Merge remote-tracking branch 'origin/dev/sumit-linux-3.10.96' into TOT-merge
[linux-3.10.git] / mm / hugetlb.c
1 /*
2  * Generic hugetlb support.
3  * (C) Nadia Yvette Chambers, April 2004
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/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/rmap.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
24 #include <linux/page-isolation.h>
25
26 #include <asm/page.h>
27 #include <asm/pgtable.h>
28 #include <asm/tlb.h>
29
30 #include <linux/io.h>
31 #include <linux/hugetlb.h>
32 #include <linux/hugetlb_cgroup.h>
33 #include <linux/node.h>
34 #include "internal.h"
35
36 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
37 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
38 unsigned long hugepages_treat_as_movable;
39
40 int hugetlb_max_hstate __read_mostly;
41 unsigned int default_hstate_idx;
42 struct hstate hstates[HUGE_MAX_HSTATE];
43
44 __initdata LIST_HEAD(huge_boot_pages);
45
46 /* for command line parsing */
47 static struct hstate * __initdata parsed_hstate;
48 static unsigned long __initdata default_hstate_max_huge_pages;
49 static unsigned long __initdata default_hstate_size;
50
51 /*
52  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
53  */
54 DEFINE_SPINLOCK(hugetlb_lock);
55
56 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
57 {
58         bool free = (spool->count == 0) && (spool->used_hpages == 0);
59
60         spin_unlock(&spool->lock);
61
62         /* If no pages are used, and no other handles to the subpool
63          * remain, free the subpool the subpool remain */
64         if (free)
65                 kfree(spool);
66 }
67
68 struct hugepage_subpool *hugepage_new_subpool(long nr_blocks)
69 {
70         struct hugepage_subpool *spool;
71
72         spool = kmalloc(sizeof(*spool), GFP_KERNEL);
73         if (!spool)
74                 return NULL;
75
76         spin_lock_init(&spool->lock);
77         spool->count = 1;
78         spool->max_hpages = nr_blocks;
79         spool->used_hpages = 0;
80
81         return spool;
82 }
83
84 void hugepage_put_subpool(struct hugepage_subpool *spool)
85 {
86         spin_lock(&spool->lock);
87         BUG_ON(!spool->count);
88         spool->count--;
89         unlock_or_release_subpool(spool);
90 }
91
92 static int hugepage_subpool_get_pages(struct hugepage_subpool *spool,
93                                       long delta)
94 {
95         int ret = 0;
96
97         if (!spool)
98                 return 0;
99
100         spin_lock(&spool->lock);
101         if ((spool->used_hpages + delta) <= spool->max_hpages) {
102                 spool->used_hpages += delta;
103         } else {
104                 ret = -ENOMEM;
105         }
106         spin_unlock(&spool->lock);
107
108         return ret;
109 }
110
111 static void hugepage_subpool_put_pages(struct hugepage_subpool *spool,
112                                        long delta)
113 {
114         if (!spool)
115                 return;
116
117         spin_lock(&spool->lock);
118         spool->used_hpages -= delta;
119         /* If hugetlbfs_put_super couldn't free spool due to
120         * an outstanding quota reference, free it now. */
121         unlock_or_release_subpool(spool);
122 }
123
124 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
125 {
126         return HUGETLBFS_SB(inode->i_sb)->spool;
127 }
128
129 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
130 {
131         return subpool_inode(file_inode(vma->vm_file));
132 }
133
134 /*
135  * Region tracking -- allows tracking of reservations and instantiated pages
136  *                    across the pages in a mapping.
137  *
138  * The region data structures are protected by a combination of the mmap_sem
139  * and the hugetlb_instantion_mutex.  To access or modify a region the caller
140  * must either hold the mmap_sem for write, or the mmap_sem for read and
141  * the hugetlb_instantiation mutex:
142  *
143  *      down_write(&mm->mmap_sem);
144  * or
145  *      down_read(&mm->mmap_sem);
146  *      mutex_lock(&hugetlb_instantiation_mutex);
147  */
148 struct file_region {
149         struct list_head link;
150         long from;
151         long to;
152 };
153
154 static long region_add(struct list_head *head, long f, long t)
155 {
156         struct file_region *rg, *nrg, *trg;
157
158         /* Locate the region we are either in or before. */
159         list_for_each_entry(rg, head, link)
160                 if (f <= rg->to)
161                         break;
162
163         /* Round our left edge to the current segment if it encloses us. */
164         if (f > rg->from)
165                 f = rg->from;
166
167         /* Check for and consume any regions we now overlap with. */
168         nrg = rg;
169         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
170                 if (&rg->link == head)
171                         break;
172                 if (rg->from > t)
173                         break;
174
175                 /* If this area reaches higher then extend our area to
176                  * include it completely.  If this is not the first area
177                  * which we intend to reuse, free it. */
178                 if (rg->to > t)
179                         t = rg->to;
180                 if (rg != nrg) {
181                         list_del(&rg->link);
182                         kfree(rg);
183                 }
184         }
185         nrg->from = f;
186         nrg->to = t;
187         return 0;
188 }
189
190 static long region_chg(struct list_head *head, long f, long t)
191 {
192         struct file_region *rg, *nrg;
193         long chg = 0;
194
195         /* Locate the region we are before or in. */
196         list_for_each_entry(rg, head, link)
197                 if (f <= rg->to)
198                         break;
199
200         /* If we are below the current region then a new region is required.
201          * Subtle, allocate a new region at the position but make it zero
202          * size such that we can guarantee to record the reservation. */
203         if (&rg->link == head || t < rg->from) {
204                 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
205                 if (!nrg)
206                         return -ENOMEM;
207                 nrg->from = f;
208                 nrg->to   = f;
209                 INIT_LIST_HEAD(&nrg->link);
210                 list_add(&nrg->link, rg->link.prev);
211
212                 return t - f;
213         }
214
215         /* Round our left edge to the current segment if it encloses us. */
216         if (f > rg->from)
217                 f = rg->from;
218         chg = t - f;
219
220         /* Check for and consume any regions we now overlap with. */
221         list_for_each_entry(rg, rg->link.prev, link) {
222                 if (&rg->link == head)
223                         break;
224                 if (rg->from > t)
225                         return chg;
226
227                 /* We overlap with this area, if it extends further than
228                  * us then we must extend ourselves.  Account for its
229                  * existing reservation. */
230                 if (rg->to > t) {
231                         chg += rg->to - t;
232                         t = rg->to;
233                 }
234                 chg -= rg->to - rg->from;
235         }
236         return chg;
237 }
238
239 static long region_truncate(struct list_head *head, long end)
240 {
241         struct file_region *rg, *trg;
242         long chg = 0;
243
244         /* Locate the region we are either in or before. */
245         list_for_each_entry(rg, head, link)
246                 if (end <= rg->to)
247                         break;
248         if (&rg->link == head)
249                 return 0;
250
251         /* If we are in the middle of a region then adjust it. */
252         if (end > rg->from) {
253                 chg = rg->to - end;
254                 rg->to = end;
255                 rg = list_entry(rg->link.next, typeof(*rg), link);
256         }
257
258         /* Drop any remaining regions. */
259         list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
260                 if (&rg->link == head)
261                         break;
262                 chg += rg->to - rg->from;
263                 list_del(&rg->link);
264                 kfree(rg);
265         }
266         return chg;
267 }
268
269 static long region_count(struct list_head *head, long f, long t)
270 {
271         struct file_region *rg;
272         long chg = 0;
273
274         /* Locate each segment we overlap with, and count that overlap. */
275         list_for_each_entry(rg, head, link) {
276                 long seg_from;
277                 long seg_to;
278
279                 if (rg->to <= f)
280                         continue;
281                 if (rg->from >= t)
282                         break;
283
284                 seg_from = max(rg->from, f);
285                 seg_to = min(rg->to, t);
286
287                 chg += seg_to - seg_from;
288         }
289
290         return chg;
291 }
292
293 /*
294  * Convert the address within this vma to the page offset within
295  * the mapping, in pagecache page units; huge pages here.
296  */
297 static pgoff_t vma_hugecache_offset(struct hstate *h,
298                         struct vm_area_struct *vma, unsigned long address)
299 {
300         return ((address - vma->vm_start) >> huge_page_shift(h)) +
301                         (vma->vm_pgoff >> huge_page_order(h));
302 }
303
304 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
305                                      unsigned long address)
306 {
307         return vma_hugecache_offset(hstate_vma(vma), vma, address);
308 }
309
310 /*
311  * Return the size of the pages allocated when backing a VMA. In the majority
312  * cases this will be same size as used by the page table entries.
313  */
314 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
315 {
316         struct hstate *hstate;
317
318         if (!is_vm_hugetlb_page(vma))
319                 return PAGE_SIZE;
320
321         hstate = hstate_vma(vma);
322
323         return 1UL << (hstate->order + PAGE_SHIFT);
324 }
325 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
326
327 /*
328  * Return the page size being used by the MMU to back a VMA. In the majority
329  * of cases, the page size used by the kernel matches the MMU size. On
330  * architectures where it differs, an architecture-specific version of this
331  * function is required.
332  */
333 #ifndef vma_mmu_pagesize
334 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
335 {
336         return vma_kernel_pagesize(vma);
337 }
338 #endif
339
340 /*
341  * Flags for MAP_PRIVATE reservations.  These are stored in the bottom
342  * bits of the reservation map pointer, which are always clear due to
343  * alignment.
344  */
345 #define HPAGE_RESV_OWNER    (1UL << 0)
346 #define HPAGE_RESV_UNMAPPED (1UL << 1)
347 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
348
349 /*
350  * These helpers are used to track how many pages are reserved for
351  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
352  * is guaranteed to have their future faults succeed.
353  *
354  * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
355  * the reserve counters are updated with the hugetlb_lock held. It is safe
356  * to reset the VMA at fork() time as it is not in use yet and there is no
357  * chance of the global counters getting corrupted as a result of the values.
358  *
359  * The private mapping reservation is represented in a subtly different
360  * manner to a shared mapping.  A shared mapping has a region map associated
361  * with the underlying file, this region map represents the backing file
362  * pages which have ever had a reservation assigned which this persists even
363  * after the page is instantiated.  A private mapping has a region map
364  * associated with the original mmap which is attached to all VMAs which
365  * reference it, this region map represents those offsets which have consumed
366  * reservation ie. where pages have been instantiated.
367  */
368 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
369 {
370         return (unsigned long)vma->vm_private_data;
371 }
372
373 static void set_vma_private_data(struct vm_area_struct *vma,
374                                                         unsigned long value)
375 {
376         vma->vm_private_data = (void *)value;
377 }
378
379 struct resv_map {
380         struct kref refs;
381         struct list_head regions;
382 };
383
384 static struct resv_map *resv_map_alloc(void)
385 {
386         struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
387         if (!resv_map)
388                 return NULL;
389
390         kref_init(&resv_map->refs);
391         INIT_LIST_HEAD(&resv_map->regions);
392
393         return resv_map;
394 }
395
396 static void resv_map_release(struct kref *ref)
397 {
398         struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
399
400         /* Clear out any active regions before we release the map. */
401         region_truncate(&resv_map->regions, 0);
402         kfree(resv_map);
403 }
404
405 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
406 {
407         VM_BUG_ON(!is_vm_hugetlb_page(vma));
408         if (!(vma->vm_flags & VM_MAYSHARE))
409                 return (struct resv_map *)(get_vma_private_data(vma) &
410                                                         ~HPAGE_RESV_MASK);
411         return NULL;
412 }
413
414 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
415 {
416         VM_BUG_ON(!is_vm_hugetlb_page(vma));
417         VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
418
419         set_vma_private_data(vma, (get_vma_private_data(vma) &
420                                 HPAGE_RESV_MASK) | (unsigned long)map);
421 }
422
423 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
424 {
425         VM_BUG_ON(!is_vm_hugetlb_page(vma));
426         VM_BUG_ON(vma->vm_flags & VM_MAYSHARE);
427
428         set_vma_private_data(vma, get_vma_private_data(vma) | flags);
429 }
430
431 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
432 {
433         VM_BUG_ON(!is_vm_hugetlb_page(vma));
434
435         return (get_vma_private_data(vma) & flag) != 0;
436 }
437
438 /* Decrement the reserved pages in the hugepage pool by one */
439 static void decrement_hugepage_resv_vma(struct hstate *h,
440                         struct vm_area_struct *vma)
441 {
442         if (vma->vm_flags & VM_NORESERVE)
443                 return;
444
445         if (vma->vm_flags & VM_MAYSHARE) {
446                 /* Shared mappings always use reserves */
447                 h->resv_huge_pages--;
448         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
449                 /*
450                  * Only the process that called mmap() has reserves for
451                  * private mappings.
452                  */
453                 h->resv_huge_pages--;
454         }
455 }
456
457 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
458 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
459 {
460         VM_BUG_ON(!is_vm_hugetlb_page(vma));
461         if (!(vma->vm_flags & VM_MAYSHARE))
462                 vma->vm_private_data = (void *)0;
463 }
464
465 /* Returns true if the VMA has associated reserve pages */
466 static int vma_has_reserves(struct vm_area_struct *vma)
467 {
468         if (vma->vm_flags & VM_MAYSHARE)
469                 return 1;
470         if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
471                 return 1;
472         return 0;
473 }
474
475 static void copy_gigantic_page(struct page *dst, struct page *src)
476 {
477         int i;
478         struct hstate *h = page_hstate(src);
479         struct page *dst_base = dst;
480         struct page *src_base = src;
481
482         for (i = 0; i < pages_per_huge_page(h); ) {
483                 cond_resched();
484                 copy_highpage(dst, src);
485
486                 i++;
487                 dst = mem_map_next(dst, dst_base, i);
488                 src = mem_map_next(src, src_base, i);
489         }
490 }
491
492 void copy_huge_page(struct page *dst, struct page *src)
493 {
494         int i;
495         struct hstate *h = page_hstate(src);
496
497         if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES)) {
498                 copy_gigantic_page(dst, src);
499                 return;
500         }
501
502         might_sleep();
503         for (i = 0; i < pages_per_huge_page(h); i++) {
504                 cond_resched();
505                 copy_highpage(dst + i, src + i);
506         }
507 }
508
509 static void enqueue_huge_page(struct hstate *h, struct page *page)
510 {
511         int nid = page_to_nid(page);
512         list_move(&page->lru, &h->hugepage_freelists[nid]);
513         h->free_huge_pages++;
514         h->free_huge_pages_node[nid]++;
515 }
516
517 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
518 {
519         struct page *page;
520
521         list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
522                 if (!is_migrate_isolate_page(page))
523                         break;
524         /*
525          * if 'non-isolated free hugepage' not found on the list,
526          * the allocation fails.
527          */
528         if (&h->hugepage_freelists[nid] == &page->lru)
529                 return NULL;
530         list_move(&page->lru, &h->hugepage_activelist);
531         set_page_refcounted(page);
532         h->free_huge_pages--;
533         h->free_huge_pages_node[nid]--;
534         return page;
535 }
536
537 static struct page *dequeue_huge_page_vma(struct hstate *h,
538                                 struct vm_area_struct *vma,
539                                 unsigned long address, int avoid_reserve)
540 {
541         struct page *page = NULL;
542         struct mempolicy *mpol;
543         nodemask_t *nodemask;
544         struct zonelist *zonelist;
545         struct zone *zone;
546         struct zoneref *z;
547         unsigned int cpuset_mems_cookie;
548
549 retry_cpuset:
550         cpuset_mems_cookie = get_mems_allowed();
551         zonelist = huge_zonelist(vma, address,
552                                         htlb_alloc_mask, &mpol, &nodemask);
553         /*
554          * A child process with MAP_PRIVATE mappings created by their parent
555          * have no page reserves. This check ensures that reservations are
556          * not "stolen". The child may still get SIGKILLed
557          */
558         if (!vma_has_reserves(vma) &&
559                         h->free_huge_pages - h->resv_huge_pages == 0)
560                 goto err;
561
562         /* If reserves cannot be used, ensure enough pages are in the pool */
563         if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
564                 goto err;
565
566         for_each_zone_zonelist_nodemask(zone, z, zonelist,
567                                                 MAX_NR_ZONES - 1, nodemask) {
568                 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask)) {
569                         page = dequeue_huge_page_node(h, zone_to_nid(zone));
570                         if (page) {
571                                 if (!avoid_reserve)
572                                         decrement_hugepage_resv_vma(h, vma);
573                                 break;
574                         }
575                 }
576         }
577
578         mpol_cond_put(mpol);
579         if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
580                 goto retry_cpuset;
581         return page;
582
583 err:
584         mpol_cond_put(mpol);
585         return NULL;
586 }
587
588 static void update_and_free_page(struct hstate *h, struct page *page)
589 {
590         int i;
591
592         VM_BUG_ON(h->order >= MAX_ORDER);
593
594         h->nr_huge_pages--;
595         h->nr_huge_pages_node[page_to_nid(page)]--;
596         for (i = 0; i < pages_per_huge_page(h); i++) {
597                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
598                                 1 << PG_referenced | 1 << PG_dirty |
599                                 1 << PG_active | 1 << PG_reserved |
600                                 1 << PG_private | 1 << PG_writeback);
601         }
602         VM_BUG_ON(hugetlb_cgroup_from_page(page));
603         set_compound_page_dtor(page, NULL);
604         set_page_refcounted(page);
605         arch_release_hugepage(page);
606         __free_pages(page, huge_page_order(h));
607 }
608
609 struct hstate *size_to_hstate(unsigned long size)
610 {
611         struct hstate *h;
612
613         for_each_hstate(h) {
614                 if (huge_page_size(h) == size)
615                         return h;
616         }
617         return NULL;
618 }
619
620 static void free_huge_page(struct page *page)
621 {
622         /*
623          * Can't pass hstate in here because it is called from the
624          * compound page destructor.
625          */
626         struct hstate *h = page_hstate(page);
627         int nid = page_to_nid(page);
628         struct hugepage_subpool *spool =
629                 (struct hugepage_subpool *)page_private(page);
630
631         set_page_private(page, 0);
632         page->mapping = NULL;
633         BUG_ON(page_count(page));
634         BUG_ON(page_mapcount(page));
635
636         spin_lock(&hugetlb_lock);
637         hugetlb_cgroup_uncharge_page(hstate_index(h),
638                                      pages_per_huge_page(h), page);
639         if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
640                 /* remove the page from active list */
641                 list_del(&page->lru);
642                 update_and_free_page(h, page);
643                 h->surplus_huge_pages--;
644                 h->surplus_huge_pages_node[nid]--;
645         } else {
646                 arch_clear_hugepage_flags(page);
647                 enqueue_huge_page(h, page);
648         }
649         spin_unlock(&hugetlb_lock);
650         hugepage_subpool_put_pages(spool, 1);
651 }
652
653 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
654 {
655         INIT_LIST_HEAD(&page->lru);
656         set_compound_page_dtor(page, free_huge_page);
657         spin_lock(&hugetlb_lock);
658         set_hugetlb_cgroup(page, NULL);
659         h->nr_huge_pages++;
660         h->nr_huge_pages_node[nid]++;
661         spin_unlock(&hugetlb_lock);
662         put_page(page); /* free it into the hugepage allocator */
663 }
664
665 static void prep_compound_gigantic_page(struct page *page, unsigned long order)
666 {
667         int i;
668         int nr_pages = 1 << order;
669         struct page *p = page + 1;
670
671         /* we rely on prep_new_huge_page to set the destructor */
672         set_compound_order(page, order);
673         __SetPageHead(page);
674         for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
675                 __SetPageTail(p);
676                 set_page_count(p, 0);
677                 p->first_page = page;
678         }
679 }
680
681 /*
682  * PageHuge() only returns true for hugetlbfs pages, but not for normal or
683  * transparent huge pages.  See the PageTransHuge() documentation for more
684  * details.
685  */
686 int PageHuge(struct page *page)
687 {
688         compound_page_dtor *dtor;
689
690         if (!PageCompound(page))
691                 return 0;
692
693         page = compound_head(page);
694         dtor = get_compound_page_dtor(page);
695
696         return dtor == free_huge_page;
697 }
698 EXPORT_SYMBOL_GPL(PageHuge);
699
700 /*
701  * PageHeadHuge() only returns true for hugetlbfs head page, but not for
702  * normal or transparent huge pages.
703  */
704 int PageHeadHuge(struct page *page_head)
705 {
706         compound_page_dtor *dtor;
707
708         if (!PageHead(page_head))
709                 return 0;
710
711         dtor = get_compound_page_dtor(page_head);
712
713         return dtor == free_huge_page;
714 }
715 EXPORT_SYMBOL_GPL(PageHeadHuge);
716
717 pgoff_t __basepage_index(struct page *page)
718 {
719         struct page *page_head = compound_head(page);
720         pgoff_t index = page_index(page_head);
721         unsigned long compound_idx;
722
723         if (!PageHuge(page_head))
724                 return page_index(page);
725
726         if (compound_order(page_head) >= MAX_ORDER)
727                 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
728         else
729                 compound_idx = page - page_head;
730
731         return (index << compound_order(page_head)) + compound_idx;
732 }
733
734 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
735 {
736         struct page *page;
737
738         if (h->order >= MAX_ORDER)
739                 return NULL;
740
741         page = alloc_pages_exact_node(nid,
742                 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
743                                                 __GFP_REPEAT|__GFP_NOWARN,
744                 huge_page_order(h));
745         if (page) {
746                 if (arch_prepare_hugepage(page)) {
747                         __free_pages(page, huge_page_order(h));
748                         return NULL;
749                 }
750                 prep_new_huge_page(h, page, nid);
751         }
752
753         return page;
754 }
755
756 /*
757  * common helper functions for hstate_next_node_to_{alloc|free}.
758  * We may have allocated or freed a huge page based on a different
759  * nodes_allowed previously, so h->next_node_to_{alloc|free} might
760  * be outside of *nodes_allowed.  Ensure that we use an allowed
761  * node for alloc or free.
762  */
763 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
764 {
765         nid = next_node(nid, *nodes_allowed);
766         if (nid == MAX_NUMNODES)
767                 nid = first_node(*nodes_allowed);
768         VM_BUG_ON(nid >= MAX_NUMNODES);
769
770         return nid;
771 }
772
773 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
774 {
775         if (!node_isset(nid, *nodes_allowed))
776                 nid = next_node_allowed(nid, nodes_allowed);
777         return nid;
778 }
779
780 /*
781  * returns the previously saved node ["this node"] from which to
782  * allocate a persistent huge page for the pool and advance the
783  * next node from which to allocate, handling wrap at end of node
784  * mask.
785  */
786 static int hstate_next_node_to_alloc(struct hstate *h,
787                                         nodemask_t *nodes_allowed)
788 {
789         int nid;
790
791         VM_BUG_ON(!nodes_allowed);
792
793         nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
794         h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
795
796         return nid;
797 }
798
799 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
800 {
801         struct page *page;
802         int start_nid;
803         int next_nid;
804         int ret = 0;
805
806         start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
807         next_nid = start_nid;
808
809         do {
810                 page = alloc_fresh_huge_page_node(h, next_nid);
811                 if (page) {
812                         ret = 1;
813                         break;
814                 }
815                 next_nid = hstate_next_node_to_alloc(h, nodes_allowed);
816         } while (next_nid != start_nid);
817
818         if (ret)
819                 count_vm_event(HTLB_BUDDY_PGALLOC);
820         else
821                 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
822
823         return ret;
824 }
825
826 /*
827  * helper for free_pool_huge_page() - return the previously saved
828  * node ["this node"] from which to free a huge page.  Advance the
829  * next node id whether or not we find a free huge page to free so
830  * that the next attempt to free addresses the next node.
831  */
832 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
833 {
834         int nid;
835
836         VM_BUG_ON(!nodes_allowed);
837
838         nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
839         h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
840
841         return nid;
842 }
843
844 /*
845  * Free huge page from pool from next node to free.
846  * Attempt to keep persistent huge pages more or less
847  * balanced over allowed nodes.
848  * Called with hugetlb_lock locked.
849  */
850 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
851                                                          bool acct_surplus)
852 {
853         int start_nid;
854         int next_nid;
855         int ret = 0;
856
857         start_nid = hstate_next_node_to_free(h, nodes_allowed);
858         next_nid = start_nid;
859
860         do {
861                 /*
862                  * If we're returning unused surplus pages, only examine
863                  * nodes with surplus pages.
864                  */
865                 if ((!acct_surplus || h->surplus_huge_pages_node[next_nid]) &&
866                     !list_empty(&h->hugepage_freelists[next_nid])) {
867                         struct page *page =
868                                 list_entry(h->hugepage_freelists[next_nid].next,
869                                           struct page, lru);
870                         list_del(&page->lru);
871                         h->free_huge_pages--;
872                         h->free_huge_pages_node[next_nid]--;
873                         if (acct_surplus) {
874                                 h->surplus_huge_pages--;
875                                 h->surplus_huge_pages_node[next_nid]--;
876                         }
877                         update_and_free_page(h, page);
878                         ret = 1;
879                         break;
880                 }
881                 next_nid = hstate_next_node_to_free(h, nodes_allowed);
882         } while (next_nid != start_nid);
883
884         return ret;
885 }
886
887 static struct page *alloc_buddy_huge_page(struct hstate *h, int nid)
888 {
889         struct page *page;
890         unsigned int r_nid;
891
892         if (h->order >= MAX_ORDER)
893                 return NULL;
894
895         /*
896          * Assume we will successfully allocate the surplus page to
897          * prevent racing processes from causing the surplus to exceed
898          * overcommit
899          *
900          * This however introduces a different race, where a process B
901          * tries to grow the static hugepage pool while alloc_pages() is
902          * called by process A. B will only examine the per-node
903          * counters in determining if surplus huge pages can be
904          * converted to normal huge pages in adjust_pool_surplus(). A
905          * won't be able to increment the per-node counter, until the
906          * lock is dropped by B, but B doesn't drop hugetlb_lock until
907          * no more huge pages can be converted from surplus to normal
908          * state (and doesn't try to convert again). Thus, we have a
909          * case where a surplus huge page exists, the pool is grown, and
910          * the surplus huge page still exists after, even though it
911          * should just have been converted to a normal huge page. This
912          * does not leak memory, though, as the hugepage will be freed
913          * once it is out of use. It also does not allow the counters to
914          * go out of whack in adjust_pool_surplus() as we don't modify
915          * the node values until we've gotten the hugepage and only the
916          * per-node value is checked there.
917          */
918         spin_lock(&hugetlb_lock);
919         if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
920                 spin_unlock(&hugetlb_lock);
921                 return NULL;
922         } else {
923                 h->nr_huge_pages++;
924                 h->surplus_huge_pages++;
925         }
926         spin_unlock(&hugetlb_lock);
927
928         if (nid == NUMA_NO_NODE)
929                 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
930                                    __GFP_REPEAT|__GFP_NOWARN,
931                                    huge_page_order(h));
932         else
933                 page = alloc_pages_exact_node(nid,
934                         htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
935                         __GFP_REPEAT|__GFP_NOWARN, huge_page_order(h));
936
937         if (page && arch_prepare_hugepage(page)) {
938                 __free_pages(page, huge_page_order(h));
939                 page = NULL;
940         }
941
942         spin_lock(&hugetlb_lock);
943         if (page) {
944                 INIT_LIST_HEAD(&page->lru);
945                 r_nid = page_to_nid(page);
946                 set_compound_page_dtor(page, free_huge_page);
947                 set_hugetlb_cgroup(page, NULL);
948                 /*
949                  * We incremented the global counters already
950                  */
951                 h->nr_huge_pages_node[r_nid]++;
952                 h->surplus_huge_pages_node[r_nid]++;
953                 __count_vm_event(HTLB_BUDDY_PGALLOC);
954         } else {
955                 h->nr_huge_pages--;
956                 h->surplus_huge_pages--;
957                 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
958         }
959         spin_unlock(&hugetlb_lock);
960
961         return page;
962 }
963
964 /*
965  * This allocation function is useful in the context where vma is irrelevant.
966  * E.g. soft-offlining uses this function because it only cares physical
967  * address of error page.
968  */
969 struct page *alloc_huge_page_node(struct hstate *h, int nid)
970 {
971         struct page *page;
972
973         spin_lock(&hugetlb_lock);
974         page = dequeue_huge_page_node(h, nid);
975         spin_unlock(&hugetlb_lock);
976
977         if (!page)
978                 page = alloc_buddy_huge_page(h, nid);
979
980         return page;
981 }
982
983 /*
984  * Increase the hugetlb pool such that it can accommodate a reservation
985  * of size 'delta'.
986  */
987 static int gather_surplus_pages(struct hstate *h, int delta)
988 {
989         struct list_head surplus_list;
990         struct page *page, *tmp;
991         int ret, i;
992         int needed, allocated;
993         bool alloc_ok = true;
994
995         needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
996         if (needed <= 0) {
997                 h->resv_huge_pages += delta;
998                 return 0;
999         }
1000
1001         allocated = 0;
1002         INIT_LIST_HEAD(&surplus_list);
1003
1004         ret = -ENOMEM;
1005 retry:
1006         spin_unlock(&hugetlb_lock);
1007         for (i = 0; i < needed; i++) {
1008                 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1009                 if (!page) {
1010                         alloc_ok = false;
1011                         break;
1012                 }
1013                 list_add(&page->lru, &surplus_list);
1014         }
1015         allocated += i;
1016
1017         /*
1018          * After retaking hugetlb_lock, we need to recalculate 'needed'
1019          * because either resv_huge_pages or free_huge_pages may have changed.
1020          */
1021         spin_lock(&hugetlb_lock);
1022         needed = (h->resv_huge_pages + delta) -
1023                         (h->free_huge_pages + allocated);
1024         if (needed > 0) {
1025                 if (alloc_ok)
1026                         goto retry;
1027                 /*
1028                  * We were not able to allocate enough pages to
1029                  * satisfy the entire reservation so we free what
1030                  * we've allocated so far.
1031                  */
1032                 goto free;
1033         }
1034         /*
1035          * The surplus_list now contains _at_least_ the number of extra pages
1036          * needed to accommodate the reservation.  Add the appropriate number
1037          * of pages to the hugetlb pool and free the extras back to the buddy
1038          * allocator.  Commit the entire reservation here to prevent another
1039          * process from stealing the pages as they are added to the pool but
1040          * before they are reserved.
1041          */
1042         needed += allocated;
1043         h->resv_huge_pages += delta;
1044         ret = 0;
1045
1046         /* Free the needed pages to the hugetlb pool */
1047         list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1048                 if ((--needed) < 0)
1049                         break;
1050                 /*
1051                  * This page is now managed by the hugetlb allocator and has
1052                  * no users -- drop the buddy allocator's reference.
1053                  */
1054                 put_page_testzero(page);
1055                 VM_BUG_ON(page_count(page));
1056                 enqueue_huge_page(h, page);
1057         }
1058 free:
1059         spin_unlock(&hugetlb_lock);
1060
1061         /* Free unnecessary surplus pages to the buddy allocator */
1062         if (!list_empty(&surplus_list)) {
1063                 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1064                         put_page(page);
1065                 }
1066         }
1067         spin_lock(&hugetlb_lock);
1068
1069         return ret;
1070 }
1071
1072 /*
1073  * When releasing a hugetlb pool reservation, any surplus pages that were
1074  * allocated to satisfy the reservation must be explicitly freed if they were
1075  * never used.
1076  * Called with hugetlb_lock held.
1077  */
1078 static void return_unused_surplus_pages(struct hstate *h,
1079                                         unsigned long unused_resv_pages)
1080 {
1081         unsigned long nr_pages;
1082
1083         /* Uncommit the reservation */
1084         h->resv_huge_pages -= unused_resv_pages;
1085
1086         /* Cannot return gigantic pages currently */
1087         if (h->order >= MAX_ORDER)
1088                 return;
1089
1090         nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1091
1092         /*
1093          * We want to release as many surplus pages as possible, spread
1094          * evenly across all nodes with memory. Iterate across these nodes
1095          * until we can no longer free unreserved surplus pages. This occurs
1096          * when the nodes with surplus pages have no free pages.
1097          * free_pool_huge_page() will balance the the freed pages across the
1098          * on-line nodes with memory and will handle the hstate accounting.
1099          */
1100         while (nr_pages--) {
1101                 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1102                         break;
1103                 cond_resched_lock(&hugetlb_lock);
1104         }
1105 }
1106
1107 /*
1108  * Determine if the huge page at addr within the vma has an associated
1109  * reservation.  Where it does not we will need to logically increase
1110  * reservation and actually increase subpool usage before an allocation
1111  * can occur.  Where any new reservation would be required the
1112  * reservation change is prepared, but not committed.  Once the page
1113  * has been allocated from the subpool and instantiated the change should
1114  * be committed via vma_commit_reservation.  No action is required on
1115  * failure.
1116  */
1117 static long vma_needs_reservation(struct hstate *h,
1118                         struct vm_area_struct *vma, unsigned long addr)
1119 {
1120         struct address_space *mapping = vma->vm_file->f_mapping;
1121         struct inode *inode = mapping->host;
1122
1123         if (vma->vm_flags & VM_MAYSHARE) {
1124                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1125                 return region_chg(&inode->i_mapping->private_list,
1126                                                         idx, idx + 1);
1127
1128         } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1129                 return 1;
1130
1131         } else  {
1132                 long err;
1133                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1134                 struct resv_map *reservations = vma_resv_map(vma);
1135
1136                 err = region_chg(&reservations->regions, idx, idx + 1);
1137                 if (err < 0)
1138                         return err;
1139                 return 0;
1140         }
1141 }
1142 static void vma_commit_reservation(struct hstate *h,
1143                         struct vm_area_struct *vma, unsigned long addr)
1144 {
1145         struct address_space *mapping = vma->vm_file->f_mapping;
1146         struct inode *inode = mapping->host;
1147
1148         if (vma->vm_flags & VM_MAYSHARE) {
1149                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1150                 region_add(&inode->i_mapping->private_list, idx, idx + 1);
1151
1152         } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1153                 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
1154                 struct resv_map *reservations = vma_resv_map(vma);
1155
1156                 /* Mark this page used in the map. */
1157                 region_add(&reservations->regions, idx, idx + 1);
1158         }
1159 }
1160
1161 static struct page *alloc_huge_page(struct vm_area_struct *vma,
1162                                     unsigned long addr, int avoid_reserve)
1163 {
1164         struct hugepage_subpool *spool = subpool_vma(vma);
1165         struct hstate *h = hstate_vma(vma);
1166         struct page *page;
1167         long chg;
1168         int ret, idx;
1169         struct hugetlb_cgroup *h_cg;
1170
1171         idx = hstate_index(h);
1172         /*
1173          * Processes that did not create the mapping will have no
1174          * reserves and will not have accounted against subpool
1175          * limit. Check that the subpool limit can be made before
1176          * satisfying the allocation MAP_NORESERVE mappings may also
1177          * need pages and subpool limit allocated allocated if no reserve
1178          * mapping overlaps.
1179          */
1180         chg = vma_needs_reservation(h, vma, addr);
1181         if (chg < 0)
1182                 return ERR_PTR(-ENOMEM);
1183         if (chg)
1184                 if (hugepage_subpool_get_pages(spool, chg))
1185                         return ERR_PTR(-ENOSPC);
1186
1187         ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1188         if (ret) {
1189                 hugepage_subpool_put_pages(spool, chg);
1190                 return ERR_PTR(-ENOSPC);
1191         }
1192         spin_lock(&hugetlb_lock);
1193         page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
1194         if (page) {
1195                 /* update page cgroup details */
1196                 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h),
1197                                              h_cg, page);
1198                 spin_unlock(&hugetlb_lock);
1199         } else {
1200                 spin_unlock(&hugetlb_lock);
1201                 page = alloc_buddy_huge_page(h, NUMA_NO_NODE);
1202                 if (!page) {
1203                         hugetlb_cgroup_uncharge_cgroup(idx,
1204                                                        pages_per_huge_page(h),
1205                                                        h_cg);
1206                         hugepage_subpool_put_pages(spool, chg);
1207                         return ERR_PTR(-ENOSPC);
1208                 }
1209                 spin_lock(&hugetlb_lock);
1210                 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h),
1211                                              h_cg, page);
1212                 list_move(&page->lru, &h->hugepage_activelist);
1213                 spin_unlock(&hugetlb_lock);
1214         }
1215
1216         set_page_private(page, (unsigned long)spool);
1217
1218         vma_commit_reservation(h, vma, addr);
1219         return page;
1220 }
1221
1222 int __weak alloc_bootmem_huge_page(struct hstate *h)
1223 {
1224         struct huge_bootmem_page *m;
1225         int nr_nodes = nodes_weight(node_states[N_MEMORY]);
1226
1227         while (nr_nodes) {
1228                 void *addr;
1229
1230                 addr = __alloc_bootmem_node_nopanic(
1231                                 NODE_DATA(hstate_next_node_to_alloc(h,
1232                                                 &node_states[N_MEMORY])),
1233                                 huge_page_size(h), huge_page_size(h), 0);
1234
1235                 if (addr) {
1236                         /*
1237                          * Use the beginning of the huge page to store the
1238                          * huge_bootmem_page struct (until gather_bootmem
1239                          * puts them into the mem_map).
1240                          */
1241                         m = addr;
1242                         goto found;
1243                 }
1244                 nr_nodes--;
1245         }
1246         return 0;
1247
1248 found:
1249         BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1250         /* Put them into a private list first because mem_map is not up yet */
1251         list_add(&m->list, &huge_boot_pages);
1252         m->hstate = h;
1253         return 1;
1254 }
1255
1256 static void prep_compound_huge_page(struct page *page, int order)
1257 {
1258         if (unlikely(order > (MAX_ORDER - 1)))
1259                 prep_compound_gigantic_page(page, order);
1260         else
1261                 prep_compound_page(page, order);
1262 }
1263
1264 /* Put bootmem huge pages into the standard lists after mem_map is up */
1265 static void __init gather_bootmem_prealloc(void)
1266 {
1267         struct huge_bootmem_page *m;
1268
1269         list_for_each_entry(m, &huge_boot_pages, list) {
1270                 struct hstate *h = m->hstate;
1271                 struct page *page;
1272
1273 #ifdef CONFIG_HIGHMEM
1274                 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1275                 free_bootmem_late((unsigned long)m,
1276                                   sizeof(struct huge_bootmem_page));
1277 #else
1278                 page = virt_to_page(m);
1279 #endif
1280                 __ClearPageReserved(page);
1281                 WARN_ON(page_count(page) != 1);
1282                 prep_compound_huge_page(page, h->order);
1283                 prep_new_huge_page(h, page, page_to_nid(page));
1284                 /*
1285                  * If we had gigantic hugepages allocated at boot time, we need
1286                  * to restore the 'stolen' pages to totalram_pages in order to
1287                  * fix confusing memory reports from free(1) and another
1288                  * side-effects, like CommitLimit going negative.
1289                  */
1290                 if (h->order > (MAX_ORDER - 1))
1291                         totalram_pages += 1 << h->order;
1292         }
1293 }
1294
1295 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1296 {
1297         unsigned long i;
1298
1299         for (i = 0; i < h->max_huge_pages; ++i) {
1300                 if (h->order >= MAX_ORDER) {
1301                         if (!alloc_bootmem_huge_page(h))
1302                                 break;
1303                 } else if (!alloc_fresh_huge_page(h,
1304                                          &node_states[N_MEMORY]))
1305                         break;
1306         }
1307         h->max_huge_pages = i;
1308 }
1309
1310 static void __init hugetlb_init_hstates(void)
1311 {
1312         struct hstate *h;
1313
1314         for_each_hstate(h) {
1315                 /* oversize hugepages were init'ed in early boot */
1316                 if (h->order < MAX_ORDER)
1317                         hugetlb_hstate_alloc_pages(h);
1318         }
1319 }
1320
1321 static char * __init memfmt(char *buf, unsigned long n)
1322 {
1323         if (n >= (1UL << 30))
1324                 sprintf(buf, "%lu GB", n >> 30);
1325         else if (n >= (1UL << 20))
1326                 sprintf(buf, "%lu MB", n >> 20);
1327         else
1328                 sprintf(buf, "%lu KB", n >> 10);
1329         return buf;
1330 }
1331
1332 static void __init report_hugepages(void)
1333 {
1334         struct hstate *h;
1335
1336         for_each_hstate(h) {
1337                 char buf[32];
1338                 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
1339                         memfmt(buf, huge_page_size(h)),
1340                         h->free_huge_pages);
1341         }
1342 }
1343
1344 #ifdef CONFIG_HIGHMEM
1345 static void try_to_free_low(struct hstate *h, unsigned long count,
1346                                                 nodemask_t *nodes_allowed)
1347 {
1348         int i;
1349
1350         if (h->order >= MAX_ORDER)
1351                 return;
1352
1353         for_each_node_mask(i, *nodes_allowed) {
1354                 struct page *page, *next;
1355                 struct list_head *freel = &h->hugepage_freelists[i];
1356                 list_for_each_entry_safe(page, next, freel, lru) {
1357                         if (count >= h->nr_huge_pages)
1358                                 return;
1359                         if (PageHighMem(page))
1360                                 continue;
1361                         list_del(&page->lru);
1362                         update_and_free_page(h, page);
1363                         h->free_huge_pages--;
1364                         h->free_huge_pages_node[page_to_nid(page)]--;
1365                 }
1366         }
1367 }
1368 #else
1369 static inline void try_to_free_low(struct hstate *h, unsigned long count,
1370                                                 nodemask_t *nodes_allowed)
1371 {
1372 }
1373 #endif
1374
1375 /*
1376  * Increment or decrement surplus_huge_pages.  Keep node-specific counters
1377  * balanced by operating on them in a round-robin fashion.
1378  * Returns 1 if an adjustment was made.
1379  */
1380 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
1381                                 int delta)
1382 {
1383         int start_nid, next_nid;
1384         int ret = 0;
1385
1386         VM_BUG_ON(delta != -1 && delta != 1);
1387
1388         if (delta < 0)
1389                 start_nid = hstate_next_node_to_alloc(h, nodes_allowed);
1390         else
1391                 start_nid = hstate_next_node_to_free(h, nodes_allowed);
1392         next_nid = start_nid;
1393
1394         do {
1395                 int nid = next_nid;
1396                 if (delta < 0)  {
1397                         /*
1398                          * To shrink on this node, there must be a surplus page
1399                          */
1400                         if (!h->surplus_huge_pages_node[nid]) {
1401                                 next_nid = hstate_next_node_to_alloc(h,
1402                                                                 nodes_allowed);
1403                                 continue;
1404                         }
1405                 }
1406                 if (delta > 0) {
1407                         /*
1408                          * Surplus cannot exceed the total number of pages
1409                          */
1410                         if (h->surplus_huge_pages_node[nid] >=
1411                                                 h->nr_huge_pages_node[nid]) {
1412                                 next_nid = hstate_next_node_to_free(h,
1413                                                                 nodes_allowed);
1414                                 continue;
1415                         }
1416                 }
1417
1418                 h->surplus_huge_pages += delta;
1419                 h->surplus_huge_pages_node[nid] += delta;
1420                 ret = 1;
1421                 break;
1422         } while (next_nid != start_nid);
1423
1424         return ret;
1425 }
1426
1427 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1428 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
1429                                                 nodemask_t *nodes_allowed)
1430 {
1431         unsigned long min_count, ret;
1432
1433         if (h->order >= MAX_ORDER)
1434                 return h->max_huge_pages;
1435
1436         /*
1437          * Increase the pool size
1438          * First take pages out of surplus state.  Then make up the
1439          * remaining difference by allocating fresh huge pages.
1440          *
1441          * We might race with alloc_buddy_huge_page() here and be unable
1442          * to convert a surplus huge page to a normal huge page. That is
1443          * not critical, though, it just means the overall size of the
1444          * pool might be one hugepage larger than it needs to be, but
1445          * within all the constraints specified by the sysctls.
1446          */
1447         spin_lock(&hugetlb_lock);
1448         while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1449                 if (!adjust_pool_surplus(h, nodes_allowed, -1))
1450                         break;
1451         }
1452
1453         while (count > persistent_huge_pages(h)) {
1454                 /*
1455                  * If this allocation races such that we no longer need the
1456                  * page, free_huge_page will handle it by freeing the page
1457                  * and reducing the surplus.
1458                  */
1459                 spin_unlock(&hugetlb_lock);
1460                 ret = alloc_fresh_huge_page(h, nodes_allowed);
1461                 spin_lock(&hugetlb_lock);
1462                 if (!ret)
1463                         goto out;
1464
1465                 /* Bail for signals. Probably ctrl-c from user */
1466                 if (signal_pending(current))
1467                         goto out;
1468         }
1469
1470         /*
1471          * Decrease the pool size
1472          * First return free pages to the buddy allocator (being careful
1473          * to keep enough around to satisfy reservations).  Then place
1474          * pages into surplus state as needed so the pool will shrink
1475          * to the desired size as pages become free.
1476          *
1477          * By placing pages into the surplus state independent of the
1478          * overcommit value, we are allowing the surplus pool size to
1479          * exceed overcommit. There are few sane options here. Since
1480          * alloc_buddy_huge_page() is checking the global counter,
1481          * though, we'll note that we're not allowed to exceed surplus
1482          * and won't grow the pool anywhere else. Not until one of the
1483          * sysctls are changed, or the surplus pages go out of use.
1484          */
1485         min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1486         min_count = max(count, min_count);
1487         try_to_free_low(h, min_count, nodes_allowed);
1488         while (min_count < persistent_huge_pages(h)) {
1489                 if (!free_pool_huge_page(h, nodes_allowed, 0))
1490                         break;
1491                 cond_resched_lock(&hugetlb_lock);
1492         }
1493         while (count < persistent_huge_pages(h)) {
1494                 if (!adjust_pool_surplus(h, nodes_allowed, 1))
1495                         break;
1496         }
1497 out:
1498         ret = persistent_huge_pages(h);
1499         spin_unlock(&hugetlb_lock);
1500         return ret;
1501 }
1502
1503 #define HSTATE_ATTR_RO(_name) \
1504         static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1505
1506 #define HSTATE_ATTR(_name) \
1507         static struct kobj_attribute _name##_attr = \
1508                 __ATTR(_name, 0644, _name##_show, _name##_store)
1509
1510 static struct kobject *hugepages_kobj;
1511 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1512
1513 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
1514
1515 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
1516 {
1517         int i;
1518
1519         for (i = 0; i < HUGE_MAX_HSTATE; i++)
1520                 if (hstate_kobjs[i] == kobj) {
1521                         if (nidp)
1522                                 *nidp = NUMA_NO_NODE;
1523                         return &hstates[i];
1524                 }
1525
1526         return kobj_to_node_hstate(kobj, nidp);
1527 }
1528
1529 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
1530                                         struct kobj_attribute *attr, char *buf)
1531 {
1532         struct hstate *h;
1533         unsigned long nr_huge_pages;
1534         int nid;
1535
1536         h = kobj_to_hstate(kobj, &nid);
1537         if (nid == NUMA_NO_NODE)
1538                 nr_huge_pages = h->nr_huge_pages;
1539         else
1540                 nr_huge_pages = h->nr_huge_pages_node[nid];
1541
1542         return sprintf(buf, "%lu\n", nr_huge_pages);
1543 }
1544
1545 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
1546                         struct kobject *kobj, struct kobj_attribute *attr,
1547                         const char *buf, size_t len)
1548 {
1549         int err;
1550         int nid;
1551         unsigned long count;
1552         struct hstate *h;
1553         NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
1554
1555         err = kstrtoul(buf, 10, &count);
1556         if (err)
1557                 goto out;
1558
1559         h = kobj_to_hstate(kobj, &nid);
1560         if (h->order >= MAX_ORDER) {
1561                 err = -EINVAL;
1562                 goto out;
1563         }
1564
1565         if (nid == NUMA_NO_NODE) {
1566                 /*
1567                  * global hstate attribute
1568                  */
1569                 if (!(obey_mempolicy &&
1570                                 init_nodemask_of_mempolicy(nodes_allowed))) {
1571                         NODEMASK_FREE(nodes_allowed);
1572                         nodes_allowed = &node_states[N_MEMORY];
1573                 }
1574         } else if (nodes_allowed) {
1575                 /*
1576                  * per node hstate attribute: adjust count to global,
1577                  * but restrict alloc/free to the specified node.
1578                  */
1579                 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
1580                 init_nodemask_of_node(nodes_allowed, nid);
1581         } else
1582                 nodes_allowed = &node_states[N_MEMORY];
1583
1584         h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
1585
1586         if (nodes_allowed != &node_states[N_MEMORY])
1587                 NODEMASK_FREE(nodes_allowed);
1588
1589         return len;
1590 out:
1591         NODEMASK_FREE(nodes_allowed);
1592         return err;
1593 }
1594
1595 static ssize_t nr_hugepages_show(struct kobject *kobj,
1596                                        struct kobj_attribute *attr, char *buf)
1597 {
1598         return nr_hugepages_show_common(kobj, attr, buf);
1599 }
1600
1601 static ssize_t nr_hugepages_store(struct kobject *kobj,
1602                struct kobj_attribute *attr, const char *buf, size_t len)
1603 {
1604         return nr_hugepages_store_common(false, kobj, attr, buf, len);
1605 }
1606 HSTATE_ATTR(nr_hugepages);
1607
1608 #ifdef CONFIG_NUMA
1609
1610 /*
1611  * hstate attribute for optionally mempolicy-based constraint on persistent
1612  * huge page alloc/free.
1613  */
1614 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
1615                                        struct kobj_attribute *attr, char *buf)
1616 {
1617         return nr_hugepages_show_common(kobj, attr, buf);
1618 }
1619
1620 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
1621                struct kobj_attribute *attr, const char *buf, size_t len)
1622 {
1623         return nr_hugepages_store_common(true, kobj, attr, buf, len);
1624 }
1625 HSTATE_ATTR(nr_hugepages_mempolicy);
1626 #endif
1627
1628
1629 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1630                                         struct kobj_attribute *attr, char *buf)
1631 {
1632         struct hstate *h = kobj_to_hstate(kobj, NULL);
1633         return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1634 }
1635
1636 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1637                 struct kobj_attribute *attr, const char *buf, size_t count)
1638 {
1639         int err;
1640         unsigned long input;
1641         struct hstate *h = kobj_to_hstate(kobj, NULL);
1642
1643         if (h->order >= MAX_ORDER)
1644                 return -EINVAL;
1645
1646         err = kstrtoul(buf, 10, &input);
1647         if (err)
1648                 return err;
1649
1650         spin_lock(&hugetlb_lock);
1651         h->nr_overcommit_huge_pages = input;
1652         spin_unlock(&hugetlb_lock);
1653
1654         return count;
1655 }
1656 HSTATE_ATTR(nr_overcommit_hugepages);
1657
1658 static ssize_t free_hugepages_show(struct kobject *kobj,
1659                                         struct kobj_attribute *attr, char *buf)
1660 {
1661         struct hstate *h;
1662         unsigned long free_huge_pages;
1663         int nid;
1664
1665         h = kobj_to_hstate(kobj, &nid);
1666         if (nid == NUMA_NO_NODE)
1667                 free_huge_pages = h->free_huge_pages;
1668         else
1669                 free_huge_pages = h->free_huge_pages_node[nid];
1670
1671         return sprintf(buf, "%lu\n", free_huge_pages);
1672 }
1673 HSTATE_ATTR_RO(free_hugepages);
1674
1675 static ssize_t resv_hugepages_show(struct kobject *kobj,
1676                                         struct kobj_attribute *attr, char *buf)
1677 {
1678         struct hstate *h = kobj_to_hstate(kobj, NULL);
1679         return sprintf(buf, "%lu\n", h->resv_huge_pages);
1680 }
1681 HSTATE_ATTR_RO(resv_hugepages);
1682
1683 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1684                                         struct kobj_attribute *attr, char *buf)
1685 {
1686         struct hstate *h;
1687         unsigned long surplus_huge_pages;
1688         int nid;
1689
1690         h = kobj_to_hstate(kobj, &nid);
1691         if (nid == NUMA_NO_NODE)
1692                 surplus_huge_pages = h->surplus_huge_pages;
1693         else
1694                 surplus_huge_pages = h->surplus_huge_pages_node[nid];
1695
1696         return sprintf(buf, "%lu\n", surplus_huge_pages);
1697 }
1698 HSTATE_ATTR_RO(surplus_hugepages);
1699
1700 static struct attribute *hstate_attrs[] = {
1701         &nr_hugepages_attr.attr,
1702         &nr_overcommit_hugepages_attr.attr,
1703         &free_hugepages_attr.attr,
1704         &resv_hugepages_attr.attr,
1705         &surplus_hugepages_attr.attr,
1706 #ifdef CONFIG_NUMA
1707         &nr_hugepages_mempolicy_attr.attr,
1708 #endif
1709         NULL,
1710 };
1711
1712 static struct attribute_group hstate_attr_group = {
1713         .attrs = hstate_attrs,
1714 };
1715
1716 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
1717                                     struct kobject **hstate_kobjs,
1718                                     struct attribute_group *hstate_attr_group)
1719 {
1720         int retval;
1721         int hi = hstate_index(h);
1722
1723         hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
1724         if (!hstate_kobjs[hi])
1725                 return -ENOMEM;
1726
1727         retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
1728         if (retval)
1729                 kobject_put(hstate_kobjs[hi]);
1730
1731         return retval;
1732 }
1733
1734 static void __init hugetlb_sysfs_init(void)
1735 {
1736         struct hstate *h;
1737         int err;
1738
1739         hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1740         if (!hugepages_kobj)
1741                 return;
1742
1743         for_each_hstate(h) {
1744                 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
1745                                          hstate_kobjs, &hstate_attr_group);
1746                 if (err)
1747                         pr_err("Hugetlb: Unable to add hstate %s", h->name);
1748         }
1749 }
1750
1751 #ifdef CONFIG_NUMA
1752
1753 /*
1754  * node_hstate/s - associate per node hstate attributes, via their kobjects,
1755  * with node devices in node_devices[] using a parallel array.  The array
1756  * index of a node device or _hstate == node id.
1757  * This is here to avoid any static dependency of the node device driver, in
1758  * the base kernel, on the hugetlb module.
1759  */
1760 struct node_hstate {
1761         struct kobject          *hugepages_kobj;
1762         struct kobject          *hstate_kobjs[HUGE_MAX_HSTATE];
1763 };
1764 struct node_hstate node_hstates[MAX_NUMNODES];
1765
1766 /*
1767  * A subset of global hstate attributes for node devices
1768  */
1769 static struct attribute *per_node_hstate_attrs[] = {
1770         &nr_hugepages_attr.attr,
1771         &free_hugepages_attr.attr,
1772         &surplus_hugepages_attr.attr,
1773         NULL,
1774 };
1775
1776 static struct attribute_group per_node_hstate_attr_group = {
1777         .attrs = per_node_hstate_attrs,
1778 };
1779
1780 /*
1781  * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
1782  * Returns node id via non-NULL nidp.
1783  */
1784 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1785 {
1786         int nid;
1787
1788         for (nid = 0; nid < nr_node_ids; nid++) {
1789                 struct node_hstate *nhs = &node_hstates[nid];
1790                 int i;
1791                 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1792                         if (nhs->hstate_kobjs[i] == kobj) {
1793                                 if (nidp)
1794                                         *nidp = nid;
1795                                 return &hstates[i];
1796                         }
1797         }
1798
1799         BUG();
1800         return NULL;
1801 }
1802
1803 /*
1804  * Unregister hstate attributes from a single node device.
1805  * No-op if no hstate attributes attached.
1806  */
1807 static void hugetlb_unregister_node(struct node *node)
1808 {
1809         struct hstate *h;
1810         struct node_hstate *nhs = &node_hstates[node->dev.id];
1811
1812         if (!nhs->hugepages_kobj)
1813                 return;         /* no hstate attributes */
1814
1815         for_each_hstate(h) {
1816                 int idx = hstate_index(h);
1817                 if (nhs->hstate_kobjs[idx]) {
1818                         kobject_put(nhs->hstate_kobjs[idx]);
1819                         nhs->hstate_kobjs[idx] = NULL;
1820                 }
1821         }
1822
1823         kobject_put(nhs->hugepages_kobj);
1824         nhs->hugepages_kobj = NULL;
1825 }
1826
1827 /*
1828  * hugetlb module exit:  unregister hstate attributes from node devices
1829  * that have them.
1830  */
1831 static void hugetlb_unregister_all_nodes(void)
1832 {
1833         int nid;
1834
1835         /*
1836          * disable node device registrations.
1837          */
1838         register_hugetlbfs_with_node(NULL, NULL);
1839
1840         /*
1841          * remove hstate attributes from any nodes that have them.
1842          */
1843         for (nid = 0; nid < nr_node_ids; nid++)
1844                 hugetlb_unregister_node(node_devices[nid]);
1845 }
1846
1847 /*
1848  * Register hstate attributes for a single node device.
1849  * No-op if attributes already registered.
1850  */
1851 static void hugetlb_register_node(struct node *node)
1852 {
1853         struct hstate *h;
1854         struct node_hstate *nhs = &node_hstates[node->dev.id];
1855         int err;
1856
1857         if (nhs->hugepages_kobj)
1858                 return;         /* already allocated */
1859
1860         nhs->hugepages_kobj = kobject_create_and_add("hugepages",
1861                                                         &node->dev.kobj);
1862         if (!nhs->hugepages_kobj)
1863                 return;
1864
1865         for_each_hstate(h) {
1866                 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
1867                                                 nhs->hstate_kobjs,
1868                                                 &per_node_hstate_attr_group);
1869                 if (err) {
1870                         pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
1871                                 h->name, node->dev.id);
1872                         hugetlb_unregister_node(node);
1873                         break;
1874                 }
1875         }
1876 }
1877
1878 /*
1879  * hugetlb init time:  register hstate attributes for all registered node
1880  * devices of nodes that have memory.  All on-line nodes should have
1881  * registered their associated device by this time.
1882  */
1883 static void hugetlb_register_all_nodes(void)
1884 {
1885         int nid;
1886
1887         for_each_node_state(nid, N_MEMORY) {
1888                 struct node *node = node_devices[nid];
1889                 if (node->dev.id == nid)
1890                         hugetlb_register_node(node);
1891         }
1892
1893         /*
1894          * Let the node device driver know we're here so it can
1895          * [un]register hstate attributes on node hotplug.
1896          */
1897         register_hugetlbfs_with_node(hugetlb_register_node,
1898                                      hugetlb_unregister_node);
1899 }
1900 #else   /* !CONFIG_NUMA */
1901
1902 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
1903 {
1904         BUG();
1905         if (nidp)
1906                 *nidp = -1;
1907         return NULL;
1908 }
1909
1910 static void hugetlb_unregister_all_nodes(void) { }
1911
1912 static void hugetlb_register_all_nodes(void) { }
1913
1914 #endif
1915
1916 static void __exit hugetlb_exit(void)
1917 {
1918         struct hstate *h;
1919
1920         hugetlb_unregister_all_nodes();
1921
1922         for_each_hstate(h) {
1923                 kobject_put(hstate_kobjs[hstate_index(h)]);
1924         }
1925
1926         kobject_put(hugepages_kobj);
1927 }
1928 module_exit(hugetlb_exit);
1929
1930 static int __init hugetlb_init(void)
1931 {
1932         /* Some platform decide whether they support huge pages at boot
1933          * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1934          * there is no such support
1935          */
1936         if (HPAGE_SHIFT == 0)
1937                 return 0;
1938
1939         if (!size_to_hstate(default_hstate_size)) {
1940                 default_hstate_size = HPAGE_SIZE;
1941                 if (!size_to_hstate(default_hstate_size))
1942                         hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1943         }
1944         default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
1945         if (default_hstate_max_huge_pages)
1946                 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1947
1948         hugetlb_init_hstates();
1949         gather_bootmem_prealloc();
1950         report_hugepages();
1951
1952         hugetlb_sysfs_init();
1953         hugetlb_register_all_nodes();
1954         hugetlb_cgroup_file_init();
1955
1956         return 0;
1957 }
1958 module_init(hugetlb_init);
1959
1960 /* Should be called on processing a hugepagesz=... option */
1961 void __init hugetlb_add_hstate(unsigned order)
1962 {
1963         struct hstate *h;
1964         unsigned long i;
1965
1966         if (size_to_hstate(PAGE_SIZE << order)) {
1967                 pr_warning("hugepagesz= specified twice, ignoring\n");
1968                 return;
1969         }
1970         BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
1971         BUG_ON(order == 0);
1972         h = &hstates[hugetlb_max_hstate++];
1973         h->order = order;
1974         h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1975         h->nr_huge_pages = 0;
1976         h->free_huge_pages = 0;
1977         for (i = 0; i < MAX_NUMNODES; ++i)
1978                 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1979         INIT_LIST_HEAD(&h->hugepage_activelist);
1980         h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
1981         h->next_nid_to_free = first_node(node_states[N_MEMORY]);
1982         snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1983                                         huge_page_size(h)/1024);
1984
1985         parsed_hstate = h;
1986 }
1987
1988 static int __init hugetlb_nrpages_setup(char *s)
1989 {
1990         unsigned long *mhp;
1991         static unsigned long *last_mhp;
1992
1993         /*
1994          * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
1995          * so this hugepages= parameter goes to the "default hstate".
1996          */
1997         if (!hugetlb_max_hstate)
1998                 mhp = &default_hstate_max_huge_pages;
1999         else
2000                 mhp = &parsed_hstate->max_huge_pages;
2001
2002         if (mhp == last_mhp) {
2003                 pr_warning("hugepages= specified twice without "
2004                            "interleaving hugepagesz=, ignoring\n");
2005                 return 1;
2006         }
2007
2008         if (sscanf(s, "%lu", mhp) <= 0)
2009                 *mhp = 0;
2010
2011         /*
2012          * Global state is always initialized later in hugetlb_init.
2013          * But we need to allocate >= MAX_ORDER hstates here early to still
2014          * use the bootmem allocator.
2015          */
2016         if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2017                 hugetlb_hstate_alloc_pages(parsed_hstate);
2018
2019         last_mhp = mhp;
2020
2021         return 1;
2022 }
2023 __setup("hugepages=", hugetlb_nrpages_setup);
2024
2025 static int __init hugetlb_default_setup(char *s)
2026 {
2027         default_hstate_size = memparse(s, &s);
2028         return 1;
2029 }
2030 __setup("default_hugepagesz=", hugetlb_default_setup);
2031
2032 static unsigned int cpuset_mems_nr(unsigned int *array)
2033 {
2034         int node;
2035         unsigned int nr = 0;
2036
2037         for_each_node_mask(node, cpuset_current_mems_allowed)
2038                 nr += array[node];
2039
2040         return nr;
2041 }
2042
2043 #ifdef CONFIG_SYSCTL
2044 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2045                          struct ctl_table *table, int write,
2046                          void __user *buffer, size_t *length, loff_t *ppos)
2047 {
2048         struct hstate *h = &default_hstate;
2049         unsigned long tmp;
2050         int ret;
2051
2052         tmp = h->max_huge_pages;
2053
2054         if (write && h->order >= MAX_ORDER)
2055                 return -EINVAL;
2056
2057         table->data = &tmp;
2058         table->maxlen = sizeof(unsigned long);
2059         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2060         if (ret)
2061                 goto out;
2062
2063         if (write) {
2064                 NODEMASK_ALLOC(nodemask_t, nodes_allowed,
2065                                                 GFP_KERNEL | __GFP_NORETRY);
2066                 if (!(obey_mempolicy &&
2067                                init_nodemask_of_mempolicy(nodes_allowed))) {
2068                         NODEMASK_FREE(nodes_allowed);
2069                         nodes_allowed = &node_states[N_MEMORY];
2070                 }
2071                 h->max_huge_pages = set_max_huge_pages(h, tmp, nodes_allowed);
2072
2073                 if (nodes_allowed != &node_states[N_MEMORY])
2074                         NODEMASK_FREE(nodes_allowed);
2075         }
2076 out:
2077         return ret;
2078 }
2079
2080 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2081                           void __user *buffer, size_t *length, loff_t *ppos)
2082 {
2083
2084         return hugetlb_sysctl_handler_common(false, table, write,
2085                                                         buffer, length, ppos);
2086 }
2087
2088 #ifdef CONFIG_NUMA
2089 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2090                           void __user *buffer, size_t *length, loff_t *ppos)
2091 {
2092         return hugetlb_sysctl_handler_common(true, table, write,
2093                                                         buffer, length, ppos);
2094 }
2095 #endif /* CONFIG_NUMA */
2096
2097 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
2098                         void __user *buffer,
2099                         size_t *length, loff_t *ppos)
2100 {
2101         proc_dointvec(table, write, buffer, length, ppos);
2102         if (hugepages_treat_as_movable)
2103                 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
2104         else
2105                 htlb_alloc_mask = GFP_HIGHUSER;
2106         return 0;
2107 }
2108
2109 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2110                         void __user *buffer,
2111                         size_t *length, loff_t *ppos)
2112 {
2113         struct hstate *h = &default_hstate;
2114         unsigned long tmp;
2115         int ret;
2116
2117         tmp = h->nr_overcommit_huge_pages;
2118
2119         if (write && h->order >= MAX_ORDER)
2120                 return -EINVAL;
2121
2122         table->data = &tmp;
2123         table->maxlen = sizeof(unsigned long);
2124         ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2125         if (ret)
2126                 goto out;
2127
2128         if (write) {
2129                 spin_lock(&hugetlb_lock);
2130                 h->nr_overcommit_huge_pages = tmp;
2131                 spin_unlock(&hugetlb_lock);
2132         }
2133 out:
2134         return ret;
2135 }
2136
2137 #endif /* CONFIG_SYSCTL */
2138
2139 void hugetlb_report_meminfo(struct seq_file *m)
2140 {
2141         struct hstate *h = &default_hstate;
2142         seq_printf(m,
2143                         "HugePages_Total:   %5lu\n"
2144                         "HugePages_Free:    %5lu\n"
2145                         "HugePages_Rsvd:    %5lu\n"
2146                         "HugePages_Surp:    %5lu\n"
2147                         "Hugepagesize:   %8lu kB\n",
2148                         h->nr_huge_pages,
2149                         h->free_huge_pages,
2150                         h->resv_huge_pages,
2151                         h->surplus_huge_pages,
2152                         1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2153 }
2154
2155 int hugetlb_report_node_meminfo(int nid, char *buf)
2156 {
2157         struct hstate *h = &default_hstate;
2158         return sprintf(buf,
2159                 "Node %d HugePages_Total: %5u\n"
2160                 "Node %d HugePages_Free:  %5u\n"
2161                 "Node %d HugePages_Surp:  %5u\n",
2162                 nid, h->nr_huge_pages_node[nid],
2163                 nid, h->free_huge_pages_node[nid],
2164                 nid, h->surplus_huge_pages_node[nid]);
2165 }
2166
2167 void hugetlb_show_meminfo(void)
2168 {
2169         struct hstate *h;
2170         int nid;
2171
2172         for_each_node_state(nid, N_MEMORY)
2173                 for_each_hstate(h)
2174                         pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2175                                 nid,
2176                                 h->nr_huge_pages_node[nid],
2177                                 h->free_huge_pages_node[nid],
2178                                 h->surplus_huge_pages_node[nid],
2179                                 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2180 }
2181
2182 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2183 unsigned long hugetlb_total_pages(void)
2184 {
2185         struct hstate *h;
2186         unsigned long nr_total_pages = 0;
2187
2188         for_each_hstate(h)
2189                 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2190         return nr_total_pages;
2191 }
2192
2193 static int hugetlb_acct_memory(struct hstate *h, long delta)
2194 {
2195         int ret = -ENOMEM;
2196
2197         spin_lock(&hugetlb_lock);
2198         /*
2199          * When cpuset is configured, it breaks the strict hugetlb page
2200          * reservation as the accounting is done on a global variable. Such
2201          * reservation is completely rubbish in the presence of cpuset because
2202          * the reservation is not checked against page availability for the
2203          * current cpuset. Application can still potentially OOM'ed by kernel
2204          * with lack of free htlb page in cpuset that the task is in.
2205          * Attempt to enforce strict accounting with cpuset is almost
2206          * impossible (or too ugly) because cpuset is too fluid that
2207          * task or memory node can be dynamically moved between cpusets.
2208          *
2209          * The change of semantics for shared hugetlb mapping with cpuset is
2210          * undesirable. However, in order to preserve some of the semantics,
2211          * we fall back to check against current free page availability as
2212          * a best attempt and hopefully to minimize the impact of changing
2213          * semantics that cpuset has.
2214          */
2215         if (delta > 0) {
2216                 if (gather_surplus_pages(h, delta) < 0)
2217                         goto out;
2218
2219                 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2220                         return_unused_surplus_pages(h, delta);
2221                         goto out;
2222                 }
2223         }
2224
2225         ret = 0;
2226         if (delta < 0)
2227                 return_unused_surplus_pages(h, (unsigned long) -delta);
2228
2229 out:
2230         spin_unlock(&hugetlb_lock);
2231         return ret;
2232 }
2233
2234 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2235 {
2236         struct resv_map *reservations = vma_resv_map(vma);
2237
2238         /*
2239          * This new VMA should share its siblings reservation map if present.
2240          * The VMA will only ever have a valid reservation map pointer where
2241          * it is being copied for another still existing VMA.  As that VMA
2242          * has a reference to the reservation map it cannot disappear until
2243          * after this open call completes.  It is therefore safe to take a
2244          * new reference here without additional locking.
2245          */
2246         if (reservations)
2247                 kref_get(&reservations->refs);
2248 }
2249
2250 static void resv_map_put(struct vm_area_struct *vma)
2251 {
2252         struct resv_map *reservations = vma_resv_map(vma);
2253
2254         if (!reservations)
2255                 return;
2256         kref_put(&reservations->refs, resv_map_release);
2257 }
2258
2259 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2260 {
2261         struct hstate *h = hstate_vma(vma);
2262         struct resv_map *reservations = vma_resv_map(vma);
2263         struct hugepage_subpool *spool = subpool_vma(vma);
2264         unsigned long reserve;
2265         unsigned long start;
2266         unsigned long end;
2267
2268         if (reservations) {
2269                 start = vma_hugecache_offset(h, vma, vma->vm_start);
2270                 end = vma_hugecache_offset(h, vma, vma->vm_end);
2271
2272                 reserve = (end - start) -
2273                         region_count(&reservations->regions, start, end);
2274
2275                 resv_map_put(vma);
2276
2277                 if (reserve) {
2278                         hugetlb_acct_memory(h, -reserve);
2279                         hugepage_subpool_put_pages(spool, reserve);
2280                 }
2281         }
2282 }
2283
2284 /*
2285  * We cannot handle pagefaults against hugetlb pages at all.  They cause
2286  * handle_mm_fault() to try to instantiate regular-sized pages in the
2287  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
2288  * this far.
2289  */
2290 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2291 {
2292         BUG();
2293         return 0;
2294 }
2295
2296 const struct vm_operations_struct hugetlb_vm_ops = {
2297         .fault = hugetlb_vm_op_fault,
2298         .open = hugetlb_vm_op_open,
2299         .close = hugetlb_vm_op_close,
2300 };
2301
2302 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
2303                                 int writable)
2304 {
2305         pte_t entry;
2306
2307         if (writable) {
2308                 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
2309                                          vma->vm_page_prot)));
2310         } else {
2311                 entry = huge_pte_wrprotect(mk_huge_pte(page,
2312                                            vma->vm_page_prot));
2313         }
2314         entry = pte_mkyoung(entry);
2315         entry = pte_mkhuge(entry);
2316         entry = arch_make_huge_pte(entry, vma, page, writable);
2317
2318         return entry;
2319 }
2320
2321 static void set_huge_ptep_writable(struct vm_area_struct *vma,
2322                                    unsigned long address, pte_t *ptep)
2323 {
2324         pte_t entry;
2325
2326         entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
2327         if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
2328                 update_mmu_cache(vma, address, ptep);
2329 }
2330
2331 static int is_hugetlb_entry_migration(pte_t pte)
2332 {
2333         swp_entry_t swp;
2334
2335         if (huge_pte_none(pte) || pte_present(pte))
2336                 return 0;
2337         swp = pte_to_swp_entry(pte);
2338         if (non_swap_entry(swp) && is_migration_entry(swp))
2339                 return 1;
2340         else
2341                 return 0;
2342 }
2343
2344 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
2345 {
2346         swp_entry_t swp;
2347
2348         if (huge_pte_none(pte) || pte_present(pte))
2349                 return 0;
2350         swp = pte_to_swp_entry(pte);
2351         if (non_swap_entry(swp) && is_hwpoison_entry(swp))
2352                 return 1;
2353         else
2354                 return 0;
2355 }
2356
2357 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
2358                             struct vm_area_struct *vma)
2359 {
2360         pte_t *src_pte, *dst_pte, entry;
2361         struct page *ptepage;
2362         unsigned long addr;
2363         int cow;
2364         struct hstate *h = hstate_vma(vma);
2365         unsigned long sz = huge_page_size(h);
2366         unsigned long mmun_start;       /* For mmu_notifiers */
2367         unsigned long mmun_end;         /* For mmu_notifiers */
2368         int ret = 0;
2369
2370         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
2371
2372         mmun_start = vma->vm_start;
2373         mmun_end = vma->vm_end;
2374         if (cow)
2375                 mmu_notifier_invalidate_range_start(vma, mmun_start,
2376                                                     mmun_end, MMU_MIGRATE);
2377
2378         for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
2379                 src_pte = huge_pte_offset(src, addr);
2380                 if (!src_pte)
2381                         continue;
2382                 dst_pte = huge_pte_alloc(dst, addr, sz);
2383                 if (!dst_pte) {
2384                         ret = -ENOMEM;
2385                         break;
2386                 }
2387
2388                 /* If the pagetables are shared don't copy or take references */
2389                 if (dst_pte == src_pte)
2390                         continue;
2391
2392                 spin_lock(&dst->page_table_lock);
2393                 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
2394                 entry = huge_ptep_get(src_pte);
2395                 if (huge_pte_none(entry)) { /* skip none entry */
2396                         ;
2397                 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
2398                                     is_hugetlb_entry_hwpoisoned(entry))) {
2399                         swp_entry_t swp_entry = pte_to_swp_entry(entry);
2400
2401                         if (is_write_migration_entry(swp_entry) && cow) {
2402                                 /*
2403                                  * COW mappings require pages in both
2404                                  * parent and child to be set to read.
2405                                  */
2406                                 make_migration_entry_read(&swp_entry);
2407                                 entry = swp_entry_to_pte(swp_entry);
2408                                 set_huge_pte_at(src, addr, src_pte, entry);
2409                         }
2410                         set_huge_pte_at(dst, addr, dst_pte, entry);
2411                 } else {
2412                         if (cow) {
2413                                 huge_ptep_set_wrprotect(src, addr, src_pte);
2414                                 mmu_notifier_invalidate_range(src, mmun_start,
2415                                                                    mmun_end);
2416                         }
2417                         entry = huge_ptep_get(src_pte);
2418                         ptepage = pte_page(entry);
2419                         get_page(ptepage);
2420                         page_dup_rmap(ptepage);
2421                         set_huge_pte_at(dst, addr, dst_pte, entry);
2422                 }
2423                 spin_unlock(&src->page_table_lock);
2424                 spin_unlock(&dst->page_table_lock);
2425         }
2426
2427         if (cow)
2428                 mmu_notifier_invalidate_range_end(vma, mmun_start,
2429                                                   mmun_end, MMU_MIGRATE);
2430
2431         return ret;
2432 }
2433
2434 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
2435                             unsigned long start, unsigned long end,
2436                             struct page *ref_page)
2437 {
2438         int force_flush = 0;
2439         struct mm_struct *mm = vma->vm_mm;
2440         unsigned long address;
2441         pte_t *ptep;
2442         pte_t pte;
2443         struct page *page;
2444         struct hstate *h = hstate_vma(vma);
2445         unsigned long sz = huge_page_size(h);
2446         const unsigned long mmun_start = start; /* For mmu_notifiers */
2447         const unsigned long mmun_end   = end;   /* For mmu_notifiers */
2448
2449         WARN_ON(!is_vm_hugetlb_page(vma));
2450         BUG_ON(start & ~huge_page_mask(h));
2451         BUG_ON(end & ~huge_page_mask(h));
2452
2453         tlb_start_vma(tlb, vma);
2454         mmu_notifier_invalidate_range_start(vma, mmun_start,
2455                                             mmun_end, MMU_MIGRATE);
2456 again:
2457         spin_lock(&mm->page_table_lock);
2458         for (address = start; address < end; address += sz) {
2459                 ptep = huge_pte_offset(mm, address);
2460                 if (!ptep)
2461                         continue;
2462
2463                 if (huge_pmd_unshare(mm, &address, ptep))
2464                         continue;
2465
2466                 pte = huge_ptep_get(ptep);
2467                 if (huge_pte_none(pte))
2468                         continue;
2469
2470                 /*
2471                  * Migrating hugepage or HWPoisoned hugepage is already
2472                  * unmapped and its refcount is dropped, so just clear pte here.
2473                  */
2474                 if (unlikely(!pte_present(pte))) {
2475                         huge_pte_clear(mm, address, ptep);
2476                         continue;
2477                 }
2478
2479                 page = pte_page(pte);
2480                 /*
2481                  * If a reference page is supplied, it is because a specific
2482                  * page is being unmapped, not a range. Ensure the page we
2483                  * are about to unmap is the actual page of interest.
2484                  */
2485                 if (ref_page) {
2486                         if (page != ref_page)
2487                                 continue;
2488
2489                         /*
2490                          * Mark the VMA as having unmapped its page so that
2491                          * future faults in this VMA will fail rather than
2492                          * looking like data was lost
2493                          */
2494                         set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
2495                 }
2496
2497                 pte = huge_ptep_get_and_clear(mm, address, ptep);
2498                 tlb_remove_tlb_entry(tlb, ptep, address);
2499                 if (huge_pte_dirty(pte))
2500                         set_page_dirty(page);
2501
2502                 page_remove_rmap(page);
2503                 force_flush = !__tlb_remove_page(tlb, page);
2504                 if (force_flush)
2505                         break;
2506                 /* Bail out after unmapping reference page if supplied */
2507                 if (ref_page)
2508                         break;
2509         }
2510         spin_unlock(&mm->page_table_lock);
2511         mmu_notifier_invalidate_range_free_pages(vma, start, address);
2512         /*
2513          * mmu_gather ran out of room to batch pages, we break out of
2514          * the PTE lock to avoid doing the potential expensive TLB invalidate
2515          * and page-free while holding it.
2516          */
2517         if (force_flush) {
2518                 force_flush = 0;
2519                 tlb_flush_mmu(tlb);
2520                 if (address < end && !ref_page)
2521                         goto again;
2522         }
2523         mmu_notifier_invalidate_range_end(vma, mmun_start,
2524                                           mmun_end, MMU_MIGRATE);
2525         tlb_end_vma(tlb, vma);
2526 }
2527
2528 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
2529                           struct vm_area_struct *vma, unsigned long start,
2530                           unsigned long end, struct page *ref_page)
2531 {
2532         __unmap_hugepage_range(tlb, vma, start, end, ref_page);
2533
2534         /*
2535          * Clear this flag so that x86's huge_pmd_share page_table_shareable
2536          * test will fail on a vma being torn down, and not grab a page table
2537          * on its way out.  We're lucky that the flag has such an appropriate
2538          * name, and can in fact be safely cleared here. We could clear it
2539          * before the __unmap_hugepage_range above, but all that's necessary
2540          * is to clear it before releasing the i_mmap_mutex. This works
2541          * because in the context this is called, the VMA is about to be
2542          * destroyed and the i_mmap_mutex is held.
2543          */
2544         vma->vm_flags &= ~VM_MAYSHARE;
2545 }
2546
2547 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
2548                           unsigned long end, struct page *ref_page)
2549 {
2550         struct mm_struct *mm;
2551         struct mmu_gather tlb;
2552
2553         mm = vma->vm_mm;
2554
2555         tlb_gather_mmu(&tlb, mm, start, end);
2556         __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
2557         tlb_finish_mmu(&tlb, start, end);
2558 }
2559
2560 /*
2561  * This is called when the original mapper is failing to COW a MAP_PRIVATE
2562  * mappping it owns the reserve page for. The intention is to unmap the page
2563  * from other VMAs and let the children be SIGKILLed if they are faulting the
2564  * same region.
2565  */
2566 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
2567                                 struct page *page, unsigned long address)
2568 {
2569         struct hstate *h = hstate_vma(vma);
2570         struct vm_area_struct *iter_vma;
2571         struct address_space *mapping;
2572         pgoff_t pgoff;
2573
2574         /*
2575          * vm_pgoff is in PAGE_SIZE units, hence the different calculation
2576          * from page cache lookup which is in HPAGE_SIZE units.
2577          */
2578         address = address & huge_page_mask(h);
2579         pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
2580                         vma->vm_pgoff;
2581         mapping = file_inode(vma->vm_file)->i_mapping;
2582
2583         /*
2584          * Take the mapping lock for the duration of the table walk. As
2585          * this mapping should be shared between all the VMAs,
2586          * __unmap_hugepage_range() is called as the lock is already held
2587          */
2588         mutex_lock(&mapping->i_mmap_mutex);
2589         vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
2590                 /* Do not unmap the current VMA */
2591                 if (iter_vma == vma)
2592                         continue;
2593
2594                 /*
2595                  * Shared VMAs have their own reserves and do not affect
2596                  * MAP_PRIVATE accounting but it is possible that a shared
2597                  * VMA is using the same page so check and skip such VMAs.
2598                  */
2599                 if (iter_vma->vm_flags & VM_MAYSHARE)
2600                         continue;
2601
2602                 /*
2603                  * Unmap the page from other VMAs without their own reserves.
2604                  * They get marked to be SIGKILLed if they fault in these
2605                  * areas. This is because a future no-page fault on this VMA
2606                  * could insert a zeroed page instead of the data existing
2607                  * from the time of fork. This would look like data corruption
2608                  */
2609                 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
2610                         unmap_hugepage_range(iter_vma, address,
2611                                              address + huge_page_size(h), page);
2612         }
2613         mutex_unlock(&mapping->i_mmap_mutex);
2614
2615         return 1;
2616 }
2617
2618 /*
2619  * Hugetlb_cow() should be called with page lock of the original hugepage held.
2620  * Called with hugetlb_instantiation_mutex held and pte_page locked so we
2621  * cannot race with other handlers or page migration.
2622  * Keep the pte_same checks anyway to make transition from the mutex easier.
2623  */
2624 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
2625                         unsigned long address, pte_t *ptep, pte_t pte,
2626                         struct page *pagecache_page)
2627 {
2628         struct hstate *h = hstate_vma(vma);
2629         struct page *old_page, *new_page;
2630         int avoidcopy;
2631         int outside_reserve = 0;
2632         unsigned long mmun_start;       /* For mmu_notifiers */
2633         unsigned long mmun_end;         /* For mmu_notifiers */
2634
2635         old_page = pte_page(pte);
2636
2637 retry_avoidcopy:
2638         /* If no-one else is actually using this page, avoid the copy
2639          * and just make the page writable */
2640         avoidcopy = (page_mapcount(old_page) == 1);
2641         if (avoidcopy) {
2642                 if (PageAnon(old_page))
2643                         page_move_anon_rmap(old_page, vma, address);
2644                 set_huge_ptep_writable(vma, address, ptep);
2645                 return 0;
2646         }
2647
2648         /*
2649          * If the process that created a MAP_PRIVATE mapping is about to
2650          * perform a COW due to a shared page count, attempt to satisfy
2651          * the allocation without using the existing reserves. The pagecache
2652          * page is used to determine if the reserve at this address was
2653          * consumed or not. If reserves were used, a partial faulted mapping
2654          * at the time of fork() could consume its reserves on COW instead
2655          * of the full address range.
2656          */
2657         if (!(vma->vm_flags & VM_MAYSHARE) &&
2658                         is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
2659                         old_page != pagecache_page)
2660                 outside_reserve = 1;
2661
2662         page_cache_get(old_page);
2663
2664         /* Drop page_table_lock as buddy allocator may be called */
2665         spin_unlock(&mm->page_table_lock);
2666         new_page = alloc_huge_page(vma, address, outside_reserve);
2667
2668         if (IS_ERR(new_page)) {
2669                 long err = PTR_ERR(new_page);
2670                 page_cache_release(old_page);
2671
2672                 /*
2673                  * If a process owning a MAP_PRIVATE mapping fails to COW,
2674                  * it is due to references held by a child and an insufficient
2675                  * huge page pool. To guarantee the original mappers
2676                  * reliability, unmap the page from child processes. The child
2677                  * may get SIGKILLed if it later faults.
2678                  */
2679                 if (outside_reserve) {
2680                         BUG_ON(huge_pte_none(pte));
2681                         if (unmap_ref_private(mm, vma, old_page, address)) {
2682                                 BUG_ON(huge_pte_none(pte));
2683                                 spin_lock(&mm->page_table_lock);
2684                                 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2685                                 if (likely(pte_same(huge_ptep_get(ptep), pte)))
2686                                         goto retry_avoidcopy;
2687                                 /*
2688                                  * race occurs while re-acquiring page_table_lock, and
2689                                  * our job is done.
2690                                  */
2691                                 return 0;
2692                         }
2693                         WARN_ON_ONCE(1);
2694                 }
2695
2696                 /* Caller expects lock to be held */
2697                 spin_lock(&mm->page_table_lock);
2698                 if (err == -ENOMEM)
2699                         return VM_FAULT_OOM;
2700                 else
2701                         return VM_FAULT_SIGBUS;
2702         }
2703
2704         /*
2705          * When the original hugepage is shared one, it does not have
2706          * anon_vma prepared.
2707          */
2708         if (unlikely(anon_vma_prepare(vma))) {
2709                 page_cache_release(new_page);
2710                 page_cache_release(old_page);
2711                 /* Caller expects lock to be held */
2712                 spin_lock(&mm->page_table_lock);
2713                 return VM_FAULT_OOM;
2714         }
2715
2716         copy_user_huge_page(new_page, old_page, address, vma,
2717                             pages_per_huge_page(h));
2718         __SetPageUptodate(new_page);
2719
2720         mmun_start = address & huge_page_mask(h);
2721         mmun_end = mmun_start + huge_page_size(h);
2722         mmu_notifier_invalidate_range_start(vma, mmun_start, mmun_end,
2723                                             MMU_MIGRATE);
2724         /*
2725          * Retake the page_table_lock to check for racing updates
2726          * before the page tables are altered
2727          */
2728         spin_lock(&mm->page_table_lock);
2729         ptep = huge_pte_offset(mm, address & huge_page_mask(h));
2730         if (likely(pte_same(huge_ptep_get(ptep), pte))) {
2731                 /* Break COW */
2732                 huge_ptep_clear_flush(vma, address, ptep);
2733                 mmu_notifier_invalidate_range_free_pages(vma, mmun_start,
2734                                                          mmun_end);
2735                 set_huge_pte_at(mm, address, ptep,
2736                                 make_huge_pte(vma, new_page, 1));
2737                 page_remove_rmap(old_page);
2738                 hugepage_add_new_anon_rmap(new_page, vma, address);
2739                 /* Make the old page be freed below */
2740                 new_page = old_page;
2741         }
2742         spin_unlock(&mm->page_table_lock);
2743         mmu_notifier_invalidate_range_end(vma, mmun_start, mmun_end,
2744                                           MMU_MIGRATE);
2745         /* Caller expects lock to be held */
2746         spin_lock(&mm->page_table_lock);
2747         page_cache_release(new_page);
2748         page_cache_release(old_page);
2749         return 0;
2750 }
2751
2752 /* Return the pagecache page at a given address within a VMA */
2753 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
2754                         struct vm_area_struct *vma, unsigned long address)
2755 {
2756         struct address_space *mapping;
2757         pgoff_t idx;
2758
2759         mapping = vma->vm_file->f_mapping;
2760         idx = vma_hugecache_offset(h, vma, address);
2761
2762         return find_lock_page(mapping, idx);
2763 }
2764
2765 /*
2766  * Return whether there is a pagecache page to back given address within VMA.
2767  * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
2768  */
2769 static bool hugetlbfs_pagecache_present(struct hstate *h,
2770                         struct vm_area_struct *vma, unsigned long address)
2771 {
2772         struct address_space *mapping;
2773         pgoff_t idx;
2774         struct page *page;
2775
2776         mapping = vma->vm_file->f_mapping;
2777         idx = vma_hugecache_offset(h, vma, address);
2778
2779         page = find_get_page(mapping, idx);
2780         if (page)
2781                 put_page(page);
2782         return page != NULL;
2783 }
2784
2785 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2786                         unsigned long address, pte_t *ptep, unsigned int flags)
2787 {
2788         struct hstate *h = hstate_vma(vma);
2789         int ret = VM_FAULT_SIGBUS;
2790         int anon_rmap = 0;
2791         pgoff_t idx;
2792         unsigned long size;
2793         struct page *page;
2794         struct address_space *mapping;
2795         pte_t new_pte;
2796
2797         /*
2798          * Currently, we are forced to kill the process in the event the
2799          * original mapper has unmapped pages from the child due to a failed
2800          * COW. Warn that such a situation has occurred as it may not be obvious
2801          */
2802         if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
2803                 pr_warning("PID %d killed due to inadequate hugepage pool\n",
2804                            current->pid);
2805                 return ret;
2806         }
2807
2808         mapping = vma->vm_file->f_mapping;
2809         idx = vma_hugecache_offset(h, vma, address);
2810
2811         /*
2812          * Use page lock to guard against racing truncation
2813          * before we get page_table_lock.
2814          */
2815 retry:
2816         page = find_lock_page(mapping, idx);
2817         if (!page) {
2818                 size = i_size_read(mapping->host) >> huge_page_shift(h);
2819                 if (idx >= size)
2820                         goto out;
2821                 page = alloc_huge_page(vma, address, 0);
2822                 if (IS_ERR(page)) {
2823                         ret = PTR_ERR(page);
2824                         if (ret == -ENOMEM)
2825                                 ret = VM_FAULT_OOM;
2826                         else
2827                                 ret = VM_FAULT_SIGBUS;
2828                         goto out;
2829                 }
2830                 clear_huge_page(page, address, pages_per_huge_page(h));
2831                 __SetPageUptodate(page);
2832
2833                 if (vma->vm_flags & VM_MAYSHARE) {
2834                         int err;
2835                         struct inode *inode = mapping->host;
2836
2837                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
2838                         if (err) {
2839                                 put_page(page);
2840                                 if (err == -EEXIST)
2841                                         goto retry;
2842                                 goto out;
2843                         }
2844
2845                         spin_lock(&inode->i_lock);
2846                         inode->i_blocks += blocks_per_huge_page(h);
2847                         spin_unlock(&inode->i_lock);
2848                 } else {
2849                         lock_page(page);
2850                         if (unlikely(anon_vma_prepare(vma))) {
2851                                 ret = VM_FAULT_OOM;
2852                                 goto backout_unlocked;
2853                         }
2854                         anon_rmap = 1;
2855                 }
2856         } else {
2857                 /*
2858                  * If memory error occurs between mmap() and fault, some process
2859                  * don't have hwpoisoned swap entry for errored virtual address.
2860                  * So we need to block hugepage fault by PG_hwpoison bit check.
2861                  */
2862                 if (unlikely(PageHWPoison(page))) {
2863                         ret = VM_FAULT_HWPOISON |
2864                                 VM_FAULT_SET_HINDEX(hstate_index(h));
2865                         goto backout_unlocked;
2866                 }
2867         }
2868
2869         /*
2870          * If we are going to COW a private mapping later, we examine the
2871          * pending reservations for this page now. This will ensure that
2872          * any allocations necessary to record that reservation occur outside
2873          * the spinlock.
2874          */
2875         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
2876                 if (vma_needs_reservation(h, vma, address) < 0) {
2877                         ret = VM_FAULT_OOM;
2878                         goto backout_unlocked;
2879                 }
2880
2881         spin_lock(&mm->page_table_lock);
2882         size = i_size_read(mapping->host) >> huge_page_shift(h);
2883         if (idx >= size)
2884                 goto backout;
2885
2886         ret = 0;
2887         if (!huge_pte_none(huge_ptep_get(ptep)))
2888                 goto backout;
2889
2890         if (anon_rmap)
2891                 hugepage_add_new_anon_rmap(page, vma, address);
2892         else
2893                 page_dup_rmap(page);
2894         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2895                                 && (vma->vm_flags & VM_SHARED)));
2896         set_huge_pte_at(mm, address, ptep, new_pte);
2897
2898         if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
2899                 /* Optimization, do the COW without a second fault */
2900                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2901         }
2902
2903         spin_unlock(&mm->page_table_lock);
2904         unlock_page(page);
2905 out:
2906         return ret;
2907
2908 backout:
2909         spin_unlock(&mm->page_table_lock);
2910 backout_unlocked:
2911         unlock_page(page);
2912         put_page(page);
2913         goto out;
2914 }
2915
2916 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2917                         unsigned long address, unsigned int flags)
2918 {
2919         pte_t *ptep;
2920         pte_t entry;
2921         int ret;
2922         struct page *page = NULL;
2923         struct page *pagecache_page = NULL;
2924         static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2925         struct hstate *h = hstate_vma(vma);
2926
2927         address &= huge_page_mask(h);
2928
2929         ptep = huge_pte_offset(mm, address);
2930         if (ptep) {
2931                 entry = huge_ptep_get(ptep);
2932                 if (unlikely(is_hugetlb_entry_migration(entry))) {
2933                         migration_entry_wait_huge(mm, ptep);
2934                         return 0;
2935                 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
2936                         return VM_FAULT_HWPOISON_LARGE |
2937                                 VM_FAULT_SET_HINDEX(hstate_index(h));
2938         }
2939
2940         ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2941         if (!ptep)
2942                 return VM_FAULT_OOM;
2943
2944         /*
2945          * Serialize hugepage allocation and instantiation, so that we don't
2946          * get spurious allocation failures if two CPUs race to instantiate
2947          * the same page in the page cache.
2948          */
2949         mutex_lock(&hugetlb_instantiation_mutex);
2950         entry = huge_ptep_get(ptep);
2951         if (huge_pte_none(entry)) {
2952                 ret = hugetlb_no_page(mm, vma, address, ptep, flags);
2953                 goto out_mutex;
2954         }
2955
2956         ret = 0;
2957
2958         /*
2959          * If we are going to COW the mapping later, we examine the pending
2960          * reservations for this page now. This will ensure that any
2961          * allocations necessary to record that reservation occur outside the
2962          * spinlock. For private mappings, we also lookup the pagecache
2963          * page now as it is used to determine if a reservation has been
2964          * consumed.
2965          */
2966         if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
2967                 if (vma_needs_reservation(h, vma, address) < 0) {
2968                         ret = VM_FAULT_OOM;
2969                         goto out_mutex;
2970                 }
2971
2972                 if (!(vma->vm_flags & VM_MAYSHARE))
2973                         pagecache_page = hugetlbfs_pagecache_page(h,
2974                                                                 vma, address);
2975         }
2976
2977         /*
2978          * hugetlb_cow() requires page locks of pte_page(entry) and
2979          * pagecache_page, so here we need take the former one
2980          * when page != pagecache_page or !pagecache_page.
2981          * Note that locking order is always pagecache_page -> page,
2982          * so no worry about deadlock.
2983          */
2984         page = pte_page(entry);
2985         get_page(page);
2986         if (page != pagecache_page)
2987                 lock_page(page);
2988
2989         spin_lock(&mm->page_table_lock);
2990         /* Check for a racing update before calling hugetlb_cow */
2991         if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2992                 goto out_page_table_lock;
2993
2994
2995         if (flags & FAULT_FLAG_WRITE) {
2996                 if (!huge_pte_write(entry)) {
2997                         ret = hugetlb_cow(mm, vma, address, ptep, entry,
2998                                                         pagecache_page);
2999                         goto out_page_table_lock;
3000                 }
3001                 entry = huge_pte_mkdirty(entry);
3002         }
3003         entry = pte_mkyoung(entry);
3004         if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3005                                                 flags & FAULT_FLAG_WRITE))
3006                 update_mmu_cache(vma, address, ptep);
3007
3008 out_page_table_lock:
3009         spin_unlock(&mm->page_table_lock);
3010
3011         if (pagecache_page) {
3012                 unlock_page(pagecache_page);
3013                 put_page(pagecache_page);
3014         }
3015         if (page != pagecache_page)
3016                 unlock_page(page);
3017         put_page(page);
3018
3019 out_mutex:
3020         mutex_unlock(&hugetlb_instantiation_mutex);
3021
3022         return ret;
3023 }
3024
3025 /* Can be overriden by architectures */
3026 __attribute__((weak)) struct page *
3027 follow_huge_pud(struct mm_struct *mm, unsigned long address,
3028                pud_t *pud, int write)
3029 {
3030         BUG();
3031         return NULL;
3032 }
3033
3034 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3035                          struct page **pages, struct vm_area_struct **vmas,
3036                          unsigned long *position, unsigned long *nr_pages,
3037                          long i, unsigned int flags)
3038 {
3039         unsigned long pfn_offset;
3040         unsigned long vaddr = *position;
3041         unsigned long remainder = *nr_pages;
3042         struct hstate *h = hstate_vma(vma);
3043
3044         spin_lock(&mm->page_table_lock);
3045         while (vaddr < vma->vm_end && remainder) {
3046                 pte_t *pte;
3047                 int absent;
3048                 struct page *page;
3049
3050                 /*
3051                  * Some archs (sparc64, sh*) have multiple pte_ts to
3052                  * each hugepage.  We have to make sure we get the
3053                  * first, for the page indexing below to work.
3054                  */
3055                 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3056                 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3057
3058                 /*
3059                  * When coredumping, it suits get_dump_page if we just return
3060                  * an error where there's an empty slot with no huge pagecache
3061                  * to back it.  This way, we avoid allocating a hugepage, and
3062                  * the sparse dumpfile avoids allocating disk blocks, but its
3063                  * huge holes still show up with zeroes where they need to be.
3064                  */
3065                 if (absent && (flags & FOLL_DUMP) &&
3066                     !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3067                         remainder = 0;
3068                         break;
3069                 }
3070
3071                 /*
3072                  * We need call hugetlb_fault for both hugepages under migration
3073                  * (in which case hugetlb_fault waits for the migration,) and
3074                  * hwpoisoned hugepages (in which case we need to prevent the
3075                  * caller from accessing to them.) In order to do this, we use
3076                  * here is_swap_pte instead of is_hugetlb_entry_migration and
3077                  * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3078                  * both cases, and because we can't follow correct pages
3079                  * directly from any kind of swap entries.
3080                  */
3081                 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3082                     ((flags & FOLL_WRITE) &&
3083                       !huge_pte_write(huge_ptep_get(pte)))) {
3084                         int ret;
3085
3086                         spin_unlock(&mm->page_table_lock);
3087                         ret = hugetlb_fault(mm, vma, vaddr,
3088                                 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3089                         spin_lock(&mm->page_table_lock);
3090                         if (!(ret & VM_FAULT_ERROR))
3091                                 continue;
3092
3093                         remainder = 0;
3094                         break;
3095                 }
3096
3097                 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3098                 page = pte_page(huge_ptep_get(pte));
3099 same_page:
3100                 if (pages) {
3101                         pages[i] = mem_map_offset(page, pfn_offset);
3102                         get_page(pages[i]);
3103                 }
3104
3105                 if (vmas)
3106                         vmas[i] = vma;
3107
3108                 vaddr += PAGE_SIZE;
3109                 ++pfn_offset;
3110                 --remainder;
3111                 ++i;
3112                 if (vaddr < vma->vm_end && remainder &&
3113                                 pfn_offset < pages_per_huge_page(h)) {
3114                         /*
3115                          * We use pfn_offset to avoid touching the pageframes
3116                          * of this compound page.
3117                          */
3118                         goto same_page;
3119                 }
3120         }
3121         spin_unlock(&mm->page_table_lock);
3122         *nr_pages = remainder;
3123         *position = vaddr;
3124
3125         return i ? i : -EFAULT;
3126 }
3127
3128 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3129                 unsigned long address, unsigned long end, pgprot_t newprot)
3130 {
3131         struct mm_struct *mm = vma->vm_mm;
3132         unsigned long start = address;
3133         pte_t *ptep;
3134         pte_t pte;
3135         struct hstate *h = hstate_vma(vma);
3136         unsigned long pages = 0;
3137
3138         BUG_ON(address >= end);
3139         flush_cache_range(vma, address, end);
3140
3141         mmu_notifier_invalidate_range_start(vma, start, end, MMU_MPROT);
3142         mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
3143         spin_lock(&mm->page_table_lock);
3144         for (; address < end; address += huge_page_size(h)) {
3145                 ptep = huge_pte_offset(mm, address);
3146                 if (!ptep)
3147                         continue;
3148                 if (huge_pmd_unshare(mm, &address, ptep)) {
3149                         pages++;
3150                         continue;
3151                 }
3152                 if (!huge_pte_none(huge_ptep_get(ptep))) {
3153                         pte = huge_ptep_get_and_clear(mm, address, ptep);
3154                         pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3155                         pte = arch_make_huge_pte(pte, vma, NULL, 0);
3156                         set_huge_pte_at(mm, address, ptep, pte);
3157                         pages++;
3158                 }
3159         }
3160         spin_unlock(&mm->page_table_lock);
3161         /*
3162          * Must flush TLB before releasing i_mmap_mutex: x86's huge_pmd_unshare
3163          * may have cleared our pud entry and done put_page on the page table:
3164          * once we release i_mmap_mutex, another task can do the final put_page
3165          * and that page table be reused and filled with junk.
3166          */
3167         flush_tlb_range(vma, start, end);
3168         mmu_notifier_invalidate_range(mm, start, end);
3169         mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
3170         mmu_notifier_invalidate_range_end(vma, start, end, MMU_MPROT);
3171
3172         return pages << h->order;
3173 }
3174
3175 int hugetlb_reserve_pages(struct inode *inode,
3176                                         long from, long to,
3177                                         struct vm_area_struct *vma,
3178                                         vm_flags_t vm_flags)
3179 {
3180         long ret, chg;
3181         struct hstate *h = hstate_inode(inode);
3182         struct hugepage_subpool *spool = subpool_inode(inode);
3183
3184         /*
3185          * Only apply hugepage reservation if asked. At fault time, an
3186          * attempt will be made for VM_NORESERVE to allocate a page
3187          * without using reserves
3188          */
3189         if (vm_flags & VM_NORESERVE)
3190                 return 0;
3191
3192         /*
3193          * Shared mappings base their reservation on the number of pages that
3194          * are already allocated on behalf of the file. Private mappings need
3195          * to reserve the full area even if read-only as mprotect() may be
3196          * called to make the mapping read-write. Assume !vma is a shm mapping
3197          */
3198         if (!vma || vma->vm_flags & VM_MAYSHARE)
3199                 chg = region_chg(&inode->i_mapping->private_list, from, to);
3200         else {
3201                 struct resv_map *resv_map = resv_map_alloc();
3202                 if (!resv_map)
3203                         return -ENOMEM;
3204
3205                 chg = to - from;
3206
3207                 set_vma_resv_map(vma, resv_map);
3208                 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
3209         }
3210
3211         if (chg < 0) {
3212                 ret = chg;
3213                 goto out_err;
3214         }
3215
3216         /* There must be enough pages in the subpool for the mapping */
3217         if (hugepage_subpool_get_pages(spool, chg)) {
3218                 ret = -ENOSPC;
3219                 goto out_err;
3220         }
3221
3222         /*
3223          * Check enough hugepages are available for the reservation.
3224          * Hand the pages back to the subpool if there are not
3225          */
3226         ret = hugetlb_acct_memory(h, chg);
3227         if (ret < 0) {
3228                 hugepage_subpool_put_pages(spool, chg);
3229                 goto out_err;
3230         }
3231
3232         /*
3233          * Account for the reservations made. Shared mappings record regions
3234          * that have reservations as they are shared by multiple VMAs.
3235          * When the last VMA disappears, the region map says how much
3236          * the reservation was and the page cache tells how much of
3237          * the reservation was consumed. Private mappings are per-VMA and
3238          * only the consumed reservations are tracked. When the VMA
3239          * disappears, the original reservation is the VMA size and the
3240          * consumed reservations are stored in the map. Hence, nothing
3241          * else has to be done for private mappings here
3242          */
3243         if (!vma || vma->vm_flags & VM_MAYSHARE)
3244                 region_add(&inode->i_mapping->private_list, from, to);
3245         return 0;
3246 out_err:
3247         if (vma)
3248                 resv_map_put(vma);
3249         return ret;
3250 }
3251
3252 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
3253 {
3254         struct hstate *h = hstate_inode(inode);
3255         long chg = region_truncate(&inode->i_mapping->private_list, offset);
3256         struct hugepage_subpool *spool = subpool_inode(inode);
3257
3258         spin_lock(&inode->i_lock);
3259         inode->i_blocks -= (blocks_per_huge_page(h) * freed);
3260         spin_unlock(&inode->i_lock);
3261
3262         hugepage_subpool_put_pages(spool, (chg - freed));
3263         hugetlb_acct_memory(h, -(chg - freed));
3264 }
3265
3266 #ifdef CONFIG_MEMORY_FAILURE
3267
3268 /* Should be called in hugetlb_lock */
3269 static int is_hugepage_on_freelist(struct page *hpage)
3270 {
3271         struct page *page;
3272         struct page *tmp;
3273         struct hstate *h = page_hstate(hpage);
3274         int nid = page_to_nid(hpage);
3275
3276         list_for_each_entry_safe(page, tmp, &h->hugepage_freelists[nid], lru)
3277                 if (page == hpage)
3278                         return 1;
3279         return 0;
3280 }
3281
3282 /*
3283  * This function is called from memory failure code.
3284  * Assume the caller holds page lock of the head page.
3285  */
3286 int dequeue_hwpoisoned_huge_page(struct page *hpage)
3287 {
3288         struct hstate *h = page_hstate(hpage);
3289         int nid = page_to_nid(hpage);
3290         int ret = -EBUSY;
3291
3292         spin_lock(&hugetlb_lock);
3293         if (is_hugepage_on_freelist(hpage)) {
3294                 /*
3295                  * Hwpoisoned hugepage isn't linked to activelist or freelist,
3296                  * but dangling hpage->lru can trigger list-debug warnings
3297                  * (this happens when we call unpoison_memory() on it),
3298                  * so let it point to itself with list_del_init().
3299                  */
3300                 list_del_init(&hpage->lru);
3301                 set_page_refcounted(hpage);
3302                 h->free_huge_pages--;
3303                 h->free_huge_pages_node[nid]--;
3304                 ret = 0;
3305         }
3306         spin_unlock(&hugetlb_lock);
3307         return ret;
3308 }
3309 #endif