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