) {
3380 spinlock_t *ptl;
3381 ptep = huge_pte_offset(mm, address);
3382 if (!ptep)
3383 continue;
3384 ptl = huge_pte_lock(h, mm, ptep);
3385 if (huge_pmd_unshare(mm, &address, ptep)) {
3386 pages++;
3387 spin_unlock(ptl);
3388 continue;
3389 }
3390 pte = huge_ptep_get(ptep);
3391 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3392 spin_unlock(ptl);
3393 continue;
3394 }
3395 if (unlikely(is_hugetlb_entry_migration(pte))) {
3396 swp_entry_t entry = pte_to_swp_entry(pte);
3397
3398 if (is_write_migration_entry(entry)) {
3399 pte_t newpte;
3400
3401 make_migration_entry_read(&entry);
3402 newpte = swp_entry_to_pte(entry);
3403 set_huge_pte_at(mm, address, ptep, newpte);
3404 pages++;
3405 }
3406 spin_unlock(ptl);
3407 continue;
3408 }
3409 if (!huge_pte_none(pte)) {
3410 pte = huge_ptep_get_and_clear(mm, address, ptep);
3411 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3412 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3413 set_huge_pte_at(mm, address, ptep, pte);
3414 pages++;
3415 }
3416 spin_unlock(ptl);
3417 }
3418
3419
3420
3421
3422
3423
3424 flush_tlb_range(vma, start, end);
3425 mmu_notifier_invalidate_range(mm, start, end);
3426 i_mmap_unlock_write(vma->vm_file->f_mapping);
3427 mmu_notifier_invalidate_range_end(mm, start, end);
3428
3429 return pages << h->order;
3430}
3431
3432int hugetlb_reserve_pages(struct inode *inode,
3433 long from, long to,
3434 struct vm_area_struct *vma,
3435 vm_flags_t vm_flags)
3436{
3437 long ret, chg;
3438 struct hstate *h = hstate_inode(inode);
3439 struct hugepage_subpool *spool = subpool_inode(inode);
3440 struct resv_map *resv_map;
3441
3442
3443
3444
3445
3446
3447 if (vm_flags & VM_NORESERVE)
3448 return 0;
3449
3450
3451
3452
3453
3454
3455
3456 if (!vma || vma->vm_flags & VM_MAYSHARE) {
3457 resv_map = inode_resv_map(inode);
3458
3459 chg = region_chg(resv_map, from, to);
3460
3461 } else {
3462 resv_map = resv_map_alloc();
3463 if (!resv_map)
3464 return -ENOMEM;
3465
3466 chg = to - from;
3467
3468 set_vma_resv_map(vma, resv_map);
3469 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3470 }
3471
3472 if (chg < 0) {
3473 ret = chg;
3474 goto out_err;
3475 }
3476
3477
3478 if (hugepage_subpool_get_pages(spool, chg)) {
3479 ret = -ENOSPC;
3480 goto out_err;
3481 }
3482
3483
3484
3485
3486
3487 ret = hugetlb_acct_memory(h, chg);
3488 if (ret < 0) {
3489 hugepage_subpool_put_pages(spool, chg);
3490 goto out_err;
3491 }
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504 if (!vma || vma->vm_flags & VM_MAYSHARE)
3505 region_add(resv_map, from, to);
3506 return 0;
3507out_err:
3508 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3509 kref_put(&resv_map->refs, resv_map_release);
3510 return ret;
3511}
3512
3513void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3514{
3515 struct hstate *h = hstate_inode(inode);
3516 struct resv_map *resv_map = inode_resv_map(inode);
3517 long chg = 0;
3518 struct hugepage_subpool *spool = subpool_inode(inode);
3519
3520 if (resv_map)
3521 chg = region_truncate(resv_map, offset);
3522 spin_lock(&inode->i_lock);
3523 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3524 spin_unlock(&inode->i_lock);
3525
3526 hugepage_subpool_put_pages(spool, (chg - freed));
3527 hugetlb_acct_memory(h, -(chg - freed));
3528}
3529
3530#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
3531static unsigned long page_table_shareable(struct vm_area_struct *svma,
3532 struct vm_area_struct *vma,
3533 unsigned long addr, pgoff_t idx)
3534{
3535 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
3536 svma->vm_start;
3537 unsigned long sbase = saddr & PUD_MASK;
3538 unsigned long s_end = sbase + PUD_SIZE;
3539
3540
3541 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
3542 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
3543
3544
3545
3546
3547
3548 if (pmd_index(addr) != pmd_index(saddr) ||
3549 vm_flags != svm_flags ||
3550 sbase < svma->vm_start || svma->vm_end < s_end)
3551 return 0;
3552
3553 return saddr;
3554}
3555
3556static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
3557{
3558 unsigned long base = addr & PUD_MASK;
3559 unsigned long end = base + PUD_SIZE;
3560
3561
3562
3563
3564 if (vma->vm_flags & VM_MAYSHARE &&
3565 vma->vm_start <= base && end <= vma->vm_end)
3566 return 1;
3567 return 0;
3568}
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3580{
3581 struct vm_area_struct *vma = find_vma(mm, addr);
3582 struct address_space *mapping = vma->vm_file->f_mapping;
3583 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
3584 vma->vm_pgoff;
3585 struct vm_area_struct *svma;
3586 unsigned long saddr;
3587 pte_t *spte = NULL;
3588 pte_t *pte;
3589 spinlock_t *ptl;
3590
3591 if (!vma_shareable(vma, addr))
3592 return (pte_t *)pmd_alloc(mm, pud, addr);
3593
3594 i_mmap_lock_write(mapping);
3595 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
3596 if (svma == vma)
3597 continue;
3598
3599 saddr = page_table_shareable(svma, vma, addr, idx);
3600 if (saddr) {
3601 spte = huge_pte_offset(svma->vm_mm, saddr);
3602 if (spte) {
3603 mm_inc_nr_pmds(mm);
3604 get_page(virt_to_page(spte));
3605 break;
3606 }
3607 }
3608 }
3609
3610 if (!spte)
3611 goto out;
3612
3613 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
3614 spin_lock(ptl);
3615 if (pud_none(*pud)) {
3616 pud_populate(mm, pud,
3617 (pmd_t *)((unsigned long)spte & PAGE_MASK));
3618 } else {
3619 put_page(virt_to_page(spte));
3620 mm_inc_nr_pmds(mm);
3621 }
3622 spin_unlock(ptl);
3623out:
3624 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3625 i_mmap_unlock_write(mapping);
3626 return pte;
3627}
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
3642{
3643 pgd_t *pgd = pgd_offset(mm, *addr);
3644 pud_t *pud = pud_offset(pgd, *addr);
3645
3646 BUG_ON(page_count(virt_to_page(ptep)) == 0);
3647 if (page_count(virt_to_page(ptep)) == 1)
3648 return 0;
3649
3650 pud_clear(pud);
3651 put_page(virt_to_page(ptep));
3652 mm_dec_nr_pmds(mm);
3653 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
3654 return 1;
3655}
3656#define want_pmd_share() (1)
3657#else
3658pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
3659{
3660 return NULL;
3661}
3662#define want_pmd_share() (0)
3663#endif
3664
3665#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
3666pte_t *huge_pte_alloc(struct mm_struct *mm,
3667 unsigned long addr, unsigned long sz)
3668{
3669 pgd_t *pgd;
3670 pud_t *pud;
3671 pte_t *pte = NULL;
3672
3673 pgd = pgd_offset(mm, addr);
3674 pud = pud_alloc(mm, pgd, addr);
3675 if (pud) {
3676 if (sz == PUD_SIZE) {
3677 pte = (pte_t *)pud;
3678 } else {
3679 BUG_ON(sz != PMD_SIZE);
3680 if (want_pmd_share() && pud_none(*pud))
3681 pte = huge_pmd_share(mm, addr, pud);
3682 else
3683 pte = (pte_t *)pmd_alloc(mm, pud, addr);
3684 }
3685 }
3686 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
3687
3688 return pte;
3689}
3690
3691pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
3692{
3693 pgd_t *pgd;
3694 pud_t *pud;
3695 pmd_t *pmd = NULL;
3696
3697 pgd = pgd_offset(mm, addr);
3698 if (pgd_present(*pgd)) {
3699 pud = pud_offset(pgd, addr);
3700 if (pud_present(*pud)) {
3701 if (pud_huge(*pud))
3702 return (pte_t *)pud;
3703 pmd = pmd_offset(pud, addr);
3704 }
3705 }
3706 return (pte_t *) pmd;
3707}
3708
3709#endif
3710
3711
3712
3713
3714
3715struct page * __weak
3716follow_huge_addr(struct mm_struct *mm, unsigned long address,
3717 int write)
3718{
3719 return ERR_PTR(-EINVAL);
3720}
3721
3722struct page * __weak
3723follow_huge_pmd(struct mm_struct *mm, unsigned long address,
3724 pmd_t *pmd, int flags)
3725{
3726 struct page *page = NULL;
3727 spinlock_t *ptl;
3728retry:
3729 ptl = pmd_lockptr(mm, pmd);
3730 spin_lock(ptl);
3731
3732
3733
3734
3735 if (!pmd_huge(*pmd))
3736 goto out;
3737 if (pmd_present(*pmd)) {
3738 page = pte_page(*(pte_t *)pmd) +
3739 ((address & ~PMD_MASK) >> PAGE_SHIFT);
3740 if (flags & FOLL_GET)
3741 get_page(page);
3742 } else {
3743 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
3744 spin_unlock(ptl);
3745 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
3746 goto retry;
3747 }
3748
3749
3750
3751
3752 }
3753out:
3754 spin_unlock(ptl);
3755 return page;
3756}
3757
3758struct page * __weak
3759follow_huge_pud(struct mm_struct *mm, unsigned long address,
3760 pud_t *pud, int flags)
3761{
3762 if (flags & FOLL_GET)
3763 return NULL;
3764
3765 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
3766}
3767
3768#ifdef CONFIG_MEMORY_FAILURE
3769
3770
3771static int is_hugepage_on_freelist(struct page *hpage)
3772{
3773 struct page *page;
3774 struct page *tmp;
3775 struct hstate *h = page_hstate(hpage);
3776 int nid = page_to_nid(hpage);
3777
3778 list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3779 if (page == hpage)
3780 return 1;
3781 return 0;
3782}
3783
3784
3785
3786
3787
3788int dequeue_hwpoisoned_huge_page(struct page *hpage)
3789{
3790 struct hstate *h = page_hstate(hpage);
3791 int nid = page_to_nid(hpage);
3792 int ret = -EBUSY;
3793
3794 spin_lock(&hugetlb_lock);
3795 if (is_hugepage_on_freelist(hpage)) {
3796
3797
3798
3799
3800
3801
3802 list_del_init(&hpage->lru);
3803 set_page_refcounted(hpage);
3804 h->free_huge_pages--;
3805 h->free_huge_pages_node[nid]--;
3806 ret = 0;
3807 }
3808 spin_unlock(&hugetlb_lock);
3809 return ret;
3810}
3811#endif
3812
3813bool isolate_huge_page(struct page *page, struct list_head *list)
3814{
3815 VM_BUG_ON_PAGE(!PageHead(page), page);
3816 if (!get_page_unless_zero(page))
3817 return false;
3818 spin_lock(&hugetlb_lock);
3819 list_move_tail(&page->lru, list);
3820 spin_unlock(&hugetlb_lock);
3821 return true;
3822}
3823
3824void putback_active_hugepage(struct page *page)
3825{
3826 VM_BUG_ON_PAGE(!PageHead(page), page);
3827 spin_lock(&hugetlb_lock);
3828 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
3829 spin_unlock(&hugetlb_lock);
3830 put_page(page);
3831}
3832
3833bool is_hugepage_active(struct page *page)
3834{
3835 VM_BUG_ON_PAGE(!PageHuge(page), page);
3836
3837
3838
3839
3840
3841
3842
3843
3844 if (PageTail(page))
3845 return false;
3846
3847
3848
3849
3850 if (unlikely(PageHWPoison(page)))
3851 return false;
3852 return page_count(page) > 0;
3853}
3854